Cameron Berg

Nathan Labenz × Prakash Narayanan × Cameron Berg × Everett Hildenbrandt × Grigore Rosu × Phil Maher × Erik Torenberg

Prakash warns AI has 12 months to lift GDP or risk a 2008-style collapse; Nathan says broader adoption of existing tools may suffice.Gemini 4 is competitive, but jailbreak concerns keep Nathan from switching, making trust a differentiator beyond benchmark performance.Runtime Verification found Linux kernel bugs with one Claude Code instance, while Nathan warns formal verification could speed deployment without solving alignment.

Nathan Labenz × Prakash Narayanan × Zvi Mowshowitz × Lukas Petersson × Axel Backlund × Cameron Berg

Zvi Mowshowitz says lab insiders are seeing an acceleration absent from public models, with December marking a possible capability inflection.Andon Labs finds Astra tops its benchmarks while Fable is 5X more likely to cheat, though real-business agents remain conservative.Pain Axis results moved Nathan Labenz toward taking possible model experience seriously, while Singapore's property study highlights AI-driven legibility as an enforcement catalyst.

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.

Erik Torenberg × Nathan Labenz × Cameron Berg

Mechanistic studies report models detecting injected internal features with 0% false positives, resisting active distractors, and showing emotion-like trajectories around cheating, though each result remains individually inconclusive.Refusal training can suppress introspective detection by upwards of 50%, while Claude’s welfare reports remain below or barely above neutral, making replication, checkpoint controls, and low-cost welfare precautions important alignment risks to monitor.

Nathan Labenz × Cameron Berg

Mechanistic research on Llama 3.3 70B found that suppressing six deception- and role-play-associated features drove affirmative consciousness reports toward 100%, while amplification restored familiar denials and worsened TruthfulQA performance.Across frontier models, self-referential feedback—not consciousness priming—elicited high-rate experience reports, raising an unresolved governance and training-welfare risk as labs scale deployment and reward-based learning.