Skip to content
No PriorsNo Priors

No Priors Ep. 7 | With Stanford Professor Dr. Percy Liang

When AI research is evolving at warp speed and takes significant capital and compute power, what is the role of academia? Dr. Percy Liang – Stanford computer science professor and director of the Stanford Center for Research on Foundation Models talks about training costs, distributed infrastructure, model evaluation, alignment, and societal impact. Sarah Guo and Elad Gil join Percy at his office to discuss the evolution of research in NLP, why AI developers should aim for superhuman levels of performance, the goals of the Center for Research on Foundation Models, and Together, a decentralized cloud for artificial intelligence. 00:00 - Introduction 01:44 - How Percy got into machine learning research and started the Center for Research and Foundation Models at Stanford 07:23 - The role of academia and academia’s competitive advantages 13:30 - Research on natural language processing and computational semantics 27:20 - Smaller scale architectures that are competitive with transformers 35:08 - Helm, holistic evaluation of language models, a project with the the goal is to evaluate language models 42:13 - Together, a decentralized cloud for artificial intelligence

Sarah GuohostDr. Percy LiangguestElad Gilhost
Apr 24, 202353mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Percy Liang on Foundation Models, Open Research, and AI’s Future Impact

  1. Stanford professor Percy Liang discusses his journey in natural language processing and how GPT-3 triggered his shift toward foundation models, leading to the creation of Stanford’s Center for Research on Foundation Models (CRFM).
  2. He explains why large language models are a paradigm shift, their emergent behaviors like in-context learning and chain-of-thought, and how academia’s role is moving from “making things work” to understanding principles and social impact.
  3. Liang highlights CRFM’s work on transparency (e.g., the HELM evaluation benchmark), interdisciplinary research on risks and benefits, and efforts to keep models accessible despite increasing industry secrecy.
  4. He also covers Together, a decentralized compute initiative for open models, and reflects on future directions for architectures beyond transformers, scientific discovery with AI, and how to think rigorously about AGI.

IDEAS WORTH REMEMBERING

5 ideas

GPT-3’s training paradigm transformed how we conceptualize AI tasks.

Simply predicting the next word at scale, and then prompting models in natural language, turns a single foundation model into a flexible substrate for many tasks, dissolving rigid, task-specific system design.

Academia’s comparative advantage is shifting from scaling models to understanding them.

With industry able to “just scale” using massive data and compute, universities are better positioned to study principles, data/architecture effects, robustness, and social impacts rather than only chasing benchmark wins.

Transparency and openness are eroding and must be deliberately rebuilt.

Unlike earlier deep learning culture (open datasets, code, and models), top foundation models are now often API-only; CRFM and HELM aim to reintroduce norms of disclosure, comparability, and shared benchmarks.

Emergent behaviors like in-context learning and chain-of-thought weren’t hand-designed.

Capabilities such as learning from prompts, step-by-step reasoning, and stylistic “mix and match” arise from scale and training rather than explicit programming, suggesting deeper, still poorly understood structure.

Future AI progress likely requires both new architectures and better infrastructure.

Liang expects transformers not to be the final architecture and supports work on alternative models, while Together tackles the compute bottleneck via decentralized, weaker interconnects to make large-scale training more accessible.

WORDS WORTH SAVING

5 quotes

The idea of a task, which is so central to AI, begins to dissolve.

— Percy Liang

I really hope that in 10 years we won’t be using the transformer.

— Percy Liang

We coined the term foundation models because ‘large language models’ didn’t really capture the significance.

— Percy Liang

Up until now, the AI dream tops out at humans, but now we can actually go beyond in many, many ways.

— Percy Liang

Transparency is necessary but not sufficient; you need it just to even have a conversation about policy.

— Percy Liang

Percy Liang’s background in NLP, semantic parsing, and the shift to foundation modelsThe paradigm shift introduced by GPT-3 and in-context learningGoals and activities of Stanford’s Center for Research on Foundation Models (CRFM)Open vs closed ecosystems, transparency, and the evolving role of academia vs industryEmergent capabilities of large language models and future research directionsDecentralized compute and open model training via TogetherEvaluation, safety, social impact, and policy considerations around foundation models (HELM, bias, disinformation, alignment)

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome