Noam Shazeer leaves Google for OpenAI

Good morning 👋 The co-author of the transformer paper just announced he's joining OpenAI. Everything else you build is downstream of that paper.

In today's issue:

• 🔭 Noam Shazeer leaves Google for OpenAI
• 🔬 Research: your AGENTS.md is probably wrong, and LLM personality scores aren't real
• 🛠️ Tools & how-to: write an AGENTS.md your coding agent will actually use
• 🔗 SpaceX buys Cursor for $60B, emergent misalignment has a surgical fix

🔭 THE ONE THING

🧠 Noam Shazeer leaves Google for OpenAI

Shazeer announced Tuesday on X that he's joining OpenAI as Lead for AI Architecture Research. He's one of the eight co-authors of "Attention Is All You Need," the 2017 paper that gave every current LLM its spine. He'd co-founded Character.AI after leaving Google the first time; Google paid approximately $2.7 billion to bring him back in 2024 as VP of Engineering and co-lead of Gemini. He lasted under two years. Sam Altman's reply: "noam is one of the people I have most wanted to work with since the very beginning of openai. only took 10 years. i think it will be worth the wait!"

What his day-to-day looks like isn't confirmed beyond the title. What is clear: Google's Gemini just lost a co-lead in the run-up to OpenAI's reported September IPO, and the person now overseeing OpenAI's model architecture is the one who helped write the blueprint everything is built from. The $2.7B retention package bought under two years.

🔬 RESEARCH HIGHLIGHTS

• Probe-and-Refine Tuning for Coding Agents: an automated method for improving AGENTS.md files from the bottom up. The agent probes the codebase, identifies where its own understanding breaks down, then rewrites the guidance doc. Tested on SWE-bench-style evals; improved files produced measurably better task success. If your coding agent keeps getting the same subsystem wrong, the file is the problem. arxiv.org/abs/2606.20512

• LLM Psychological Profiles Are Largely Measurement Artifacts: formal psychometric analysis applied to the tools used to assign LLMs stable "personality" scores found the profiles weren't consistent in ways that would make them meaningful. Product teams using personality benchmarks to evaluate or fine-tune model character should treat those numbers skeptically until somebody builds instruments designed for LLMs rather than for humans. arxiv.org/abs/2606.20205

🛠️ TOOLS & HOW-TO

Write an AGENTS.md your coding agent will actually use

The Probe-and-Refine paper's core finding: most AGENTS.md files are written top-down from how the author thinks about the codebase, not bottom-up from where agents actually get confused. Flip the process:

1. Run your agent on 5 representative tasks. Log every wrong file guess or subsystem confusion.

2. Add a one-line directory index for each area it got wrong: src/auth/ = session state only; src/users/ = profile data.

3. Document the test command and specify which failure types are expected noise vs. a genuine regression.

4. Add a "traps" section: "Fixing X in module A broke B three times. Always check both."

5. After every future failed run, ask what was missing from the file, not what was wrong with the agent.

Prompt: Review this AGENTS.md and identify: (1) subsystem boundaries an agent would likely confuse, (2) missing test guidance, (3) historical pitfalls not yet captured. Suggest 3 specific additions with exact wording.

5 trending right now

• 🤖 OpenMythos: first-principles reconstruction of the Claude Mythos architecture from published research. github.com/kyegomez/OpenMythos
• 📄 OfficeCLI: read, edit, automate Word/Excel/PowerPoint for AI agents; no Office install required; single open-source binary. github.com/iOfficeAI/OfficeCLI
• 🌐 ai-website-cloner-template: clone any website with one command using AI coding agents; 17K stars. github.com/JCodesMore/ai-website-cloner-template
• 🎬 OpenMontage: open-source agentic video production system with 12 pipelines, 52 tools, and 500+ agent skills. github.com/calesthio/OpenMontage
• 📚 learn-harness-engineering: zero-to-shipped harness engineering tutorial. github.com/walkinglabs/learn-harness-engineering

🔗 QUICK LINKS

• SpaceX acquires Cursor — in case you missed it last week: $60B all-stock deal, largest VC-backed startup acquisition on record. Anysphere becomes a wholly owned SpaceX subsidiary; deal closes Q3. Disclosed in a SpaceX 8-K filed June 16, four days after SpaceX's Nasdaq IPO.

• Emergent misalignment has a surgical fix. New paper finds a shared activation-space direction across four model families (Qwen, Llama, Mistral, Falcon) induced by fine-tuning on insecure code. Steering against that direction reduces harmful outputs without full retraining. arxiv.org/abs/2606.20225

See you tomorrow.

Pradeep Perugu