o1 Is Slow and Hard to Use. Percy Liang Still Calls It a Turning Point

When OpenAI released o1, the internet treated it as a breakthrough. Stanford computer science professor Percy Liang sat down on Redpoint Ventures’ podcast *Unsupervised Learning* and offered a rarer split verdict. As a product, he found it slow and hard to use for many of the things he wanted. As a research signal, it marks a change he expects the field to keep walking toward.

That assessment carries weight. Liang co-founded the AI company Together AI, and his lab has shipped pieces of infrastructure the industry actually runs on: generative agents — a *Sims*-like virtual world where agents live, talk, and gossip — the HELM evaluation framework, and this year’s cybersecurity benchmark Cybench. When a frontier lab ships a new model, it often has to sit his exam. Host Jacob Effron opened with the question everyone had that week: was o1 a crazy breakthrough? Liang did not hedge.

Liang’s core claim is that test-time compute has been around; o1 is the moment it reached a scale that made the idea sink in. The old picture of a language model was simple: a prompt goes in, a response comes out, the whole thing finishes in a few seconds, and labs compete on tokens per second. o1 breaks that rhythm. It will spend minutes reasoning on a single problem, the way a student allowed scratch paper scores differently on hard questions. Directionally, he wants systems that can take on projects measured in days, weeks, or months — inventing a drug, starting a research program. o1 is a small step, but you can see the road.

He also sees a cycle coming back around. A decade ago, in the AlphaGo era, reinforcement learning was the rage. For about five years after that the field almost forgot it, because large language models and next-token prediction ate all the attention. Agents are bringing the word back. The way to interpret a language model’s generation, Liang argues, is not necessarily as text but as actions in some space. Once an agent is embedded in a longer task, it can accumulate experience; with an appropriate reward signal, it can learn and iterate. Generation is being re-read as doing, not writing.

When people say benchmarks are saturated, he points to Cybench, a capture-the-flag cybersecurity suite his group just released. The exercises are hard enough that he does not think he can solve most of them. The hardest one takes a team of human competitors more than 24 hours. Current models can solve the challenges that a person or team first cracked in around 11 minutes. There is still a large gap. Dropping o1 in as a replacement, though, did not lift the overall score. It improved some subtasks, then ignored the team’s own templates for reflection and planning and just generated an answer. The new engine was not compatible with a chassis built for ordinary language models.

That is a cautionary tale about raw scores, and a second one about systems: a better model that does not fit will not deliver the gains. Jacob asked whether the prompt chains and planning loops built around GPT-4 still matter. Liang was blunt — nobody should be proud of or attached to that scaffolding; it is “very dispensable.” o1 internalizes the reasoning and hides it, which from a research angle is annoying: when things go wrong you get no stack trace, and developers lose a layer of customizability. He is honest about his own use. Over a weekend he tried o1 on a React app he had not touched in years; it helped, but it was slow and annoying, so he switched back to GPT-4o. Cost, speed, accuracy, task fit, and customizability all matter. Calling progress one-dimensional is an oversimplification. o1 is stronger at math and coding, where chains help and supervision is easier; on other tasks it can recede, and the black box is getting larger.

Academia has a collective anxiety: OpenAI-class labs hold tens of thousands of GPUs, so what is a professor supposed to compete on? Liang’s answer is not to play that game. Going head-to-head with a frontier lab is pointless. The move is to pick a different one: research that is orthogonal to frontier models. The test he gives himself and his students is simple. On the day GPT-5 comes out, the project should either be enhanced by it or be roughly irrelevant to it. Do not pick a direction the next model will casually crush.

His own projects follow that rule. Generative agents get more realistic as the underlying models improve; a new release is ammunition, not a threat. MLAgentBench, which asks whether a model can complete machine learning engineering tasks on its own, also gets more interesting every time a stronger model lands. Jacob notes that startups use the same move, and that Sam Altman has said you should back companies that are rooting for the models to get better. Liang agrees: the advice is not specific to universities.

Beyond topic choice, he carves out two other jobs for academia. One is open science. Frontier labs hold a lot of knowledge that is not in the public domain; even reinventing those ideas and publishing them has value, because the broader community can build on them. Work on data quality and how to weight pre-training data is an example: maybe the labs already do something sophisticated, but putting the ideas in the open is itself a contribution. The other job is auditing and transparency. Universities have no commercial interest, so they can benchmark providers and score their transparency in ways that are extremely hard, almost impossible, inside a company. His group also collaborates with Stanford Law School on the policy problems that show up in an academic setting.

