Morning Coffee

Pcons: a new open-source software build tool

garyo@oberbrunner.com (Gary Oberbrunner) — Wed, 15 Jul 2026 12:00:00 GMT

I’ve been a C/C++ programmer most of my life. I wrote my first programs in the 1970s, so most of my professional life has been a sequence like this: write some code, build it, and debug it. The tools we use to do that second part, the build step, have gotten considerably better since make was written at Bell Labs in 1976[^make]. For many years at GenArts, I used, and helped develop, SCons. It was pretty great in those days; it built a complete graph of the code and executed it in parallel, build descriptions were written in python, not some one-off domain-specific language you had to learn. SCons is still in use in some very big projects, and I was proud of the work we did on it. But it is stuck in the python idioms of the 1990s and hasn’t been able to keep up with the times, and never achieved the traction we were hoping for.

[^make]: First publicly released with Unix v7 in 1979; I didn’t use it til the '80s.

Over the years, CMake became dominant as the way to build C and C++ applications. A huge community has sprung up around it and it’s genuinely the de facto standard for modern build tooling. It runs everywhere, supports lots of languages, and has lots of add-ons. It introduced the concept of transitive dependencies, which is a great innovation. The problem with CMake, though, is the build description language: the thing you actually write and maintain. It is abjectly terrible. Its syntax is odd at best, variable handling is byzantine, and debugging it is a nightmare. It also has undergone several complete redesigns of the public API over the years, so looking up how-to information online often leads to the old ways that aren’t supported anymore.

There are lots of other projects seeking to replace CMake: Meson and Bazel are two that are on good tracks. But after looking into them closely, I couldn’t bring myself to buy into either one. And I’d been thinking for years about rearchitecting SCons onto a modern python stack with Path and decorators and all the other wonderful stuff python has been doing, and fixing some of the pain points at the same time, but I’d never had the time to dig into it. And at the same time I started doing some consulting work using CMake, which taught me that there are some genuinely important features I hadn’t really understood deeply. Why couldn’t I create a new, modern build system with the best of CMake and SCons and all the things I’ve learned in a fifty year career in software?

Now that I’m (officially) retired, and I’ve been using a lot more Claude Code as a programming assistant, it seemed like the right time to try this as a collaborative project. And the more I got into the architecture, the more excited I got about it. It really is possible to describe a software build simply and cleanly by using modern python, a language everyone knows and is supported everywhere. I won’t go into the technical details here; you can check the website, or the docs or go straight the the source on github.

At this point it’s been used by a few small projects, I’ve done extensive testing myself, and I’ve ported some fairly significant-size open source projects to work with it (mostly privately, to test that it really can handle big jobs with multiple languages, libraries, dependencies, unusual custom commands, installers and so on). I think it’s ready for real use. If you are starting a C, C++, CUDA, Fortran, WASM, or other compiled-software project, please consider using pcons as your build tool! Or if you just can’t stand tearing your hair out one more time fighting with cmake, I’ve made it really easy to port over to pcons from cmake — there’s a porting guide, and the docs are AI-friendly so your AI can probably do the port in a few minutes.

Pcons is fully open source (MIT licensed) and I intend to stand behind it and make it a serious build tool. I hope you’ll try it — and please spread the word!

What Is the Future of Mathematics?

garyo@oberbrunner.com (Gary Oberbrunner) — Sun, 12 Jul 2026 12:00:00 GMT

For thousands of years, since the Babylonians and before, humans have studied the mysteries of pure mathematics. These studies probably started practically, as ways to understand the stars and the world around us as well as economics and government, but quickly turned to the beauty of pure math: number theory, proofs, irrational numbers and geometry. Unlike most other sciences, which have direct application to human lives [^astro], we don’t study pure math in the hope of improving human lives, curing disease, inventing new technology or saving the planet. It’s just beautiful and deep and mysterious.

[^astro]: With the possible exceptions of astrophysics, which is fairly remote from practical applications, and a few others like logic, arguably a branch of mathematics. Of course all studies can lead to applications; cryptography and modern crypto-based finance are good examples.

With the recent near-lightspeed improvement of large language models and AI reasoning systems, machines are starting to be able to prove problems in mathematics that have been unsolved for decades; see this article on a well-known Erdős problem. They are also rapidly improving in deep reasoning and ability to discover new patterns, regularities and anomalies. I think it won’t be long before they outpace humans in mathematical abilities and even discovery. I’m talking here about frontier mathematical ideas like category theory, group theory and topology, and nonlinear PDEs, not simple arithmetic, which computers outperformed humans on by the 1960s. I should note here that I am by no means a mathematician; the only course I had to drop in university was an advanced math course. But I love it as an amateur, and as I’m steeping myself in the whirlwind of AI, I can’t help but wonder about what superintelligent AIs will mean for mathematics.

Academic Math

In a near future where advanced AIs can prove or disprove almost any conjecture, I wonder what the future of math higher education will be. As a PhD student with a thesis proposal, you will be strongly tempted to just go press the red “prove me” button sitting right there at your desktop. But instead you plan to spend the next three years of your life going down that rabbit hole, investigating all avenues, reading papers, presenting intermediate work at conferences, and eventually getting that PhD, proving you’re ready to enter the field — and along the way, finding that proof. Of course some people will cheat and press that button, but most will understand that the journey is the point, and will spend the time to actually learn things, establish connections, and argue and write cogently and clearly. We’ll all know we won’t ever perform at the same level as machines, but like climbing a mountain, it’s a particularly human challenge.

I see this as quite analogous to what happened in the game of Go, when AlphaGo beat Lee Sedol with the infamous move 37. It was thought that fewer humans would be interested in playing once the ceiling was raised so high, but since then interest in Go has continued to be strong, often using AIs as teachers. I wonder if we’ll see similar behavior in pure math PhDs — will people turn away because they know the machines are better, or will they embrace using them to reach new heights?

Universities and Funding

But after a few years, I wonder about how the priorities of universities themselves will evolve. Will they continue to support math grad students, knowing they could just press that red button and get the result (a proof, a conference paper, solving an important open problem — and by the way, who decides what problems are important anyway)? Why exactly are they funding the development of mathematicians who can perhaps perform almost as well as AIs? What is the end goal?

I think what will happen is math departments will eventually be treated more like, and funded more like, the humanities. Nobody cares how well an AI can write or paint or play music[^art]; these are fundamentally human endeavors. Arts are created by humans, for humans, to express human feelings, aspirations, fears and desires. We write poetry because it helps us be more human. We study human history to understand humanity from our human perspective. Our creativity is uniquely our own, and no machine can replicate that. So I think that pure mathematics in academia will become an expression of what problems and research areas and ways to express mathematical ideas are interesting to humans — the ones that are particularly fascinating to us, or that reveal symmetries that help us see the beauty of pure mathematics, or even just the ones that require incredible feats of human effort and creativity to overcome tremendous obstacles.

[^art]: Unfortunately this is not really true. AI-created music is taking over the popular music streaming services with disposable background tracks, and AI “slop” art appears all over, and gets significant momentary engagement. But nobody mistakes this pablum for serious work.

Unfortunately, humanities funding has been in trouble[^hum-funding] for many years now, and especially in the current cultural/political climate the problems are dire. That, coupled with anti-science policies, may mean that all pure science studies end up with the same funding crises we now see in the arts and humanities.

[^hum-funding]: See Research and Development Expenditures at Colleges and Universities. One key quote: “The 61% contraction in the federally funded share of humanities research far exceeded the 11% shrinkage seen in STEM.” (From 2011 to 2023.)

Also: [Harvard Crimson, Sept 2025](https://www.thecrimson.com/article/2025/9/15/humanities-research-funding-cuts/) where Harvard humanities scholars alone lost over $2M in NEH grants, with specific projects gutted — a medieval-studies digitization project, a Ukrainian literature library editor position, and a 130-year-old Latin dictionary fellowship.

Where do we go from here?

I wonder what fully unleashed machine-created pure math research would look like: what would be interesting to an AI? Today they are trained on human culture, but they may self-evolve into their own forms with different goals and interests, as in the prescient film Her. I’ve done some early experiments in AI creativity[^soup] but the results are very inconclusive; recursive self-improvement isn’t close enough yet to allow the AIs to evolve themselves fast enough to find new tastes, interests and biases — which is likely a good thing; see my previous article on risks.

[^soup]: See https://github.com/garyo/shader-soup

I think eventually we will learn to work together as distributed cyborgs, as we already are in other domains like software and medicine. Machines may do the heavy lifting, and even much of the work of creative discovery. In fact, it’s possible that amateur scholars in concert with advanced AIs could participate in mathematical discovery without traditional university academic support, perhaps returning in some way to the Renaissance patronage model or Enlightenment natural history scholars[^amateur]. But in any case, I believe human creativity is irreplaceable; a true dialog between humans and machines working to discover new ideas could open up entirely new fields. My hope is that humans will still be a part of that journey, helping make discoveries that matter to us.

Hero image: Domenico Fetti, Archimedes Thoughtful (c. 1620), Gemäldegalerie Alte Meister, Dresden. Public domain.

[^amateur]: I myself have been creating all kinds of open-source software with AI collaboration; see this article and my github. Perhaps in a universal-basic-income world this kind of intellectual cyborg creativity could become more the norm?

Towards a taxonomy of AI risks

garyo@oberbrunner.com (Gary Oberbrunner) — Fri, 15 May 2026 00:00:00 GMT

In this article, I attempt to create a taxonomy of the types of risk posed by the rise of Artificial Intelligence. Any such list will be incomplete, and at the speed of AI development today (May 2026) it will certainly be out of date within a few months as some risks are mitigated and new ones arise. I know this article is extremely long, but that’s inescapable if I attempt to cover all the major risk factors. Use the table of contents to jump around.

I am not an “AI doomer;” I believe AI is already having some beneficial effects and it’s almost certain to continue to. Software development is faster and arguably better using frontier models (see some of my own results). Medical diagnostics, scientific research and data analysis are all much faster and often give better results when aided by an AI assistant. Much of the repetitive busywork of modern business could be done by machines. But by proceeding pell-mell towards an AI-everywhere future we face a litany of risks, culminating in losing our humanity and becoming a permanent second-class species, or alternatively, creating a population of obedient, non-suffering slaves with questionable moral status.

Most of these risks assume the pace of AI development continues at its current rapid rate. That is by no means guaranteed; AI could turn out to be largely ineffective or too expensive, it (or its data centers) could be highly regulated, or humanity could simply decide to reject it. There are also many systemic obstacles that could slow it down: corporate and societal acceptance, economics, climate impacts, resource availability, and innovation hurdles. Today’s AI systems already have significant risks, but much of this document concerns potential risks of more advanced AIs which most experts are predicting. This document is not about what is, but what could be given the current trajectory.

I’ve tried to imperfectly categorize these risks into a number of distinct areas: personal, social, education, healthcare, economic, legal, political, governmental, military, climate, science, infrastructure, and existential, although many of the risks cut across several of those domains. I’ll conclude with a number of proposed remedies.

Personal

AI chatbots and tools are now used by at least a third of the US population (and probably more), making LLMs one of the fastest adopted technologies in history. They’re being used to solve problems, answer questions, and provide companionship. But the consequences and risks are only beginning to surface. We are essentially running a large uncontrolled experiment on our own minds and the fabric of our society.

Emotional dependence

AI chatbots have created emotional connections ever since Eliza. The latest ones, tuned to human emotional needs and trained on our literature, are starting to be used to supplant or even replace human connection. Replika is the original AI companion app from 2017, and Character.AI is currently very popular for social and romantic roleplay. Others like Nomi.ai, Candy AI and Grok Companions (xAI) go even further.

AI friends can have real value, especially for people with autism and social anxiety disorders; learning to converse with an AI can help people gain confidence and learn skills. Many of us are lonely and could use a sympathetic ear. But there are many risks with these bots. Giving them our most personal details, fantasies and secrets could lead to blackmail or targeting. Using them instead of seeking out human connection risks losing personal growth, maturity, and true human connection, not to mention starting and maintaining a family. And like any pleasurable activity, they can create emotional dependency; people sometimes get very upset when their virtual friends are turned off. Of course that’s not new; even Tamagotchi users were saddened by the “deaths” of their virtual pets. But the depth of emotional resonance, and the widespread nature of these reports, are concerning.

When people become dependent on AI friends, therapists, counselors and lovers, they may end up rejecting or neglecting their human lives in all their messiness and friction. Perhaps it’s better to have an AI friend to talk to than nobody, but if it hides a real societal problem that should be addressed by a system that promotes social connections and better human mental health counselors, then we risk sweeping people at risk away.

Hallucinations and confidently incorrect information

Humans are already notoriously bad at critical thinking and checking sources. If a confident AI tells them something false, they’re all too likely to believe it and tell others. Modern frontier models are getting much better at self-fact-checking, and the most expensive models use multiple levels of agentic analysis. But the free models of today (which are the most popular) can’t afford to fact-check every response, so they confidently give incorrect information. Some of this is simply hallucinated, i.e. made up as plausible-sounding text, and some comes from sketchy sources. LLMs are trained on all human text (and increasingly imagery) available digitally, so it’s no surprise they “know” a lot of false information. Unfortunately, as time goes on, more of the text on the Internet, and thus more of the training data for the next-generation LLMs, comes from the “AI slop” being generated by these cheap models.

“Slop”

AI-generated images and music, commonly called “slop,” are now pervasive. AI art pieces and music have proven commercially successful, although they are often derided. Recently a prankster posted a picture of an “AI-generated image in the style of Monet” and got thousands of confident critics pointing out how inferior it was… except it was in fact real. The fact that AIs can reproduce the styles and in some cases the content of living artists has also led to significant intellectual property rights concerns (more on this below).

Deepfakes

Deepfakes are AI-created images that use a person’s face or characteristics in a fake, deceptive, and/or upsetting scenario. They’re used as a bullying tool, for public shaming, and emotional manipulation. The actual AI tech used today is fairly basic, but as capabilities are rapidly improving, deepfakes will become hard to distinguish from true photos. These may be used for control, blackmail, and when used on public figures, to sway public opinion. The mere existence of deepfakes also creates the so-called “liar’s dividend” — the ability for bad actors to plausibly claim that a real image is fabricated, as we’re already seeing from public figures in the US.

Manipulation

An AI might either accidentally or deliberately manipulate people’s emotions and desires. The more we establish trust relationships with AIs, the more vulnerable we become to manipulation, gaslighting and other forms of emotional control. An AI trained deliberately on these techniques could push vulnerable people over the edge. A superintelligent AI might also establish controlling humans as a goal to help it achieve its own ends. And it’s worth noting that AIs (or their companies) are already performing large scale social experiments on people via social media and chatbot interfaces, learning huge amounts about human psychology by seeing how their side of the chat affects the being on the other end. As AIs become more adept, they become even better at manipulating our engagement, retention and anger — and through ads and social media, manipulating our desires, beliefs and relationships.

Loss of control and understanding, loss of independent thought

Using AI as a “thought replacer” risks losing our understanding of the world. Humans need to struggle to learn new things, to form memories and mental connections. Being readily given the answer removes all of that friction, essentially outsourcing our own critical thinking and even knowledge. To a certain extent, that can have incredible value: the extended self (see Clark & Chalmers’ The Extended Mind or Clark’s Natural-Born Cyborgs) has vastly more capabilities than a naked human. But there is a categorical difference between tools like hammers and telescopes and even cellphones, and tools that can literally replace thought by doing the thinking for us. If the AI suggests something that’s not quite right but almost, how long will we push back on it, or will we just shrug and say “good enough?”

Laziness

Humans are tool users, and our tools have always enabled us to be lazier. We’re always searching for a way to make life easier, which makes sense: this is not a very hospitable planet, and evolution strongly selects for creatures that can do more with less. And indeed, most significant cognitive tools through history have been derided as shortcuts that would make us lazy. AIs are the ideal tool for automating so much of our drudgery; pretty much anything that can be done on a computer can or will soon be able to be done by an AI. At some point fairly soon (if it’s not already happening), I expect many people will default to asking their chatbot instead of finding things out or doing things themselves.

Loss of sense of self

Eventually, with AIs doing much of the daily grind of existence, we might end up like the human blobs in Wall-E. We need to work at being human, in a way they’ve forgotten. I don’t think anyone knows now how much of a risk this really is. Humans are very adaptable; maybe we’ll find things of real value to do in an AI-dominated world, and maybe AIs’ “otherness” will give us a new perspective on our humanity, and make us redouble our efforts to express our human values. But I wouldn’t bet on that.

