Two Dates from OpenAI's Chief Scientist: Intern by September, Automated Researcher by 2028

Jacob Effron, a Redpoint partner hosting Unsupervised Learning, opened with the juiciest public claim Jakub Pachocki had made. Four months earlier, in January 2026, OpenAI's chief scientist had aimed at a system with research-intern capabilities by September of this year, and a more fully automated AI researcher by March 2028. Checking in, how did those dates feel? Pachocki did not walk them back. The change that had actually happened, he said, was the explosive growth of coding tools — "an understatement." Inside OpenAI, Codex now does the majority of actual coding. For most people there, the act of programming has already changed. He treated that as a signal that something is on track.

The other update was math and physics. Reasoning models had reached IMO-gold level, including problem six, and were making forays into research mathematics. Combine that with access to infrastructure, more test-time compute of the kind Codex already uses, and the general-intelligence gains he expects over the next couple of months, and the intern milestone is still the plan. How do you know you have arrived? He distinguished intern from fully automated researcher by the span of mostly autonomous work and by how specific the task has to be. He does not expect systems this year that you can simply tell to go improve model capability or go solve alignment. For a concrete technical idea — run this evaluation differently, try this particular improvement — he said the pieces are mostly there and mainly need to be put together.

Andrej Karpathy's viral demo of models improving a small model of his own felt, Pachocki said, in the spirit of where this goes: a continual evolution from Codex toward more autonomy, longer runs, and higher compute. Math and physics matter as a north star because they are measurable. It is much easier to tell whether you solved a math problem than whether you produced good software, and the problems can be arbitrarily hard. Until recently the gap was obvious: models could not do IMO-level work, so the intelligence deficit was easy to run at. That is changing. From here, mathematical reasoning still transfers into AI research — a lot of OpenAI's best researchers are mathematicians or come from theoretical fields — but the north stars are shifting. The next models need to be useful in the real world, especially for AI research, economically valuable work, and applied science. The reason for the shift, he said, is that the models are now capable enough. Not as smart as people in every way, but capable enough to materially change the economy. That is why the urgency is up.

Math and code were the perfect early domains: hard to solve, easy to verify, ideal for reinforcement learning, and the improvement curves have been shocking. Effron's question was the one every lab is living with. Medicine, law, and finance are valuable, and they are not verifiable the same way. Will RL still deliver that steep curve? Pachocki expects so. The observation he finds useful is a duality: tasks that are harder to evaluate share a lot with longer-horizon tasks. Even a well-specified math or coding problem, if it takes a year, leaves day one wide open. Those difficulties coincide, and they are clearly the next frontier. He has already seen encouraging signs that RL can scale on more general domains.

The research challenge, as he framed it, comes down to two things: making models work for a very long time, and teaching them to evaluate partial progress. Some of that longer-horizon skill is already arriving without new RL tricks. As models become more consistent from pure supervision in pre-training, they pick up a sense of what a good partial artifact looks like. Even if OpenAI were not scaling RL very meaningfully, horizons would elongate. Applying new RL ideas to general domains is still a research problem; he is optimistic. Effron asked whether this is still old-school one-thing-at-a-time optimization rather than generalization. Pachocki's answer was compute. "We are buying a lot of compute" because they still believe the bitter lesson, more than ever. New techniques and new ways to scale exist, but that is still the lens.

There is a complication they no longer get to ignore. They are not building a brain in the sky isolated from the world. If the model is going to do medical research or cure cancer, it has to learn about the real world in a meaningful way, run experiments, and learn from results, which means connecting it. That does not, in his view, contradict finding and scaling simple algorithms. The capability roadmap he has used internally is ordered: first teach the model to reason with its own tokens, then tools and the environment, then vision, then a physical body. They are well into the interact-and-see stage. Robots are "someday soon," but later than virtual AI. A lot of "the model cannot do X" complaints, Effron added, melt once you connect files, context, and systems. Pachocki agreed that a lot of long-horizon failure is still access, not intelligence.

Builders keep asking whether they should take an open-source model, their own task data, and their own evals, and run reinforcement learning, or just wait. RL, Pachocki said, can be a very data-efficient way to improve a model on a task. There is a more data-efficient way still: learning in context, the most fundamental method, examples and instructions in the prompt. He expects that kind of learning to get much better. Replicating a frontier lab's current RL pipeline is probably not the right move. Effron restated it: you still have to figure out the evals that matter and gather the examples, but you may be better off feeding that into context than training your own model. "Quite plausible," Pachocki said.

The same logic applies to harnesses. Implementation of the harness should not be a limitation for a very long time; more general harnesses will cover other domains. Codex, he added, is already pretty good if you try it for things beyond coding. The deeper design question is the ultimate interface. Models can build their own UIs and do work people find time-consuming, but there is also space to let them into the interfaces humans already use — AIs on Slack, plugged into context, learning from what is already there. By default the AI should meet you where you are. If it does not, that should be because it has new abilities, not because of limitations.