Something else has shifted in Liang’s thinking over the last few years: he has learned to see AI’s role more holistically. Researchers from a technical background tend to treat the model as the central object. Safety then gets construed as making sure it does not respond to harmful requests. That picture is too small. The model is one piece of a larger ecosystem of actors with different incentives. What you are trying to keep safe is the whole system, not a particular model. RLHF-ing the model to death will not get you there.

His phishing example is concrete. A bad actor will not tell a model “I want to write a phishing email” and declare intent. They will ask for a personalized note — here are the top ten research papers you should read — and then swap every link. Malicious intent, decomposed into harmless subtasks, walks around the guardrails. That is harder to police than people give credit for. Defense has to be funded the way it is for email and the internet: spam filters, fraud detectors, the analog of those tools for AI. Gating access to models is a losing battle, because they get cheaper and more widespread. Fatalism is also wrong. There are measures worth taking.

On regulation he flags that he is not an expert, then makes two clear bets. It is early, and a lot is not understood; the desire not to repeat social media is real, but some bills can still be harmful. He names California’s SB 1047. He is strongly in favor of rules that emphasize transparency and disclosure, because everything is so closed off that policymakers, researchers, and third-party auditors cannot see what is happening. On upstream versus downstream, sectoral rules in finance and healthcare already exist and are more effective, because you can see the harms. Heavy-handed upstream regulation of foundation-model developers is a blunt instrument. What is reasonable is nutrition-label and spec-sheet obligations so downstream product developers know what they are being handed.

Generative agents, the project with student Joon Sung Park and collaborator Michael Bernstein, came out last year — “AI dog years,” Liang jokes, which makes it feel older. The question that triggered it was simple: language models generate text, but what else can you generate? An agent? A society of agents? They built a set of agents, each powered by a language model and a stack of prompts, grounded in a virtual environment where they could move around and talk. It was pure exploration. Emergent social dynamics showed up anyway: information diffusion, one agent announcing a run for mayor and trying to convince the others.

The first paper was about *believable* simulations: they look right. The next keyword is *valid*. If the simulations reflect something in reality, a lot of new work unlocks. Believability is enough for games. Validity would give you a digital twin of society. You could try a COVID mask policy, or a new law, and watch what the agents do. You still should not fully trust the results, but if you bet on models getting better, you might eventually simulate many of the decisions humans make. Social science is an obvious customer: recruiting participants is slow, expensive, and often limited to college kids. A demographically diverse agent population could be a first pass before the real study.

The detail he likes most is that the same agent can take both the treatment and the control. Give it condition A, wipe its memory, give it B. You cannot do that with people. Medicine is already experimenting with digital-twin clinical trials, simulating what would have happened to a patient who stayed in the control group. Company org design could be simulated the same way. He distinguishes two kinds of agents: o1-like systems that perform hard tasks, and simulators that mimic what a human would do. The second is less studied and has a pile of unused applications. Classical simulations — weather, disease — are governed by equations or highly stylized models. Language models let you simulate in much greater detail than was possible. Asked whether people will run these before major life decisions, he laughs about investments, practicing a podcast, or rehearsing a date, then adds the cold water: we still do not know how close current simulations are to the real thing.

Switch to evaluations and Liang’s words are “exciting but also a huge mess.” It used to be easier: a train set, a test set, a number, done. The hanging problem now is train-test overlap. Nobody knows what is actually in the training data, and the companies will not tell. Every evaluation he runs, he wonders whether he can trust the number — even if the model did not see the exact item, it may have seen the same type of word problem. Existing benchmarks also fail to capture the new things people want models to do as they get better.

That is why he is excited about using language models to benchmark language models. Coverage is the issue: these systems claim they can follow any instruction, so sitting down to list all the tasks is nearly impossible, and user data may not stress the tails. AutoBencher generates inputs using side information the test-taker does not have, so the evaluation is not pure self-grading. He also wants rubrics instead of superficial A-versus-B judgments. When you grade exams, you publish a standard; a model response should be held to the same kind of anchor. Two years ago his group released HELM, a holistic evaluation of language models. It has since become a framework with vertical leaderboards: safety, other languages, medical, finance. In any given vertical there is quite a bit to do.

