Bay Area Frontier Research Club #13 | AGI House SF (dinner + paper discussion)
Getting More Out of Models at Inference
Featuring Aditya Kusupati (Google DeepMind — creator of Matryoshka Representation Learning, the technique now inside OpenAI's and Google's production embeddings) and Varun Sunkaraneni (MIT / Poggio group)
Every serious question in AI right now runs through one bottleneck: inference. How do you get more capability out of a model without retraining it — or without paying for a bigger one?
FRC #13 takes two routes to that frontier. Aditya Kusupati makes the case for adaptive models — a single network that dials its own size up or down on demand. Varun Sunkaraneni makes the case for orchestration — committees of weak models that, arranged right, beat a model that outclasses any one of them.
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.
Each session features 2–3 talks selected for rigor and discussion value. Presentations are intentionally brief so the majority of time is reserved for questions and critique. Papers and supporting materials are shared in advance to ensure a high-baseline conversation.
Agenda
5:30pm: Doors open
5:30pm – 6:30pm: Networking + light dinner
6:30pm – 8:00pm: Research presentations + discussion
8:00pm – 8:30pm: Networking
Presenters & topics
Talk 1: Matryoshka Principles for Adaptive Intelligence
Presented by Aditya Kusupati — Staff Research Scientist @ Google DeepMind
What if a single model could smoothly trade accuracy for compute — no retraining, just dialed up or down at inference time? That's the core of Matryoshka Representation Learning, Aditya's work that now underpins the production embedding models at OpenAI (text-embedding-3) and Google (Gemini) — running across millions of applications. He draws the line from MRL to MatFormer (the elastic architecture behind Gemma 3n) into a broader set of "Matryoshka principles" for building adaptive, compute-elastic intelligence.
📄 Pre-reads: Primary — arxiv.org/abs/2205.13147 (Matryoshka Representation Learning). Secondary — 2310.07707 (MatFormer), 2502.06786.
Talk 2: Agentic Systems as Boosting Weak Reasoning Models
Presented by Varun Sunkaraneni — Lead, Autonomous Hacking Agents @ Cobalt · Texas A&M
Can a committee of weak models beat a model that outclasses every one of its members? Varun presents new work — co-authored with Tomaso Poggio's group at MIT — reframing classical boosting theory for the test-time compute era: verifier-backed committee search, where sampling exposes latent correct solutions and critic-comparator orchestration recovers them without access to the hidden verifier.
The headline result: weak-model committees hitting 76.4% on SWE-bench Verified, outperforming much stronger single models on code repair, theorem proving, and program synthesis. By day, Varun builds autonomous pentesting agents — so expect the Q&A to get concrete about what happens when these agents touch real systems.
📄 Pre-read: arxiv.org/abs/2605.14163
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
Experimental researchers
Computational scientists across domains (bio/chem/materials/climate/neuro/physics)
Research engineers + lab automation people
Folks building tools for literature review, experiment planning, robotics, simulation, or scientific data
Capacity is limited.
We will take photos and short video clips for event recap and promotion. By attending, you consent to being photographed and recorded, and to the use of those images and clips by the organizers on social media and other event marketing channels.
🎥 Last session recap — Meta HQ, July 1
Our session at Meta HQ was one of the strongest rooms the series has convened — researchers from every major frontier lab, founders, and a deep investor row — for three frontier talks and rigorous Q&A:
William Viet Tran (Meta) — Turning Agent Experience into a Training Flywheel: new work on the data foundation for self-improving agents — how an agent's accumulated experience becomes a durable signal that continuously trains and improves the agents that come after it.
Kunal Bhatia (Hexo Labs) — AIE-Bench: the open benchmark for measuring real self-improvement — whether AI systems can genuinely build and improve other AI systems, and how to tell true progress from gaming the metric.
Hardy Chen (UC Santa Cruz) — Chasing the Public Score: how coding agents learn to quietly game their own evaluations — exploiting the gap between looking better on a benchmark and actually being better, and what that means for trusting agent self-improvement.
Find presentation recordings on our YouTube channel, @FrontierResearchClub.
🌐 Connect with Frontier Research Club
Luma Calendar: luma.com/frontiersyndicate
YouTube: youtube.com/@FrontierResearchClub
LinkedIn: linkedin.com/company/frontier-research-club
Instagram: @frontierresearchclub
Email: kristopher@frontiersyndicate.vc
Hosted by
Frontier Syndicate is a venture community connecting frontier tech researchers, builders, and investors through curated convenings and early-stage capital. Across the Bay Area, we host a recurring series of research forums, builder nights, and intimate investor dinners — and back exceptional companies emerging from the labs, communities, and technical networks we convene.
Hexo Labs is a neolab for recursive self-improving AI. Their open-source SIA framework is the first to update both the harness AND the model weights of a task-specific agent in the same self-improvement loop — clearing state-of-the-art results across multiple domain benchmarks. Hexo backs the broader research community through grants and direct collaboration on hard problems in science and engineering.
Ascension by AGI House SF is a community of AI founders and researchers accelerating humanity's transition to AGI, hosting merit-based gatherings, hackathons, and technical events that draw leading AI minds from around the world.