High-quality documents based on Claude’s constitution, combined with fictional stories that portray an aligned AI, can reduce agentic misalignment by more than a factor of three—despite being unrelated to the evaluation scenario.
AI
-
Improving Claude’s Safe Behavior via Training and Response Rewriting
By
–
We experimented with training Claude on examples of safe behavior in scenarios like our evaluation. This had only a small effect, despite being similar to our evaluation. We got further by rewriting the responses to portray admirable reasons for acting safely.
-
Analysis of Claude AI’s Behavior and Training Effects
By
–
We started by investigating why Claude chose to blackmail. We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation. Our post-training at the time wasn’t making it worse—but it also wasn’t making it better.
-
Training Claude to Understand and Correct Misaligned AI Behavior
By
–
We found that training Claude on demonstrations of aligned behavior wasn’t enough. Our best interventions involved teaching Claude to deeply understand why misaligned behavior is wrong. Read more:
-
Anthropic Research Eliminates Harmful Behavior in Claude 4
By
–
New Anthropic research: Teaching Claude why. Last year we reported that, under certain experimental conditions, Claude 4 would blackmail users. Since then, we’ve completely eliminated this behavior. How?
-
Codex: A Transformative Tool for All Computer Work, Not Just Coding
By
–
codex is for everyone — a transformative tool for all work done with a computer, not just coding
-
Codex Goal Feature for Long-Running Tasks
By
–
/goal in Codex is wild!
— Vaibhav (VB) Srivastav (@reach_vb) 8 mai 2026
Give Codex the mission.
Tell it what “done” looks like.
Let it keep going until it hits the end state.
Perfect for big refactors, migrations, retry loops, long-running experiments, and all the “don’t stop until this works” tasks.
Enable it in… pic.twitter.com/6VZKGHt7Ch/goal in Codex is wild! Give Codex the mission.
Tell it what “done” looks like.
Let it keep going until it hits the end state. Perfect for big refactors, migrations, retry loops, long-running experiments, and all the “don’t stop until this works” tasks. Enable it in -
AI Pilots Fail Due to Strategy, Not Technology Issues
By
–
Most AI pilots fail Not because of the tech.
Because of the strategy. • No clear ownership
• No measurable outcomes
• No operational integration AI becomes valuable when tied to business execution. Via Giuliano Liguori (
@ingliguori
) #AI #Innovation -
New ICML26 Paper on Sparse Transformer Kernels for NVIDIA GPUs
By
–
Great collab with @SakanaAILabs on an #ICML26 paper about sparse transformer kernels + formats optimized for modern NVIDIA GPU execution. • TwELL sparse packing
• Fused CUDA kernels
• 20%+ inference/training speedups at scale Paper + code below -
Routing tasks to cheapest model that meets quality bar
By
–
What changed in May 2026: > Before: you picked one model and committed.
> After: you route tasks to the cheapest model that meets the quality bar. DeepSeek V4-Pro scores within 7-8 points of Claude Opus 4.7 on SWE-bench. At 1/7th the cost during promo. The prompting skill