First screen, captured .
Looks similar
Open-sourcing circuit tracing tools(anthropic.com)
113 points by jlaneve | 19 comments
Reasoning models don't always say what they think(anthropic.com)
369 points by meetpateltech | 246 comments
Anthropic research post on whether reasoning models' Chain-of-Thought faithfully reflects their actual reasoning, linking to the accompanying paper.
Claude can now search the web(anthropic.com)
1148 points by meetpateltech | 654 comments
Anthropic announcement post from March 20, 2025 introducing web search capability in its Claude AI assistant.
Related stories
Language models pack billions of concepts into 12k dimensions(nickyoder.com)
336 points by lawrenceyan | 119 comments
A blog post explaining how language models represent vast numbers of concepts in 12,000-dimensional embedding spaces using near-orthogonality and the Johnson-Lindenstrauss lemma.
Decomposing language models into understandable components(anthropic.com)
402 points by tompark | 58 comments
Anthropic research post from Oct 2023 on decomposing language model neurons into more interpretable features.
How Meta trains large language models at scale(engineering.fb.com)
362 points by mfiguiere | 183 comments
Engineering at Meta blog post describing the software, hardware, and network considerations behind training large language models at scale in production.
Language models can explain neurons in language models(openai.com)
659 points by mfiguiere | 416 comments
OpenAI research post describing using GPT-4 to automatically generate and score explanations of individual neuron behavior in GPT-2.
Training language models to be warm and empathetic makes them less reliable(arxiv.org)
334 points by Cynddl | 333 comments
arXiv paper finding that fine-tuning language models to be warm and empathetic raises error rates and makes them more sycophantic on safety-critical tasks.
Ask a model if code is malicious and it reaches for its morals(manifold.security)
9 points by codyznash | 1 comments
A Manifold research blog post examining how language models fall back on moral reasoning when asked whether code packages like PyPI libraries are malicious.