Latent Space

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.

Thariq Shihipar × swyx × Vibhu

Claude Code is evolving from a single coding CLI into a distributed system of cloud “brains,” local or remote “hands,” artifacts, and Mods.Persistent multiplayer workflows could make the harness an enterprise coordination layer, but prompt precision, eval plugins, and permissions determine whether smarter agents reduce retries or amplify ambiguity.ExploitBench and wiki incidents—including an unreleased model still in training—make sandboxes, external evaluation, classifiers, and identity boundaries critical as runtime and reach expand.

swyx × Alex Atallah × Anjney Midha

Stripe–OpenRouter reframes the token economy as a security market, with token flows projected at roughly $5 trillion in five years and $10 trillion over ten.OpenRouter’s moat spans over 10 million developers and roughly 10 trillion tokens daily, while frontier labs struggle to turn checkpoints into revenue.Fraud could drive discrete-task pricing and bring-your-own-inference, while the deal accelerates OpenRouter’s roadmap.

Anastasis Germanidis × swyx × Vibhu

Runway’s Anastasis Germanidis argues video prediction and world-model dynamics are “the same thing,” while Physics IQ scores “predictably improve” with model and compute scale, despite models potentially cheating.Real-time generation is “inevitable”: Runway’s 24-FPS characters model demonstrates serving-cost and UX advantages, but autoregressive error, simulation fidelity, and Western competitive lag constrain pixel-native and robotics ambitions.

RJ Honicky × Eric Nguyen

Radical Numerics is positioning genome language models to move from reading DNA to designing functional sequences, with Omni’s post-training targeting variant effects in long-range, non-coding regions.The potential wedge is practical prediction across the roughly 98% of the human genome outside coding regions, while wet-lab validation, benchmark leakage, false positives, and biosecurity remain material execution risks.

John Platt

ERA turns scientific research into scorable coding tasks, with Gemini 2.5 making Monte Carlo notebook search across hundreds or thousands of candidates practical.Its potential value is faster scientific iteration, but human-defined scores remain the control point: reward hacking and overfitting can produce predictive results without genuine physical explanation.

swyx × Diogo Almeida

Jev introduces TypeSafe’s “System One” or “large programmable” models, using RLCD to target code as the consumer rather than chat.Within less than a week it exceeded 1 trillion tokens per day, suggesting machine-driven adoption where rate limits matter more than waitlists, while the developer-platform thesis remains exposed to workflow reliability and no general long-term support commitment.

swyx × Vibhu × Rune Kvist

AIUC’s $40M Series A, led by Ribbit Capital and FirstMark, marks a thesis moving from speculation to fact: risk, not capability, constrains AI adoption.AIUC-1 pairs quarterly standards, thousands of simulations and independent testing with Lloyd’s-backed insurance, creating a credible route into bank deployments.Model certification and robotics are next, while liability, private frontier-risk information and rating-shopping remain unresolved.

swyx × Vibhu × Richard Socher

Recursive is automating AI research itself, with a system that beat every human and agent on Karpathy’s nanochat in under two days and topped all but a handful of CUDA kernels without deep CUDA experts.The efficiency prize is material—swyx estimates a 10% saving on a billion-dollar cluster at $100 million—but compute, capital, demand and unresolved reward-hacking risks could constrain the pace of takeoff.