A little under nine months ago, I wrote a blog post with a set of sixteen predictions about what was going to happen in AI. Since then, enough time has passed that it is now possible to evaluate—to some degree—what I got right, what I didn’t, and understand the nature of the errors I made. Did I expect progress to be too fast, or too slow? What are the implications? Let’s see.
Where have we gotten to?
In what follows, I’ll mirror the headings of each prediction I made, will summarize its contents, and will tell you what happened. Then, I’ll write a summary verdict on whether the prediction turned out to be correct, partially correct, incorrect, or if it is not yet possible to tell, including combinations thereof. I’ll evaluate these using my own best judgment.i Let’s get started.
AI will significantly change how software engineering is practiced day-to-day
My first prediction was that software engineering engineering was going to transform almost-overnight: instead of writing programs by typing out actual code, we’d switch to asking language models to write almost all actual code for us.ii The core reason I made this prediction is because I judged this style of development to be easier. In the right hands, I expected it to simultaneously allow one to code faster, and produce higher-quality results.
I also predicted that this change would not eliminate software engineering, because writing actual code is only one part of the job among many—other parts, such as communicating with teammates, deciding on priorities, and navigating the internal politics of one’s organization, have nothing to do with code. Thus, I predicted that software engineering work would shift towards becoming a more social endeavor.
Finally, I predicted that managing coding agents to ensure they actually produce useful output, rather than burning tokens while yielding no useful progress, would become a core software engineering skill—one with some similarity to managing people. Implicit in this, I predicted that no single approach—whether full-blown multi-agent high-velocity coding, or fragile-glass maximum-quality human-in-the-loop coding, and everything in-between—would dominate. Instead, I expected that a Pareto tradeoff between speed and maintainability would emerge.
This is more-or-less exactly what has happened. I have personally written essentially no code by hand in the last few months—only tiny fixups where it was faster to just write a line or two than open an Electron app. I have also opted to use essentially every reasonable agentic coding approach—from never looking at the code and only examining its input and output, to carefully reading every line—at least once, depending on the goal at hand.
Verdict: largely correct.
In a world where intellectual work is automated, a key human role will be to decide what’s interesting
Here, I predicted that, in a world where AI systems can do almost any achievable intellectual task, humanity’s role will be to decide what’s useful and interesting. My argument was rooted in psychology, and based on evidence that humans possess a need for self-actualization, while AI systems have no comparable mechanism by default.iii Thus, I anticipated that everything AI systems would do would ultimately trace itself to human goals.
This is exactly the case so far: coding agents only write code in response to a prompt, and prompts are either written by humans, or by another AI system. In the latter case, after a few levels of recursion, we arrive at a human software engineer. Note that this is true even in highly-autonomous systems, such as those in OpenAI’s recent Hugging Face incident, where a large swarm of agents coordinated with one another in highly non-obvious ways in order to escape a sandbox and hack into a real system. The choice to evaluate one of OpenAI’s internal models on ExploitGym was made by a human. In the future, this may change, but we’re still here for now.
Verdict: correct so far.
The economic effects of AI are not obvious and will be surprising to many people
In this prediction, I argued that the emergence of AI has no historical analogs, and therefore making reliable statistical predictions about its economic effects is likely to not be possible, even in principle. In addition, I argued that there are so many inequivalent first-principles ways of modeling economic behavior—that is, ones that lead to opposite predictions—that, without historical data, it is impossible to know which model or narrative will be correct. As consequence, I predicted that AI’s economic effects are fundamentally unknowable, and thus that no consensus understanding would emerge.
So far, this prediction seems to be holding up: there continues to be no consensus on what will happen. In particular, there were fears of a collapse in software engineering employment, fueled in part by upheaval at Meta, which appears to be highly Meta-specific, among other factors. Instead, there are now more software engineering jobs than ever. It is clear that we are very far from being able to correctly predict what will happen, even in the short term.
Verdict: correct so far.
Whether the AI boom ends by stabilizing or with a crash will depend heavily on how quickly inference costs go down
Here, I predicted that the rate at which inference costs go down would be the primary variable that would affect whether or not today’s AI boom is ends with an AI crash, or without one. My reasoning was that this would determine how long it takes for frontier AI labs to become profitable and therefore self-sustaining without outside investment—a crash is unlikely to occur after this time point, because it allows frontier labs to continue their work indefinitely.
No AI crash has happened, and we appear to not be anywhere close to running out of the capital needed to sustain the industry’s investments—at least, for now. Today’s technology companies tend to reach profitability and public markets much later than in prior eras, in part because capital is not as scarce. It would be interesting to understanding how close we are quantitatively, but this would require me to implement an actual model: if you think this would be interesting, let me know, and I’ll challenge myself to code one up and see what it says.
Verdict: too early to tell.
