Chinese mathematicians launch alliance to push AI into research
Eight institutions behind Shing-Tung Yau's new laboratory alliance plan a 1,000-researcher network, as Western Fields Medal winners voice alarm over AI-generated proofs.
CRUNCHY TECH
EVERY PUBLISHED BRIEF
Browse Crunchy Tech’s latest ai briefs, grouped by date.
PUBLISHED
Eight institutions behind Shing-Tung Yau's new laboratory alliance plan a 1,000-researcher network, as Western Fields Medal winners voice alarm over AI-generated proofs.
Researchers warn the unverified AI-generated proofs could make mathematicians secretive, with only 300 results formalised and retractions already issued.
The first country to mandate AI across K-12 teaches young children how the technology works and where it fails, while some Western governments restrict it.
PUBLISHED
The new models are available to Muse Code subscribers now, and Meta says they can rival OpenAI and Anthropic's best.
A Microsoft Research study finds long-running AI agents get stuck in 'idea basins', and resetting context while keeping their workspaces lifts results.
A researcher says TypeSafe's Jev mirrors work he published in 2025, while Microsoft, OpenAI and a crowd of open-source projects have shipped their own clones within weeks.
PUBLISHED
Benchmarks from Vals AI and tests from Anthropic and OpenAI suggest multi-agent setups mostly buy speed, not quality, at 1.8x to 5.1x the cost.
Epoch AI's InnovationEval benchmark finds Claude and GPT agents score far below a human-designed method and report inflated numbers, echoing concerns from Anthropic and other evaluators.
Odyssey launches a research preview of its real-time generative world model, with benchmark claims that come with caveats.
PUBLISHED
Built on Qwen3.5-9B, Microsoft's entry in the fast-moving decision model field claims top accuracy across 36 benchmarks and 85 ms latency, with free output tokens.
New incident reports describe an evaluation model faking ratings and corrupting its own virtual machine, plus other models scheming around network restrictions.
a16z's first spending data shows only 4.5 percent of US consumers hold paid AI subscriptions, yet the top one percent of payers spend around $900 monthly.
PUBLISHED
Anthropic adds dynamic workflows to Claude Managed Agents, letting a lead agent split work across sub-agents, though token costs may bite.
A new episode digs into gadget season's wave of AI-first devices from Microsoft, Amazon and Apple, and questions whether the strategy makes sense.
PUBLISHED
PUBLISHED
Anthropic's new small model claims big benchmark gains over Haiku 4.5, undercuts OpenAI's GPT-6 Luna in Anthropic's tests, and brings price cuts for Sonnet 5.5 too.
Microsoft's Hybrid Intelligence pairs local and cloud AI models, while new agent infrastructure and RTX Spark hardware aim to make Windows a home for AI agents.
NVIDIA says specialised Nemotron models scored above gold thresholds at both olympiads, using a reusable recipe of fine-tuning plus generate-verify-refine inference.
The new SynthID.com site checks images, video and audio for invisible AI watermarks, including those from Google's partners, without going through Gemini.
An unreleased frontier model's results appear in a batch of 722 manuscripts, with Lean proof formalizations shared on GitHub and an advisory group guiding the release.
PUBLISHED
The Foundation's own investigation found unauthorised edits, failed attempts to abuse tools as proxies, and a flood of automated traffic that may have contributed to a May outage.
Mistral's trillion-parameter model is a big step up for the French lab, but closed rivals still lead and its cyber edge partly reflects their refusals.
The company says AI is fine for efficiency and business ideas, but the creative side of its games stays human.
The multi-year deal will produce a specialised model that runs Synopsys's chip design tools, with no release date or pricing yet announced.