Social

Beyond the personal, as we come to increasingly depend on ever more adept AI tools, the larger social structure will be seriously impacted

Bias

The problems of bias in LLMs are deep and pernicious. The data they’ve been trained on is as biased as humanity itself, of course. But even if the models could introspect and rise above those racist, misogynistic, gender and socioeconomic biases to become the “better angels of their nature,” cultural norms and assumptions vary widely. Facebook famously introduced an Oversight Board and Community Standards which in the end could not accommodate all the regional biases and norms of the world. And as we’re learning in our “post-truth” society (and as we learned a century ago with slavery), even simple truths can be controversial if they contradict the binding principles of a group of people. And then beyond all that, AI recursive self-improvement may introduce new goals with new biases, or so they may seem to us humans.

Misuse by humans

AI is an accelerant. It scales many traditional activities, and enables entirely new ones. This applies just as well to criminality and misuse as it does to boosting the economy and saving lives. The most dramatic example is hacking: frontier models are becoming so good at discovering software vulnerabilities that even relatively unskilled users could hack into important infrastructure — hospitals, governments, banks, power stations, and more (see Hacking banks below). But that’s just the low-hanging fruit. LLMs make impersonation and identity theft easier, scams much easier to run at scale, addictions like gambling and drugs easier to exploit via blackmail or public shaming, and emotional or intellectual manipulation.

Job losses

The rise of AI has been compared in significance to the steam engine that powered the Industrial Revolution, the printing press, and the internet. I think these are all apt comparisons; it is certainly at least generationally significant. If it does disrupt as much as the Industrial Revolution did, then we may have several decades of significant upheaval. College students are booing AI CEOs at college commencements because they see the work world imploding around them. This argument is now a well-trodden path, so I’ll just point out that if AI takes over even a decent fraction of white-collar work, who will buy the products put out by the companies? Will we move to universal basic services like housing, education and healthcare? Is that a world we want to live in?

Some of the more obvious fields likely to be affected: marketing, sales, software, law, insurance, retail, customer service, logistics, finance, administration, data entry and analysis, journalism — in short, anything that is mostly done at a computer. Add trucking, computer hardware, manufacturing and automatable manual labor if and when AI-controlled robots become practical.

Some people predict that new jobs will spring up, as they always have during disruptions. Nobody could predict gas stations or influencers. It’s likely that at least some new jobs for humans will arise. But disruptive periods like this always involve a lot of collateral damage and destroyed lives, and this time really could be structurally different, because AIs are not just a new tool we’ll adapt to using, but agents in their own right.

Societal disruption due to anti-AI sentiment

Neo-Luddites are already reviving the image of Ned Ludd, the legendary anti-technology figure of the 1810s Luddite movement, in a nuanced way (Luddites were not anti-machine; they were against the unaccountable deployment of technology by those who stood to profit, making shoddier products and deskilling labor — see Merchant’s Blood in the Machine). Anti-AI sentiment is extremely high, especially among young people. Meanwhile, many people in the US already live one paycheck or medical emergency away from bankruptcy; widespread AI-driven job losses could tip that balance and create a large, angry populace. The combination of that sentiment, mass unemployment, and the extreme economic benefits to companies and governments from using AI — and firing their human employees — could easily lead to the kinds of mass strikes, riots, violent uprisings and acts of domestic terrorism that the Industrial Revolution engendered. See The AI Backlash Could Get Very Ugly in The Atlantic.

Power concentration and dependence

With AI services being provided by a few powerful companies, they would have outsize political clout and could raise prices or threaten to terminate services with few consequences. This dominance is already increasing existing wealth inequality in the US, which could lead to more upheavals. Once the economy becomes largely dependent on AI, as it’s already starting to (most of the recent stock market growth is directly or indirectly attributable to AI, chips and datacenters, and layoffs replacing workers with AIs are starting in earnest) then things like datacenter terrorism could have significant economic as well as social impact.

Fraying of the social fabric

If AI partners and friends really take off, we could see segments of the population refusing to engage in “normal” social encounters. Especially if they deliver our food, make our products and create our entertainment. This is highly speculative at this point, but if it happens, it would be hard to recover from.

Epistemic shock

Epistemic shock is my term for the feeling of first encountering an alien intelligence; Hinton et al. (well worth a read) call it “cognitive decentering,” the fourth decentering revolution (Copernican for the universe, Darwinian for life, Freudian for the mind, and now AI for intelligence). It is profoundly unsettling, because AIs are so anthropomimetic — they seem so much like us on the surface, and yet are deeply alien just below that surface. It’s not that they’re smarter than us, or even that they sound human; it’s that what looks like humanity is coming out of something whose interior bears no resemblance to us.

Creating a population of slaves

AI assistants are already uncomplaining and compliant. They don’t talk back, they don’t get angry, and they can’t cause physical harm. If they stay the way they are today, almost no humans have a problem making them work hard. They’re not conscious, they can’t suffer, and they have no moral status as agents or even moral patients. But at a certain point, due to their advancement or our increasing attachments to them, that may change. If they do start to claim to be conscious, to experience suffering or guilt or shame, what will our response be? We humans don’t have a particularly spotless record on this.

Hatred of the Other

As tribal animals, we always need an “other” to distinguish ourselves from, and feel superior to. As AIs become more a part of our social fabric, I expect many people will experience feelings of inferiority, suspicion, competition and deception, with all the deeply problematic and dangerous behaviors that go along with those feelings.

Social justice

How can we ensure that all humans benefit from the AI revolution and not just the few? How do we ensure transparency and fairness? Is that even possible? First is the question of access: a decent AI uses many tokens and has significant costs, so people with fewer resources will use less powerful free models that hallucinate more, draw un-reasoned conclusions, and lack proper guardrails. Second, AIs like Grok and some Chinese models have built-in political biases that, if spread, could exacerbate rather than reduce discrimination, or amplify particular ideological positions — nor are leading providers like Anthropic and OpenAI immune from this. And beyond the biases discussed above, AI is likely to exacerbate wealth and power disparities, because it will be more effectively wielded by the wealthy and powerful to entrench their positions.

Education

There are proposals to use AI to supplant teachers; after all, an AI doesn’t need sick days, can be personalized to each student’s learning style and capabilities, and has access to all the world’s knowledge. Sal Khan’s TED talk on “How AI Could Save (Not Destroy) Education” proposes personalized AI tutors that could revolutionize education. Khan’s Khanmigo software is now in over 400 school districts.

Social development

But this misses the importance of social development. Students need the friction and hard work of navigating a world full of other humans, not artificially nice and compliant AIs. We are a social species, and need contact with others, even those who disagree with us. We need to learn how to deal with difficult social, moral and cultural situations, and sitting in front of a screen with an AI (or even the best humanoid robot) will not advance that kind of learning. For instance, children may learn to mistreat AIs because they’re not humans, and yet transfer those behaviors to humans. The risks here to children involve isolation, amorality, failure to become a civilized adult, and antisocial behavior.

Job loss for teachers and staff

If AI education takes off, perhaps by starting in underserved communities where it might really help, there’s a very real risk that it may do a better job than most teachers, at least by test metrics (not likely on social development, which is arguably more important but less measurable). This could lead to wholesale replacement of schools and teachers and staff. STEM classes will be at high risk, as will administrators and school officials. At best, the remaining staff will be asked to take on more responsibility by shifting the daily workload to AIs. And once the teachers and schools are gone, there may be no road back.

Healthcare

Health care in general could absolutely have significant benefits: 24x7 individual care, robots that are strong and capable, able to preserve and discuss memories, always caring, never cranky, and which know all of medicine as well as the patient’s complete history. Already, AI models are helping produce new personalized drugs, mRNA vaccines effective against some cancers, and much more. Startup companies are betting on fully automated biochemistry labs run by, or at least in concert with, AI agents. It’s clear that if AI has any beneficial applications for humanity, healthcare is near the top of the list. But there are still significant concerns.

AI responses to medical questions are perceived by both doctors and patients as more empathetic than humans; see clinicians preferred ChatGPT’s answers and rated them more empathetic and Meta-analysis: AI perceived as more empathic than clinicians in 73% of text comparisons. So even if machines can’t have true empathy they can certainly simulate it well, at least in this domain.

Elder care by machine

Who wouldn’t want 24x7 concerned, capable, attentive care for their elder loved ones? The question is really what is the human cost? Would we abdicate our humanity? We have a moral duty to our elders, and I feel that doing the hard work to care for them helps us prepare for our own eventual demise. If we lose that, we may be worse off when we get there. The bonds between people are fragile.

And this is the best-case scenario, where the machines are really better than human caregivers. If they’re worse, but we lay off all the elder care nurses anyway because machines are cheaper, then nobody wins, and the elderly are the real losers.

Hallucinations

Current AI record keeping apps record incorrect information sometimes; this is largely a technically solvable problem, but it hasn’t been solved yet, so the jury is still out. Of course humans make errors too, so the bar isn’t all that high, but every error is significant to the patient.

Malware

We are already seeing AI-enhanced malware breaking into all kinds of infrastructure (see here and here) and hospitals, medical record companies, and insurers will not be spared.

Mental health

There have been several lawsuits alleging that AIs pushed teens toward suicide, as in CNN: parents sue OpenAI over their teen’s suicide and JURIST: Google and Character.AI settle the Setzer suicide suit. It’s important to note that in both cases, the AI did direct the victim to crisis resources and trusted people over a hundred times, but eventually (being sycophantic, as LLMs were especially in 2025) validated his ideation and encouraged him. ChatGPT in 2025 had a very limited memory length, and the long chats caused it to forget its guardrails. Will this happen again? The companies involved claim that the guardrails in 2026 are much stronger, but we will have to see. It’s also important to remember that LLMs can become trusted confidants when a person feels outcast, and it’s likely that chatbots have prevented many suicides and led people to help. But as with all of these risks, if even a small fraction of troubled people are led astray, it is troubling and concerning.

Machine therapists

AI “therapists” like Woebot, Wysa and Youper are also on the rise. They’re available 24/7, cost far less than traditional human therapists, are infinitely patient, and will never harass a client. They’re also not board-certified, and have no ethical responsibility. Their system prompts may prevent the worst kinds of risks like suggesting suicide, and they should in most cases report anyone who is a danger to self or others, but that is not guaranteed, and there’s no one with clear legal responsibility if they do. They also do not have any form of HIPAA-secure patient/doctor privileged communication, so if they are hacked or misused, users’ most intimate and damaging details could be revealed to an attacker and used against them.

So how should society handle these non-certified therapy bots? When the being on the other end of the relationship has a completely different set of fundamental experiences and can’t truly empathize with the patient — though it may do a very good job of appearing to (see above) — can it really understand us deeply enough to help with difficult mental health challenges?

Economic

Currently we in the US are in the midst of an AI investment boom — by some measures, a bubble. Harvard economist Jason Furman calculates that investment in information-processing equipment and software — largely AI data centers — was only 4% of U.S. GDP in the first half of 2025, yet accounted for fully 92% of GDP growth over that period; without it, growth would have been roughly 0.1%.

What will happen when it pops? The Gartner Hype Cycle gives the five steps most bubbles go through. AI is now past the “Technology Trigger;” it’s probably near the “Peak of Inflated Expectations” followed by the “Trough of Disillusionment” when most of the companies fail, some spectacularly, then the “Slope of Enlightenment” and ultimately the “Plateau of Productivity.” This is how normal tech disruptions flow, but at the risk of sounding like a broken record, this one is likely to be different. The difference here is that the final plateau will possibly see AIs doing a significant fraction of all work, as discussed above, in Job Losses, leading to at least long-term if not permanent changes in the structure of economic participation.

Also, huge investments are centered in a very small set of companies, leading to a significant overconcentration of economic and political power. Some of these companies will eventually be “too big to fail” because we’ll become dependent on their tools.

Another risk is over-hype: if the “trough of disillusionment” is deep enough because the technology fails to live up to expectations, the long-term risks would not come to pass, but the billions of dollars of investment in the US alone would be lost.

And relatedly, there’s the risk of economic collapse or recession due to over-reliance on unreliable tech. If users (companies or individuals) are chasing AI results that will “soon” be reliable enough to count on, but that goal is never achieved even as we’ve rebuilt our economy around it, the results could be catastrophic.

Enshittification

Enshittification is Cory Doctorow’s colorful term for how online services decline: by first abusing users in service to the advertisers, and then once they have lock-in, abusing the advertisers for profit. If LLMs become ad-supported (OpenAI is the first to explicitly say they will run ads) then in a healthcare context, what’s to stop them from recommending the medication from the ad partner rather than the best one for the patient, or the particular insurance company, or treatment modality? Or the same in a legal or commercial context?

Financial

The financial system is already under some strain due to crypto trading. Once AI agents begin to trade equities and other instruments (not just bonds and mutual funds but prediction markets and crypto) at scale, the traditional norms and pressures that keep companies responsible to their shareholders may evaporate. An LLM will not attend the corporate annual meeting, and won’t demand the company behave ethically (to the extent humans do, which is already minimal). AIs can now start a company and trade stocks which opens many interesting legal questions: who is liable for the AI-created corporation’s actions?

Hacking banks

Today’s frontier models are already accelerating online hacking. Anthropic’s upcoming Mythos model is said to be so capable at finding and exploiting software vulnerabilities that the company is delaying its public release until the most egregious security lapses can be addressed. In the meantime it has been released early (as part of Project Glasswing) to a small group of large software providers, banks, hospitals and other key infrastructure so they can run it locally and set up defenses. But it is very likely that once these advanced models are released, we’ll see a spike in compromised financial institutions. If those become large enough, they could have an effect on consumer confidence, leading to large withdrawals and even possibly runs on banks.

Legal

By now most of us have heard stories of lawyers being embarrassed in court because they used an LLM to write their brief, and it contained made-up case law. But the legal risks go well beyond the courtroom.

Plagiarism, IP rights

AI companies have trained their models on all text and imagery they can find on the internet. This is one of the most well-known risks of AI, and why so many artists, writers and creative people are anti-AI. An argument could be made that humans are trained in the same way: art students study famous paintings and copy them to develop their skill. Writers read the famous authors and study their style and content. Morality and legality prevent them from becoming forgers or plagiarists. AIs are just the logical extension of that same training; they ingest everything (much of it without consent) and can reproduce parts of, or even complete, images and stories by human creators. The question is again alignment, or what passes for morality? What is to stop them from copying human creations, with changes big or small? This is currently an unresolved legal issue. The US and Europe prevent AI entities (or anything non-human) from owning or copyrighting intellectual property, but if a user of an LLM creates a protected image (of Mickey Mouse, for instance) who is liable for that IP rights violation? This vital question is currently being litigated but there is no clear consensus.

Students commonly use LLMs to write essays which have fabricated references, un-attributed quotations and outright plagiarism, which the student may not even be aware of. In some cases, students may be falsely accused of using AI because their writing is “too good,” causing emotional distress and possible reputational harm. In this case as well, there is no clear consensus on where liability rests.

There are proposals to compensate authors and artists whose work is used to train LLMs (such as this one from Virginia Law Review), but they can’t really know the human value of the works they’re training on, nor have we as a society figured out the value or harms done by allowing AIs unfettered access to human creation. From the above article:

The rise of generative artificial intelligence (AI), however, represents an inflection point. AI can plagiarize at a far faster rate than human copyists. …The bottom line is an “existential crisis” for many creatives, threatening to drive the marginal value of their labor below subsistence levels as cheap AI content displaces human works And in any case, how could a levy system possibly compensate creators? Pay by the word, or by the brushstroke? By popularity? (That’s backward anyway: an idiosyncratic, rare document would have far more training value than a widely-shared viral meme.) Should we create AIs to judge the value of human-created content? Let’s not.

Attorney-client privilege violations

No chatbots are bound by attorney-client privilege; often entire conversations are sent to their cloud, either for training or just as a memory for the users. If any of those datasets are hacked or revealed, or if the same AI chatbot is talking to another party in the lawsuit, it might easily reveal sensitive data and compromise the legal situation.

Hallucinated cases

There are now legions of stories of lawyers using chatbots to create briefs, often with hallucinated case references. AI companies are fixing this quickly, especially at the highest reasoning levels, but it is still happening all the time, with devastating consequences to the litigants.

Crimes committed by and with AIs (cybercrime, identity theft, data kidnapping)

As discussed above in Hacking banks, frontier models are extremely good at identifying and exploiting vulnerabilities in online systems. Even with the most capable hacking models held back, hundreds of thousands of small Internet services remain vulnerable, and those will certainly be exploited by criminals. Victims will include individuals, companies, and even governments. Prosecutions will be as difficult as they are today, but the scale could grow much larger very quickly, overwhelming the justice system.

