The brief · 18 May 2026

The AI brief, 2026-05-18

A lab teaches an LLM to see in 3D, Cursor beefs up its own coding model, and Brussels starts putting its big AI rules rollback in writing.

1 min read J / K to navigate

Berkeley Lab teaches an LLM to read 3D material structures

MatterChat pairs a standard LLM with an interatomic potential model to predict properties like bandgap, useful for designing microelectronics or batteries. GPT-4 alone fails at this kind of calculation because it's never seen a molecule's 3D geometry, only text. This has nothing to do with your daily ChatGPT use, but it points to where the real breakthrough is headed: pairing a general-purpose LLM with a specialized scientific model instead of expecting one model to do everything.

practitioners › Published in Nature Machine Intelligence, trained on Materials Project data, announced May 18, 2026.

Berkeley Lab →

Cursor ships Composer 2.5, its in-house coding model

Cursor is weaning itself off full dependence on Claude and GPT, fine-tuning its own model post-trained on Moonshot's Kimi K2.5. If you code with Cursor, expect more reliable long-running tasks and instruction-following, with a less verbose communication style.

practitioners › Composer 2.5 is live in Cursor IDE now, with usage-based pricing tiers.

Cursor →

The EU's Digital Omnibus Is Finally Taking Concrete Shape

The provisional May 7 agreement is coming into focus: some high-risk obligations get pushed back to December 2027 or August 2028, but new bans on non-consensual intimate deepfakes and CSAM kick in as early as December 2026. If you're deploying AI in Europe, the broader timeline keeps shifting, but the tolerance window on generated intimate content is closing fast.

practitioners › Reference text: eur-lex CONSIL:ST_9247_2026_INIT, bans take effect December 2026.