Sakana AI's ICML double: multi-agent coordination without message-passing, and an optimizer merger nobody saw coming

Good morning 👋 Sakana AI landed two papers at ICML 2026 this week, and the multi-agent one solves a coordination problem every agent-swarm builder keeps hitting.

In today's issue:

• 🔭 Sakana AI's ICML double: multi-agent coordination without message-passing, and an optimizer merger nobody saw coming
• 🧠 DeepMind buys a seat at A24
• 🔬 Two papers: optimizer swaps for atoms, patching CLIP's blind spot
• 🔗 Verses AI kills its R&D program, Clarivate trims $600M, Csquare eyes Nasdaq

🔭 THE ONE THING

🧩 Sakana AI's ICML double: multi-agent coordination without message-passing, and an optimizer merger nobody saw coming

Sakana AI got two papers into ICML 2026, and the more useful one solves a problem every agent-swarm builder has hit: agents solve their own slice of a task, then negotiate only where the pieces meet, instead of passing messages back and forth forever. On multi-agent Sudoku that's a 93% solve rate against 11% for the message-passing baseline. It also holds 86% accuracy under domain shift where a standard CNN drops to 11%, which is the number that actually matters if you're deploying anything outside a lab. The sleeper paper, same conference, same team, found that evolution strategies and consensus-based optimization are the same update rule wearing different clothes, and used that to build hybrid optimizers aimed at LLM merging. Builder read: if your agent swarms fall apart under distribution shift, read the Sheaf-ADMM paper before you reach for more message-passing bandwidth.

🧠 MODELS & RELEASES

• 🎬 Google DeepMind took an equity stake in A24 and called it a "first-of-its-kind" research partnership: researchers and filmmakers building workflows together. No films, no tools, no timeline yet. Read it as DeepMind buying a seat at the table before anyone else writes the norms for AI in film. Announcement
• 🎮 NVIDIA's "model" news this week is a GeForce NOW games list, twelve titles landing in July with a DLSS/Reflex mention for Ultimate subscribers. It got tagged as AI news by the feed. It isn't one. GeForce NOW update

🔬 RESEARCH HIGHLIGHTS

• Ditch Adam for interatomic potentials. A new study training NequIP and Allegro models found that matrix-structured optimizers SOAP and SOAP-Muon beat plain Adam on both convergence speed and final accuracy, with the gap widening when force labels are sparse. Optimizer choice has been an afterthought in molecular-simulation AI. This says it shouldn't be. Paper

• CLIP's typography blind spot, patched with no retraining. Write "banana" on a photo of an apple and a CLIP-based model will often agree with the text over its own eyes. Researchers traced the failure to specific attention heads that favor lexical signal over visual signal, then fixed it by intervening on just those heads. Cheap insurance if you're shipping a CLIP classifier anyone could scribble on. Paper

🔗 QUICK LINKS

• Verses AI confirmed it's shutting down its AI R&D program entirely, per an 8-K filed this week. Filing
• Clarivate is selling its Life Sciences & Healthcare segment to Altaris for $600 million, a trim aimed at freeing up the rest of the company to chase AI-driven research intelligence. Filing
• Csquare, the Brookfield-backed data center operator built for GPU workloads and edge inference, amended its S-1 again ahead of a NYSE listing under CSQR. Filing
• Unum Group filed a fresh investor-presentation exhibit with the SEC. No headline attached, just the deck. Filing

Go read the Sudoku numbers. A 93 vs. 11 split is the kind of gap hype normally can't produce.

Pradeep Perugu