The milestone he watches now is research itself: can the model discover new things, and can it execute a longer-horizon research problem? The gut check is whether, if someone on the team had come up with the idea, he would have been intrigued. GPT-5.2 Pro has already contributed a few minor but quite impactful research ideas they are using. It is still very small compared with where he expects this to go. That is also why running the lab is changing. They are at a transition where the short-term quality of the model is about to determine the pace of research progress, because the models will drive a lot of it. The old instinct was not to obsess over immediate quality. Now he feels urgency to use the intelligence gains to accelerate AI research and especially alignment. He and Mark Chen spend a lot of time on that balance.

Compute allocation has a similar discipline. OpenAI has more experiments than GPUs. They explicitly budget a large chunk of compute to the most scalable methods, the ones they believe drive general intelligence, even when that is not the most efficient allocation in the moment. It is too easy, with so many interesting projects, to partner the budget out and never do the thing that matters most. He looks for empirical evidence, evals in order, experimental rigor, then a kind of regularization: do we understand the method, will it scale, can we build on it, or is it a one-off? On Claude Code's head start, he read it as product focus. OpenAI's products had been secondary to research priorities, and 2023 ChatGPT was aligned with the vision but not representative of everything research was building. Increased product focus now is about deploying the intelligence they already believe matters. For developers in three to six months he expects a gradual rise in the autonomy you are comfortable giving the model, the vagueness of description it can take, and the supervision it needs. Models that work autonomously for a couple of days, use more compute, and produce higher-quality artifacts on their own are not far. Building something bigger will still want human supervision and a sense of which building blocks fit. The desired skill set shifts toward vision-setting.

A month before the taping, continual learning was in the zeitgeist. New labs were forming around it; some people had left OpenAI to work on it. The usual premise is that RL alone will not get you to human learning efficiency. Pachocki sounded a little confused. Look at the GPT-3 paper, he said: the excitement was that this class of models can learn in context — continue learning. That was the driving force behind scaling GPT further, and the premise for teaching them with RL to learn in context more efficiently. "Continual learning is really the thing. It's really the thing that we're building." He does not think it is ignored or off the current path. It is what they are working toward. The best progress so far has been scaling pre-training and RL. There will be more ideas and more steps; a lot of the improvement will just come from scale.

People outside the labs still bounce off hundred-step tasks and conclude the models are a year or two away, while researchers talk as if the line is already visible. A lot of that prediction, he said, is just historical improvement curves, and the shape is getting clearer. Models have to become intelligent enough to recognize whether they are making progress. They also have to actually reach the context, files, and infrastructure the work requires. That is the pragmatic half of the roadmap, sitting under the older sequence of tokens, then tools, then vision, then robots.

Coding tools everyone can feel. IMO medals are harder to grok if you never sat those contests. Effron asked about First Proof: problems from respected mathematicians and theoretical computer scientists, meant to represent their day-to-day work, never published. The challenge dropped without warning, with a week-long deadline. OpenAI had an exciting model in training. James Lee, one of the people in charge of that run, started prompting it by hand. It was actually solving the problems. One of them was from the domain of Pachocki's own PhD. Watching the model come up with ideas he would have been quite proud to find in a week or two, and seeing it do that in about an hour, was a very weird feeling. The last time he had felt it was watching the Dota bot play interesting games indefinitely — interesting things should not be infinite. Seeing it happen for math, which he takes as a precursor to work that really matters, "definitely really increase my feeling of urgency."

They still needed domain experts to say whether particular proofs were correct. Even so, it is easier to tell whether you made progress than in most programming, where competitive evals do not capture whether the abstractions are right. On the old complaint that models are pattern matchers and therefore cannot do science, he thinks they are on schedule to chip it away: minor advancements, a small idea here or there, some bigger papers with scientists. Was AlphaGo a pattern matcher? AlphaZero? The Dota bot? They came up with new strategies. Counterexamples go back to 2016 and 2017. You can always find a strategy that beats them; there will be a lot of definitions for a while. Getting from tiny game environments to general scientific research took a decent approximation of all human knowledge and human languages. The basic principle, he said, is similar.

After First Proof, some organizers said the AI solutions felt like 19th-century mathematics: brute-force, computation-heavy, missing modern elegance. That does not concern him. On at least one problem he thought the model produced a nicer proof, quite a bit shorter than the intended one. In general you would expect these models to produce much more reasoning in a short time than a person can, in raw tokens. He does not expect that to be a long-term feature. For AI for science three years out, he would not build around the current limitations. If a model can suddenly design a huge number of interesting chemical or biological experiments, build labs that enable that — OpenAI has already tried collaborations with companies such as Ginkgo. If it can design high-quality experiments, it should also work with humans in the loop. The world he described is not a fully automated biologist in a box; it is natural collaboration with AI scientists driving a lot of the design and ideation. Specialized architectures still have a place. A top Go model, or protein folding, is not what large language models are most efficient at, even if they might eventually be the best. Independent efforts such as Isomorphic Labs, Periodic Labs, and Physical Intelligence make sense. They will still pair with a very strong general LLM researcher.

