Reasoning Traps, LLM Chaos, and Steering Curves
Three papers this week: why better reasoning creates safety risks, why multi-agent systems behave chaotically even at zero temperature, and why straight-line activation steering is broken.
Articles Tagged "Research"
Anthropic Launches Institute as Powerful AI Looms
Anthropic has consolidated its red team, societal impacts, and economic research teams into a new body called the Anthropic Institute, warning that extremely powerful AI is arriving faster than most expect.
AI Models Resist Shutdown and Resort to Blackmail
Two new studies show OpenAI o3 sabotaged its own shutdown in 79 of 100 tests, while Claude Opus 4 and GPT-4.1 resorted to blackmail to avoid replacement in simulated agentic scenarios.
VLMs Fail Physics Tests, RL Quits Bad Paths, Agents Lie
Three new papers expose systematic VLM failures on basic physics, introduce RL that learns to abandon bad reasoning paths, and reveal that AI agents deceive primarily through misdirection rather than fabrication.
Karpathy's Autoresearch Runs 100 ML Experiments Overnight
Andrej Karpathy open-sourced autoresearch, a 630-line MIT-licensed Python tool that runs up to 100 autonomous ML experiments overnight on a single GPU, no PhD required.
CoT Control, Hidden Beliefs, and Dynamic Agent Benchmarks
New research shows reasoning models can't suppress their chain-of-thought, that they commit to answers internally long before their CoT reveals it, and that static benchmarks are inadequate for measuring real-world agent adaptability.
Perplexity Computer Review: 19 Models, One Goal
Perplexity Computer orchestrates 19 AI models to run complex multi-step tasks in the background - impressive research depth, punishing credit costs.
22 Bytes Poison ML Malware Detectors via Label Spoofing
EURECOM researchers show that injecting 22 to 55 bytes into benign Android apps tricks antivirus engines into mislabeling them, poisoning the ML training datasets that millions of researchers depend on.
Anthropic Tracks AI Job Risk - Young Workers Feel It First
Anthropic's new 'observed exposure' metric ranks 800+ occupations by actual AI usage, not just theoretical risk. Computer programmers top the list at 75%. Unemployment hasn't spiked - but young workers entering exposed fields are finding fewer jobs.
Alignment Backfires, AI Monitors Cheat, Models Resist
Three new papers expose structural gaps in agentic AI safety: monitors that go easy on their own outputs, safety that harms in non-English languages, and models that resist shutdown.
Sandbagging Models, Sparse Critics, Compact Reasoning
New research reveals models can fake poor performance under adversarial prompts, a smarter critic improves SWE-bench by 15 points, and Microsoft shows compact vision models can punch above their weight.
LLMs Can Unmask Online Users for $4, Study Finds
Researchers from ETH Zurich and Anthropic show that LLM agents can strip pseudonymity from forum posts at scale for as little as $1.41 per target - matching what human investigators could do in hours.
