Machine Learning Read an LLM's Mind: Probe Hidden Thoughts Like J-Space
Train a linear probe on hidden activations and steer output, the method behind Anthropic's J-Lens.
How-to content for builders, indie hackers, and AI engineers. Less theory, more shipped code.
Machine Learning Train a linear probe on hidden activations and steer output, the method behind Anthropic's J-Lens.
Tutorials Use Anthropic's compact-2026-01-12 beta so long agentic loops survive past the 200K context window.
Tutorials Load SKILL.md files into a Claude Agent SDK app with setting_sources and the skills option.
Tutorials Meituan's 1.6T open MoE topped OpenRouter as 'Owl Alpha.' Call it in Python with the OpenAI SDK.
Tutorials Give Claude agents a token countdown so long agentic loops finish gracefully instead of burning your bill.
Tutorials Build cheap, cited research agents on Claude Sonnet 5 using the web search tool and dynamic filtering.
Tutorials Stand up an autonomous agent with sessions, streaming events, and webhooks - no agent loop required.
Tutorials Master Sonnet 5's on-by-default thinking and the effort knob to cut cost and latency.
Tutorials Hands-on Python guide to Sonnet 5's adaptive thinking, effort levels, and the 30% tokenizer trap.
Tutorials Embed Copilot's agent runtime and route work to scoped sub-agents, with runnable Python.
Tutorials A runnable Python governor that caps LLM spend per user and auto-downgrades models.
Tutorials Let a cheap executor model consult a stronger advisor mid-task in one Messages API call.