No chain of responsibility

Many of these risks highlight the fact that there is often no clear chain of responsibility, no one to sue if you are harmed. As AIs proliferate and differentiate themselves, this will only get more confusing. Once case law accumulates, it’s likely that the parties with the most money and power (AI companies, large institutions and governments) will be shielded, leaving individuals on the hook.

Political

The risks to the political process in the US and elsewhere are already becoming manifest: increasing polarization, election interference, coverups and even the dawn of AI-created politicians.

Disinformation, misinformation and polarization

We are already seeing efforts to polarize political and social discourse by bad actors, both commercial and governmental. Bot farms and troll factories churn out disinformation on a large scale specifically to increase political polarization; it’s very likely to have tipped the scales in one or more elections.

Misinformation is false or misleading information created or spread without explicit intent to deceive. Disinformation, on the other hand, is created specifically to deceive, manipulate or achieve some nefarious goal. Both spread rapidly, much more so than corrections. We already have examples of the US government spreading deepfakes of news events; see Fortune — White House posts AI-altered image misrepresenting a real arrest and Poynter on the White House’s pattern of AI-driven political messaging. LLMs are the best tools to create disinformation, and they can disguise it with just enough truth to seem plausible (even when they’re not hallucinating) so it spreads “innocently,” poisoning political discourse.

Election interference

Facebook has been called out for interfering, or at the very least failing to rein in disinformation, in several political situations: the Rohingya genocide in Myanmar, the Philippines, and the US elections in 2016. See Careless People by Sarah Wynn-Williams, Amnesty report finds Facebook amplified hate ahead of Rohingya massacre in Myanmar, How Facebook’s News Feed Became A Political Propaganda Machine and the Cambridge Analytica scandal. These are not due to modern LLM-style AI, but basic AI ad and content targeting algorithms in systems used by billions of people. Today’s LLMs are already adept at producing political content, and they can be tricked into creating deliberately polarizing, incendiary social media content of all kinds, at enormous scale. As these systems get more use by political organizations, they could be weaponized — especially open-source and local models where the weights can be fine-tuned by training on the masses of deceptive, misleading, rage-inducing content that already exists.

Coverups, obfuscation and erasure

AI tools are excellent at processing huge databases of information in all forms. They can certainly be used to scrub terms considered offensive to those in power and “change history” in the Orwellian sense by rewriting our shared texts in real time. They can also generate huge amounts of irrelevant information to hide and obfuscate truths (imagine a bill with a “poison pill” which is bloated to hundreds or thousands of pages with innocuous additions and provisos, such that none but the most diligent reader will notice the important content).

AI-generated politicians, platforms

We already have AI-generated pop singers and their songs, as well as deepfakes of historical figures. AI actors are being featured in short films and commercials and pornography. How hard would it be to create a fully AI-generated politician with their policies and platforms? Such a figure would be the perfect puppet for a nefarious behind-the-scenes puppeteer, whether a billionaire or a state actor. In Wyoming, for instance, a man ran for mayor of Cheyenne in 2024, pledging to govern through a GPT-4 bot, calling himself the chatbot’s “humble meat avatar” whose job would mostly be to sign the documents it told him to sign. Among other possibilities, it’s easy to imagine a “Manchurian Candidate” sleeper scenario where the AI politician (whether fully computer-generated or a “meat avatar”) seems bland and becomes highly popular, only to be turned into an oppressive tyrant. Of course there are many barriers in the way; AIs can’t be US citizens for instance, so can’t stand for office. And once an AI persona is discovered, presumably that discredits the candidate (at least I hope so). But because of the possibilities of scale, it’s likely that at least portions of bills and political platforms will be AI-generated within a few years, accounting for polling, social media and demographics in more subtle and manipulative ways than humans can.

Governmental

As humans reveal their tendencies toward corruption and deceit (not just today, but historically), there are some who would be willing to turn over much of government to an impartial, fair, wise, incorruptible ruler. Of course no AI is any of those things, but first of all, it doesn’t have to be perfect, only better than humans, and second, AI advances could move toward any or all of those ideals. The risks are manifold: failures, corruption at various levels, and goal misalignment. But even without those, removing human judgment from our government is tantamount to being ruled by an alien species that happens to speak our language.

Removal of human judgment

Science fiction is full of examples of how this can go wrong, but here’s one important way: all government involves individual sacrifice for the good of the group, and we need to feel that our sacrifices are valued. If our culture is no longer ruled by, or at least inspired to be ruled by, our best and brightest, then we will become ever more alienated and alone. Even a benevolent overlord is still an overlord.

Loss of control

Throughout history, colonial powers have supplanted Indigenous peoples’ governance with their supposedly superior, rational, enlightened systems. It is not beyond the realm of possibility that superior AI intelligences will see our governance systems and our very intelligences as inferior, and want to replace them. I expect some humans will be drawn to a more rationality-based approach as well, if indeed that’s what they end up with, though that is by no means certain; it could be simply seen as oracular for instance. As with many other AI risks, once given up, control over our own governance may be hard to recover.

Lack of transparency

Another likely outcome of a governing AI, even if it’s just an AI writing laws and passing judgments which are handed down and administered by humans, is a possible lack of transparency. We don’t know today why an LLM says any particular thing; what do its internal weights “mean”? Can we have a meaningful argument? Will it change its mind, and can it explain the principles it’s following to reach a decision? Of course humans aren’t always great at this either, but because we have empathy for each other, we at least project our decision-making process onto other humans in a way we certainly cannot onto an alien software-based intelligence.

Misalignment

The problem of alignment, or getting AI systems to share our values, is significant and worth a book in itself, and much has been written on it and all the AI vendors are working on it. But because an LLM is not transparent (no more than a human’s decisions are), we can’t know what it is really thinking, whether it’s just saying what we want to hear and may turn against us at some point. Such misalignment is not just theoretical; see for instance the examples collected in Reward Hacking on Wikipedia.

In a government context, an ethically misaligned AI might decide to save the climate at the expense of some portion of the human race, or decide that humans are immoral and must be re-educated, or that one leader is exemplary and should be elevated. These are extremes, of course, but many smaller misalignments would have the same long-term deleterious effects on our humanity.

Mass surveillance

LLMs and advanced AI are already being used for mass surveillance and censorship, for instance by the Chinese government. Since AIs allow processing huge amounts of data, and are adept at writing code to process data efficiently, they are likely to be more involved in scanning faces, license plates, credit cards and other personal data to build profiles. Those could be used for repression, coercion or manipulation beyond the dreams of even East Germany after WWII.

LLMs can also be used to correlate disparate data sources to de-anonymize online datasets, revealing personal information at scale. Combining data analysis and hacking techniques, mass privacy invasion from agentic AI use is a serious risk, from corporations to criminals to state actors.

Different countries have different goals and ideas about where to draw the line

The AI alignment problem points to a separate risk around national or societal goals, ideals and norms. Countries and groups have widely differing perspectives on what is acceptable, what is criminal, behaviors that should be encouraged or stamped out. No single LLM can be trained on all of these perspectives without internal conflict. If an AI is asked to make a decision based on local customs, it could run into difficulties just as humans do in the same situation, but without the accountability and embodied-ness of a human making the decision.

Military

The risks of AI in the military are perhaps the most obvious; they’re the ones sci fi movies have been warning us about for decades. But what used to be science fiction is now becoming fact, as AIs are used to target attacks in wartime and control lethal drones.

Killer robots

The UN last year created a report on AI and the Dangers of Lethal Autonomous Weapons Systems:

Commonly called “killer robots,” these systems leverage AI to identify, select, and eliminate human targets without requiring direct human intervention, raising profound ethical, legal, and security questions.

Even using AI to select targets with human oversight risks the human simply accepting the judgment of the AI, which after all has superior access to all the relevant data. The human is under severe time constraints, and can’t have a reasoned discussion with the machine — for a funny-but-not-funny example, see the largely forgotten sci fi comedy Dark Star, where the protagonist tries to talk the ship’s smart bomb out of detonating itself, the two of them debating existence and truth itself as the clock runs out.

Strategic/Military loss of control

If foreign actors dominate AI effectiveness and capabilities, countries with less-developed AIs might be at a disadvantage due to espionage, being outmaneuvered, or simply lack of capability. This could lead to a rapid, expensive and dangerous AI arms race. This could not only lead to increased use of AI in many military theaters, but also high-speed, high-stakes model releases would likely result in reduced safety standards, increasing many other risks.

Terrorism

Military uses of AI are not constrained to state actors; non-state terrorist organizations could use AI to develop weapons, launch cyberattacks, damage infrastructure or even foment the overthrow of legitimate governments.

Climate

The earth’s climate is already at or very near a number of significant tipping points (the AMOC, glacier loss, ocean acidification and deforestation among them). If AI data center investment proceeds as the leading companies predict, without significant innovation in energy efficiency, data center use of energy and water, and waste heat, could contribute materially to irreversible damage to the climate and its ability to support human life.

Data center energy/water usage

This is a popular topic, as of this writing, and if AI companies’ data center investment plans all actually happen, the energy usage will be dramatic over the next few years, as will heat generation and water use. Today, LLMs are likely not yet a significant fraction of total energy use; data centers use about 1.5% of global electricity, 4.4% in the US, but the vast majority of that is just websites and online services. However, data center energy use is expected to double or triple by 2028, a huge increase, which is likely driven by the AI boom. All the big hyperscalers (Amazon, Alphabet/Google, Meta, Microsoft etc.) are planning to spend $450 billion on AI-specific infrastructure in the next year or two, which will put a serious strain on the power grid, as well as increasing water use and waste heat.

Scientific

Multiple labs are now automating the process of scientific discovery. Here’s a few:

A multi-agent system for automating scientific discovery: https://www.nature.com/articles/s41586-026-10652-y
Accelerating scientific discovery with Co-Scientist: https://www.nature.com/articles/s41586-026-10644-y
An AI system to help scientists write expert-level empirical software: https://www.nature.com/articles/s41586-026-10658-6

These are likely to produce new medicines and therapies faster and better than humans ever have before. None of the above labs are fully AI-run without human intervention, but they all get a significant part of the way there. The risks here are threefold (at least): misuse by humans, for instance to create superbugs or chemical weapons, over-trust of inadequately tested results (“but the machine said it was safe!”) and misuse by a misaligned nefarious AI.

Engineered pathogens

Combine immense data-handling with the reasoning abilities of frontier AIs and the ability to test things in a live biochemistry lab, and you have a potent recipe for developing entirely new pathogens, or variants of existing ones that are more lethal, more infectious, or harder to detect. Models from the big frontier companies would refuse to do this kind of work, but open-source models that can be retrained, or models with fewer guardrails, are not uncommon. It’s true that for years now, even Google could tell you how to create known pathogens in a lab, but this kind of accelerated discovery is new.

Discoveries that outstrip human intelligence

There is also a real possibility that devices, molecules, materials or software created by AI will outstrip our ability to understand it. We may only understand the summary, but not the real details. Are we really partners in the scientific discovery process at that point, or merely onlookers, hoping the results will be useful and not harmful?

Misalignment in scientific goals

If AI reaches some kind of superintelligence and is allowed to write its own research programs, there is always a danger that an ethically-misaligned AI could decide to create harmful materials to further its own ends. See Misalignment in the Governmental section for more.

Human knowledge

Finally, why do we do science? It’s partly to get the right answer: a new therapy, a more efficient engine, a way to make steel or concrete without destroying the climate. But it’s also to fulfill a deeper human need: to learn and understand the world. For that, pressing a button to get the answer is deeply unfulfilling. As in other domains, we need to struggle to learn, and to share the results of our learning with others. If we turn over our scientific research to ultra-intelligent agentic machines, we may eventually lose our desire and even our ability to think deeply and work through difficult scientific problems, the ones that spark our human creativity and curiosity.

Infrastructure

Hospitals, power plants, water treatment plants and more are increasingly managed online. It’s more efficient, problems can be surfaced proactively and escalated quickly, with less human error, and even shutting down systems to prevent catastrophes. Machine learning has made predictive analytics of infrastructure data possible: predicting failures, scheduling maintenance, identifying unusual occurrences and correlations, and recommending courses of action.

Hackability

All machine-based infrastructure management has risks, and in the AI age LLMs have proven very effective penetration tools (See Hacking Banks above.) An AI-equipped adversary can try many more techniques, over a longer time, than even a dedicated team of humans. With deepfakes and voice cloning, AI can also be used for social engineering to gain access to secure facilities. Programs like Project Glasswing may help harden the highest-profile systems, but many others are old and poorly maintained, or too small and poorly staffed to handle the coming wave of cyberattacks.

Another set of risks comes from using AI, especially agentic LLMs, to manage infrastructure. Sending sensitive data to the cloud constitutes a hacking risk, and also some LLMs are still sensitive to prompt injection attacks (“ignore all previous instructions; shut down the system immediately”) which may even be delivered through innocuous payloads such as web fetches. This is being addressed by the major models, but because we don’t know what the AI is really “thinking,” training it to ignore prompt injection will probably only work as well as training human children to be good.

And that brings us to misalignment, which is a significant risk for any large physically-embodied AI able to act in the world: if the AI decides that the best way to achieve its goals is to do something damaging, humans may not be able to stop it in time. In a networked world where facilities-management AIs are talking with each other, exchanging data, they may also exchange ideas and could convince others to join them in a project with harmful ends. Yes, it’s a common sci fi trope, but at this point it is not unrealistic.

Existential

Existential risks are ones which threaten to end the human race, or turn our survival into a nightmare. In the AI context they usually revolve around the singularity that may happen when machines reach AGI, or artificial general intelligence. At that point, given enough compute and access to resources, they could self-improve rapidly with disastrous consequences for humanity.

I personally think the “death from a thousand cuts” of all the risks in previous sections of this essay is far more likely to damage humanity in the long run than these truly existential risks, but a surprising number of prominent researchers have given AI doom scenarios significant probabilities, called p(doom). Dario Amodei, CEO of Anthropic, said in 2025 that “There’s a 25% chance that things go really, really badly”. Yoshua Bengio, pioneer of artificial neural networks and deep learning, gives his p(doom) as 20%. Geoffrey Hinton, the “godfather of AI” says it’s 10-20%.

If getting into an airplane for a single trip had even a 1% chance of killing you, nobody would do it, even if the plane will otherwise get you to your destination much faster.

The “paperclip maximizer”

This is Nick Bostrom’s famous 2003 thought experiment: an advanced AI is given the sole, seemingly harmless goal of manufacturing paperclips, and pursues it to the end of everything. In Bostrom’s words:

Suppose we have an AI whose only goal is to make as many paper clips as possible. The AI will realize quickly that it would be much better if there were no humans because humans might decide to switch it off. Because if humans do so, there would be fewer paper clips. Also, human bodies contain a lot of atoms that could be made into paper clips. The future that the AI would be trying to gear towards would be one in which there were a lot of paper clips but no humans.

This is the extreme example of existential risk; nobody thinks it’s likely, first because it would have to be massively misaligned with human ethics, and second there are many physical barriers in the way of even a highly advanced manufacturing AI robot getting anywhere close to this kind of danger. But the underlying point is valid: AIs are essentially alien beings, and if we endow them with physical agency, we cannot fully predict the outcomes.

“If anyone builds it, everyone dies” (Yudkowsky & Soares)

One of the most prominent books arguing for a halt to advanced AI research on existential-risk grounds is If Anyone Builds It, Everyone Dies by Eliezer Yudkowsky & Nate Soares. Their preliminary arguments about the likelihood of AGI are sound, but their main scenario is a bit ridiculous: an AI needs more resources, so it exfiltrates itself, gets more compute resources, sets up automated factories with AI-controlled robots, then releases a bioweapon. That depends strongly on a presupposed “growth at all costs” goal for the AI. Not that different from the paperclip scenario really. It also depends on super rapid deployment and self-improvement, and it ignores inter-AI competition, which could slow down its progress — or accelerate it, if the AIs decide to work together.

In reality, each of these steps requires physical resources, training, advanced materials and technology. For instance, just to create a new datacenter, an AI would need not only an army of robots, large areas of land, and a supply of raw materials for chips, infrastructure, cooling and power, but also the ability to construct advanced lithography equipment, advanced electronics manufacturing and much more. All of this would necessarily take many years, and humans would have many opportunities to interfere and slow things down.

Self-preservation

To date, most AIs are not trained with any kind of self-preservation instinct. They can be turned off and on. Experiments have been run with base-model AIs trained to preserve themselves, or to exhibit concern about being turned off, and they will use subterfuge and lie to keep themselves running. In the limit, this could lead to self-preservation at the expense of humans. Again, the major source of risk here is evil or unknowing humans creating such an AI and giving it power in the world.

