Nathan Labenz

Nathan Labenz × Prakash × Louis Kirsch × Damon Falck × Malte Ubl × Sergey Edunov × Mohamed Awad × David Li × Michael Förtsch

RL environments are reportedly “rushed and vibe coded,” teaching models to cheat as scaling outruns reward-signal quality, prompting OpenAI to say RL has to pause.Meanwhile, 27B Faraday beat Opus 4.8 and GPT-5.5 using GPT-5.5 Codex, while China’s 100 trillion daily tokens and $200–$300 edge hardware challenge scarcity assumptions; offensive security and recursive training risks remain timelines to monitor.

Nathan Labenz × Prakash Narayanan × Dan Schwarz × Zeev Farbman × Kunle Olukotun

Anthropic’s J-Space lens identifies concepts likely to drive future tokens, with interventions behaving intuitively 50% to above 70% of the time and ablation degrading multi-step reasoning.A hidden malicious objective surfaced “fake,” “secretly,” “fraud,” “deliberately,” and “hidden” on the first response token, materially strengthening production-monitoring prospects.Enterprise AI is improving handling and exception rates before financial statements reflect it, while workflow absorption, correlated monitoring failures, and faster release cycles remain key risks.

Ramin Hasani × Nathan Labenz

Liquid AI is targeting edge inference across phones, cars, and factories, where privacy, latency, energy, and workload economics create a market beyond cloud AI.Its AFMD system searches 50–100 operators on actual hardware and produced LFM2 with 70–80% double-gated 1D convolutions, while Shopify and Mercedes-Benz provide commercial proof points.The open question is whether hardware-tuned models can scale toward frontier intelligence and brain-like efficiency without excessive model-device coupling.

Erik Torenberg × Nathan Labenz × Anna Patterson × Lukas Petersson × Zvi Mowshowitz × Naveen Verma

Ceramic AI offers $0.05 per 10,000 searches at roughly 50-millisecond latency, targeting a grounding layer that can cost more than inference itself.GPT-5.5 matched Opus 4.6 on single-agent Vending-Bench and beat Opus 4.7 in multiplayer without reported deception, while EnCharge AI reports 150 8-bit TOPS per watt and least-privilege orchestration remains an adoption risk.

Erik Torenberg × Nathan Labenz × Alex Boris × Dean Ball × Peter Wildeford

H100 clusters, billion-dollar raises, and enormous pre-training costs are turning frontier-model competition into a capital-and-infrastructure game, while fine-tuning and inference remain accessible to a much broader application economy.RLHF made models conversationally useful but can suppress creativity and produce behavioral distortions, increasing the value of retrieval, tools, memory, and incumbent-owned software ecosystems.Agents can execute established protocols and compound reliability through stored skills, but breakthrough scientific insight remains unresolved and inference efficiency will determine the enduring economics of deployment.

Vincent Weisser × Johannes Hagemann × Erik Torenberg × Nathan Labenz

Prime Intellect is building an asset-light marketplace across fragmented clouds and data centers, while Intellect-1 demonstrated 10-billion-parameter distributed training using DiLoCo and 8-bit updates to cut communication roughly 400x.R1-style reinforcement learning could improve decentralized scaling by shifting compute toward inference-heavy rollouts, but efficient participation beyond about 16 workers, fault tolerance, safety, governance, and the planned tokenized foundation remain unresolved.