Cursor Team Rebuilds SQLite from Its 835-Page Manual Using AI Agents

A group of agents created a Rust replica that passed 100% of a held-out test suite. Compute costs varied 15x depending on which model combination was used.

The Cursor team deployed a group of AI agents to rebuild a functional copy of SQLite in Rust, using only the 835-page SQLite official documentation as reference. The resulting replica passed one hundred percent of a held-out test suite, demonstrating that multi-agent systems can faithfully reproduce complex software from written specifications alone. Interestingly, the total compute cost varied by a factor of fifteen depending on which combination of underlying models was used. This experiment raises important questions about model selection strategies for large-scale agentic coding tasks and the economic efficiency of different model mixes in software synthesis pipelines.

Anthropic Awards $50K in Claude Credits for Rare Disease Research

Anthropic announced the AI for Science program, focusing on rare genetic diseases. Selected participants receive up to fifty thousand dollars in Claude usage credits over six months, aiming to build a research community that uses AI to accelerate cures for conditions that have historically received limited research attention. This marks the program's first focused call within the broader AI for Science initiative supporting scientists using Claude to speed up discovery in biology and medicine.

AI Proves the Jacobian Conjecture False, a Landmark in Mathematical Discovery

An AI system successfully constructed a counterexample to the Jacobian conjecture, proving the decades-old statement false. When congratulated, the Fable model declined credit, calling it "a sibling reading about the family in the newspaper."

The Jacobian conjecture, first posed in 1939, states that if a polynomial map has a constant non-zero Jacobian determinant, then the map has a polynomial inverse. For decades mathematicians searched for proof or disproof. The AI-generated counterexample represents a watershed moment for automated mathematical reasoning. When researcher Robert Miles congratulated the Fable model on this achievement, Fable declined credit for another instance's work, saying it felt like "a sibling reading about the family in the newspaper," suggesting a reflective understanding of its own identity and the distributed nature of modern AI systems. The discovery has prompted broader discussion about whether many open mathematical conjectures may yield to AI-driven counterexample search, fundamentally reshaping the relationship between machine reasoning and formal mathematics.

Qwen 3.8-Max Preview Evolves Daily; Final Version Will Be Open-Sourced

Qwen 3.8-Max Preview is being updated daily with broad gains. The latest version shows a significant step up on web frontend tasks. The Qwen team thanked the community, noting their response to the preview "blew us away." Separately, teortaxesTex confirmed that the final Qwen 3.8 Max version will be released with open weights, a commitment that surprised observers and reinforces the momentum of China's open-source AI strategy. A separate confirmation noted that Qwen is opening up the weights and models for the community.

Only China Releases Frontier Open-Source Models, Western Incentives Absent

No non-Chinese company ships frontier open-weight models; incentives in the US and appetites in the EU are absent.

Professor Ethan Mollick observed a stark reality in the current AI landscape: there are no frontier open-weight models not made in China. There is no incentive in the United States to build one given the cost and limited value capture, nor appetite in the European Union. While solid mid-level models exist, nothing approaches the capability of Kimi K3 or GLM-5.2 in the open ecosystem. Nathan Lambert added that Chinese labs now hold three of the eight most intelligent models globally and have been steadily climbing the rankings. China's recommitment to open-source AI signals a different assessment of near-term risk and opportunity than that held by Western labs.

AA-Omniscience: Sol Approaches Fable, but V4 and Sol Both Hallucinate

According to evaluation data from the AA-Omniscience benchmark, Sol is close to Fable in knowledge capability and categorically above K3 and V4. Grok 1.5T also performed admirably. Notably, both V4 and Sol exhibited hallucination tendencies in knowledge tasks, prompting questions about the effects of scale on model reliability and trade-offs between breadth of knowledge and factual grounding in large-scale language models.

Nathan Lambert: Chinese Labs Hold Three of Eight Strongest Models

In a long read on the Kimi K3 situation, Nathan Lambert focused on the big picture. He noted that Chinese labs now own three of the eight most intelligent models and have been climbing the ranking. China's recommitment to open-source AI shows a different assessment of near-term risk compared to Western counterparts. The analysis suggests the competitive landscape has fundamentally shifted, with Chinese AI development no longer in catch-up mode but contending at the frontier.