5 Papers That Show Where AI Research Is Heading Right Now

Why It Matters

Demonstrating that scaling protein language models yields genuine structural insights bridges AI and biotechnology, while emphasizing sample‑ and energy‑efficiency drives more sustainable, high‑impact AI deployments.

Key Takeaways

• Scaling protein language models mirrors language model scaling laws.
• Larger ESM models improve unsupervised contact prediction accuracy.
• Self‑play and memory efficiency are emerging research frontiers.
• Intelligence per sample and per watt remain critical bottlenecks.
• Bio‑AI benchmarks could accelerate cross‑disciplinary collaboration across industry and academia.

Summary

The club talk highlighted five recent papers that illustrate where AI research is heading, ranging from bio‑AI and protein language models to self‑play for large language models (LLMs) and memory‑centric architectures. Speakers such as Yas Beg, Luke from Tatsu’s lab, and Arnob presented work that pushes AI into biology, explores AlphaZero‑style self‑play for LLMs, and investigates real‑time voice agents, underscoring a shift toward more applied, cross‑domain projects.

A central insight was that scaling laws observed in natural‑language models also hold for protein‑sequence models. The new ESM‑Cranberry family, spanning 300 M to 6 B parameters, showed a clean log‑linear improvement in unsupervised long‑range contact prediction, suggesting that larger models can infer structural biology without hand‑engineered features. Parallel discussions on memory research—mem‑zero, recursive language models, dynamic chunking—and on intelligence per sample versus per watt highlighted persistent efficiency challenges.

Notable examples included Luke’s AlphaZero‑style self‑play framework for LLMs, which aims to eliminate human bias in training, and the use of internal model representations to predict protein contacts (P@L metric) as an emergent structural signal. Speakers also cited the “bitter lesson” from Sutton, arguing that general scaling beats domain‑specific engineering, and called for community‑wide benchmarks and open‑source challenges to accelerate progress.

The implications are clear: as AI models grow, their applicability to scientific domains like protein design will expand, but breakthroughs will depend on improving sample‑efficiency and energy‑efficiency. Establishing shared benchmarks and fostering interdisciplinary collaboration can translate these advances into tangible biotech innovations and more sustainable AI systems.

Original Description

It's hard to keep up with the latest AI research. That's why we started YC Paper Club — a small group of researchers, engineers, and founders who meet every two weeks at our Mountain View office to present and discuss new papers together. In this session, we cover whether scaling laws hold for protein biology, AlphaZero-style self-play for language models, streaming RAG for real-time voice agents, formal verification with Lean, and why one founder thinks programming with agents is exactly like playing a real-time strategy game. Stay tuned for more.

00:00 — Introduction by Francois Chaubard

05:47 — Yasa Baig: A World Model of Protein Biology (Evolutionary Scale Models)(https://biohub.ai/esm/protein/about)

25:38 — Luke Bailey: Scaling Self-Play with Self-Guidance (AlphaZero-style Self-play for LLMs) (https://arxiv.org/pdf/2604.20209)

37:51 — Arnab Maiti: Stream RAG: Instant and Accurate Spoken Dialogue Systems with Streaming Tool Usage (https://arxiv.org/pdf/2510.02044)

47:40 — Robert George: Lean for Science: How Formal Proofs Can Change Mathematics, AI, and Scientific Computing (https://arxiv.org/abs/2602.22631)

58:52 — Luke Orthwein: Founder AI Hacks: Programming is an RTS Game Now

1:16:07 — Closing Remarks