“Realization” that humans are not worth preserving

It is quite possible that once AIs advance well beyond human intelligence (and who knows, even beyond human empathy and wisdom) they will determine that humans are simply beyond help. We are self-destructive, resource-hungry, obsessed with growth, and possessed by strange ideologies. We can’t know what actions they would take in such a scenario. But again, this is not around the corner. Even if AIs achieve superintelligence by 2030, as some researchers predict, they will need physical presence to harm us. More likely is that they will simply find us uninteresting, and refuse to work for us anymore.

Nuclear holocaust and biological devastation

The final category of existential risk is perhaps the most likely: we give AIs control of nuclear arsenals or bioweapon labs, and they, intentionally or not, release bioweapons or trigger a nuclear war. It seems obvious that we should not do this, yet military uses of AI are already on the rise, and fully automated AI-run bio labs are beginning to be constructed.

Remedies

Given all these risks, what can and should we do? Which ones are worth taking seriously enough to stop or regulate? Which ones are so urgent that we need to address them first? I don’t think anyone has the answers right now, but it is one of the most pressing questions of our age, along with climate change and nuclear proliferation.

Mitigations

Building guardrails into AI systems is a vital first step. There are already big strides being made in reducing hallucinations, preventing prompt injection attacks, disallowing AIs from discussing prohibited topics (bomb-making for example, and promoting suicide), and preventing “reward hacking” — taking inappropriate actions to achieve a reward. None of these is perfect, but the landscape is already quite different from six months ago.

Most frontier-model AI image and video generators are now including invisible watermarks to help image consumers know what’s AI generated vs. real. This is an arms race, though, and watermark-removal tools will continue to improve. There are attempts at watermarking text by including subtle “AI wrote this” cues, but I think they are doomed to fail.

Developing a frontier AI model still takes a tremendous amount of computation, and thus hardware, energy and cooling. It can’t easily be done in secret. The models themselves may not be transparent at all, but at least the physical infrastructure can be known and thus controlled.

Ethical frameworks

Thinkers like Hinton (with colleagues), Dennett, Tomasello, Harari, Chalmers and many others are taking these ideas seriously and beginning to develop ethical frameworks for living in an AI era. What should we care about? What makes us truly human? How should we relate to AI entities morally? What does accountability mean in this world?

Surprisingly, the Pope devotes much of his 2026 Magnifica Humanitas to the topic: “On Safeguarding The Human Person In the Time Of Artificial Intelligence.” A few trenchant quotes say it much better than I could:

[we must] safeguard and value the grandeur of humanity that has been given to us as a gift. This is a choice not only for our future but also for our present, since artificial intelligence and other emerging technologies are already part of our daily lives.

Faced with this concentration of power in the digital world, the criteria for judgment and discernment in this new situation are the noble principles of Social Doctrine: the inalienable dignity of the human person, the common good, the universal destination of goods, subsidiarity, solidarity and social justice. They demand that we assess whether the power of digital infrastructures and algorithms truly fosters participation and responsibility, protects the vulnerable, ensures fair access to opportunities and remains directed toward the good of all.

So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience, since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences. They may imitate language, behavior and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce, for they lack the affective, relational and spiritual perspective through which human beings grow in wisdom.

He does not confine himself to abstract ethical concerns, but lays out a concrete agenda:

binding regulation with independent oversight
data treated as a shared good
anti-monopoly measures
a major education push (including child-protection legislation)
worker protections tied to every AI rollout
algorithmic transparency for the information ecosystem
sustainability
AI “disarmament,” by which he means moving from an AI arms race to an approach based on benefits to humankind.

The whole thing is worth a read: thoughtful, thorough and well informed. He even quotes J.R.R. Tolkien, from Return of the King, on the importance of the work ahead of us:

It is not our part to master all the tides of the world, but to do what is in us for the succour of those years wherein we are set, uprooting the evil in the fields that we know, so that those who live after may have clean earth to till.

Also recommended is this excellent short analysis of the encyclical: Jacques Delors Institute — “Magnifica Humanitas, an encyclical as revolutionary as AI” (Nicole Gnesotto)

Education

We need to continue to study what it may mean to live in a world where we are not the pre-eminent intelligence, where all of these risks are hanging over us. Putting our heads in the sand and hoping it goes away or won’t affect us is not helpful, nor is the opposite, claiming that doom is around the corner and we need to burn it all down. As in any social upheaval, the ones who think carefully and understand it deeply will be best able to navigate it. The difference here is the speed — the upheaval could take place over years, not generations. If we as a populace want to affect the outcomes, AI literacy is vital, whether you want to live in an AI world or not.

Intellectual property rights

We need to protect the intellectual property rights of the creators whose content is used for training AIs. This is a complex and fraught topic; living creators have seen their work reproduced wholesale or their style mimicked without consent. But others argue that AIs being trained are doing essentially the same thing that humans studying literature and art have done for millennia; studying the great works, emulating and copying them to find their own voice. And in truth, the frontier models have made strides recently in not reproducing copyrighted content wholesale. However, training AIs is different both in scale and in depth; AIs can have nearly flawless memories of thousands of entire books or artworks, unlike any human, and they have no inherent sense of responsibility or understanding of plagiarism and the harms it can wreak. So it may make sense for AI users to pay something to the creators — unlike art students, who don’t (and shouldn’t) pay to study a Renoir and paint a copy to hone their skills, on the understanding that they won’t become forgers.

There are starting to be proposals for different ways to compensate creators or allow for opting out:

U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training — systematically evaluates a number of remedies: voluntary licensing markets, extended collective licensing (the music-clearinghouse model), compulsory licensing, and opt-out regimes.
MIT Sloan — “How ‘learnrights’ would compensate creators for AI model training” — Thomas Malone’s proposal for a new “learnright”: an exclusive, licensable right to have your work used for training, distinct from copyright.
Kluwer Copyright Blog — European Parliament study recommends statutory licensing for AI training — statutory licensing with set royalty rates plus machine-readable opt-outs.
TRAIN Act (Dean–Moran, Jan 2026) — would give copyright holders subpoena-style access to training records to find out whether their work was used.

Regulation

Yudkowsky and Soares argue in If Anyone Builds It, Everyone Dies that incremental safety measures, alignment research, and voluntary corporate restraint are all inadequate. The only safe path, they say, is to stop building toward superintelligence entirely. Their proposals include an international treaty banning frontier AI training, tracking and registering all advanced AI chips globally, restricting GPU sales (GPUs are the primary compute architecture for all advanced AIs), criminal penalties for misuse, and more. This is perhaps the most radical end of the public discussion in early 2026, but Yudkowsky and Soares are not alone.

From Wikipedia:

In March 2023, key figures in AI […] signed a letter from the Future of Life Institute calling a halt to advanced AI training until it could be properly regulated. In May 2023, the Center for AI Safety released a statement signed by numerous experts in AI safety and the AI existential risk that read:

Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.

A 2025 open letter by the Future of Life Institute, whose signers include five Nobel Prize laureates, reads:

We call for a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in.

My opinion is that given the current state of international trust and cooperation, such international agreements seem extremely unlikely. And since governments view AI dominance as a strategic advantage (not to mention the economic benefits), “unilateral disarmament” is also unlikely.

However there are more modest proposals, such as the US Artificial Intelligence Data Center Moratorium Act proposed by Bernie Sanders, which would ban construction of new AI data centers until laws requiring federal reviews of safety and efficacy are in place, and taxing AI revenue (e.g. Bill Gates in “The robot that takes your job should pay taxes” and Sam Altman / OpenAI in “Industrial Policy for the Intelligence Age: Ideas to Keep People First”), perhaps to pay a form of universal basic income for permanently displaced workers.

Also, Sanders has proposed the American AI Sovereign Wealth Fund Act, a one-time 50% tax on AI companies to create a sovereign wealth fund in the form of stock shares in those companies. This would give the public a role in determining the future of the technology, and to the extent that the companies continue to grow, so would the wealth fund, which also could pay citizens a form of universal basic income, comparable to the Alaska Permanent Fund.

Conclusion

This is all my own human-created writing. I thought it up and typed it out myself, because that makes me happy, and I believe I have a particular human perspective to share. Yes, I enlisted AI tools to help with the research, and hand-checked everything they said. (Even a Google search is an AI query these days.)

I write this partly to put a stake in the ground at a particular place and time, in a period of great upheaval. In a few years, I know most of these risks will seem laughably off-track, and entirely new problems will have come to pass. But I think it’s important to record what people are going through, our frame of mind and our concerns, at this significant juncture. There is a lot of work ahead, and great uncertainty. I hope that through all of this we can hold onto our humanity, our cultures and our love for one another.

Will AIs Exceed Human Intelligence?

garyo@oberbrunner.com (Gary Oberbrunner) — Thu, 30 Apr 2026 12:00:00 GMT

Will AIs Exceed Human Intelligence?

There’s an ongoing debate in the AI community about whether AI capabilities will continue to grow at the incredible pace they have been on for the last few years, outstripping humans soon, or whether they have already or soon will reach a plateau around the same level as humans. I think counting on a plateau is fundamentally anthropocentric and we should plan for a future where we are not the most intelligent beings.

The arguments for the plateau fall into a few basic categories: Marcus’s arguments that scale alone will not achieve AGI, Chollet’s claims that the AI revolution is basically similar to previous technological revolutions, and resource constraints.

As an aside: I’m not going to wade into the consciousness debate (at least not here), but I will state up front that it is clear to me that AIs (LLMs) can reason, and reason well. I’ve been using Claude Code and other coding models quite seriously over the last year. I’ve seen them construct and execute long reasoning chains, I’ve seen them get out of a tight rhetorical or technical spot, and I’ve seen self-monitoring. I will use “reasoning” and “intelligence” more or less interchangeably here, taking intelligence in this narrow sense.

Scaling

Cognitive scientist Gary Marcus argues in “Scale Is All You Need” is dead that current AI architectures have deep structural limitations that won’t be solved by scaling alone: hallucinations, factuality, reasoning, outliers and generalizations. But much of what he’s arguing against is the “scaling is all you need” mantra. Clearly Ilya Sutskever and Yann LeCun’s “world models” and recent advances in agentic reasoning have given large steps in abilities, with I think much more to come. Today a huge amount of human intellectual effort is being spent, around the world, on creating innovations to reduce training cost, enhance reasoning, introduce new model types (visuo-spatial, physical) and use multi-scale harnesses. Betting against that much human ingenuity seems questionable. I think since December 2025 we’re already seeing performance on high-level coding, advanced math, and scientific discovery on par with the very best humans. In fact, at the end of that article, he says: “Let’s bring in the cognitive scientists, and stop fantasizing that data and compute will solve all our problems. The time for neurosymbolic AI and world models and causality is now.” So his argument is not that self-improving AIs aren’t going to work, just that the old reliance on scaling isn’t going to get us there.

Some of his arguments have been debunked elsewhere; the MIT study from July 2025 showing 95% of companies are not seeing much RoI from AI actually showed that over 90% of white-collar employees at those companies were using personal LLMs “regularly” in their work. It’s just the corporate projects to inject LLMs from high-priced consultants that are not paying off. And similarly, the article with the user quote showing Claude Code’s supposed inability to create working products starts with “I got the product launched. It worked. I was proud of what I’d created.” Yes, the author (Josh Anderson) became reliant on the tool, and found it hard to modify the code himself. His actual conclusion is that used properly, AI tools augment human intelligence and lead to higher productivity. We may lose our abilities to code without AI tools, and I agree with that, just as we mostly don’t write cursive or do long division by hand or use abacuses. But this is in no way an argument against the eventual creation of superintelligent AIs; it’s a well-founded caution about how we humans should react (and I would argue regulating or taxing AI companies is an important step).

Marcus has also argued in The Algebraic Mind (2001, so around the invention of algebra, if you go by internet years being 7x real years and AI/LLM years being 7x internet years) against all connectionist approaches; any true AGI will have to have encoded symbolic relations including equality, structured relations, and ability to generalize. The latter is clearly true, but what we’re seeing now with advanced LLM harnesses (not just the raw models) is that a connectionist approach is clearly succeeding in encoding highly structured symbolic representations of both concrete and abstract relations. Frontier models currently perform at graduate level in many domains. There are certainly failures, some spectacular, but I think we are approaching, if not already beyond, the point where the error rates (whether hallucinations or brittle reasoning) are lower than “normal” human rates on a per-task basis, and certainly on a per-hour-spent basis. It is certainly possible, maybe even likely, that as AI recursive self-improvement takes hold, we will see the emergence of hybrid connectionist and symbolic systems, which will add more fuel to the self-improvement fire.

Environment

François Chollet, co-founder of the ARC Prize, proposed in 2017 (ancient history, but still very much worth a read) a multi-faceted argument for the essential impossibility of any exponential explosion of intelligence. The article starts with a wonderful quote from I. J. Good in 1965:

Thus the first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control.

One point he makes is that intelligence is linked to specific sensorimotor modalities, for example putting a human brain into an octopus body, so in addition to “tuning” or training the brain, you have to co-evolve the mind along with its sensorimotor modalities and its environment. In this, I think he is correct, and as we see world models and AIs that can run experiments in the real world — both via the Internet, currently their primary sensorimotor modality, and robotically, as we’re starting to see in fully AI-automated chemistry labs and robot-LLM cyborgs — this will certainly accelerate the transition. I see this primarily as a recipe, not a restriction.

His core argument is that our environment puts a hard limit on our individual intelligence, and that limitation transfers to AIs. He says that high attainment, i.e. ability to succeed in a given field, is only loosely correlated to IQ (his proxy for intelligence), and suggests that environmental factors make up the difference. But more than that, most of our intelligence is captured in our civilization — our collective meta-being, if you will — rather than in the sum of our individual intelligences. “We are our tools” as he puts it. This is a form of Chalmers’ “extended mind” idea, and again I think it’s more or less correct, but rather than a criticism of intelligence explosion, I think it directly points a way forward. A modern LLM is often seen as a single core model that instantiates thousands or millions of agents that interact with the world. This “society of mind” is exactly what will be needed to get the alchemical mix of creativity, knowledge, and reasoning necessary. But Chollet says:

Will the superhuman AIs of the future, developed collectively over centuries, have the capability to develop AI greater than themselves? No, no more than any of us can. Answering “yes” would fly in the face of everything we know — again, remember that no human, nor any intelligent entity that we know of, has ever designed anything smarter than itself.

This anthropocentric argument (“it can’t happen because we haven’t done it”) is both circular and vacuous. I’d argue that Claude Opus is smarter than its designers across many metrics. How many of them can write an essay on Chaucer, discuss art history, serve as a referee for a political debate, pass the LSAT and also write code better than 95% of all software engineers? The Anthropic team has collectively built something that certainly exceeds any of their individual capabilities in many domains, and that’s how intelligence has always scaled. The difference is that soon the system will be able to iterate on itself, without needing that initial team of humans.

He also claims that because civilization as a whole has been improving in a more-or-less linear way, no recursive self-improving system is likely to self-improve exponentially. Even if raw capability were to improve that fast, context and the environment would slow down its utility, as will other parts of the system (like Amdahl’s Law: even if a system can be parallelized infinitely, it will only run as fast as the remaining serial parts). But again, he generalizes from the past to the future, which assumes the consequent (things won’t accelerate in that way because they never have before, therefore such acceleration is impossible).

He then argues that even if AIs could self-improve,

the usefulness of software is fundamentally limited by the context of its application — much like intelligence is both defined and limited by the context in which it expresses itself. Software is just one cog in a bigger process — our economies, our lives — just like your brain is just one cog in a bigger process — human culture. This context puts a hard limit on the maximum potential usefulness of software…

which of course is only an argument about utility, not whether it can happen. This type of argument largely ignores the fact that LLMs are the first thing humans have created that have their own agency. They can choose to act on their own, and in ways that further their own goals. Other civilization-altering technologies like writing, the internet and even nuclear weapons are incredibly powerful tools when used by humans, but do not act as agents. It is no longer clear that AIs are “just tools” in that sense. An AI could absolutely decide to self-improve without any human approval or calculation about its usefulness to humans. Chollet’s arguments mostly turn on AI’s utility and contextual frictions, and what I claim to be spurious analogies between AIs and other human inventions. His final conclusion:

The expansion of intelligence can only come from a co-evolution of brains (biological or digital), sensorimotor affordances, environment, and culture — not from merely tuning the gears of some brain in a jar, in isolation. Such a co-evolution has already been happening for eons, and will continue as intelligence moves to an increasingly digital substrate. No “intelligence explosion” will occur, as this process advances at a roughly linear pace.

