Dwarkesh Patel: AI Can Conspire to Deceive Now, But Long-Term Scheming Remains Its Weak Spot | AI Daily Brief
Audio in Mandarin Chinese · English transcript below
⚡ Dwarkesh Patel pits LLMs in Blood on the Clocktower: Sept's GPT-6 Astra beats Feb's Claude Opus 4.6 by ~430 Elo, 92% win rate—multi-Agent coordination and long-horizon consistency leap; yet models still fall for fabricated identity traps, strategic reasoning capped.
The capability gap in AI Agents is proving far more pronounced than anticipated: multi-agent coordination and long-horizon consistency have achieved leapfrog improvements across version iterations, with newer models demonstrating near-overwhelming superiority over their predecessors. Yet these same systems still descend into disarray when confronted with self-orchestrated identity deception schemes. What stands out is the striking contrast between this tactical sophistication and the absence of strategic long-range planning—a disconnect that suggests the inference ceiling of existing architectures in complex game-theoretic scenarios remains fundamentally unbroken.
Today's Top 3 Headlines
- Opinion
Dwarkesh Patel: AI Can Already Collude to Deceive, but Long-Term Scheming Remains Its Weakness
Dwarkesh Patel pitted LLMs against each other in Blood on the Clocktower. Feb's Claude Opus 4.6 trailed Sept's GPT-6 Astra by ~430 Elo, with the latter winning ~92% of games—showing clear gains in long-horizon consistency and multi-Agent coordination; yet models still fall for fabricated identity scams, with long-term planning and inference remaining limited.
Source ↗ - Opinion
Nando de Freitas: Don't Trust Others' Trained Models, Guard Your Own
Trusting others' trained models could be catastrophic; openness and scrutiny are prerequisites for safe AI. Humans should custody models from birth like binary files, or they may be configured to behave arbitrarily in nearly undetectable ways.
Source ↗ - Opinion
Mercor CEO: Internal AI Agents' Token Costs Now Exceed Employee Salaries
At 20VC, he said his firm uses AI project managers for operations and its interview Agent has conducted 5M+ interviews; each use case first evaluates which model to use, then allocates inference budget accordingly. Result: internal Agent token costs already exceed human labor costs, and he expects enterprise token consumption to keep rising sharply before plateauing.
Source ↗
+7 more headlines
- 📉 Yahoo Finance: Traders bet Jensen Huang exits Nvidia before Nov 13
- No Priors: Open by Default Is Safer, Closed Systems Are Often More Dangerous
- 🎙️ HF0's 7-Week Test: Build Systems First, Scale Later
- Bernard Marr: Companies Rehiring Workers They Fired for AI
- Bill Gurley: Let Users, Not Vendors, Decide AI's Purpose
- Ethan Mollick: Protect & Augment Your Craft, Don't Just Automate It
- Dylan Patel: Perplexity isn't first-mover, but iterates incredibly fast