30 models, 42 scenarios, 7 metrics: Percy Liang's public checkup for language models

Investors Sarah Guo and Elad Gil record *No Priors* in Percy Liang's Stanford office in March 2023. Liang has worked in machine learning and natural language processing for more than twenty years. As an undergraduate at MIT he was drawn to theory and to a simple mystery: humans hear speech and see strings of text and somehow acquire syntax and a sophisticated picture of the world, largely unsupervised. His dream was to get computers to do the same. He went to Berkeley for graduate school, then joined Stanford, and has been pursuing systems that truly understand language ever since.

The last four years, he says, have made that once-upon-a-time dream take off — not quite in the way he expected. Large language models such as GPT-3 capture a startling amount of linguistic and world structure. When he first started in NLP he was already training language models, but of a very different kind: hidden Markov models whose goal was to discover hidden structure in text, teasing apart city names from days of the week. The new systems, he says, are on a completely different level.

In 2021 he and colleagues coined the term *foundation models* and founded the Stanford Center for Research on Foundation Models, housed under the university's Human-Centered AI Institute. The name was deliberate. "Large language model" did not capture what was happening: images and multimodality were arriving, and the phenomenon was more general than language. CRFM's mission, as Liang puts it, is to increase transparency and accessibility as these models become ubiquitous and, at the same time, more closed.

Guo asks whether there was a moment when Liang decided to focus on foundation models. There was, and it was decisive: GPT-3, in the middle of the pandemic. It was not so much the model's capabilities that shocked him as the way it was trained. Take a massive amount of text, ask the model to predict the next word over and over, billions of times, and from that simple objective rose not only fluent generation but in-context learning. You can prompt the model with instructions, give it a few examples, and it figures out the task on the fly.

That, in his view, was a paradigm shift. Machine learning and NLP had been a pile of bespoke systems — one trained for question answering, another for translation. GPT-3 suggested a general substrate you can ask to do various things. The idea of a *task*, so central to AI, begins to dissolve. He finds that extremely exciting, and it is why the center exists.

The open culture that powered the last decade of deep learning — PyTorch, TensorFlow, open datasets, papers — is in retreat, he argues. Models are increasingly reachable only through APIs, with the secret sauce hidden. His diagnosis is unsurprising: these systems take enormous capital, they are competitive advantages, and there are genuine safety reasons to think twice before anyone-anything-goes release. The tension between those incentives and CRFM's transparency brief runs through the rest of the conversation.

Gil notes that innovation has historically ping-ponged between academia and industry. Now some labs publish less and ship only APIs. Liang tells his students to work on what leans on academia's competitive advantage, and he says that advantage has changed. When he was in graduate school, speech, vision, and question answering basically did not work. Academia's job was to make things work at all. Decades of that grinding, from both universities and companies, are what industry is scaling today.

The dynamic is different now. If you have data and compute, you can blast through a lot of barriers. Getting things to work is no longer uniquely academia's job. What remains, Liang argues, is *understanding*. For all their impressive feats, nobody really knows how these models work: how training data and architecture shape behavior, how to weight data, what the right objective is. Those questions benefit from rigorous analysis. The other kind of understanding is social impact — and that is why CRFM has more than thirty faculty across ten Stanford departments.

The center is looking at copyright and legality, social biases, and homogenization: what it means when a single central model makes decisions for a person across many aspects of life. Other researchers are studying disinformation and how persuasive these tools are becoming, including the risk that foreign state actors will use them. Faculty in medicine are exploring deployment in clinical practice, something Stanford can do because it has a hospital attached. Gil recalls MYCIN, a 1970s Stanford expert system that outperformed medical-school staff at predicting infectious disease and still never made it into routine care. He worries some industries simply resist change.

Liang is calm about the gap. Healthcare still has to solve privacy, robustness, and hallucination — models invent facts and sound confident doing it. Guo notes that some doctors do that too. He laughs, then makes a sharper point. For decades, human-level performance has been AI's North Star. On many axes, he thinks the target should now be superhuman. We should define objective measures of what we actually want: reliability, grounding, statistical evidence. He often wants more statistics from doctors and fails to get them. Chasing a human mimic is the wrong brief, he says, because we already have a lot of humans.

Liang's earlier research treated language as a programming language: semantic parsing maps a question into a formal object a machine can execute, such as a SQL query. The virtue is that you compute an answer rather than retrieving or inventing one. The problem is that the world is not a database. Only a small part of it is structured, so he moved to open-domain question answering and helped build SQuAD, a benchmark that, with others, fueled BERT, RoBERTa, and ELMo around 2018 and then the 2020 wave of foundation models. Tool use — handing arithmetic to a calculator instead of doing it "in the Transformer's head" — still sits awkwardly next to the informal reasoning language models now do natively. Neural versus symbolic is the old debate, back because planning and reasoning matter again.

Asked what surprised him, he starts with in-context learning. Designers did not write code to teach the model to latch onto a few labeled examples; it emerged, and it gets more reliable with better models and better data. "The real magic is you didn't have to do that, and it still does something." Chain-of-thought is the second surprise. If you let the model answer immediately, it may be right or wrong; if you ask it to generate an explanation, thinking out loud, it is much more likely to get the answer right. Humans work that way too. The capability still emerged.

The wildest example, for him, is mix-and-match. Ask a model to explain quicksort in the style of Shakespeare and it will produce something semantically on point and stylistically better than many people could write. That means it has learned the concepts "Shakespeare" and "quicksort" and fused them. Critics say large models only memorize. This text does not exist in the training data; the model has to invent it. He treats that as a glimpse of creativity, with implications for scientific discovery. Until now the AI dream topped out at humans. AlphaGo found tactics people missed after thousands of years of play. Going beyond humans, he says, unlocks a lot of possibilities.