Interpretability is in a worse spot. When Liang’s group worked on it in 2017, they at least had the weights and the training data, and it was still hard. Closed models now give you neither. Influence functions ask a counterfactual: if you removed this training example, would the prediction change? The idea is useful and has been adapted to language models, but it is difficult to scale, and private data makes it radioactive — you do not want to tell a patient they were diagnosed because of a random Reddit thread. Chain-of-thought explanations are another tack, and research already shows they may not match what the model is actually doing, “kind of like people.” His position is direct: to do serious interpretability you need weights and training data back, at least 2017-era access.

Historically, LSTMs, convolutional nets, and Transformers came from intuition, thinking about gradients, and experimentation. Mamba, the state-space line of work, is the exception. Albert Gu and Chris Ré started from a math question: given a sequence of points, how do you do an online update of the best polynomial fit? They solved that, then figured out how to make it work at scale inside a neural net. Liang does not work on architectures himself, but he would bet Transformers are not what the field keeps using forever. Whether the next change is huge remains open. The more likely places for something new are video, where a giant Transformer is a naive fit, and agentic settings, where the architecture has to sit inside a larger system. New architectures, he argues, are forced by new problems. Simple image classification did not need Transformers; machine translation did.

On Together AI and the inference market, he treats inference as a low-level primitive that has to be robust and cheap. Training needs forward passes, agents need inference, synthetic data needs inference: everything sits on that block. FlashAttention-style optimizations remain relevant because of that. Today’s inference market is mostly serving Llama 3, which functions as a stamp of approval that the model is not a dud. People do not actually care that it is Llama 3; they care that it works for their use case. Customizing for a narrow job — a call-center application you run a lot — has much more room to squeeze performance than making Llama 3 405B ten times faster at everything. Agentic workflows also open optimizations, for example high-throughput generation of many possibilities.

Meaningful near-term milestones, for him, are still the hard benches: Cybench for cybersecurity, MLAgentBench for ML research tasks. Both sit well below what they could be. Farther out, the interesting line is anything that extends human knowledge — solving an open math problem, creating new research, finding a zero-day. Catching up with humans, even expert humans, is still mimicry; IMO problems were written by people. Asked whether capabilities are plateauing, he says things are still moving up quickly. Locally, everything looks linear, a calculus joke the hosts later pick up. There will be qualitative changes too — o1 is one — and chips will get more powerful and cheaper.

Robotics is not at a ChatGPT moment. It is closer to a BERT moment: vision-language models, fine-tuned as robotics foundation models, work for narrow tasks, but the policies are brittle in ways language is not. It will be a few more years. He sees no fundamental obstacle, only missing data — there is no internet of robot experience, so you have to do it the hard way. The consolation is that robotics can borrow language and vision recipes for architecture, data, and compute, and that many supposed robotics problems are actually language and vision problems. You should not need robot data to learn what a cup is; you need it to grasp the cup.

Music is the topic Liang takes personally. The same giant-Transformer-plus-data recipe works, as companies like Suno and Udio show, but copyright is a much bigger hurdle, and control is the deeper one. Unconditional or text-to-music generation does not give artists enough say. A former postdoc in his lab, John Thickstun, built the Anticipatory Music Transformer: a generalized infilling model that can condition on any subset of events in the score — melody to harmony, one part to the rest of a section. The direction is a copilot for composers, closer to GitHub Copilot than a button that spits out a song. Years ago Liang was competing in piano competitions while doing a PhD. He did not have time to practice; he knew how he wanted to sound and his fingers would not do it. Tools that help realize that vision would be an augmentation, not an excuse to skip practice. Classical music, with its subtleties and thin data, is far from one-click generation.

He is extremely bullish on AI as a teacher and coach. Jacob already uses it to explain concepts he understands to a five-year-old. Breaking complicated things into simple ones is, Liang says, very impressive. On the quickfire round, the overhyped thing in AI is agents. The underhyped thing is also agents. The field has been through a full hype cycle; the direction has not changed. How far are we from ML agents contributing novel research? Running an ablation to answer a question is already possible, at the level of a very junior student. Coming up with new experiments and taking on new directions is something he is optimistic can happen meaningfully in the next few years. Coding is a simpler version of the same shift. The underexplored applications sitting on top of these models are fundamental science and researcher productivity; everyone else is crowded into RAG, question answering, and summarization.