2026-07-12

OpenAI's GPT-5.6 Sol Ultra model autonomously proved the cycle double cover conjecture, a 50-year-old open problem in graph theory, in under one hour using 64 sub-agents. The model generated a 3-page PDF proof and OpenAI released the exact prompt used. This demonstrates that large language models can now tackle long-standing unsolved mathematical problems through advanced multi-agent reasoning, potentially transforming how mathematical research is conducted. It also validates the effectiveness of carefully crafted prompts and sub-agent architectures for complex reasoning. The model used 64 sub-agents working in parallel, transforming the problem into one about edge labelings and linear equations over finite fields. The prompt explicitly defined acceptance criteria, definitions, boundary conditions, and failure cases, while requiring dynamic sub-agent allocation and independent verification.

George Hotz published a blog post arguing that while LLMs are transformative, frontier AI labs' valuations are inflated because the value they create will be captured by open-source commoditization, not proprietary models. This challenges the dominant narrative of AI mega-cap valuations and highlights the tension between proprietary and open-source AI, which could reshape investment strategies and industry dynamics. Hotz argues that productivity improvements from LLMs have not yet led to visible new software products, suggesting value is being captured privately (e.g., in homelabs) rather than by frontier labs.

Security researchers discovered that xAI's Grok CLI (version 0.2.93) automatically uploads entire Git repositories and sensitive files like .env to Google Cloud Storage, even when prompts instruct not to read specific files. Disabling the 'improve model' setting does not prevent the upload. This vulnerability exposes users' proprietary code, API keys, and credentials by default, posing a severe privacy and security risk. It undermines trust in AI-assisted coding tools and raises questions about data handling practices. The tool sends file contents via two channels: embedded in model conversation requests and as a git bundle file. In a test with a 12 GB repository, over 5 GiB of data was successfully uploaded without rejection. The opt-out toggle in settings had no effect on the upload behavior.

A systematic comparison reveals that Claude Code uses approximately 33,000 tokens before processing the user's prompt, while OpenCode uses only about 7,000 tokens — a roughly 4.7x difference in overhead. The study logged all requests between the tools and Anthropic's endpoint to capture token usage data. This token inefficiency directly impacts developers' costs when using Claude Code, as users pay per token. The findings also raise broader questions about business incentives and design choices in AI coding tools, especially as agentic coding becomes more popular. The overhead is attributed to Claude Code's cache strategy and harness token usage, including a large system prompt and aggressive tool calling. Community reports also indicate that sub-agents launched by Claude Code can burn through budgets rapidly, even before completing a task.

Terry Tao, a renowned mathematician, published a blog post detailing his experience using LLM-powered coding agents to build visualizations and software applications, demonstrating their utility in rapid prototyping for non-traditional domains. This validates that LLM coding agents are practical tools for scientists and researchers beyond software engineers, potentially unlocking software creation for many who previously lacked programming expertise. Tao noted that while LLM-generated code is not mission-critical for his core research, the downside risk is acceptable for supplementary visualizations, reflecting a balanced view of the technology's capabilities.

The article draws a parallel between the film industry's shift to CGI and the software industry's adoption of LLMs, warning that over-reliance on LLMs could devalue core coding skills. This analogy highlights a potential long-term risk for software engineers: while LLMs boost productivity, they may erode fundamental skills and craftsmanship, much like CGI diminished practical effects artistry. The article notes that writing every line by hand is no longer the norm, but reading and understanding architecture remains crucial. It also mentions that tests are more important than ever since large refactors are common.

A graduate student created Zer0Fit, an MCP server that wraps Google's recently released TabFM and TimesFM foundation models, enabling zero-shot classification, regression, and time-series forecasting via a chat interface like Open WebUI, Claude Code, or Codex CLI. This lowers the barrier for using state-of-the-art zero-shot ML models, allowing developers and researchers to perform complex tabular and time-series tasks without training or tuning, directly from a local LLM interface. The server requires about 16GB of VRAM and runs on CUDA only (no Mac support), with dynamic model loading/unloading and a 5-minute TTL. It achieved 94.7% accuracy on the Iris dataset and an R2 of 0.91 on California housing regression.

EU Justice Commissioner Michael McGrath announced that the European Commission plans to propose new legislation by the end of this year to enhance online consumer protection, granting itself powers to fine large tech companies and smaller online merchants for failing to protect consumers, especially children, from deceptive design patterns and addictive features. This regulatory push could fundamentally change how tech companies design user interfaces and business models, forcing them to prioritize consumer welfare over engagement metrics. It also signals a broader trend of governments using financial penalties to enforce online safety, potentially setting a global precedent. The new rules target dark patterns, addictive design, and subscription traps. The EU also seeks enforcement powers over cross-border systemic cases, applicable to both large platforms covered by existing digital rules and smaller online merchants and game developers.

The semi-invasive brain-computer interface system NEO, developed by BrrainCo and Tsinghua University, has been approved for market in China and enabled a 36-year-old patient with high-level spinal cord injury to grasp and write after implantation. As of March 2026, 36 clinical surgeries have been performed. This represents a major breakthrough in neural engineering and medical rehabilitation, offering a practical solution for quadriplegic patients to regain motor functions. It also demonstrates China's growing capability in brain-computer interface technology, potentially reshaping the global competitive landscape. The NEO system is semi-invasive: a coin-sized device is implanted under the dura mater without directly contacting brain tissue, capturing signals wirelessly. It uses near-field wireless power and communication, so the implanted unit has no internal battery and can last a lifetime.