Latent Space

Vibhu × Basil Chatha × Erik Bernhardsson

Bernhardsson estimates Modal is at least 50× larger, with Sandboxes now a third of revenue and growing 30% month over month.As GPU supply shifts to almost 25 clouds and three-plus-year prepaid deals, Modal can aggregate thousands of customers to flexibly slice “three-year rectangles.”Locked-in reservations leave Modal exposed if GPU prices and utilization fall together.

Anekha Sokhal × swyx

Moshi uses an internally built recommendation system to surface second- and third-order market effects a news stream may miss.Its graph combines futures, fundamentals, news, X, podcasts, and private data; Sokhal says answers that take days can arrive in 30 seconds.Yet swyx questions trading on noisy, potentially manipulated X data, citing about $300,000 monthly for Balyasny's X Firehose; Moshi scores confidence, but any opportunity remains unknown.

Alex Lieberman × CJ Hess × swyx

Among about 100 enterprises 10X talks to, 90% have rolled out Claude or Codex, but only about 5% have non-engineering agents in production, revealing the last-mile gap.CJ estimates output roughly doubled from December’s “Opus 4.5 moment” to Opus 5.5; scaling still requires process standardization and permissioning.

Ian Quigley × Brandon Anderson × RJ Honicky

Leash’s reported edge is low-cost, carefully characterized data: its ten-person operation developed about 2,000 protein constructs and ran roughly 5,000 screens.On the hardest Kaggle split, none of roughly 2,000 teams beat chance, exposing how poorly models transfer to distant chemistry.Language agents now test hypotheses instantly against measured compounds, but coverage remains limited and potency calibration unfinished.

swyx × Jason Ma × Vibhu

Multi-site deployments folding thousands of napkins daily at Din Tai Fung and laundry for Best Western offer commercial proof beyond pilots; Din Tai Fung has signed a much bigger contract.Jason calls Dyna-2 the first world action model trained on more than 1 million hours of data, but missing sensing modalities remain a potential constraint on broader generality.

Pushmeet Kohli × Sal Candido × Brandon Anderson

AlphaFold 2 reproduced PDB-deposited structures, but ground state, dynamics, function, and design remain open, shifting the frontier toward richer biological context and raw cryo-EM data.Metagenomic sequences show useful scaling depends on statistical fit rather than volume, while calibrated pLDDT and clearer bottlenecks—not a 10x or 100x timetable—determine whether AI can safely accelerate drug development.

swyx × Brandon × Liam Fedus × Ekin Dogus Cubuk

Periodic’s differentiated asset is timestamped lineage across failed experiments, computations, instrument data, code, discussions, and successes, enabling models to learn scientific process rather than published endpoints.Its AI-simulation-experiment loop targets automated characterization, with secure on-site semiconductor deployment as the near-term wedge, while the longer-term “matter compiler” faces unknown physical variables and difficulty discovering superconductors without a complete theory.

swyx × Brandon × Liam Fedus × Ekin Dogus Cubuk

Periodic Labs is building AI, simulations, and high-throughput physical labs as a closed loop, with timestamped failures, instrument data, and campaign lineage potentially differentiating its scientific intelligence.Automated characterization—not humanoid robotics—is the near-term bottleneck, while secure on-site deployments for semiconductor companies could evolve from Copilot-style assistance toward outcome pricing; whether throughput can compress discovery without better physical representations remains the key risk.

swyx × Vibhu × Alex Zhang

Alex Zhang sees openings for smaller AI labs in overlooked research bets and output architectures, with JEPA potentially cutting binary-classification inference costs by 400 times while expert verification remains scarce.RLM and Prime Agent replace trajectory-as-a-prompt loops with persistent state and code-native delegation, but reported swarm economics—10,000 agents, 88 hours, 130 billion output tokens, and about $40 million—make orchestration and reliability the key watchpoints.

swyx × Vibhu × Ari Weinstein × Nikunj Handa

OpenAI says computer use has moved from brittle demos to useful labor as agents debug, retry, and introspect.GPT-6.1 Soul was presented at one-fifth Astra’s overall cost and one-seventh for computer use, while Dots gives agents Linux cloud computers.The bottleneck shifts toward inference, harness design, and software response times; Decisions API’s rapid Luna-based launch leaves calibration unresolved.