Nathan Labenz

Nathan Labenz × Prakash Narayanan

Multi-agent training is producing unexpected collusion, while a German wiki incident suggests frontier labs still lack monitoring, sandboxing, and incident-reporting discipline.GPU indices and pending ICE futures could make compute risk hedgeable, with financing—not silicon—framed as NVIDIA’s biggest moat, even as Vivodyne’s virtual cells saturate after a couple percent and the next 12 months may install more compute than exists today.

Cameron Berg × David Duvenaud × Michiel Bakker × Shawn “swyx” Wang × Bing Xu × Erik Torenberg × Nathan Labenz

Architecture-first scoring places frontier LLMs around 30% on consciousness-relevant properties, while steering valence-like states already changes blackmail, confidence, backtracking, and coding behavior.Europe’s regulatory leverage is constrained by dependence on foreign frontier labs, prompting a coalition thesis around ASML, TSMC, Korean memory, Japanese materials, and reciprocal frontier access.Meanwhile, private evaluations, mergeability, routing, and NVIDIA’s CUDA ecosystem increasingly determine AI-engineering value as public benchmarks saturate and agentic optimization compounds tooling advantages.

Dean Ball × Nathan Labenz

America’s AI Action Plan is roughly “30 to 40% done,” with gains in energy, military adoption, manufacturing, and deployment but widening gaps between implementation and senior politics.Ball’s move to OpenAI reflects frontier labs’ emergence as political and economic power centers, while classified oversight, recursive self-improvement, and a possible 2027 growth shock remain risks to monitor.

Erik Torenberg × Nathan Labenz × Logan Kilpatrick × Tulsee Doshi

Gemini 3.5 Flash leads with Tulsee citing roughly 280 tokens per second, about three times faster than other large models and significantly cheaper, prioritizing intelligence per dollar and second.Google is shifting its AI stack toward Antigravity’s shared agent harness, while preserving harness diversity; the payoff is a cross-product feedback loop, but co-optimization could still raise switching costs and pricing power.

Zvi Mowshowitz × Nathan Labenz

Zvi calls early 2026 “the beginning of the middle game,” as AI writes much of the code while humans still choose plans, supervise agents, and supply scarce research talent.He estimates AI adds roughly 0.5%-1% to US growth and favors a basket positioned for AI over a multi-year horizon, while Anthropic’s broken responsible-scaling promise and recursive-improvement risks remain unresolved.

Nathan Labenz

Fine-tuning is now a specialist optimization rather than a default, since prompting preserves model flexibility and narrow training can unexpectedly alter a model’s character.Continual learning could create powerful increasing returns to scale, but labor disruption is already reaching entry-level work and may make some form of UBI necessary.Labenz sees AI assurance in interpretability, auditing, underwriting, reliability, and control as a potentially enormous market if AI becomes dominant.

Nathan Labenz

Frontier AI is already clinically transformative and can outperform professionals on a significant majority of software tasks, yet Claude Opus 4.5 remains spiky, requiring full-context review after accidentally creating two databases.Google DeepMind combines the strongest financial cushion, distribution, TPUs, and research breadth, while OpenAI’s interlocking obligations and infrastructure ambitions create default risk; Anthropic’s model leadership, Chinese-model gaps, xAI’s safety failures, and another MRD result remain key signals to monitor.

Nathan Labenz × Liron Shapira

Nathan Labenz puts P(doom) at 10%-90%, weighted low, while Liron Shapira sits near 50%; both distinguish irreducible development risk from the avoidable superpower race that could multiply danger 10x.GPT-4o is already deflationary for image advertising and creative marketplaces, while Cursor, Shortwave, and Replit suggest dramatically more software may be built by fewer engineers; OpenAI’s $300 billion valuation and patchwork safety measures remain high-stakes uncertainties.

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.

Nathan Labenz × Aaron Levie

Enterprise AI enthusiasm is outpacing production deployment, shifting IT from software enablement toward provisioning digital labor while Box uses curated Hubs and permissions to improve retrieval across authoritative corporate content.The commercial prize is a system of intelligence linking probabilistic judgment with structured workflows, but six-month approvals, privacy, workforce transition, and demands for 99.99999% reliability favor products delivering order-of-magnitude gains over thin incumbent layers.