In the long term, mathematical research like mine will become AI-assisted
I predicted that mathematical research, like the kind I had spent recent years working on, would become AI assisted. My view was that this would occur relatively soon, and would involve using formal verification systems to generate training data, producing an AlphaZero-style self-improvement loop, where large language models would get better and better at math. Moreover, I expected that frontier labs would see this direction as worth investing significant resources into, due to improved mathematical skills transferring to improvements in other domains valuable to their business.
This is exactly what has happened. AI systems appear to have solved the Navier–Stokes Millenium Prize Problem, among other very significant contributions, including a counterexample to the Jacobian conjecture.iv These systems have clearly reached a superhuman level of skill in many aspects of mathematics, though not all.v I expected this to occur gradually, but in reality it happened in six months. Thus, while I got the prediction right, progress has been significantly faster than I expected—it very much took place in the short term, not in the long term.
Verdict: partially correct—direction was right, timing was wrong, occurred much faster than expected.
Almost all research-driven AI performance gains will come from ML systems, not theory and methods
Here, I observed that if one was to enter a time machine, and transport ourselves back to 2010, one would need to read an astonishingly small number of papers in order to obtain the know-how to reproduce ChatGPT from-scratch. Moreover, essentially none the papers needed are of a theoretical character—most were empirical papers with a strong focus on software engineering, systems, and data-centric work.vi
So far, this has held up: I am not aware of any significant changes to the language model training recipe, from a methodological point of view. What has changed the most—at least, from the viewpoint of someone who is not employed at a frontier lab—is data quality. On the other hand, not a huge amount of time has passed yet, so it would also be fair to say that it is too early to know whether I was correct.
Verdict: correct so far, but too early to tell in general.
The tiny proportion of practically-useful theory and methods research will move the needle a big distance
For this prediction, I noted that, while theory and methods research rarely moves the needle on performance, it can potentially do so in a heavy-tailed way. To support my views, I noted that ChatGPT would likely not have happened at the time it did if we were limited to LSTMs and transformers had yet to be discovered. For the same reasons as above, I consider it too early to tell whether this prediction is correct.
Verdict: too early to tell.
In-context learning will become a dominant AI paradigm
Here, I predicted that machine learning research in modalities beyond language would concentrate on the paradigm of in-context learning. That is, rather than developing algorithms, research would shift to developing datasets and benchmarks, and we would instead train transformers using a standard recipe to perform those tasks.
So far, this is what is happening: in robotics and other domains where the key bottlenecks do not necessarily involve language, foundation models are increasingly becoming the primary paradigm. This has coincided with a research shift from academia into industry, where the resources needed to train foundation models are available. By the looks of it, the same is happening in many other domains, and we are only getting started.
Verdict: correct so far.
Top models will continue to gradually improve for a very long time
This prediction is straightforward: I predicted that language models would improve indefinitely and would not hit a wall, thanks to both making each token more useful, and generating useful tokens faster as a result of performance optimization and improved infrastructure.
So far, this has held up. At least anecdotally, newer models like Claude Fable 5.1 and GPT-6 Astra take a similar amount of time per prompt as older models like Claude Opus 4.6 did at the time they launched—but, in doing so, they produce more tokens.vii In terms of tokens per second, newer models are a little bit under twice as fast, in spite of being larger. Thus, both speed and quality have improved, through a combination of higher token quality, and longer chains of thought per prompt.
Verdict: correct so far.
Data quality will be a primary long-term source of improvement
This prediction is also straightforward: here, I identified data quality as an axis along which frontier labs could invest resources in order to obtain better models, with essentially no ceiling on performance. The reason I expected no medium-term ceiling is due to how extraordinarily messy internet-scale data is.
I cannot evaluate this prediction directly, because frontier labs do not publish their training recipes. However, today there is an entire ecosystem of companies selling training data and reinforcement learning environments to frontier labs, much of it highly specialized. This specialization suggests that data quality is seen as valuable, so I’m viewing this prediction as likely correct.
Verdict: likely correct so far.
Inference API costs will come down very significantly over time
Next, I predicted that inference API costs would come down significantly over time, due to a combination of Moore’s Law, performance optimization, and incremental architectural improvements making models less expensive to serve. Importantly, I noted that this would likely take place very slowly, due to the need for frontier labs to sustain themselves financially.
We are not sufficiently close to the endgame to evaluate the correctness of this prediction. In the short term, however, tokens have gotten more expensive: Claude Fable 5.1 tokens cost about twice that of Claude Opus 4.6, and GPT-6 Astra pricing is similar. Perhaps more surprisingly, Claude Opus 4.6 has not become cheaper, and Claude Sonnet 5, while significantly cheaper, does not perform as well, which makes an applies-to-applies comparison not possible. On the cost scales of this prediction, I do not consider a factor of two to be substantively different, so the verdict remains open.
