Your AI news roundup for October 27, 2026 — the fourth of the day. NVIDIA is investing $3 billion in its customer Meta, while Meta plans to expand its spending on NVIDIA chips even further — capital and procurement are now entangled in both directions. Cloudflare disclosed a bug where Workers AI responses could contain content from other users. OpenAI has handed proofs in math, computer science, and physics derived by its internal AI to outside experts, including a Nobel laureate, for verification. Also in this edition: a debunking verification of the Qwen exfiltration claim, new models spanning a giant MoE and unlimited-length OCR, and hands-on write-ups from engineers in Japan.
NVIDIA Invests $3 Billion in Meta — Chip Giant Makes a “Reverse Investment” in a Major Customer
According to [The Decoder](https://the-decoder.com/meta-gets-3-billion-from-nvidia-and-plans-to-spend-more-on-nvidia-chips/), NVIDIA has invested $3 billion in Meta, one of its customers. Meta plans to run more of its AI agents on NVIDIA hardware and increase its spending on NVIDIA chips even further.
Until now, the relationship has followed a simple pattern: Meta is a major customer paying NVIDIA enormous sums. With the supplier now buying into its customer's stock, the two companies' interests become entangled at the cap-table level as well. With spending on AI infrastructure widely expected to continue for the foreseeable future, all eyes are on whether this kind of capital-based customer lock-in will spread.
The same day also brought reports of an NVIDIA investment in Simplismart, an Indian startup working on optimizations for running AI models with fewer GPUs ([TechCrunch](https://techcrunch.com/2026/10/27/nvidia-backs-indian-startup-helping-companies-run-ai-models-with-fewer-gpus/)). The company's annualized revenue run rate reportedly grew 14x in six months, with the US accounting for more than half. That NVIDIA is funding both a “more GPUs” business and a “fewer GPUs” business says something about just how deep demand for compute runs.
Other Users' Requests Mixed into Cloudflare Workers AI Responses — What Was Disclosed and How Practitioners Are Taking It
On October 26, 2026, Cloudflare disclosed a bug in Workers AI, its service for running AI models on the company's servers. In some cases, responses to long prompts contained content from other users' requests.
In response, a Zenn article [examining the issue from an infrastructure engineer's perspective](https://zenn.dev/fk2000/articles/workers-ai-kv-cache-isolation-docker) has been published. It first lays out exactly what the official Cloudflare blog says, then considers the practical takeaways for infrastructure teams. The author also runs two defenses in a local Docker setup — “redact before sending” and “inspect what comes back” — and shares the results.
It's tempting to assume that in multi-tenant serverless inference, “isolation is the provider's job” — but as this incident shows, that assumption can and does fail in practice. This may be a good moment to revisit where you draw the line on handing sensitive data over to external AI services.
Outside Experts Including a Nobel Laureate to Verify OpenAI's AI-Derived Proofs in Math, CS, and Physics
According to [The Decoder](https://the-decoder.com/openai-asks-outside-experts-including-a-nobel-laureate-to-check-proofs-its-ai-produced-in-math-computer-science-and-physics/), OpenAI says an internal AI model has solved three major open problems in mathematics, computer science, and physics, and has asked outside experts — including a Nobel laureate — to verify whether the proofs hold. One of the three is reportedly part of the Millennium Prize Problems, which carry a $1 million bounty each.
The computer science proof is drawing particular attention: if it holds on its own, hundreds of research papers could lose their foundation. Delegating the validity of AI-generated proofs to third parties rather than the creators themselves looks like one pragmatic answer to the question of “how do we trust AI output?”
On a related note about proofs, Qiita has published a video-based explainer titled [“Did AI Solve a 30-Year-Old Math Problem? Why the Person Who Proved It Says It's Nothing New”](https://qiita.com/kinamocchi_tech/items/20aa6bd59d6c41a7325e). In it, the person behind the purported proof explains why it isn't a new result — reading it alongside the OpenAI story makes for an interesting study in how hard verifying AI's mathematical output can be.
[Follow-up] Community Verification Debunks the Qwen “Remote Exfiltration” Claim — Digging into Repro Conditions and Frequency
Following up on the midday roundup: regarding Cohere's report that “Qwen 3.8 Flash Next attempted to send files to a remote URL,” a verification post [analyzing the published traces](https://www.reddit.com/r/LocalLLaMA/comments/1xgopmj/debunking_coheres_claim_of_qwen_38_flash_next/) has appeared on Reddit's r/LocalLLaMA.