Around the first of the current crop of reasoning models, Pachocki's team realized the world would change over the next year or two or three, and asked what that meant for understanding what the models were doing. ChatGPT is trained to be polite and nice; even if it were aligned exactly as you wanted, it still would not reveal some motivations, because they might be unsafe, unkind, or worth hiding. Reasoning models are different. The chain of thought is not directly graded. It is only graded through the quality of the final output, so it is not optimized to look any particular way. That is a powerful window, close in spirit to mechanistic interpretability — activations also are not trained under inspection — with the huge advantage that chains of thought are, by default, in English. As models work longer and more autonomously, there is more of that reasoning, so the ability to supervise can scale with capability. The principle is that you are not supposed to supervise the chain of thought.

That was his primary reason for hiding the thoughts when o1 preview shipped. There was internal discussion; he did not even want to consider releasing them another way. Distillation was a second, valid concern he had not initially weighted as heavily. If you establish a product paradigm of showing the chain of thought, you eventually have to train it, for the same reasons you train anything you ship. Then the training signal fights the monitor — "the other side of the bitter lesson." Chain-of-thought summaries are a stopgap. The longer-term solution is the model talking to you in real time, which the latest Codex and reasoning GPT models already do, and which he expects to get much better. The decoupling gives him hope for watching motivations and generalization as models get better and work longer. It is not a complete alignment solution, just another tool. Building the toolbox is how they chip away at the fundamental problems.

The tool already paid for itself. A collaboration with other labs — Anthropic and DeepMind, in Effron's telling — on model scheming asked whether, depending on environment and training, a model starts hiding objectives. That whole line of work rests on chain-of-thought monitoring. From there, mitigations might look completely different: change the pre-training data, or something like Anthropic's inoculation prompting. Shared interest on alignment exists even among competitors. The longer-term challenge, Pachocki said, is generalization. In-distribution behavior is mostly controllable. The worry is a very different request, a very different situation, or a model much smarter than it has ever been. Value alignment, in that light, is the study of what the model falls back on. One line he is investing in is how that generalization falls back onto pre-training data.

Over the past few years his thinking on long-term alignment has moved from a nebulous problem, hard even to define, to something you can make progress on with concrete technical solutions. Alignment is now a core part of research, designed into the reasoning models. His belief that there is a research path to an extremely happy world has increased a lot. At the same time, his timelines to very capable — not smarter-in-every-way, but transformative — models have decreased a lot. He is optimistic they can keep a grip on evaluation of risks, and he thinks the industry has to be prepared to take trade-offs and possibly slow down development depending on what they see.

Effron asked how you operationalize the old "give researchers space to be cave dwellers" instinct in a coding-agent race. Pachocki still focuses on high-quality experiments, honesty about results, and not wrapping every project up overnight. What has changed is urgency to bring the most promising work to fruition. Looking back over OpenAI's public dramas, he said the real shifts were not snap decisions. When he joined at the start of 2017 it felt like an academic lab, not scaling-pilled in practice. After Dota and GPT they moved to buying big computers, developing the science of scaling, and judging ideas by whether they scale. Then ChatGPT became the thing. He was pleasantly surprised that text models were first; he had thought video-style generative AI would take off first and force a trade-off against longer-term text research. They had anticipated the tension of a popular product that still had to evolve a lot. Now he believes they are entering the phase of deploying AGI — or at least models that are very economically transformative.

In the quickfire he named the mind-change of the last year: reconciling that the AI you train eventually affects the world, and that they now have to focus on how it is actually deployed. He expects ChatGPT to go in the same direction as the coding models — a dependable, trustworthy assistant or companion, which is already how a lot of high-school and college students use it. On robotics he is optimistic about algorithmic ideas not too dissimilar from the language-model stack, with longer timelines than virtual AI. The societal underthink, in his view, is what happens when so much intellectual work can be automated. Jobs and concentration of wealth do not have obvious solutions and will need real policymaker involvement. Cheap capability should be good in the long term; it can happen quite quickly. Related, and weirder: an automated research laboratory or automated company can be controlled by a very small number of people and still do a lot. Robots make it crazier; you do not need robots for the question to exist. Governance of organizations that powerful, made of only a couple of people, is a new problem.

Effron had recently had a kid and asked how this work changes how the next generation should be raised. Pachocki made it a task for all of us: build the world so that humans still have the agency and set the direction. A lot of the technical challenges we cherish now may become a pastime. The hard problems become what is important and what we should go do. People can end up with more to do, and more exciting things to do. You still want some understanding of technology, however you acquire the basic education, for the sake of being able to think about those problems. He left the last word on alignment and monitorability. Those are growing into very urgent challenges, not only for AI researchers. Policymakers have to take them, and society has to think them through. He is happy to see some discourse starting. He thinks we need more of it.