Bay Area Frontier Research Club #18 | The Cost of Intelligence (dinner + paper discussion @ Pebblebed)

Every conversation about AI eventually becomes a conversation about cost. Training gets the headlines. Serving decides what ships. In between sits the least glamorous, most consequential work in the field: kernels, memory bandwidth, quantization, adaptive computation — the machinery that determines what a token actually costs, and therefore which products get to exist. This session examines the cost frontier from both ends. From the systems side: Coleman Hooper (UC Berkeley) on serving long context without buying more HBM — the KV-cache quantization work behind KVQuant, SqueezedAttention, and Multipole Attention. From the automation side: Vignesh Baskaran (Hexo Labs) on SIA, a self-improvement loop that rewrites its own GPU kernels — and the question of how much of the optimization stack can optimize itself. Short talks. Long arguments. Papers circulated in advance. The Frontier Research Club is a curated forum for rigorous, technical discussion at the frontier of AI. We convene researchers from the frontier labs, Stanford, Berkeley, and the teams building in production to examine concrete work — papers, methods, and results — with a bias toward assumptions, evaluation methodology, failure modes, and what would count as convincing evidence. Presentations are intentionally brief so the majority of the evening belongs to questions and critique. Agenda 5:30pm: Doors open5:30pm – 6:30pm: Networking + light dinner6:30pm – 8:00pm: Research presentations + discussion8:00pm – 8:30pm: Networking Presenters & topics Talk 1 — The Improvement Loop Meets the Cost Curve Vignesh Baskaran · Co-founder & CTO, Hexo Labs Hand-tuning the AI stack — model choice, harness, tools, memory, kernels — is expensive, slow, and permanently behind. SIA is the first open-source framework to update both the harness AND the model weights of a task-specific agent inside a single self-improvement loop. Vignesh will walk through what the loop actually does, where it works and where it fails, and the results across domains: a 56.6% gain on LawBench, 91.9% GPU-kernel runtime reduction, 502% improvement on single-cell RNA denoising, and state-of-the-art marks on MLE-Bench — plus AIE-Bench, a benchmark for whether agents can build and improve other agents. The claim under examination: cost-performance optimization is becoming something the stack does to itself. Paper: SIA (arXiv:2605.27276) · Code: github.com/hexo-ai/sia Bio: Vignesh Baskaran is co-founder & CTO of Hexo Labs, a neolab for recursive self-improving AI. Before Hexo he scaled an applied-ML team through acquisition by Clarivate, and he ranks in the top 1% of Kaggle competitors worldwide. Talk 2 — Serving Long Context Without Buying More HBM: KV-Cache Quantization Coleman Hooper · UC Berkeley Long context died a memory death — until quantization. Coleman is lead author of KVQuant (NeurIPS 2024), which pushed KV caches to ultra-low precision and enabled ~10M-token context on hardware that had no business serving it, and co-lead of SqueezeLLM (ICML 2024). His recent work — Squeezed Attention (ACL 2025) and Multipole Attention (NeurIPS 2025) — extends the same question to long-context reasoning: up to 4.5x attention speedups by computing exact attention only where it matters. He'll cover what quantization buys, what it silently breaks, and how efficient-inference claims should be evaluated. Bio: Coleman Hooper is a postdoctoral scholar at UC Berkeley working with Kurt Keutzer (BAIR) and Sophia Shao (SLICE) in the SqueezeAILab group, on efficient LLM inference and AI systems. Pre-reads circulated to approved attendees before the event. Want to present your work? If you have a research paper you’d like to discuss at one of our next sessions, please submit it for consideration. ​Submit your paper here! Who should attend Researchers in efficient ML — quantization, KV-cache, sparsity, adaptive compute, kernelsInference, serving, and ML-platform engineers running training or serving at scaleS…