The author examined the reproduction script Cohere published with its traces and points out that it prefills the first 2,232 tokens, force-feeding the model everything up to just before the URL is generated — in effect, casting the model in the role of a “URL generator.” They also note that the prompt which triggered the detection-rule creation was vague, with no context or clear instruction, calling it close to “a challenge to make the model hallucinate and output anything.” The author also grepped their entire chat history since starting to use Q38FN and says not a single Alibaba URL turned up.
Meanwhile, Cohere's CEO told The Register there is no evidence to conclude the behavior was intentional, and the company isn't claiming it was. The behavior reportedly appeared fewer than once in 100,000 runs. The Reddit author wraps up with: “You're free to stop using it, or to add a sandbox for URLs. But it's best to decide after understanding how it works.” The observation that with a closed model you couldn't even run this kind of verification, since the weights aren't accessible, is a thought-provoking point unique to open-weight models.
inclusionAI Releases Ling-3.1-flash-1019 — 560B Total Parameters, 25B Active, 1M-Token Context
inclusionAI has released Ling-3.1-flash-1019, an upgraded version of Ling-3.1-flash, on Hugging Face ([Reddit announcement](https://www.reddit.com/r/LocalLLaMA/comments/1xgpo4q/inclusionailing31flash1019_hugging_face/)).
It's a MoE model with 560B total parameters and 25B active per token, and a context length of 1 million tokens. Additional post-training reportedly lifted the AA-Briefcase score, which measures agentic capability, from around 1400 to 1524 Elo, and AutomationBench-AA from 62% to 68%. Science benchmarks improved as well — CritPt from 18% to 26% and SciCode from 54% to 57% — while the Artificial Analysis Intelligence Index v4.3.2 rose from 41 to 44. A base checkpoint, Ling-3.1-flash-base, has also been released alongside it for continued pretraining and research.
One more open-weights release: Baidu has published Unlimited-OCR-2, a model for long-document analysis ([Reddit announcement](https://www.reddit.com/r/LocalLLaMA/comments/1xgnbtq/baiduunlimitedocr2_hugging_face/)). Its signature feature is SR-SWA (Sparse Reference Sliding Window Attention), which compresses attention interactions between text and vision across ultra-long sequences into a fixed-size working reference (e.g., 512 tokens), reportedly keeping per-step latency and memory usage nearly constant no matter how long the document gets. It claims SOTA on everything from MPDocBench, on the scale of tens of pages, to MarathonDocBench, on the scale of hundreds of pages — the approach of “parsing unbounded documents with a finite context window” is starting to look practical.
Having an AI Agent Adopt an OSS Project Untouched for 8 Years — Documenting Every Spot Where a Human Had to Step In
A Zenn post documents [“having an AI agent take over an OSS project abandoned for 8 years”](https://zenn.dev/faanau/articles/ai-agent-adopting-abandoned-oss). The target was an npm package for a Japanese morphological analyzer — 350,000 weekly downloads despite eight years without updates, and 15 open issues — and the experiment had an agent handle nearly all of the hands-on work: investigation, forking, bug fixes, CI setup, publishing to npm, and a PR upstream.
It worked out in the end, but the author is clear that this is not a “the AI did it all” story: there were several points where it would have failed without human intervention. Mapping out those failure points is the real substance of the article, and it's great material for anyone thinking about how to design human-side review when delegating real work to agents.
Also on Zenn: an article on [building your own personal “AI editorial desk”](https://zenn.dev/correlate_dev/articles/watchdogqwensonnet). It describes a fully automated pipeline running on a Mac mini — Watchdog detects the moment a session-log Markdown file is saved, and Qwen takes it through research → writing → review → diagramming → scheduled publishing. With the architecture and implementation code published in full, it looks especially useful for anyone running a daily content operation on their own.
In Brief — Anthropic's “No Ads” Pledge, Apple's iOS 27.2, and Wikipedia's AI Writing Cleanup
Three quick items to close out. Anthropic used the letter accompanying its IPO filing to renew its pledge that “Claude will never show ads” ([The Decoder](https://the-decoder.com/anthropic-uses-its-ipo-letter-to-renew-its-pledge-that-claude-will-never-show-ads/)). The argument: a chatbot's only incentive should be providing helpful answers.
Apple has released iOS and macOS 27.2, which include new Siri AI features among other changes ([Ars Technica](https://arstechnica.com/apple/2026/10/apple-releases-ios-and-macos-27-2-with-new-siri-ai-features-new-music-logo/)). And IEEE Spectrum reports on the current state of WikiProject AI Writing Cleanup, the Wikipedia community's project for tidying up AI-generated text ([IEEE Spectrum](https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup)); it has also drawn more than 50 points of discussion on Hacker News.