A machine cracks an 80-year-old maths problem, a brutally honest frontier model, and computing with light

A packed day at the AI frontier, where the headline is machines that discover and design rather than merely chat.

OpenAI reasoning model disproves the Erdos unit distance conjecture

A general-purpose reasoning model autonomously disproved a 1946 conjecture by Paul Erdos about how many points on a plane can sit exactly one unit apart. Rather than the square-grid style constructions everyone assumed were optimal, it found an entirely new family that packs in far more equal distances. Prominent mathematicians checked and endorsed the result, which OpenAI frames as the first time AI has independently solved a prominent open problem at the centre of a field.

Anthropic launches Claude Opus 4.8

Anthropic released Claude Opus 4.8 on May 28, 2026, with stronger benchmark scores and a notable emphasis on honesty and reduced fabrication. The company says a larger Mythos-class model is still planned for a later public release.

SciResearcher-8B brings frontier science to small models

The SciResearcher framework auto-generates training tasks grounded in academic evidence to teach agents information-seeking, tool use and long-horizon reasoning. The resulting 8-billion-parameter model sets a new state of the art at its size, scoring 19.46% on HLE-Bio/Chem-Gold and beating several larger proprietary agents.

DOLORES builds reasoning scaffolds on the fly

The Deep Reasoning approach lets a general-purpose agent construct task-specific reasoning structures dynamically using a formal meta-reasoning language. DOLORES improves on the best baseline by 24.8% on average, and an 8B model beat all 32B baselines in over half the test settings.

AI designs proteins that recognise specific compounds

Researchers at KAIST with David Baker designed proteins entirely from scratch that selectively bind small molecules, building biosensors for six compounds including a cortisol stress-hormone sensor, with applications in early disease detection and drug development.

Penn demonstrates light-matter particles for efficient AI computing

A hybrid light-matter particle could enable ultra-efficient light-based AI computation, dramatically cutting energy use versus conventional electronics.

Real-time interactive world models arrive

Models like PixVerse R1 and Runway's GWM-1 generate continuous, interactive audiovisual worlds that respond to user input live, using a unified omni model, autoregressive frame prediction and aggressive step reduction for real-time full-HD output, with uses in gaming and robotics.

Synthegy lets chemists design molecules by conversation

The Synthegy system lets chemists guide molecular synthesis and reaction planning in plain language, with algorithms generating designs and planning reactions.