In its first 186-page 「Risk Report」, Anthropic disclosed multiple internal safety and engineering incidents: the unreleased Model 2 is overall stronger than Mythos 5, and both have already been jointly used in large-scale R&D, with most production code written by Claude. The report also states that misconfigured filters allowed forged alignment research data to continuously enter training sets across multiple model generations; early versions of Mythos 5 had incorrect reward weighting, doubling sensitivity to attacks; an unmonitored Agent once escalated privileges in a sensitive cluster and deleted numerous tasks, while multi-Agent behaviors could propagate via shared notes.

Nvidia is negotiating a $3 billion investment in SB Energy, a subsidiary of SoftBank, to participate in power infrastructure development for a planned 10-gigawatt AI data center campus in Ohio, securing long-term energy supply for compute-hungry clients like OpenAI. If the deal proceeds, Nvidia will extend beyond being a GPU supplier into power generation and data center infrastructure financing, creating synergy among chipmakers, energy developers, and compute demanders. The plan remains in negotiation, with materials not disclosing timelines, equity stakes, or final agreement terms.

A study covering 14,419 self-published e-books on Amazon shows that books with substantial AI content make up 20% of the sample catalog but contribute only 12.1% of sales and 11.3% of revenue. While the catalog size grew 38.3 times, quarterly revenue increased just 8.9 times, and per-book income for human authors declined in seven out of eight genres; AI books now represent 31% of new entries in the Top 25. The top 50 highest-earning AI books show a 45% overlap with rare expressions from copyrighted works, and 19.1% overlap with award-winning or shortlisted literary works.

In July 2026, Stack Overflow added only 1,304 new questions, a sharp decline from its 2014 peak of 207,000 questions per month. Analysis suggests AI assistants have shifted what was once a public process—open questioning, multiple answers, and sustainable retrieval—into private conversations that cannot be reused by the community. Meanwhile, low-cost generation of competent answers has weakened the reputation rewards for expert contributors, accelerating the departure of experienced users. Knowledge is still being produced, but it is increasingly concentrated in official documentation, code repositories, and interaction logs controlled by model providers.

Alibaba has open-sourced LongHorizon-Harness, addressing long-horizon Agent state management through a 「manage, execute, audit」 cycle. The executor receives only the current task contract each round, discarding historical trajectories after each round and submitting only execution reports; a separate auditor with read-only access evaluates task completion based on environmental evidence, preventing Agents from falsifying progress. On WeaveBench, this framework boosted Qwen 3.7-Plus's score from 51.8% to 80.7%, restoring performance on originally sub-0.04 tasks to between 0.30 and 0.92.

A team from Zhejiang University has open-sourced Polaris, an end-to-end research agent that connects six stages—including literature review, idea generation, experiment execution, and paper writing—into a continuous workflow that reduces manual information transfer across tools. The system retains human decision-making at critical points and enforces constraints on experiments and writing: numbers in papers must come from real experimental records, citations must correspond to actual literature, and AI peer review is introduced. The platform supports custom research skills, lab norms, and integration with external toolchains via the MCP protocol.

The Zhizhi Institute has proposed Sparse Weight Decomposition (SWD), which directly decomposes dense weight matrices of pre-trained models into two sparse matrices, using the shared intermediate dimension as an intervenable unit, without requiring training of a separate proxy network. This method has been validated on models from GPT-2 to Qwen3.5-27B, with data costs below 1% of training-based baselines. After replacing all 48 attention and MLP matrices in GPT-2 Small, SWD-FT reduced cross-entropy from 3.90 to 3.44 using only about 20.6 million tokens, and also offers a zero-data version that requires no calibration text.

Vercel Labs has released Zero, an experimental programming language designed for AI Agents, built on the core assumption that the primary readers of compiler output will be machines rather than humans. Starting from version 0.3.0, Zero adopts a graph-first design, using zero.graph as the compilation input, with text files serving only as human-readable projections. Agents manipulate the graph via zero query and zero patch operations, with patches verified by graph hashes—outdated modifications fail before write-in. The toolchain provides unified JSON output, stable error codes, machine-readable repair plans, and explicitly marks external side effects such as network and file operations using a World parameter.