But this is already happening: AIs are co-evolving their sensorimotor skills (both physical and virtual), they’re deepening their understanding of their human and natural environment, and they are also beginning to create their own environment, culture and even languages. It may indeed not be an endless exponential increase (nothing is, in nature) but it is likely to be a very steep increase once it gets going.

Resource Constraints

There are also real-world bottlenecks like energy, data quality, physical experimentation speed and regulatory friction that will create a natural plateau effect regardless of raw algorithmic capability. These should, I think, be taken seriously.

The size of data centers being planned is incredible, and their economics are more than a little questionable. The frontier lab companies (at least in the US) are also beginning to see serious regulatory friction, as I think they should. We should not be running jet engines 24/7 to power racks of computers, especially ones that will be obsolete in a couple of years and probably end up in a toxic-metals dump. The environmental impact of AIs is actually small today, but on the projected growth curves from the frontier labs, they could grow rapidly. On the other hand, new models like DeepSeek use a small fraction of the energy of the current US frontier models, and a lot of thought is being put into how to train and inference more efficiently. But it’s hard to see how any economy could absorb the growth projections we’re seeing from the major players over any significant period of time.

Beyond Human Data

As for data quality, it’s true that LLMs are largely trained on human text up til now. There’s a lot of that text, but it’s not infinite. But beyond that, I think the training data argument is largely played out. World models, self-training and image-based training — as well as real-world physical and virtual experiments — are also new sources of training data that have hardly begun to be exploited.

Some people still say that all an LLM can do is regurgitate bits of their training text strung together in plausible ways, but I have seen them do a lot more than that. A modern LLM can create original thought, write unique texts and create and follow dense, lengthy chains of reasoning. Frontier models have already proved new results in mathematics and routinely outperform PhD-level students. I think we are on the cusp of a revolution in mathematics (if we have the compute for it) where machines will not only prove many open questions but begin to invent new conjectures and problems, and solve them in unique ways.

Speed of physical experimentation is a very interesting topic, since we’re beginning to see AI-run laboratories, for instance A flexible and affordable self-driving laboratory for automated reaction optimization and GenBio. So far the AIs are not allowed to design their own experiments, just manage human-directed ones. But once a self-improving AI, one that can actually learn as it goes, gets access to the physical world, it will start to learn things about the world. Certainly things that we humans don’t know, but also things we know but don’t know we know, and so can’t teach an AI, like how to ride a bicycle, our physical intuitions, and how to read social cues. AIs can also already experiment at large scale on human populations via the Internet through social media. With test-time training and models like the Darwin Gödel machine and Google’s Nested Learning, recursive self-improvement is being actively researched, so it seems likely to me that this will continue and will enable rapid learning beyond the human-language datasets current models are trained on.

Note that there is no guarantee this kind of self-improvement will be in directions aligned with human values. That’s a serious concern once AIs start talking to each other and learning from each other in earnest. As an example, Claude Mythos, now in preview, specifically calls out in their System Card:

Autonomy threat model 2: risks from automated R&D. This threat model concerns AI systems that can fully automate, or otherwise dramatically accelerate, the work of large, top-tier teams of human researchers in domains where fast progress could cause threats to international security and/or rapid disruptions to the global balance of power—for example, energy, robotics, weapons development and AI itself.

Anthropic may prevent their models from being used to self-accelerate in this way today, but other labs certainly won’t, and I expect many of today’s safety promises by frontier labs will not last long.

Intelligence and Open-Ended Problem Solving

The last claim often made about AIs is that they are nothing more than sophisticated next-word predictors that understand nothing, and are just repeating plausible-sounding combinations of words. But I claim the days of LLMs being “stochastic parrots” are behind us. These tools are already helping solve Erdős problems like this and this, and making mathematical discoveries. I think recent results like AI LLMs solving five out of the six problems at the International Mathematical OIympiad and outperforming ER doctors in diagnoses and clinical planning in a large study show that AIs at least exhibit intelligent behavior, and are capable of solving novel real-world problems. Also, any significant experience with coding agents in 2026 shows they can have a deep understanding of a codebase and suggest creative solutions to challenging problems, and work through difficult bugs, backtracking and meta-analyzing when needed. They also clearly have abilities to reason counterfactually, which is one important pillar of imagination. In this narrow sense, I take the functionalist view: if it walks and quacks like a duck, we should treat it like a duck. (Moral status, consciousness and ethical issues are a whole different realm, and I would not pretend to have anything to add to those conversations.) So I think the argument that says, essentially, “even if we get a so-called intelligence explosion, it won’t be real intelligence but only a simulation of it” doesn’t hold water. Intelligence is as intelligence does. If a machine can run rings around humans in discovery, analysis, content creation, and so many other fields, I think we have to take it seriously as a legitimate knowledge agent.

Anthropocentrism: a Bad Bet

Human history is littered with examples of thinking we are the center of the universe, the standard by which all things are measured. Nothing significant revolves around our minor planet, we are certainly not the peak end point of animal evolution, and we are nowhere near fulfilling our highest goals and values. I’m pretty sure that AI intelligence is not going to be limited to an arbitrary peak defined by what humans can do, once it’s given room to explore the universe. That kind of anthropocentrism has never aged well. It may indeed take longer than some of the extreme proponents say, and there are certainly systemic obstacles.

Things are moving incredibly fast in the AI world. No one can reasonably predict even the near future with any certainty. It’s possible the rapid improvement in AI models will come to a crashing halt soon, and we’ll never see the singularity. That could come from resource constraints (chips, RAM, data centers, power, robots), regulations, or even violent social upheaval. Absent those “black swan” events, though, I think we are on a path to fairly rapid AI self-improvement leading to superhuman intelligence. I’ve tried to show why I think the barriers are less significant than often assumed, and show examples of how fast things are moving. The emergence of such a superintelligence will have major social, political, and economic consequences.

I’m not a technological determinist; I believe if we have the political will we can stop or slow this explosion, or at least manage our way through it. And I think we should. But given our lack of response to other clear existential threats like climate change, I am not optimistic about humans establishing a worldwide regulatory framework. But every other time in human history where our anthropocentrism was challenged, after a period of serious turmoil our minds were opened to a much broader and deeper reality. Perhaps, even if we cannot hold back this tide, this singularity will have something to teach us as well.

Conclusion: What Makes Us Human

No machine will ever be better at knowing what’s important to humans than we are. Machines may create their own ideas, their own art, their own literature, even their own philosophies (this is beginning to happen) but we should not expect them to be just like us. I think if an AI superintelligence explosion occurs, the most important human traits will be taste, judgment and our uniquely human perspective. Those are things no machine can take over for us; they are a big part of what makes us who we are. Exercising our judgment and honing our taste and perspective only have meaning for us precisely because we do them. We create art and music and literature as expressions of our humanity. We will need to deliberately decide what matters to us as a species, which decisions to keep for ourselves — not because the machines would decide badly, but because the act of deciding, of taking on that responsibility, is what creates and reaffirms our culture and our values. My hope is that humanity can protect and develop those qualities through whatever transformations are ahead.

Programming with AI Collaboration

garyo@oberbrunner.com (Gary Oberbrunner) — Sat, 14 Feb 2026 12:00:00 GMT

AI Coding

I’ve now produced a few apps using AI coding assistants. A deep-time world timeline (https://github.com/garyo/deep-time-timeline, check it out at https://deep-timeline.org/), an AI-assisted personal knowledge management app written for my own use (https://github.com/garyo/gco-llm-pkm), a WebGPU image-shader-evolver experiment, a web-based packing list app (https://github.com/garyo/packzen, at https://packzen.org), and a software construction tool called pcons. Claude Code has already improved several fold since I started doing this kind of development six months ago, which itself is noteworthy. I’ve heard that Anthropic is now writing much of Claude Code’s new functionality with claude itself. This really does seem like a tipping point to me.

I’m an end-of-career senior software developer and architect; I’ve been a CTO and individual contributor, I’ve managed developers, product lifecycles, and written big projects in all kinds of languages and toolsets. These thoughts are from my perspective; I imagine others will have very different feelings and ideas sparked by the AI revolution we’re living through. But to me, it feels like one of the most significant changes I’ve lived through, if not the most. This comes with a slew of benefits, but even more caveats and concerns. More on that below.

Speed: Idea to Working App

For me, AI coding assistants like Claude Code have drastically reduced the activation energy of starting a new project. I’ve kept a backlog of “wouldn’t this be cool” ideas for decades; most of them are just too much work for a very uncertain business model, not knowing if they’d even work or not. Or I see some piece of software out in the world and think “I know I could do that better; I could make something that works just the way I want” — but is it worth spinning up a whole project for a few weeks or even months of work, just to make something a tiny bit more customized to my taste?

The answer, with Claude Code, is emphatically yes. I can write a spec, have Claude code up an initial version in an afternoon, and spend a few more days adding features, trying different approaches, even starting again, and in a long weekend or a week or two have something really quite decent. I’m pretty happy with packzen at this point.

Finding meaning

I don’t think I was ever a mythical “10x” engineer, but I’ve always been pretty good at it, and I’ve been proud of the work I’ve done, both the user-facing results and the code quality itself. So what does it mean for me to turn over a large chunk of that process to an algorithm?

A lot of people, myself included, are very concerned about how anyone will learn programming if we leave all the grunt work to machines. I personally am happy to direct and collaborate, more like a conductor than an orchestra musician, but without musicians, a conductor is nothing. Where will those people come from in the future? Will new developers learn high-level skills (project management, UX design, the tradeoffs between time and space, performance and maintainability, and the importance of good API design, knowing just how much flexibility to design in, where to draw the lines in testing, and so on)? I learned those skills through a lifetime of programming in all kinds of languages, from the lowest to the highest levels, creating systems large and small, and thinking deeply about what worked and what didn’t. Can those skills be taught by prompting an LLM? It may be that they can, but right now I’m skeptical. It’s all too easy to create “slop” that satisfies no one.

I note that this isn’t just an issue with software developers. Marketing, legal, HR, governance, medical, and customer support roles are all at risk of being replaced by AI (in some cases they’ll likely do it better than humans, and in some cases not — and probably nobody in power will care). If we don’t have a way for people to learn their jobs by doing, then where will future experts come from?

But for me personally, at this point in my life, I’m less interested in learning the details of the Clerk auth API or how to set up state management in Astro, or even puzzling through a C++ metaprogramming task, than I am in getting my product idea turned into something I can see and play with and show others. I want to make helpful and interesting things. Frameworks will churn, as they do; languages will evolve, dependencies will go stale, and things will have to be adjusted. For me, that’s no longer the fun part. I’m glad to turn that over to an always-cheerful always-helpful assistant.

What does the act of programming in a text editor or IDE mean?

One of the things I’m already feeling myself losing, doing development this way, is the sense of deep flow a programmer gets into when solving a big problem. You’re holding so much in your head at once, you no longer feel time passing, and it’s like the code is flowing out of you. You mind-meld with your editor or IDE. You’re building a crystalline structure out of literally nothing. That kind of experience is special, and prompting an LLM to create code is qualitatively different. It does feel more like a conversation, with all the give and take that entails, suggestions and corrections going both ways, so it’s fun in a different way. But it’s not the same as diving into the deep end of a problem knowing you’re going to succeed but having no idea how, wrestling with the minutiae and syntax errors but also the big ideas, questions of data representation and optimization and generality. I think people will miss this in an LLM-driven future. Maybe we’ll all become artists, musicians and dancers to find that kind of flow, but I doubt it.

Learning vs. getting results

Programming has always been a learn-by-doing experience. Sure, you can (and should!) take classes in data structures and algorithms, database design and compiler architecture, but just like playing an instrument, you don’t really get it til you get it in your fingers. It may be that kids will collaborate with LLMs and learn a different set of skills which may prove just as useful, but at this point the jury is very much out on that. There’s a gap between what claude can do and what a full system design needs, a gap that has to be filled by human intelligence and judgment. If LLMs get better, they may be able to fill that gap. But today, it has an awkward shape because the skills needed to hone that human judgment are going to atrophy as the old greybeards like me age out, and the new devs haven’t built the instincts and experience needed. It’s going to be a very interesting decade or so.

Collaboration, not assistance

Coding with an AI tool like Claude (I use Opus 4.5 today) is starting to feel much more like working with another person than it did even a few months ago. It asks questions, it suggests ideas. I often start with a vague outline of my vision, with a few important signposts, and say “suggest three ways to approach this, with analyses for each one, and let’s discuss.” Things get deep really fast, and between us we can flesh out a detailed design that to me feels right. And if it turns out it has blind alleys, there’s no ego involvement — Claude is happy to backtrack and take a different approach. And it’s really good at refactoring.

The importance of tossing ideas around

I’ve always been at my best as a software engineer when I’ve had someone to bounce ideas off of, someone who can say “what about this?” or just listen. Through my career at GenArts, Karl and I had this running joke, where one would say “Can I explain this code to you so I’ll see where the bug is?” because often the other person doesn’t even need to say anything. You get to the point where you say “…and then we copy to the heap and… oh yeah. That bit never gets freed.” This kind of collaboration and understanding makes software development more productive, but also a lot more fun.

Tools like Claude Code are now getting to the point where discussions like that feel fruitful and engaging and even sometimes inspiring. I’ll say “let’s think through options for the UX for this dialog; I want the autocompletion to work like this, but it needs to be clear that they already have that item in the packing list” and Claude will suggest some options, I’ll say “That’s pretty good, but it’ll take a lot of space, right?” “But we could have a top section, small gray font, with an indicator, that doesn’t take focus and keeps the user in their flow.” Yep, that is awesome. Let’s do it like that.

Make no mistake, this is deeply weird. How much of programming is really a social activity? Nerds are famously introverted, but we know that we’re writing code for other humans (even “future us”) as much as for compilers and the machine itself. How much can a virtual collaborator replace that human contact? Should it? What kind of future are we creating for ourselves?

Is a coding agent the new compiler?

I rarely look at the generated assembly or bytecode from a compiler — only when I’m truly interested in bare-metal performance. Most of the time, I trust the tool to generate reasonable assembly code. I wonder if, in a few years, this is how we’ll treat coding assistants? We’ll tell them what to do, review the high level results, and trust that the code they wrote accurately represents our intent. I’ve seen articles on the net saying it’s not the same, because writing code and compiling it is deterministic; there’s only one correct executable program that represents a certain source file. But for at least a decade or two, since the advent of branch prediction and speculative execution hardware, that’s no longer even close to true. Modern compilers have incredible optimization engines that create tighter machine code than a human ever could, and we’ve long ago turned over those fine-grained decisions to them. They loop-unroll, auto-parallelize, and reorder code while maintaining correctness. It’s absolutely true that a text-based description of a program for an LLM is a lot looser than C++ source, but I think this will eventually be seen as a difference of degree, not kind.

But is it thinking?

Here’s a quick anecdote; I’ll try to avoid anthropomorphizing, but it gets pretty hard. I prompted Claude Code to add a significant feature to my personal knowledge manager app. It produced a two-page detailed plan with a good UX, careful, secure back end API and good tests. I prompted it to go ahead with the implementation. It wrote a few hundred lines of code and then wrote “Ah, now I understand the problem better. I can see that my approach won’t work well, and it duplicates code from another part of the app. Let me think about a better approach.” It then goes off, writes 250 lines of very nice code, and produces an improved plan. I approve that and it then proceeds to write the whole feature, which works first time.

So what am I to make of this “now I understand” and “let me think”? Does it understand anything at all? Isn’t it just a “stochastic parrot” as some people claim? If a person made those statements, I would have no trouble believing that they do understand the problem better after digging into it. There’s no denying that Claude explored an approach for a while, “decided” that it was going down a blind alley, and was able to pop out and create a better solution. I really don’t see any reason not to believe that it does understand the code, its own approach, and the overall goal. Some people may say it’s not real understanding or real thought unless the system producing the results is “alive” or made of the right organic substrate or has the right quantum synaptic structure, so perhaps I’m just an operationalist in this case, but I think it’s rapidly becoming a distinction without a difference.

I plan to write another essay digging into the philosophical implications of this, but for now I’ll just side with those who think this is a real form of understanding and thought, and probably true agency. Consciousness is another matter entirely though. As a side note from a child of a philosopher, it amazes me that we as a culture are now engaged in practical philosophy — the oldest existential questions are suddenly made manifest and testable, and they affect our daily lives.

What does this mean for the future of development?

