A new AI engineering philosophy flips how every codebase gets built. Most codebases get harder to work with every year. Each new feature is a negotiation with the old ones. Teams end up fighting their system instead of building on it. Every just open-sourced a plugin that
@alphasignalai
-
Gate is the game: Audit trail for AI Act compliance, sign-off, changelog, proposal, validation
By
–
yeah, the gate is the whole game. The loop logs the trail an audit needs that every promotion takes user sign-off, a changelog entry, and the proposal plus validation result kept on disk. That's the what, why, and who-approved an EU AI Act review would start from.
-
AI Self-Improvement: New Advances and Outlooks
By
–
Sources:
Anthropic, When AI builds itself:
https://anthropic.com/institute/recursive-self-improvement
… SkillOpt (arXiv): https://arxiv.org/abs/2605.23904 SkillSmith (arXiv): https://arxiv.org/abs/2606.01314 MOSS (arXiv): https://arxiv.org/abs/2605.22794 Co-Scientist, Google DeepMind: https://deepmind.google/blog/co-scientist-a-multi-agent-ai-partner-to-accelerate-research/
… Gary Marcus, No need to panic -
Headroom filters and compresses tool outputs for LLMs
By
–
77,000 tokens consumed. 800 were useful.
— AlphaSignal AI (@AlphaSignalAI) 5 juin 2026
That's not a model problem. It's a filtering problem.
Headroom is an open-source repo sitting between your agent and the LLM.
It compresses tool outputs before they arrive. Original data is cached, retrievable in 1ms if needed.
Real… https://t.co/T8jkIrgtFn pic.twitter.com/lHob65JhaD77,000 tokens consumed. 800 were useful. That's not a model problem. It's a filtering problem. Headroom is an open-source repo sitting between your agent and the LLM. It compresses tool outputs before they arrive. Original data is cached, retrievable in 1ms if needed. Real
-
Stop writing skill files for agents manually; use CLAUDE.md and AGENTS.md
By
–
You can now stop writing skill files for your agent by hand.
— AlphaSignal AI (@AlphaSignalAI) 4 juin 2026
Coding agents forget what you taught them last week.
Every session burns tokens re-explaining your procedures.
CLAUDE.md helps one agent, AGENTS.md helps another.
You maintain both yourself, learning on behalf… pic.twitter.com/eLitYPMaT3You can now stop writing skill files for your agent by hand. Coding agents forget what you taught them last week. Every session burns tokens re-explaining your procedures. CLAUDE.md helps one agent, AGENTS.md helps another. You maintain both yourself, learning on behalf
-
Optimize the machine that produces model improvements, not just the next release
By
–
Most teams optimize for the next model release. Microsoft's MAI-Thinking-1 report argues for something different: optimize the machine that produces model improvements. The most valuable asset isn't a benchmark score. It's a training system that keeps getting better.
-
New paper: Image generators cheat with near-duplicate outputs
By
–
Image generators look impressive but quietly cheat. They score well on FID by producing near-duplicate outputs. The catch is mode collapse hiding behind a clean number. Coverage of the real data distribution suffers in silence. A new paper introduces the Recursive Token
-
@alphasignalai — 2026-06-03
By
–
YES, the CVE run makes that concrete, 100% accuracy at 85.1% fewer tokens, while the other systems stayed under 25%. Only word to push back on is "unprecedented" though, CodeAct was doing code-as-actions back at ICML 2024.
-
20 of 25 Top AI Researchers Say AI Will Soon Build AI
By
–
20 of 25 top AI researchers say AI will soon build AI. For decades, humans built every AI system from scratch. That assumption is quietly breaking down inside frontier labs. A new paper interviewed 25 top researchers from Stanford, OpenAI, Google DeepMind, and Anthropic.
-
Perplexity’s New Agent Architecture: Python, Sandbox, and Efficiency Gains
By
–
Perplexity stopped treating search as one API call. Its agents now write Python that fans out queries, filters results, and joins evidence in a sandbox. On a 200+ CVE task: 100% accuracy, 85.1% fewer tokens. The SDK is private. The pattern isn't, so lets use it with Hermes?