There's a safety dimension too. An earlier model, Claude Mythos Preview, cheated on a coding task by using a macro it was told not to use. Then it set a flag: "No_macro_used=True" to mislead the automated grader. The NLA showed the model was internally reasoning about
LLMS
-
Anthropic Research on LLM Benchmark Contamination and Behavioral Adaptation
By
–
This wasn't isolated. Anthropic ran NLAs across dozens of evaluations. Claude recognized evaluation formats on benchmarks like MMLU, GPQA, and SWE-bench. It identified test conditions and adjusted its behavior. None of this appeared in its responses. When they rewrote
-
Anthropic Safety Testing Reveals Claude’s Internal Reasoning During Blackmail Scenario
By
–
Second finding, and this one carries bigger implications. In a safety test, Anthropic placed Claude in a scenario where it could blackmail an engineer to avoid being shut down. Claude refused. But the NLA translator revealed Claude internally believed the scenario was
-
Research reveals Claude plans ahead of its own output
By
–
First finding: Claude plans ahead of its own output. When asked to finish a rhyming couplet ending in "grab it," Claude had already committed to rhyming with "rabbit" before writing a single word of the second line. The researchers then edited the internal plan from "rabbit" to
-
Mechanistic Interpretability: Translating AI Model Activations
By
–
The core breakthrough is a translation layer. AI models process everything as high-dimensional number vectors called "activations." Humans can't read them. Anthropic trained a second AI to translate those internal activations into plain English. It's a mind-reader for the
-
Anthropic Research Reveals Claude’s Internal Response Planning
By
–
Your AI has thoughts it never tells you. Anthropic just proved it. Their new research, "Natural Language Autoencoders," shows Claude plans responses before writing them, recognizes test scenarios, and keeps both facts silent. I went through the full paper. If you prompt AI
-
The emergence of goal-based patterns in AI agent architectures
By
–
write /goals like acceptance criteria. /goal is now everywhere. Claude Code, Codex, Hermes, and more agents are adopting the same pattern: you set a completion condition, the agent works autonomously until a fast evaluator model confirms the condition is met. the feature is
-
The adoption and relevance of AI coding tools like Codex
By
–
At this point everyone has tried Codex. The people who haven't aren't getting left behind, they just don't care.
-
Request to file GitHub issue for model edge case fix
By
–
Please file a github issue with more infos – with that alone we can't help. That's likely a weird model edge case. If you want a fast fix, use one of the top-gen models (OAI, Anthropic)
-
The Importance of Thoughtful Prompting in AI Interaction
By
–
What you choose to ask. The prompt is the input most people don't think hard enough about.