AlphaSignal

The most read technical newsletter in AI. Stay on top of the latest research, repos, and models with 5 min daily summaries. Join 200,000+ AI developers.

Over 12,000 subscribers

By subscribing, you agree Substack's Terms of Use, and acknowledge its Information Collection Notice and Privacy Policy.

Claude Fable 5 didn’t lose. It got switched off.

Why a closed frontier API is now a single-vendor risk, and how open weights (GLM-5.2) and orchestration (Sakana Fugu) hedge it.

20 hrs ago•AlphaSignal AI

LatestTopDiscussions

HarnessX Turns the Agent Harness Into a Typed Object You Can Compose, Evolve, and Train

+14.5 points average across five benchmarks, up to +44.0, 14 of 15 configurations improved, model untouched

Jun 19•AlphaSignal AI

How to Let a Fixed Model Rewrite Its Own Harness, And How to Run the Loop Yourself

A propose-and-regression-gate loop lifted held-out Terminal-Bench pass rates up to 21.4 points across three models, with nothing trained.

Jun 17•AlphaSignal AI

10 Layers of Self-Improving Harness Stack

28 research papers say developers should build around the model (Harness)

Jun 16•AlphaSignal AI

Claude Fable 5 Prompt Leak Is a User Manual for Long-Running Agents

The leaked prompt points to files, tools, artifacts, checks, and work packets instead of longer one-shot prompts

Jun 12•AlphaSignal AI

How to Actually Prompt Claude Fable 5

Anthropic’s official guide swaps instruction-stacking for effort settings, verifier subagents, and boundary blocks

Jun 11•AlphaSignal AI

Claude Fable 5's Hardest Test Was Knowing the Metric Was Wrong

Three ML demos, one Opus 4.8 comparison, and the cost of trusting the wrong score

Jun 10•AlphaSignal AI

Most Developers Do Not Need Agent Loops Yet

The patterns were documented in 2024. Here’s who it pays off for, and the four conditions that decide.

Jun 8•AlphaSignal AI