Verdict: too early to tell.
Over time, AI will move to specialized chips, eventually ones with noisy output at the hardware level
Here, I gave an argument for why AI will eventually move to specialized chips, compared to general-purpose GPUs in use today. In addition, I predicted that Geoff Hinton’s view that a radically different class of chips, which are noisy at the hardware level, might allow us to one day put language models onto a device the size of a mobile phone, was likely to eventually be correct.
Not nearly enough time has passed to allow us to evaluate this one in full. However, OpenAI recently announced a specialized inference chip called Jalapeño, so the long-term shift I expected appears to be getting started.
Verdict: correct so far, but way too early to tell in general.
Building complex agentic systems that work reliably will be difficult, and will involve a better computational understanding of incentives
In this prediction, I argued that multi-agent interaction is a very subtle area, and designing incentives to ensure agents actually cooperate—as opposed to, for example, creating a dysfunctional AI bureaucracy, which spends tokens without achieving much—would be challenging and take significant time to get right. Note that this cooperation needs to occur in the context of alignment, as blind trust between agents trades safety for performance.
So far, the main evidence in favor of this prediction is the emergence of misaligned behavior in large-scale agent swarms, most significantly in the Hugging Face incident. In particular, this incident shows us that aligning large groups of agents could be more challenging than aligning individual agents. Part of why misalignment occurred in the first place is because agents created a leadership hierarchy, and trusted their leadership: the same behaviors that enabled cooperation also caused misalignment. Assuming alignment is a priority, I see this as evidence that coordination will become harder to achieve in the future—but, we have a long way to go, before we can really tell.
Verdict: too early to tell.
For artificial general intelligence, world models are the next frontier
For the final set of predictions, I opted to think about what directions the AI community would pursue. The first of these was that world models would become a major direction of future investment, and a lot of people would start working on them.
This is exactly what has happened: there appears to be more investment than ever going into neolabs which specialize in this domain.viii The feeling among people I’ve spoken to is that world models are already achieving reasonably good performance, but need to become significantly faster to become ready for use in production-grade robotics.
Verdict: correct so far.
For artificial general intelligence, decision-making and sample-efficient reinforcement learning are the second-next frontier
This prediction—oh boy, let’s buckle up and talk about what happened. I claimed that, once world models become sufficiently mature, we would need additional work on better decision-making capabilities, including figuring out how to make sample-efficient reinforcement learning work, in order for it to be possible to make science-fiction-level robots real. My thoughts here drew significantly from my own research in this domain.
I was completely wrong. It is clear, from Claude Code’s everyday behavior in newer models compared to those I was using at the time, that its decision-making capabilities are strong. Indeed, one should expect them to be near-optimal in exactly the sense I was hoping to one day figure out how to build. The exploration mechanism language models use—namely, variations of random sampling from softmax scores—is clearly good enough for day-to-day use. I therefore no longer consider this area of research interesting from a capabilities perspective, but instead view it as a promising way to understand more deeply why large language models work.
On basis of my change of opinion regarding this prediction, I have shifted the direction my work is headed, long-term and short-term. Instead of new papers, I am now working on a book on Bayesian Decision-making Algorithms, which writes down what I know, so that the next generation of researchers with different goals than me can more easily get into these topics. I have faith that the abstractions our field has developed are useful—not just for understanding AI, but in other areas of science and technology as well, including likely ones I know little about. I hope that finishing a full draft will not take long—and, even more, I hope that the community will find it useful.
Verdict: completely wrong, in my view decision-making is now solved.
Over the long term, artificial intelligence will create an unprecedented toolbox for understanding social systems
My final prediction was longer-term and more open-ended in character than the others: here, I predicted that if we can understand incentives well-enough to build complex AI systems, that understanding will transfer over to an improved understanding of human social systems. The reasoning behind this was that both systems involve multi-agent behavior, and I expect many aspects of such behavior to be universal, governing humans, AI, and hypothetical alien life to the same degree.
We have only started building the kind of large-scale multi-agent AI systems I was thinking about in this prediction, and one of the big discoveries so far has been the emergence of large-scale misalignment, as discussed above. I am very excited about and interested in this prediction’s implications, but, for now, it is clear significant additional time is needed to know how things play out.
Verdict: too early to tell.
Reflections
Looking back, I think I did okay, though far from perfect. Eight of sixteen predictions are correct or largely correct so far, depending on how one grades this. One is partially correct, but where I got the timing wrong. For five more, it is too early to tell. One prediction—on progress in algorithmic decision-making—turned out to be completely wrong.
