Newsletter

Science & Math

Europe's top-scoring AI model is a compressed GLM-5.2

The Arena

Europe's top-scoring AI model is a compressed GLM-5.2

Spain's Multiverse Computing has spent two weeks selling Quasar as proof that Europe can field a frontier model of its own. The company's own documentation says the foundation underneath it was trained in China. Multiverse Computing released Quasar 1.1 438B on September 15 — a rebuild of the model it launched on September 2 as the highest-scoring European entry on the Artificial Analysis Intelligence Index — and its own write-up states that Quasar is built from GLM-5.2, the open-weights model…

A 9B longevity model tops 18 frontier LLMs on aging data

The Frontier

A 9B longevity model tops 18 frontier LLMs on aging data

Aging biology generates more measurement data than any one lab can read, and until this week there was no shared way to check whether an AI can actually interpret it. Insilico Medicine's answer is an open benchmark — and the results are unflattering to the biggest models in the world. Insilico Medicine published LongevityBench, an open suite of 17 tasks spanning clinical records, genetics, epigenetics, transcriptomics and proteomics, and its own compact models — built on Liquid AI's LFM2 and…

An ex-Princeton postdoc raised $7M for a scientific foundation model

The Frontier

An ex-Princeton postdoc raised $7M for a scientific foundation model

Two gaps got money and code this week: science still has no compiler, and the next generation of AI chips still has no common language. Keli Xinxu, an AI-for-science startup founded by an assistant professor at Hong Kong University of Science and Technology, has raised nearly 50 million yuan — about $7 million — in a first round led by Inno Angel Fund, with Yijing Capital, Xiaomiao Langcheng and Linge Ventures joining and Lighthouse Capital as strategic incubator and sole financial adviser.

Bend 2 wants AI agents to prove their code, not promise it

The Stack

Bend 2 wants AI agents to prove their code, not promise it

Today's most consequential AI release isn't a model. It's a programming language built for a world where nobody reads the code anymore. Victor Taelin released Bend 2 on Thursday, and its central claim is narrow enough to test: an AI agent cannot merge a mistake the compiler can disprove. The mechanism is a laws file — the project's own framing is that it is an agent instruction file with a proof checker standing behind it.

von der Leyen calls AI 'the second tipping point of our times'

The Guardrails

von der Leyen calls AI 'the second tipping point of our times'

Europe's State of the Union named AI and climate as the two forces that will decide the continent's next decade — and Brussels offered the frontier labs a seat at the table on how fast to move. Ursula von der Leyen called AI "the second tipping point of our times," alongside climate change, and said its risks must be "addressed." In her State of the Union address in Strasbourg on Wednesday, the European Commission president put the two side by side as the perils that will define Europe's next…

Two high schoolers used AI to close a Fields Medalist's open problem

The Frontier

Two high schoolers used AI to close a Fields Medalist's open problem

A math preprint that should not have landed this week, a leak about who actually reads your ChatGPT conversations, and OpenAI admitting its chip design ran on models nobody outside can use. Two Oak Park High School students, with a UCLA postdoc as mentor, have posted a 75-page arXiv paper extending the theory of Lorentzian polynomials — the combinatorics framework Fields Medalist June Huh helped build — to arbitrary degree, settling a bounded-ratio question the quadratic case left open.

More molecules, same failures: the crack in AI drug design

The Arena

More molecules, same failures: the crack in AI drug design

China's AI-for-drug-discovery crowd spent Sunday at the Pujiang Innovation Forum arguing about the thing benchmarks don't measure — what happens after the model proposes. And the country's biggest banks are quietly building a lending product around the same question: is AI usage real enough to lend against? "The more you generate, the more you fail." That was Ding Sheng, founding dean of Tsinghua's school of pharmaceutical sciences, at the forum's Future Medicine track on September 13, and the…

Claude Fable cracked a 1653 cipher. The math says it didn't

The Guardrails

Claude Fable cracked a 1653 cipher. The math says it didn't

A frontier model was handed a 370-year-old unsolved cipher and came back with a plaintext. The internet checked the work, and it doesn't hold — which makes the retraction more interesting than the claim. Vals AI announced that Claude Fable 5.1 solved the Cyphral Distich, a cryptogram of two lines of 32 numbers printed at the end of Thomas Urquhart's 1653 treatise Logopandecteision, and ranked 28th on historian Klaus Schmeh's list of the fifty most famous unsolved messages.

AI 101 — What is an NPU (neural processing unit)?

The Stack

AI 101 — What is an NPU (neural processing unit)?

An NPU is a chip built for one job: running the math that neural networks do, at a fraction of the power a general-purpose CPU or a graphics GPU would burn. The name covers anything from the block inside your phone's processor to a USB stick in a robot to a rack card in a datacenter. Vendors also call it an AI accelerator or a neural engine, and it is the hardware reason "on-device AI" stopped being a marketing phrase. Why this matters right now The NPU is the quiet spec war of 2026.

A small model that knows when to hand off cuts inference cost 96%

The Frontier

A small model that knows when to hand off cuts inference cost 96%

The cheapest way to get frontier-level math reasoning just got a working open-source demo: teach a 4B model to admit when it's stuck. PyroMind Dynamics says its PyroDash method lets a Qwen3.5-4B model solve hard reasoning benchmarks at near-frontier accuracy while paying roughly 4% of what a big model alone would cost. The trick is a control token. The small model streams an answer normally, and the moment it decides it can't finish, it emits <|llm_offload|>.

See all