Instruction-following on ChatGPT, Bing Chat, and Anthropic's models looks striking, but how much is emergence versus "secret sauce" under the hood is hard to tell. Hallucination is the ugly twin. His hope — not a prediction — is that as models get better, some of it goes away. Next-word prediction sounds simple, but to do it well the model must compress syntax, speakers, and facts into a world model. A better model can start to tell reliable parts of the internet from unreliable ones. You still have to say "don't hallucinate," but control is easier if the system has a notion of what hallucination even is. Tell a weak model "Elad travel to ___" and it will fill in a destination. A stronger one, with more context, might ask where he intends to go.

Someone close to the Transformer's development told Gil the architecture is in a self-reinforcing loop: scale produces emergence, emergence justifies more scale, and other designs never get enough compute to show what they can do. Liang's reply is blunt. He really hopes that in ten years we will not still be using the Transformer. It is a good architecture, people have tried to improve it, and it is good enough that everyone presses ahead — but scientifically there is no reason to believe it is the one. His colleague Chris Ré and students have built alternatives that, at smaller scales, compete with Transformers and do not require attention. Challenging the status quo, he says, is exactly what academia is suited to do.

Gil pours cold water: LSTMs have scaling laws too, they just burn several times more compute for the same result, and everyone lives inside a compute budget. Liang agrees. Given a fixed budget, the Transformer strictly dominates the LSTM; the "what if we scaled LSTMs" question is mostly irrelevant. In academia you can train one-billion-parameter models — not cheap by university standards, but possible — and CRFM is training six- or seven-billion-parameter models. That is enough to try ideas. Because emergence tracks scale, a promising small-scale result is only a hypothesis until someone tests it farther out, or watches the gap close.

That bottleneck is why he co-founded Together. The premise is that compute is central to foundation models, and a lot of it sits idle and decentralized. Datacenter interconnect is extremely fast; a home or desktop link is a hundred times slower or worse. With Ré, Ce Zhang, and others, Together developed scheduling and compression techniques that let weakly connected machines train anyway. Folding@home is the inspiration: mid-pandemic it briefly assembled the world's largest supercomputer by FLOP count for COVID molecular-dynamics simulations. Foundation models are harder because big weights and big data have to be shuffled around. The company is building a research cloud, where academics contribute idle GPUs and draw a larger share when they need it, and a commercial platform for fine-tuning and inference on open models. OpenChatKit, built on EleutherAI's GPT-NeoX, is being released as a work in progress for the community to improve.

Liang's then-uncommon view is that the future holds many foundation models for different use cases, some derived from existing ones and some trained from scratch. The exhibit is BioMedLM, trained at CRFM with MosaicML on PubMed articles. It is not huge, yet for a while it was state of the art on the U.S. Medical Licensing Examination, until Google's model — about 200 times larger — beat it. Scale matters, he says, but cost is a constraint, and many settings should use the smaller model. Farther out, he wants systems that read the literature, generate hypotheses, propose experiments, write code, and revise their picture of the world — a scientist with a human in the loop. Class-project-level work is already in reach. A true breakthrough, like AlphaGo's alien tactics, would take a more careful setup than sampling a language model.

CRFM's flagship over the prior year is HELM, Holistic Evaluation of Language Models. The trouble with "evaluate language models," Liang says, is that it is like saying "evaluate the internet." A model takes text in and text out and can be used for a myriad of applications. The project laid out scenarios as systematically as they could and measured not just accuracy but robustness, calibration — whether the model knows what it does not know — fairness, bias, toxicity, and efficiency. They grabbed every prominent model they could reach: open ones such as OPT and BLOOM, and APIs from Cohere, AI21, OpenAI, Anthropic, and Microsoft. Overall: 30 models, 42 scenarios, seven metrics, the same protocol on all of them.

The results live on a website. You can see top-line numbers, then drill into a benchmark, a single instance, the model's prediction, and the prompt that produced it. The point is transparency, laid out scientifically. The hard part is speed. About three months after the paper, ChatGPT and LLaMA had appeared, Cohere and AI21 had updated models, and GPT-4 was rumored. HELM became a living board, refreshed every two weeks, with new scenarios for the things people actually ask — write an email, give life advice, write an essay — plus new risks. Liang wants persuasion benchmarks and security tests. He is especially worried that once models take external input and act in the world, you can jailbreak not only your own system but someone else's, setting off a cascade of errors. Multimodal evaluation is next.

Policy, for the center, starts in the same place. A few players disagree about open versus closed. Models are described as aligned to human values, but once you ask *which* humans, who determined those values, what legitimacy that has, and who is accountable, you notice the black box. Transparency is necessary but not sufficient; you need some of it even to have a policy conversation. Liang wants something like nutrition labels or the spec sheets on electronic devices: an obligation to say what the product is, what data went in, what instructions were used to align it, and where the bounds are. Together, meanwhile, is shipping a client so people can contribute compute and an API for inference and fine-tuning.

Asked how a rigorous scientist thinks about AGI, Liang says his opinions have changed. For a while the mainstream community treated it as laughable. He has been in touch for a decade with people who think about AGI and existential risk, but he sees the world differently and has focused on nearer-term robustness of ML systems. Foundation models' emergent qualities taught him to stay open-minded. Everyone, including scientists, gets drawn into a paradigm. The world is changing on the technical side, and so is how we conceive of AI — maybe even of humans. The show's name, *No Priors*, is the attitude he recommends for the next few years.