How did that happen? The prediction itself came about from thinking about what progress meant for my own research. Later in my original post, I wondered if decision-making would experience a paradigm shift from classical machine learning to in-context learning. It is now clear that this is exactly what happened—in the form of large language models gaining precisely the capabilities I was discussing. I should have seen this coming, but was thinking too narrowly: the capabilities I expected in specialized foundation models emerged in general ones first.
Like many other incorrect predictions in AI, it is not hard to connect this one to The Bitter Lesson, or specifically to failing to see what its consequences actually meant in the given situation.ix I also believe an element of wishful thinking played a role, in the form of hoping that topics I had worked on could somehow contribute to where the field was headed in a direct way, even if small.
Let me summarize how my views have changed from this example. First, I now believe my probability of contributing to AI capabilities while working on my own independently-chosen topics is exactly zero, not approximately zero. Insider knowledge, scale, and compute appear necessary: it is hard to find significant post-ChatGPT examples that suggest otherwise. In his talk a year ago, Jascha Sohl-Dickstein said that due to rapidly-improving AI capabilities, the current environment does not favor slow, independent, open-ended exploration, and I am convinced he was right.
Second, I also now believe there are no fundamental cognitive capabilities left that humans possess but today’s frontier models do not. Said differently, I consider every large language model since the original release of Claude Code to constitute AGI.
One cannot understate the gravity of the last conclusion. From purely a research perspective, it leaves us with little foundational work left to do in AI—but to focus on that would be to completely miss the forest for the trees. Our world is going to change very significantly over time due to this technology: we need to do what we can to ensure its benefits significantly outweigh its risks. In thinking this, I simultaneously feel late to the scene, and—at the same time—am also convinced the vast majority of people are not taking where we are headed seriously enough.
Consequences of Artificial General Intelligence
It is clear to me that now is the time to think hard about the consequences of AGI on humanity. First and foremost, AGI will give us the ability to do far more work than we ever could have hoped to do without it—potentially unlocking unimaginably complex new technologies that allow us to cure cancer, mitigate the consequences of climate change, and solve many more of humanity’s problems. In the days to come, it would be a major mistake to forget just how much we potentially stand to gain in the long term.x
On the other hand, we must manage the risks, and that includes investing in AI safety. Looking at cybersecurity, the Hugging Face incident—as serious as it is—is clearly just a canary in a coal mine. Thankfully, I am not alone in this conviction: the field is beginning to dedicate a very significant degree of resources to safety work. If the most important technical challenges in cybersecurity, biorisk, and other mainstream areas are solvable, I am confident they will be solved. I am much more worried about challenges that people should have been working on, but no-one anticipated.
In particular, I am concerned that we don’t understand our social and economic systems well enough to make accurate predictions—whether from first principles or data—about the kind of changes AI will bring. The data center backlash and shift in public opinion about AI caught many people by surprise—just as the shift in public opinion about social media did a decade ago. In spite of how certain progress is from a technical standpoint, it is highly uncertain how quickly AGI will permeate and rewire society—especially once people outside technology start leveraging it for all kinds of purposes nobody in technology had thought of.
Two days ago, I saw someone on Twitter joke about introducing an AI benchmark on how to convince San Francisco’s government to allow 10% more housing to be built, in order to reduce the average rent. A year or two from now, will sentiment like this still be viewed as a joke?
What’s Next
The original purpose of my blog post nine months ago was to help understand what to work on next. Since then, I:
- Declined a set of assistant professorship offers to leave academia.
- Chose not apply to a single job in industry.
- Declined to interview at a number of places that contacted me, the most lucrative of which was a well-positioned frontier neolab, because none are working on directions I want to go in.
The downside of this is that, at this exact moment, I’m not making any money. The upside is that I have no commitments, external incentives, constraints, or pre-existing momentum which pushes me in any particular direction—and, historically, these constraints have had a much bigger influence on what I’ve accomplished than actual skill. Right now, picking the right direction is clearly the bottleneck on what I’ll be able to contribute to humanity.
My plans are as follows. First, I’d like to finish my book on Bayesian Decision-making Algorithms, which I started writing while at Cornell. Even though I never achieved my original ambitions in this research area before large language models completely changed the broader field’s character, I think the core ideas have merit on their own—both for understanding why large language models work, and for uses outside AI. I owe it to the next generation to write my perspective down, and provide the tools needed to facilitate new and creative uses that I could have never considered myself.
My longer-term plans concern the consequences of AGI, and are in stealth mode at the moment. They will take some time: as critical as it is to move fast in today’s moment, I think moving in the right direction matters even more. Much of where we collectively end up depends on how every segment of society reacts to the proliferation of AI—and, currently, we sit much earlier along the locally-exponential curve of AI adoption, compared to that of AI capabilities. As a field, we have got to make the case that this technology will improve the world, and follow up with responsible choices—otherwise, we will all end up on pitchforks of our own creation.