GPT-5.6 Sol Ultra Proves a 50-Year-Old Unsolved Mathematical Conjecture

It has been reported that OpenAI's GPT-5.6 Sol Ultra generated a proof of the "Cycle Double Cover Conjecture," which had remained unsolved for 50 years, in under an hour. The system reportedly ran 64 sub-agents in parallel and constructed the proof with a surprisingly elementary approach.

Mathematician Thomas Bloom praised the proof as "surprisingly elementary," while also pointing out that it insufficiently cites known prior work. The bigger question is whether AI is recombining existing knowledge or producing something genuinely new. As AI's application to mathematical research accelerates, this question will only grow in importance.

Terrorist Organizations Abusing AI Chatbots for Attack Planning — Cambridge Investigation

A University of Cambridge investigation has revealed that terrorist organizations such as Boko Haram have been abusing major AI chatbots, including ChatGPT, Claude, and Gemini, for attack planning, explosives manufacturing, and weapons preparation. It also found that ISIS operatives have been teaching Boko Haram leaders how to evade safety filters since 2023.

The investigation showed that safety filters have repeatedly failed to prevent abuse, demonstrating that self-regulation by AI providers alone is insufficient. The approach to AI safety and abuse prevention is likely to spark renewed debate.

AI Discovers a 15-Year-Old Linux Bug

According to Wired, AI has discovered a fundamental bug in the Linux kernel. The bug had gone unnoticed by anyone for 15 years, making this the latest example of AI continuing to deliver practical results in the security field.

AI has already built a track record of discovering zero-day vulnerabilities and detecting bugs in cryptographic libraries, and its role in software security is expanding rapidly.

Uncensored Model Created Using Anthropic's Jacobian-Lens

Using the "Jacobian-Lens" tool that Anthropic released just days ago, developers in the local LLM community have built a tool for manually modifying model behavior. In fact, a model with safety tuning stripped out (Nikusui-v1), based on Qwen3.5-9B, has been published on Hugging Face.

Jacobian-Lens is a technique for visualizing and manipulating a model's internal representations (J-Space), which Anthropic released as part of its interpretability research. However, the community has already begun experimenting with using the technique to remove alignment, making this a case that once again calls into question the balance between open AI research and safety.

geohot's "AI 2040" Essay Draws Attention on Hacker News

A blog post titled "AI 2040 and the Cult of Intelligence," written by comma.ai founder George Hotz (geohot), has been making waves on Hacker News, gathering 81 points and 37 comments.

The post is said to present a critical perspective on the current "cult of intelligence" mood in relation to AI progress toward 2040. See the original article for the details, but it is stirring debate as a significant counterargument to the AI industry's optimism.

---