Start / AI Fire Daily / %f0%9f%8e%99%ef%b8%8f ep 19 ai that can think in steps inside anthropics new tool

🎙️ EP 19: AI That Can Think in Steps — Inside Anthropic’s New Tool

13 min • 30 maj 2025

Anthropic just gave us something wild — a tool that lets you see inside an AI’s brain. You can actually trace how a model makes decisions, step by step. It’s called circuit tracing. This might be the beginning of editable reasoning in LLMs.

We’ll talk about:

Anthropic’s new circuit tracing tool and how it works
Why it matters for AI safety and transparency
DeepSeek’s quiet new model that just beat Claude 3.7 in coding
Google’s AI confusion — still doesn’t know what year it is
AI browser from Opera, Odyssey’s interactive video demos, and Grammarly’s $1B raise
Plus: NASA’s GAIA AI model that can predict hurricanes using 25 years of satellite data

Keywords:

Anthropic, circuit tracing, attribution graphs, DeepSeek R1-0528, Claude 3.7, Google AI fail, Gemini, GAIA AI, AI interpretability, AI reasoning, foundation models, AI transparency, interactive video AI, Grammarly funding, AI browser, OpenAI vs creators, AI Napster moment

Links: