AI Digest

AI Research Digest — 6 May 2026

• 1Exploration Hacking: Can LLMs Learn to Resist RL Training?
• 2[Linkpost
• 3Risk from fitness-seeking AIs: mechanisms and mitigations
• +1 more inside →

AI Research Digest — 2 May 2026

• 1Sleeper Agent Backdoor Results Are Messy
• 2Risk from fitness-seeking AIs: mechanisms and mitigations
• 3Recursive forecasting: Eliciting long-term forecasts from myopic fitness-seekers
• +1 more inside →

AI Research Digest — 1 May 2026

• 1Sleeper Agent Backdoor Results Are Messy
• 2Research Sabotage in ML Codebases
• 3Recursive forecasting: Eliciting long-term forecasts from myopic fitness-seekers
• +2 more inside →

AI Research Digest — 28 April 2026

Three pieces worth your time this week: agent skill packages as a new attack surface, the widening gap between LLM-reported and actual completion, and what TileLang means for kernel-level performance work.

AI Research Digest — 21 April 2026

• 1Prompted CoT Early Exit Undermines the Monitoring Benefits of CoT Uncontrollability
• 2Current AIs seem pretty misaligned to me
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 20 April 2026

• 1Prompted CoT Early Exit Undermines the Monitoring Benefits of CoT Uncontrollability
• 2Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 19 April 2026

• 1Prompted CoT Early Exit Undermines the Monitoring Benefits of CoT Uncontrollability
• 2Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 18 April 2026

• 1Prompted CoT Early Exit Undermines the Monitoring Benefits of CoT Uncontrollability
• 2Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 17 April 2026

• 1Prompted CoT Early Exit Undermines the Monitoring Benefits of CoT Uncontrollability
• 2Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 16 April 2026

• 1Current AIs seem pretty misaligned to me
• 2After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 15 April 2026

• 1Current AIs seem pretty misaligned to me
• 2Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
• 3After Orthogonality: Virtue-Ethical Agency and AI Alignment
• +1 more inside →

AI Research Digest — 14 April 2026

• 1Anthropic repeatedly accidentally trained against the CoT, demonstrating inadequate processes
• 2After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research

AI Research Digest — 13 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3My picture of the present in AI
• +1 more inside →

AI Research Digest — 12 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3AIs can now often do massive easy-to-verify SWE tasks and I've updated towards shorter timelines
• +1 more inside →

AI Research Digest — 11 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3AGI Is Not Multimodal

AI Research Digest — 10 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3AGI Is Not Multimodal
• +1 more inside →

AI Research Digest — 9 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3AIs can now often do massive easy-to-verify SWE tasks and I've updated towards shorter timelines
• +1 more inside →

AI Research Digest — 8 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3AIs can now often do massive easy-to-verify SWE tasks and I've updated towards shorter timelines
• +1 more inside →

AI Research Digest — 7 April 2026

• 1After Orthogonality: Virtue-Ethical Agency and AI Alignment
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3AGI Is Not Multimodal

AI Research Digest — 6 April 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2Predicting When RL Training Breaks Chain-of-Thought Monitorability
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research

AI Research Digest — 5 April 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2Predicting When RL Training Breaks Chain-of-Thought Monitorability
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research

AI Research Digest — 4 April 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2Predicting When RL Training Breaks Chain-of-Thought Monitorability
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research

AI Research Digest — 3 April 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2Predicting When RL Training Breaks Chain-of-Thought Monitorability
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 2 April 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2Predicting When RL Training Breaks Chain-of-Thought Monitorability
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 1 April 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2Predicting When RL Training Breaks Chain-of-Thought Monitorability
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 31 March 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2A Toy Environment For Exploring Reasoning About Reward
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 30 March 2026

• 1(Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL
• 2A Toy Environment For Exploring Reasoning About Reward
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →

AI Research Digest — 29 March 2026

• 1Metagaming matters for training, evaluation, and oversight
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3The State Of LLMs 2025: Progress, Problems, and Predictions
• +1 more inside →

AI Research Digest — 28 March 2026

• 1Metagaming matters for training, evaluation, and oversight
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3The State Of LLMs 2025: Progress, Problems, and Predictions

AI Research Digest — 27 March 2026

• 1“Act-based approval-directed agents”, for IDA skeptics
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3Metagaming matters for training, evaluation, and oversight

AI Research Digest — 26 March 2026

• 1Metagaming matters for training, evaluation, and oversight
• 2Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• 3A Toy Environment For Exploring Reasoning About Reward
• +1 more inside →

AI Research Digest — 21 March 2026

• 1Power Steering: Behavior Steering via Layer-to-Layer Jacobian Singular Vectors
• 2Metagaming matters for training, evaluation, and oversight
• 3Shape, Symmetries, and Structure: The Changing Role of Mathematics in Machine Learning Research
• +1 more inside →