BAIR Berkeley
2026 BAIR Graduate Showcase
reasoning
What happened
The 2026 Berkeley Artificial Intelligence Research (BAIR) Graduate Showcase highlights the work of graduating PhDs. Notably, researcher Charlie Snell's work focuses on bridging the gap and trading off between different LLM scaling paradigms, specifically comparing test-time scaling (drawing long chains of inference) with pretraining scaling (compressing representations from large datasets).
Why it matters
The frontier of LLM research is shifting from pure pretraining scale to optimizing the trade-offs between pretraining and test-time compute.
The take
Charlie Snell's research focus is at the absolute center of the current paradigm shift in LLMs: test-time compute (inference-time scaling) vs. pretraining compute. Understanding how to trade off and combine these two paradigms is what makes models like o1 and DeepSeek-R1 possible and cost-effective.
Do this
Read Charlie Snell's published papers on test-time scaling to understand the theoretical limits and trade-offs of inference-time computation.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.