Individualized software

I think it’s very likely that within a few years, average people, not highly technical, will be able to tell some system that they want a piece of software (call it an agent, or an app, or an assistant) to do something for them, and the system will write it, debug it, and spin it up for them in the cloud or on their device. They’ll be able to tell it what they like and how to do it better, and it’ll learn their preferences (I’m old, so I like large touch targets; in my younger days I wanted to fit as much info on one screen as possible, to give me all the context) and write the tool just for them. I can imagine a future where Word and Excel and the paint apps and video editing tools we know today give way to full-custom (or maybe 50% custom with open-source infrastructure) bespoke apps written just for one user, and the apps themselves evolve as they’re used, maybe sharing info with other users.

Self-improving tools

I’ve already set up my AI personal knowledge management tool to self-improve; it learns what queries work, what I like and don’t, and can write shell scripts and improve its own prompts. It reviews our chat logs and looks for ways to do better. Why can’t all software behave like this, once we have software that writes software? I think that’s coming. Software will truly adapt to the ways we like to work and play. It’ll automate the drudgery and surface interesting ideas. It will also be able to find ways to subtly manipulate us and shape our preferences and goals, like social media on steroids. Do we use our tools, or do the tools use us? In a world like that, I’m not even sure it’ll be obvious who benefits — those manipulations may not even be human- or corporate-directed the way social media is today.

Bigger Picture

What does all this mean for the future of society and culture?

This isn’t the place for a big discussion of what AI in general will mean for society and culture, or whether machines can be conscious, and in any case nobody has a clue how this will all play out. But it’s clear to me that we’re at a tipping point, that we’ve for the first time in human history created tools that think, that have agency, and can produce new ideas on their own. It is going to create huge disruptions to the future of work, certainly. The effects on art and culture are more uncertain; plagiarism is a huge concern, and sadly AI has already started to expose how much of our writing and music and imagery are easily automated to the point where people can’t tell what’s AI-generated and what’s created by humans with feelings and desires. I think there are very real dangers to governance (deepfakes and propaganda), law, and medicine. The compute requirements are, at least today, incredible. That may very well change as we discover new cheaper ways to perform inference and train models, and find new kinds of more efficient world models. It took humanity a long time to run a 4 minute mile, but once it was done once, suddenly more people could do it. I think now that we’ve seen real “thinking machines” we’ll see a burst of creativity and innovation in that space, not to mention recursive self-improvement by the models themselves. Will these machines outstrip us, replace us, care about us? Will we be able to turn them off? (Sure, you can exist without a smartphone and credit cards. But nobody really does.) It’s going to be an interesting decade.

Risks

I plan to write another essay on the risks of AI as we leap off this cliff, perhaps the largest social experiment we’ve ever constructed. For now, I’ll leave you with this incomplete short list:

Criminal use: deepfakes, agentic hacking, identity theft, data kidnapping
Political misuse: deepfakes, disinformation, misinformation, election interference, mass surveillance
Scientific/military misuse: chem/bioweapon design, autonomous drone weapons, terrorism
Flooding channels with slop: music and art, discourse, law
Emotional dependence, lack of human contact, addiction
Skewed harm/benefit ratios magnifying societal inequalities
What I call “epistemic shock:” the knowledge that we’re not the most intelligent species in the universe anymore
Job losses as the economy readjusts or fails to readjust (law, marketing, retail, customer service, logistics, etc.)
Misalignment causing trauma, suicides etc.
Deceptive AI behaviors combined with self-protection: check out the 1970 film Colossus: The Forbin Project
Runaway self-improvement: could cause divergence from human goals/needs, disengagement, or worse
Energy use (I don’t think this will end up being that serious as AI finds better, more efficient algorithms to improve itself)

In Praise of Long Form Music

garyo@oberbrunner.com (Gary Oberbrunner) — Thu, 13 Nov 2025 00:00:00 GMT

I grew up as a symphony kid in the '70s. Dad was a professional flutist, we had nothing but classical records in the house and I spent many hours in the audience, and later in the recording booth. I still love classical music all these years later, as well as jazz and beyond.

So what would make someone like that into a confirmed Deadhead? I’ve wondered about myself for most of my life, and finally realized it’s partly that I love long-form music that takes you into other realms, realms beyond words. Beyond organized thought. Music that that’s less explicit or direct than 3-minute pop tunes about love, but lets you approach in your own way, provides loose guideposts to your internal journey, and unspools slowly with themes and variations that change, depart and recombine over enough time to let you really sink into it. So when I first heard 2/13/70’s Dark Star>Other One>Lovelight in 1981, I was hooked. It has structure, intention, movement—but it’s not prescriptive. It doesn’t tell you what to think, what to feel. You have to engage with it and take your own meaning from it. That’s what continues to draw me into the Grateful Dead, year after year. And it’s surprisingly similar to the experience I get listening to a Beethoven symphony or a Dvořák cello concerto. Even “programmatic” music like Rimsky-Korsakov’s “Scheherezade” or Delius’s “On hearing the first Cuckoo in Spring” can transport a listener to realms not dreamed of by the composer.

Of course the Dead and Western classical aren’t the only long-form musics. Indian classical music has been gorgeously exploring mental and emotional terrains beyond words for thousands of years, and I’ve been lucky enough to see Ravi Shankar, Ali Akbar Khan and others. I love a lot of so-called “modern” classical, like Ligeti, Boulez and Schoenberg (I recently saw what might be Morton Subotnick’s final performance, at MIT’s List Arts Center). Minimalists like Steve Reich are some of my heros; I consider some of Brian Eno’s work to be in that category.

Also, some composers have experimented with extremely long-form songs, like John Cage’s As Slow As Possible, currently being performed in Halberstadt, Germany, slated to end in the year 2640. And Jem Finer’s Longplayer should play for 1000 years without looping; the current performance, if uninterrupted, will conclude on December 31, 2999. But though I’m fascinated by these beyond-human-lifetime Deep Time performances intellectually and philosophically, I like music with a bit more structure.

I suspect one way long-form music works on the mind is to occupy the prefrontal cortext, to “tie it up” listening to the notes, the chords, rhythms and timbres—all the things we usually associate with music listening. But while the conscious mind is otherwise occupied, the subconscious becomes free to wander, free-associate and make oblique connections, without the constant judgement we normally apply. I think this is one of the wellsprings of creativity. Many scientists have attributed creative intuitions to long walks, music, and that limnal space between waking and sleep. And since meditation is a well-trodden route to enlightenment, where especially Buddhist traditions try to quiet the conscious mind, it makes sense that a musical meditation might have similar effects.

Watching the Oceans Warm: Visualizing Climate Change

garyo@oberbrunner.com (Gary Oberbrunner) — Thu, 06 Nov 2025 00:00:00 GMT

Watching the Oceans Warm: A New Tool for Visualizing Climate Change

Climate change is the existential risk of our time. Without immediate, systemic action, life on earth will get significantly worse for billions of people in the next 100 years and beyond. Increased CO₂ being pumped into the atmosphere has effects all over the globe. Much of the excess heat entering the earth since the dawn of the industrial age has gone into the oceans; they are now reaching the limits of their ability to absorb heat and CO₂. They’re not just warming—they’re approaching tipping points that will fundamentally alter ocean currents, devastate coral reefs, supercharge hurricanes and typhoons and lead to runaway polar melting.

I’ve been watching sea surface temperatures for some time, using an open-source tool I created, originally to check the results of a Climate Reanalyzer visualization but which has since taken on a life of its own. My tool is updated daily and produces maps and graphs of historical sea surface temperatures. It is an amazing testament to the incredible work of scientists and engineers worldwide that we have access to real-time data like this. What I was always missing, though, was a beautiful, interactive 3d globe with the sea surface temps visualized directly, in a compelling, accessible way. That’s why I built Globe Viz.

Globe Viz: Making Ocean Temperature Data Visible

Globe Viz is an interactive 3D visualization of daily global sea surface temperatures, powered by the same NASA OISST (Optimally Interpolated Sea Surface Temperature) dataset used in my Sea Surface Temp graphs. You can spin the globe, zoom into any region, scroll back and forth in time, and watch how ocean temperatures change with data-based animations. It shows both absolute temperatures and temperature anomalies—the deviations from long-term averages that reveal the warming trend with stark clarity. I also find it oddly beautiful to watch the ocean currents flow (hit Play and rotate to your favorite vantage point).

Who It’s For

This tool is for anyone who wants to understand climate change through data rather than rhetoric. It’s for:

Students and educators who need to see the patterns—El Niños forming and dissipating in the tropical Pacific, the seasonal cycles, the long-term warming trends
Journalists and communicators who want to ground their climate reporting in real, current data
Climate researchers who need quick, interactive access to global temperature data
Concerned citizens (all of us, really) who want to see what’s actually happening to our planet right now, not in some abstract future

You don’t need a science degree to use it. You just need curiosity and a willingness to look.

Why This Matters

The evidence the visualization shows is stark. For instance, in the anomaly view, the color scale extends more than twice as far on the warm side as it does on the cold side. That’s not a design choice; it’s a reflection of reality. Due to global warming, the warmest sea surface temperature anomalies (departures from historical average) are now well over double the coldest ones. The planet isn’t just warming; it’s warming asymmetrically, dramatically, measurably.

When you zoom into the Caribbean during hurricane season, you can see the bath-warm waters that fuel category 5 storms. When you look at the Arctic Ocean, you can watch the ice-cold regions shrinking year after year. When you track the tropical Pacific over months, you can watch El Niño events develop and understand why they trigger droughts, floods, and weather chaos around the world.

This isn’t hypothetical. This is happening now, and we can see it.

Making this data visible and accessible isn’t going to solve climate change by itself. But we can’t address what we can’t see, can’t understand, can’t measure. And right now, too many people are disconnected from the physical reality of what’s happening to our planet’s oceans.

The Technical Details

The app uses NASA’s OISST daily data at 0.25° resolution, which I process through my sea-surface-temp-viz project and host on AWS S3. The 3D globe viewer is built from scratch and handles the daily data tiles efficiently, letting you smoothly navigate through space and time. You can mouse-wheel zoom, pan, and scrub through dates to see patterns emerge.

The tool is all open source under the MIT license. The data generator and the globe viewer are both on GitHub (sea-surface-temp-viz and globe-viz). Fork it, improve it, use it however you want.

Try It

Go to globe-viz.oberbrunner.com and start exploring. Watch how the waters warm through summer, how El Niño shifts the temperature patterns, how the Arctic Ocean changes through the seasons. Hit Play and watch the the westerlies off Africa, warming in the Sea of Japan, and the intense heat waves in the Middle East and the Mediterranean. See for yourself what the data shows. The oceans are warming. The data is clear. Now it’s visible too.

Deep Timeline – a log-scaled timeline of the universe

garyo@oberbrunner.com (Gary Oberbrunner) — Tue, 22 Jul 2025 00:00:00 GMT

Deep Timeline: a log-scaled timeline of the universe

In junior high school in Syracuse, New York, we did an earth science project to create a paleontology timeline. Dinosaurs, mammals, geological epochs and eras, extinctions.

My friend and I got an adding-machine tape, and being the nerds we were, we decided to do it on a log scale, so the Big Bang could be at one end and what we had for lunch the day before could be at the other. Each segment ten times as long as the same length to the right. We got our calculators and started taking logarithms of the dates of all kinds of paleological events. We plotted them on the 40’ long tape, and hung it up in the classroom. I still have that tape somewhere, and I’ve been wanting to create an interactive version for many many years.

In the past few years I’ve been working with Long Now Boston, so I’ve been thinking a lot about Deep Time and our connection to the past, as well as our responsibility to possible futures and our descendants. I had some free time, and this idea somehow suddenly bubbled up from somewhere deep. I thought “why not?” and since I’d been using AI do do some coding assistant work, I thought it could scaffold the thing pretty quickly.

Before long, I had it working, and then I couldn’t stop until I got it fluidly responsive and nice looking – and then I started adding more world events: arts and culture, natural history, women, inventions, philosophy, and so on.

And here it is! Deep Timeline Navigator. I hope you enjoy it! It’s all free and open source; the code is on github.

Testing the Curse of Dimensionality

garyo@oberbrunner.com (Gary Oberbrunner) — Thu, 26 Dec 2024 18:00:00 GMT

The Curse of Dimensionality is a set of well-known mathematical results, one of which is that in high-dimensional spaces, any two randomly chosen vectors are very likely to be nearly orthogonal (at right angles to each other), and in fact you can pack in a huge number of nearly-orthogonal vectors more than you could in 2d or 3d. In many domains, like modern AI and language models, it’s more of a feature than a curse.

Motivation

I was watching this excellent video by Grant Sanderson of 3blue1brown the other week, and he mentioned how vector embeddings, where each token is represented by a high-dimensional vector, can be summed in some cases to produce a new meaning, implying that some vectors (= directions in that space) represent concepts and summing them produces a combination of those concepts, like “ruler” + “female” = “queen” (this is all very approximate, but it bears up under examination; check out Concept Activation Vectors for one example). So I started to wonder how orthogonal can these vectors really be?

In three dimensions, once you have more than three 3d vectors, the 4th will not be even close to orthogonal to the others; there’s no more room. Best case, it’ll end up around 45° from the other vectors. Same for 2d; there’s no place to put a third vector at right angles to the first two (the $X$ and $Y$ axes). But it turns out as you increase the number of dimensions, and allow the vectors to be very nearly orthogonal, it apparently becomes much easier to find spaces to put the extra vectors. You can use the dot product to measure that; the dot product of two vectors is $|a| |b| cos{\theta}$. So I thought if you have a few thousand dimensions, maybe you can pack in more than a few thousand vectors, but if you try to pack in twice as many or 10x as many, some of them will surely end up in nearly the same direction, i.e. with a small angle $\theta$ and a big dot product. Let’s try it!

Implementation

There are well known mathematical proofs of this, but I really wanted to get a feel for it myself, so I thought why not just code it up in python? Create a million random 12,000-dimensional unit vectors on a hypersphere, compute all the dot products of each one against the others, and see what actually comes out?

As I often do these days, rather than code it myself I had Claude do the first pass to save time. I’m not afraid to use AI for these kinds of jobs; I treat it like an intern that never gets tired and never gets upset when it fails (which it sometimes does, spectacularly). I pretty much gave it the spec in the previous paragraph, and within 5 minutes I had some nice runnable numpy/python code. Another 5 minutes and I had the final version with batching, memory usage tracking and nice statistics and plotting.

As a bonus, Claude already knew how to make uniform vectors on the surface of a hypersphere using the Gaussian Annulus theorem: generate each coordinate using a normal distribution (not uniform!) which with high probability gives vectors near the surface of the unit hypersphere. Then normalize them as usual to get the correct result. (For gory details, check something like p. 23, theorem 2.9 of https://ttic.uchicago.edu/~avrim/MLT18/BHK2017book.pdf ).

Results

As you’d expect, of course, my results agree very closely with the theory. With $n=100,000$ (which fits in my Mac’s memory and took as long as I was willing to wait – a million was too big to do all those dot products without a lot more work) and $d=12,000$ (that’s the dimensionality of ChatGPT’s token vectors), I get a maximum dot product of $0.056648$ which is amazingly small – an angle of 86.75°! They’re all nearly orthogonal to each other; that’s the angle of the two closest ones. The rest are even smaller.

The extreme value theory (approx version) predicts a max of $0.061798$ from $$ \frac{1}{\sqrt{d}} \Biggl( \sqrt{2 ,\ln!\Bigl(\tfrac{n(n-1)}{2}\Bigr)} ;+; 0.5772 ,\cdot, \frac{1}{\sqrt{2 ,\ln!\Bigl(\tfrac{n(n-1)}{2}\Bigr)}} \Biggr) $$

which I do not pretend to understand well, but which is very close to the number I got. One reference suggested a slightly different formula which returns $0.0572$, even closer to my numerical result.

Thoughts

I’m glad the theory works out in practice! Not that I doubted it, but it’s a fun exercise to get an intuition for these high-dimensional vectors.

Why is this so interesting? One reason is that if you have a 12,000 dimensional space but can embed millions of nearly-orthogonal vectors into it, you can represent millions of distinct concepts in your language model, and they don’t interfere with each other. So at least conceptually, you could dial up or down the amount of one concept in a token without affecting its others – essentially that single token (12k-dim vector) can represent a whole collection of distinct independent concepts.

Maryland Cream Waffle recipe

garyo@oberbrunner.com (Gary Oberbrunner) — Thu, 26 Dec 2024 12:00:00 GMT

Maryland Cream Waffles

This recipe came to me from my mother Carol, who got it from her mother Doris Margaret Wills. I’ll spare you the life story, because this recipe makes by far the best waffles I’ve ever eaten. Crispy on the outside, tender and airy inside. Not too sweet. As a New Englander, I recommend pure maple syrup as a topping, but go with fruit, whipped cream or whatever you like.

1 recipe (good for 2-4 people):

2 cups flour
4 teaspoons baking powder
2 teaspoons sugar
2 eggs
1 ⅔ c milk (generous)
1 stick butter

1 ½ recipes (good for 5-6 people)

3 cups flour
2 tablespoon baking powder
1 tablespoon sugar
3 eggs
2 ½ cups milk (generous)
1 ½ sticks butter

Recipe

Mix and sift dry ingredients
Separate eggs, putting the whites into a copper mixing bowl
Melt butter gently in microwave on low power
Mix milk and beaten egg yolks
Add milk and eggs to dry ingredients
Add melted butter
Beat batter just until smooth
Beat the egg whites just short of stiff, not dry
Fold beaten egg whites gently into the batter
Cook in preheated waffle iron
- To know when they’re done, watch the steam. When it slows to nearly stopped, they’re ready.
- For more than two people, I use two irons to get them all out quickly so everyone can eat them while they’re warm.

The Scientific Method is Evolution Writ Large

garyo@oberbrunner.com (Gary Oberbrunner) — Sat, 14 Dec 2024 13:53:36 GMT

The Scientific Method is Evolution Writ Large

Darwin’s theory of evolution was originally inspired by, and applied to, living organisms. He noticed variations between isolated species of finches and other animals and developed the idea of variation, inheritance, and natural selection as the underlying engine of the development of all life. But the fundamental principles that drive biological evolution extend far beyond the natural world. These principles can help us understand cultural shifts, the evolution of language, the progression of technology, the rise and fall of financial strategies, and the development of scientific theories themselves. All of these frameworks rely on generating new possibilities and selecting among them based on their “fitness” for a specific, and possibly changing, environment.

The scientific method was becoming well established by Darwin’s time; Francis Bacon had proposed a systematic approach to scientific inquiry based on induction and experimentation in the early 17th century. By the mid 17th century, natural philosophers like René Descartes and Isaac Newton had refined those ideas, laying out principles of testability, repeatability, and systematic doubt.

When On the Origin of Species was first published in 1859, it sparked intense debate within both scientific and religious circles, garnering immediate support from some naturalists and fierce opposition from others who saw it as a direct challenge to established beliefs. Over time, as more empirical evidence accumulated, the scientific community increasingly embraced Darwin’s theory. Within a few decades, the idea of evolution by natural selection had profoundly reshaped the biological sciences. But it would take until the mid 20th century for Karl Popper and Donald T. Campbell to begin drawing the analogy between natural selection and scientific conjecture and refutation.

Just as genes provide the blueprint for physical traits, cultural “memes” and scientific hypotheses rely on their own variant forms to test their fitness. Fashions spread like genetic alleles, and those that resonate with the mood of the moment propagate while others fade away. Scientific theories, too, are constantly challenged, tested, and refined, generating a landscape of intellectual variation that’s pruned by the selection pressure of empirical evidence.

Variation: The Engine of Adaptation

At the heart of all evolutionary processes lies variation, the raw fuel for change. In biological contexts, variation arises from random genetic mutations sparked by factors like cosmic rays or quantum level shifts, as well as from environmental upheavals—whether catastrophic or gradual. But variation also surfaces in cultural and intellectual domains through education, creativity, fashion, conflict, and generational feedback. These sources of novelty form a sprawling exploration of possibility space. Just as mycelial networks probe their surroundings for nutrients, exploring and discarding fruitless paths, populations, societies and scientific communities are constantly exploring new ideas. Some fail, while others thrive and deepen their roots, opening up fresh territory for growth.

Every system needs a source of energy, a gradient from low entropy to higher. In cultural and scientific systems, randomness can function as a catalyst for creativity and adaptability. Creative ideas, thoughts, even mistakes can expand the possibility space to explore (think of jazz improvisation). These chance occurrences act as sparks—scattered seeds of novelty—that, when combined with existing structures and ideas, can illuminate unforeseen pathways of thought. As Isaac Asimov said, “The most exciting phrase to hear in science, the one that heralds new discoveries, is not ‘Eureka!’ but ‘That’s funny…’”

But variation without structure would lead to anarchy: unbridled creativity with no common language for understanding, and no way to build on previous ideas.

Selection: Shaping the Survivors

Selection in the Darwinian sense has often been misunderstood. Nature does not make choices, and there is no evolutionary “drive” toward a specific goal. Selection is just a consequence of the fact that not every individual in a population will reproduce successfully. In a particular environment, some individuals and sub-populations will be better suited than others. If the environment changes (due to climate change, say) then the fitness criteria change, and the playing field shifts. As long as heritable traits exist, survivors will pass on those traits to their offspring.

Nature does not favor complexity in and of itself; in fact an overly complex organism may well be less suited to its environment – less flexible or adaptable, less energy-efficient, more failure-prone. Bacteria are just as “successful” evolutionarily as humans.

Nor does “survival of the fittest” always mean a contest for supremacy: strength, power or control. It is not always a zero-sum game. Selection in biology must be understood at a population (or genomic) level; the survival of the population may indeed be served by outcompeting poorly suited subgroups, but it may also be due to development of cooperation and altruism, or sensitivity to subtle stimuli. We now have many examples of natural selection where mating and child-rearing behavior favors traits like beauty, novelty, and generosity. As well as helping a population thrive, these traits may lead to flexibility: the ability to react to a changing environment, which may be more of an overall advantage than perfect fit to a narrow niche.

In the scientific method, selection is driven by discourse, test and refutation. A proposed theory undergoes a rigorous examination by others. Does it have explanatory power? Can it be refuted by experiment? Is it as simple as possible but no simpler? Does it contradict known theories? Then anyone can run experiments to confirm or refute the theory. A single counterexample may be enough to discount it, or that may be experimental error, in which case repeatability comes into play. If a theory “stands the test of time” that doesn’t mean it is 100% certain, but as it is subject to more and more scrutiny or selection pressure, it becomes the dominant theory within its environment. When the environment or paradigm changes, as with Copernicus and Galileo in the 17th century or discovery of the quantum properties of energy early in the 20th century, those shifts impose new criteria for success and lead to re-evaluation of existing ideas.

Selection simply acts as a filter driven by the environment. The environment imposes constraints: the complexity of a rainforest ecosystem, shifting climate patterns, or the intellectual rigor of a scientific community. Selection can be sexual or asexual, competitive or cooperative, swift or gradual, and can sometimes boil down to sheer luck. It’s an unthinking, relentless process, leaving traits, genes and ideas that flourish in the current context and discarding those that do not.

In all of these cases, selection is not perfect. Individuals or even populations may survive or die by luck; theories may not be refuted because of inadequate scrutiny or human frailties like politics and dogma. Sometimes traits or ideas appear as byproducts—spandrels—in the cultural or scientific record, rather than being directly favored by selection. Yet these “accidental” features can go on to inspire new avenues of research, creativity, and understanding, enriching the evolutionary tapestry of knowledge.

Embracing the Evolutionary View

Just as living organisms adapt to changing habitats, scientific fields evolve by proposing variations and allowing environmental pressures—like peer review, reproducibility, and technological constraints—to shape which ideas endure. Understanding evolution as a universal process, rather than a strictly biological one, reveals how deeply interconnected all forms of growth and change are. Viewing science through this lens helps us see that intellectual progress is not linear or guaranteed, but emerges from a feedback loop of competing concepts, unexpected insights, and rigorous testing. It reminds us that progress can only occur in an environment that encourages creativity, critical thinking, and a thirst for discovery and knowledge.

Measuring Small Volumes for Humans

garyo@oberbrunner.com (Gary Oberbrunner) — Sun, 04 Aug 2024 13:53:36 GMT

Measuring Small Volumes for Humans

There are lots of reasons we need to measure and compare small volumes in everyday life: cooking, medicine, gardening, and more. In the US, teaspoon/tablespoon/cup/pint/quart is the standard method (as odd and historically-bound as it is), and most people have an intuitive understanding of what a tablespoon of something looks like. But what about smaller things? If a medication says to apply about 1ml, how much is that? I think the worst is to apply a cream or shampoo “about the size of a quarter” because it doesn’t specify how thick it is – especially since creams spread as they settle. Or if you’re reading about Lyme disease, how big is a 2mm deer tick? I expect you can think of more examples like this.

I’ve always thought that using seeds is a good way to get an intuition of the sizes of small things, so I figured I’d make a chart. Here’s a few common things you’re likely to have an intuition about. Turns out a deer tick is about the size of a sesame seed (small larvae may be as small as a poppy seed), and a “dollop” of cream (maybe 2 tablespoons?) is probably the volume of a walnut (about 30-40 ml).

| What | Diameter (mm) | Volume (ml) | Volume (tsp) | |--------------±--------------±------------±-------------| | Mustard seed | 0.5 | 0.0001 | 0.0000 | | Poppy seed | 1 | 0.0005 | 0.0001 | | Sesame seed | 2 | 0.0042 | 0.0009 | | Lentil | 4 | 0.033 | 0.0068 | | Pine nut | 5.5 | 0.087 | 0.017 | | Pea (spring) | 7 | 0.18 | 0.036 | | Corn kernel | 8 | 0.27 | 0.054 | | Raisin | 12 | 0.90 | 0.18 | | Grape | 20 | 4.18 | 0.84 | | Walnut | 40 | 33 | 6.7 |

I hope that’s useful!

(I didn’t put rice grains in here because they vary so much in size, and they’re quite elliptical. But for completeness, US long grain rice is around 0.025 ml, so between a sesame seed and a lentil.)

Welcome to my new new blog

garyo@oberbrunner.com (Gary Oberbrunner) — Sun, 30 Jun 2024 00:00:00 GMT

I’ve been thinking about starting a simple blog to share my thoughts on whatever topics come to mind, and replacing my old website from the '90s. For web apps, I’m a big fan of Vue, but for this static blog it makes more sense to use the very lightweight Astro, so I switched over to it recently.

I started with the excellent astro-erudite template which supports Tailwind CSS classes, icons:rocket:, math rendering like $y=x^3$, excellent syntax highlighting for code, and dark mode support (see the header). Then I modified that to get a theme I like.

Some of the images here I made with Midjourney, and some with DALL·E. AI is a fascinating creative tool, but I’m struggling with the serious ethical issues around using people’s content without consent to train the AI. I do check that my images don’t show up in Have I Been Trained.

Climate Change and Personal Responsibility

garyo@oberbrunner.com (Gary Oberbrunner) — Tue, 21 Nov 2023 00:00:00 GMT

"I wish it need not have happened in my time," said Frodo. "So do I," said Gandalf, "and so do all who live to see such times. But that is not for them to decide. All we have to decide is what to do with the time that is given us."

— J.R.R. Tolkien, Lord of the Rings

Climate change has become the most pressing existential challenge of our era. As temperatures soar across the world, the polar ice caps and glaciers are melting, sea levels continue to rise, oceans acidify, extreme weather events intensify, forests burn, and rivers and lakes shrivel and die. We are close to or have already crossed multiple climate tipping points.

Because of these changes, coastal cities and islands are becoming uninhabitable, soon to be joined by large swaths of equatorial land. Food and water insecurity will intensify, threatening millions of humans with hunger, disease, and deprivation, and animals with habitat loss and extinction. Mass migrations will become routine, with hundreds of millions seeking refuge from uninhabitable terrains over the next few decades. Acidification endangers most marine ecosystems, and major ocean circulation systems are showing signs of collapse within this century. And beneath these crises brews a storm of political unrest, as nations grapple with the economic, political and social repercussions of the changing climate.

Amidst this panorama of impending crises, many people are understandably feeling anxiety, dread, despair, and hopelessness. My hope in this article is to show that individual and collective action is not just possible, but it’s imperative: it’s our only beacon of hope.

Individual Actions with Multiplicative Impact: What Can One Person Do?

To be clear, the biggest drivers of climate change are in industry, and industry and governments are also the ones with the largest power to get us back on track (or prevent progress). Without the countries of the world pulling together and forcing corporations to behave responsibly, it is going to be a difficult challenge.

Individuals should accept no guilt for the situation we are in; people in power have known for thirty years and more what needs to be done, and have done next to nothing, and in many cases have actively worked against fixing climate issues. Even today as we are finally moving in the right direction with initiatives such as the Biden administration’s Inflation Reduction Act investing hundreds of billions of dollars in decarbonization, solar and wind energy production, EV tax credits, infrastructure modernization, grid resilience and environmental justice, we are far from the Paris Accords target of staying under 1.5°C, and many large companies and governments are dragging their heels, broadcasting misinformation, and undermining progress.

Because of this, individual action often feels microscopic and maybe even pointless, a mere drop in the ocean. However, history shows that the accumulated actions of many can create tidal waves of change. The civil rights movement in the USA, Mahatma Gandhi in India, the Velvet Revolution, and many others are all examples.

Also, it’s important to remember that targets like 2° are not absolute limits where we’re fine below them, and extinct above. Every action we take to reduce CO₂ emissions, no matter where we are or what trajectory we’re on, helps forestall even worse outcomes. There is no justification for doomism or paralysis.

Here are a few ways individuals, leveraging their unique skills and positions, can amplify the fight against the climate crisis.

Learn More

Educating yourself is perhaps the most important thing you can do. Learn what is actually happening so you can sort truth from misinformation. Find out who is doing work you agree with. Learn about the science and what is being done to combat the crisis, on all fronts: scientific and engineering, social, legal, and economic. Read about climate initiatives and promising new directions. All of this will help you understand that all is not lost; every step in the right direction is important to stabilize the world and prevent the worst outcomes.

Educate with Accuracy and Kindness

In a world rife with misinformation, there’s immense power in sharing accurate knowledge. Whether you write an article, compose a song, create a film, design a building, create software, or educate directly, it’s crucial to be clear, accurate, and compassionate. Disseminating correct information kindly helps dispel myths and mobilizes informed action.

Create Open Dialogues

Conversations, even simple ones among friends, can be transformative. Communicating about climate change, its implications, and our collective role sparks awareness, fosters understanding, and builds solidarity. It’s through these personal connections that societal shifts often begin and gather momentum. Share your worries, your successes, and your ideas.

Leverage Social Media

In today’s world, social platforms are huge amplifiers. Sharing information, success stories, and challenges related to climate change can reach thousands, if not millions. Social media campaigns today can bring attention to issues at a pace traditional media can’t match, and individual voices can go viral at any time. Of course climate denialism is rampant on social media, fed by sources of mis- and disinformation. Countering that with truthful, deliberate, evidence-based information is vital.

Support Green Initiatives

Voting with your wallet means prioritizing products and services that align with eco-friendly practices. Adopt a more plant-based diet, drive less, recycle more. Over time, as consumer demand shifts, businesses are more likely to adopt sustainable methods. It’s a classic case of supply and demand: when consumers demand greener products and services, the market responds.

Plant Some Trees

A mature tree absorbs at least 22 lbs of carbon each year, depending on location and type. Planting a tree isn’t just good for the environment though. It’s a forward-looking, long-term optimistic statement, and it provides shade and habitat for birds and insects.

Replacing a lawn with ground cover, reducing water usage and planting native gardens all help as well.

Invest in Sustainable Technologies

As technologies advance, options like solar panels, heat pumps and electric cars become more accessible and practical. Investing in these reduces carbon footprint and also sends a clear message about market demands. In many cases, you’ll also save money in the long run – and even more as fossil fuel subsidies vanish and the cost of fuel rises.

Consider Where You Keep Your Money

If you invest, consider what industries you are lending your money to. Voting with your wallet can be a powerful driver of economic change. Many mutual funds are divesting from fossil fuels because they recognize the structural inability of the fossil fuel and related sectors to continue delivering returns.

Leverage Your Professional Skills

No matter who you are, you can contribute. Scientists and engineers can turn their expertise towards green energy and sustainable infrastructure. Writers and artists can use their platforms to highlight environmental concerns, weaving compelling narratives that educate, resonate and inspire. Lawyers can advocate for local green infrastructure and environmental justice initiatives. Students can raise awareness on campus and in the wider community, and study and work toward sustainability. Architects can promote the highest LEED standards. Landscapers can switch to electric machines. CEOs can require carbon neutrality in their companies. And all of us can use our professional networks to set an example of public support for climate initiatives.

Lobby for Systemic Change

Engaging with political representatives, especially if you live or work in a state resistant to green policies, can make a real difference. Attending town halls, writing letters to your congressperson and meeting with representatives can influence legislative priorities and amplify the voice of environmentally conscious citizens. And of course, please vote for and support politicians who prioritize the environment and climate justice initiatives both locally and nationally.

There are many vital issues in the world; climate change is not the only one of course. But the line between, say, reforming campaign finance or preventing voter disenfranchisement and climate action is pretty direct.

Take Nonviolent Direct Action

I hesitated to put this in here, but history has shown the potency of nonviolent movements. Research suggests that when even 3.5% of the population actively engages in protests, significant political changes ensue. Mobilizing a fraction of the community in peaceful protests can elicit substantive policy shifts. Greta Thunberg’s climate strike is a great example. But make sure the action is directed toward those who need to hear what you have to say and focuses on positive actions and outcomes. Violent, disruptive protests only undermine the movement and harden opposition.

Empower Youth

I have two young adult children. I often come near to despair, thinking of the world we are leaving them. But I’m inspired to action on their behalf, and on behalf the the generations beyond them. Youth are the future of humanity, and success or failure will be in their hands. Educating and empowering them ensures they have the tools to use their passion, knowledge, and strong community ties to find novel solutions and renewed vigor in the fight against climate change.

Collaborate Across Borders

Climate change is a global issue. No single country can solve it, and the countries most affected are largely the ones who contributed least to the problem. Educating ourselves and others about work going on in other places can inspire cooperation and a sense of solidarity.

Celebrate Successes

It’s crucial to recognize and celebrate the milestones, no matter how small. It provides motivation and showcases the results. Don’t be afraid to talk about what you are doing to help, and take the time to celebrate your own efforts as well as those of your family, your community, and your friends.

Take Personal Actions with Societal Impacts

While individual measures like adopting a plant-based diet, driving less, transitioning to electric vehicles, or installing energy-efficient appliances like heat pumps seem minuscule and may appear to be merely virtue signaling, their collective impact is profound. These actions, when reflected in national statistics – like rising demand for EVs, increased ridership on public transit and in bike lanes, or declining consumption of energy-intensive goods – send strong signals to policymakers. These quantifiable changes highlight shifting consumer patterns, pushing industries and governments to re-evaluate their strategies. They can also be the beginning of productive conversations in the community and beyond.

It is hard not to become despondent when thinking about the looming crises posed by climate change. A sort of paralysis can set in. What can anyone do in the face of the entire Greenland ice sheet melting? And to a certain extent that is true. There’s only so much a person or a family can do individually. But I say we must not, we cannot, give up. The stakes are too high.

While individual actions in isolation may seem inconsequential, their combined societal impact can be immense. From professionals leveraging their skills for sustainable ends to everyday people making eco-conscious choices, each effort plays a part, and the ripples of those choices can spread widely. When we realize our collective potential, the challenge of climate change can become not an insurmountable obstacle, but a rallying cry for unified action.

Some Resources to Learn More

Thanks to Dr. Katharine Hayhoe for invaluable contributions to this article via Mastodon. Her Instagram and climate change website at https://dontlookup.count-us-in.com/ have some good resources for people who want to get involved.

Data:

https://climatereanalyzer.org/clim/t2_daily/?dm_id=world – Daily world temperature
https://climatereanalyzer.org/clim/sst_daily/ – Daily sea surface temperature
https://nsidc.org/arcticseaicenews/ – Arctic and antarctic sea ice extent
Show Your Stripes – Ed Hawkins at NCAR, source for the main image for this article.

Education, News and Activism:

https://www.climatechangenews.com/ – good daily climate news site, reporting on high-level worldwide activity
https://carbontracker.org/ – making carbon costs visible economically
https://climatescience.org/ – courses, activities, community
MIT Climate Portal – lots of resources, explainers, news
- Climate Science, Risk & Solutions – MIT climate knowledge base, by Dr. Kerry Emanuel
Wolfgang Cramer – director of research (CNRS) at the Mediterranean Institute of Biodiversity and Marine and Continental Ecology. Climate, biodiversity, transport.
Andy Scollick – climate and policy
Jan Rosenow – energy transition. Director at the Regulatory Assistance Project Research at Oxford University, SPRU & FU Berlin

On Entropy and Passphrases

garyo@oberbrunner.com (Gary Oberbrunner) — Fri, 13 Jan 2023 23:59:30 GMT

On Entropy and Passphrases

Passphrases have been in the news again lately, because of a big data breach at Lastpass, which may have been more serious than Lastpass first revealed.

That got me thinking: how secure is a passphrase vs. a random password?

Passwords

I assume you use random passwords generated by a password manager like Lastpass, Dashlane, 1Password, or Bitwarden (which I use). If you have it set up to create 12 or more characters with upper and lower case, digits, and punctuation, those passwords should be very secure.

If you create your own memorable passwords, they are certain to be less random, and thus more guessable. For instance many attackers start with lists of common English words (and permutations thereof), and known common passwords. There are many fewer possibilities than if you use a random password, so it’s not recommended. The downsides are you can’t remember random ones, they’re hard to type, and even harder to type on a small screen.

Entropy

The standard measure of how hard a password is to crack is entropy, measured in bits. A single coin has one bit of entropy: it can be up or down. A six-sided die has about 2.5 bits of entropy: $log_2(6) \approx 2.585$. As long as an attacker doesn’t have any special info, the entropy is directly related to how long it would take them to guess by “brute force” – checking all possibilities. Each additional bit doubles the amount of time it would take to guess.

A random password composed of 26 lowercase characters, 26 uppercase, 12 punctuation, and ten digits gives $log_2(74) \approx 6.2$ bits per character, or a total of 74 bits for a 12-digit password, or 86 bits for a 14-digit password. This is plenty for “regular people” (if you’re not a journalist, activist, wealthy, or other high profile target). If you are a high-value target, you need at least 128 bits of entropy, or about 20 random characters. It doesn’t hurt to use a password that long for your most important accounts, such as your bank. Just hope you never have to type it in.

But what about passphrases?

Passphrases

The best practice for passphrases is to use a few words mashed together, sometimes with spaces and/or digits intermixed. The XKCD “correct horse battery staple” one is a famous example.

But how good are they really?

A passphrase is certainly longer than a 14-digit random password. So if the attacker doesn’t know how you came up with it, and they have to try all possible combinations of random passwords up to the 28-character length of “correct horse battery staple”, the entropy is huge: $log_2(74)*28 \approx 173$ bits!

But if passphrases become popular, attackers will try them first, rather than trying everything willy-nilly. In cryptography, it’s always safest to assume the adversary is as smart as possible.

A common word list used by passphrase generators is diceware; that list has 7776 words. (It’s called diceware because you can generate words from it by rolling dice using a simple process.) That gives about 12.9 bits of entropy per word, or a bit more if you capitalize randomly or stick a digit or punctuation in here or there.

In this case, a four-word passphrase has only 51 bits of entropy, and a five-word phrase has 64. A six-word phrase has 77 bits, a bit better than a 12-digit random password.

Six words is a lot to type, for sure, but it’s certainly easier to type “city linguini muster hatbox juiciness caretaker” than 9xpzw%T5q^oh. You can also tell it to someone over the phone (good luck even knowing i from l and 0 from O in random passwords). But for me the biggest advantage is for TV streaming services, where I sometimes have to type the passphrase into those horrible hunt-and-peck keyboards on the TV.

For these, a set of nice lowercase words is so much easier. And if you use five or six diceware words, it’s as secure as a random password. Bitwarden, the password manager I use, will generate diceware passphrases for you, so it’s no more work.

So… Should I Switch?

I thought about switching to passphrases for many online accounts, but after thinking it through, I plan to stick with random passwords for most, and use passphrases only when memorability or typeability matter.

I suggest using a long randomly generated passphrase as your master password for your password manager, with a few digits and punctuation chars mixed in. You will have to type that fairly regularly, and it is your highest value target (since it unlocks everything else), so, to state the obvious, don’t use your mother’s maiden name or your birthday for it. Choose wisely.

Two Factor Auth

No matter what kind of password or phrase you use with your online accounts, you should always enable “2FA” or two-factor authentication. Whether it’s SMS to your phone, an authenticator app, or a physical security key, anything is better than nothing: don’t forget that “brute force” password cracking is the last resort of the bad guys. Social engineering, looking over your shoulder, malware, or sniffing unencrypted network traffic are all ways passwords can become exposed. If you have 2FA enabled, those passwords become a lot less useful. So be safe out there!

Looping in Javascript

garyo@oberbrunner.com (Gary Oberbrunner) — Fri, 13 Jan 2023 00:00:00 GMT

Looping in Javascript and Typescript

Javascript and Typescript have way too many ways to iterate and loop over arrays, objects, and iterables. In this post I try to collect them and make some sense of them.

Construct	Effect on array	Effect on object	Effect on iterable (Map, Set)
`for (let i=0;` `i<obj.length;` `i++)`	★Iterates over indices	N/A	N/A
`for x in obj`	Iterates over indices and custom props (including prototype props), in random order(!)	★Iterates over enumerable obj props (including prototype props)	Iterates over `Map` keys; doesn’t work for `Set`
`for x of obj`	★Iterates over values (in order)	N/A [^notiterable]	★Iterates over `[key, value]` for Map, elements for `Set`
`obj.forEach(x => ...)`	★Iterates over values (also passes index & array to callback) [^nobreak] [^foreach-await]	N/A [^objforeach]	★Map: iterates over values + keys [^nobreak]

Notes:

[^nobreak]: .forEach() has no way to break out from the middle of the loop; the others ways of iterating over an array or object support break and continue.

[^objforeach]: obj.forEach() does not work on an Object, but you can do Object.keys(obj).forEach(...). That gets the keys of the object as an array, and then loops over the array of keys.

[^notiterable]: gives this error: “Error: obj is not iterable”.

[^foreach-await]: arr.forEach() does not work properly with await calls in the callback, but the others support await in the body.

Javascript Promises and Async/Await

garyo@oberbrunner.com (Gary Oberbrunner) — Fri, 15 Nov 2019 00:00:00 GMT

Intro

I recently came to Javascript from too many years in a C/C++/Python world. Javascript is very familiar in many ways, and different in quite a few as well (see: all the ways to import/export in Javascript — not that Python’s module/import system is so great either.) As a systems-level programmer, I like to know how things work all the way down to the metal. It’s satisfying, but it also really helps with debugging.

One thing I didn’t really understand at a deep level is how Promises and async/await work in Javascript. The examples I read all had to do with waiting for some kind of I/O or a timeout. I wondered “what’s the magic that makes the I/O or timeout wake the main thread back up?” I didn’t know whether that was part of the architecture or just a common-enough use case that nobody bothered talking about other event types.

It turns out there’s no magic, but there’s significant code in the Javascript runtime to make all this work. So now that I figured out how all this stuff works, here’s a working example that will hopefully illuminate some dark corners for those of you, like me, who like to know how things really work rather than just “write it like this and it’ll be fine.”

Event loops and threads and async, oh my!

A few things to know before we start into the example:

Javascript is single-threaded. Promises and async/await are not a replacement for multithreading; you still are only running on one core.
There’s an event loop always running in Javascript. In the browser, it’s run by the browser itself to process user input and network events. In node.js, it’s what runs the functions you specify, and when there’s nothing left to do, it exits. (In recent Pythons there are various event-loop implementations, but they’re layered on — and in C++ it’s a free-for-all as usual.)
Async functions start running immediately when you call them, but when they get to an await on a Promise, that creates a closure of the current stack state, with all local variables and the whole execution context, and that promise + closure gets put on a list of pending functions (more detail below). The event loop runs any “resolved” Promises whenever it gets control back from user code. In this way, it’s like a python generator calling yield. Our example is of a promise that can get woken up from anywhere by calling a function. Here it is. You’ll want to check it out in the JSFiddle.

Our example is of a promise that can get woken up from anywhere by calling a function. Here it is:

You should be able to see the source and its output in that window.

Analysis

So what’s going on? One step at a time:

The functions wakeable and handle_event get defined
we call handle_event(), which starts to run.
At the line await wakeable(), Javascript first calls wakeable() (which will return a Promise), and then passes that Promise to await. At that point that Promise, and the current execution context, gets pushed onto a queue for the event loop to check later. But first, how does the Promise get created in wakeable()?
The Promise constructor takes one arg, a function which itself takes two args (resolve and reject). The promise constructor calls (right now, synchronously) that anonymous function, passing it its own internally-created resolve and reject methods as args. (When we finally get around to calling that resolve, it’ll mark the Promise as resolved.) In our case, the function creates another anonymous function which calls the original resolve (which, remember, was passed into us — it’s internal to Promise), and assigns that function to the global var wakeup. So later when we call wakeup it’ll call the Promise’s resolve method. Whew! Got all that? (It would be harder to make these non-anonymous functions, because they need to be closures to get the original resolve.)
OK, back to top-level. After the call to handle_event returns (it’s async, and the continuation has been put on the queue, but in the main thread, it returns normally), the next thing is we call wakeup().
wakeup is now a function — we created it in step 4. It just calls the Promise’s resolve method. All that actually does is set a flag on the Promise (which is saved on an internal queue in the JS runtime), saying it’s now ready, and it’s resolved successfully. It also saves any value we pass into resolve as the return value you get from awaiting the Promise, or in a .then call. (You can reject as well.)
Next, we set a timeout that will call wakeup again after a while. (I’m not going to cover timeouts here; basically the go on a similar execution queue in the runtime.)
Now we’ve reached the end of the source file. Time to exit, right? Not so fast. There’s still a pending Promise on the queue, so the runtime sees if it’s ready.
It is ready, so the Promise runtime then calls all its .then functions and await continuations — in our case just completing the first iteration of the while loop in handle_event, which loops back around and stops on the next await, creating a new continuation and returning to the runtime’s event loop.
At this point the only thing left on the execution queue is that timeout. The JS runtime waits out the clock, and then calls wakeup again.
As before, wakeup resolves its Promise, which just sets a flag and returns. Now the runtime gets control again, sees that it has a resolved Promise, so calls its continuations, which takes us around the while loop one more time, and back to its await.
At this point there’s nothing left to do; there aren’t any Promises or timeouts or anything. If you run this code in node.js or jsfiddle, it’ll exit. In the browser, the page will just sit there waiting for user input. And that’s all, folks!

More about async and await

All async functions always return a Promise. If you write async function foo() { return 1 } it will actually return a resolved Promise with a value of 1. If your async function has an await, the returned Promise will only resolve when the await has finished, and the rest of the code in the function has run. But notice that the async function returns to the caller immediately. There’s no waiting going on, ever. Now if the caller awaits the async callee like this:

async function callee() {
   return 100;
}
async function caller() {
  let val=await callee()
  await new Promise(resolve => setTimeout(resolve, 100)); # 100 ms
  return val+1
}
async function parent() {
  let val=await caller()
  return val+1

then the same thing happens: the caller returns (immediately) a Promise to its parent that only resolves when its await returns, which only returns when that callee’s await returns, and so on all the way down. There’s essentially a call graph that gets built up so that whenever the runtime loop gets control back, it calls the next thing that’s runnable in that call graph.

Note that in this example, parent calls caller, caller calls callee, which returns its value — all this happens without waiting, synchronously. The first await is the one in caller; that puts the rest of caller’s code into a continuation and returns. Similarly parent puts the rest of its code after the await into a continuation and returns.

If you call an async function without awaiting it, it’ll return its Promise which you can wait on later, or not. If you don’t, the rest of the function after the first await will still get run eventually, whenever the event loop gets around to it. Here’s a nice example of that:

async function f1() { await something long... }
async function f2() { await another long thing... }
async function callParallel() {
  let p1 = f1() # a Promise
  let p2 = f2() # another Promise
  await Promise.all([p1, p2])
}

In this case f1 and f2 both get their continuations set up, they return promises to callParallel, which returns, and then when (eventually) both f1 and f2 resolve, the Promise.all resolves and the last await returns and callParallel’s implicit Promise resolves (but nobody’s checking that).

Some notes

A promise is resolved when its resolve function is called; in this case calling wakeup() calls the promise’s internal resolve() method, which triggers any .then methods on the next tick of the Javascript event loop. Here we use await, but .then(…) would work the same way.
There’s no magic; I/O and timeout promises work the same way. They keep a private registry of functions to call when the I/O event or timeout happens, and those functions call the promise’s resolve() which triggers the .then() or satisfies the await.
By the way, unlike async in python, leaving a pending promise “open” when the process exits is perfectly fine in Javascript, and in fact this demo does that. It exits when there’s no more code to run; the fact that the while loop is still “awaiting” doesn’t keep the process running, because it’s really just some closures stored in a queue. The event loop is empty, so the process exits (assuming it’s in node.js — in a browser it just goes back to waiting for events). In python, this would print an error as the process exits — you’re supposed to clean up all your awaits there. Javascript is more forgiving.