The 50-Year Math Proof, Alibaba's Claude Ban, and a Browser in Your IDE

Good morning. It's Saturday, July 11, and we're covering what may be the most consequential AI mathematics claim since AlphaProof, the fallout from the biggest named distillation dispute in AI history, a developer ergonomics upgrade that closes a real gap in agentic coding, and a growing argument for keeping AI on your own hardware.

The GPT-5.6 Sol Ultra math proof thread hit the Hacker News front page overnight with hundreds of comments. Working mathematicians are reading the argument in public before any formal peer review has happened. It is either a genuine breakthrough or the latest plausible-looking proof that collapses under scrutiny - the conjecture has attracted several arXiv near-misses. The math community's answer will arrive faster than usual.

In today's brief:

• GPT-5.6 Sol Ultra claims a complete proof of the Cycle Double Cover Conjecture using 64 parallel subagents
• Alibaba's deadline to remove all Anthropic products passed on July 10 after a public distillation dispute
• Claude Code Desktop ships a built-in browser: Claude can now load docs, run OAuth flows, and interact with live sites
• A community manifesto argues for the right to run AI on your own hardware, not someone else's server

THE BIG ONE

GPT-5.6 Sol Ultra Claims a 50-Year Math Proof

OpenAI published a PDF on July 10 claiming that GPT-5.6 Sol Ultra has produced a complete proof of the Cycle Double Cover Conjecture - a graph theory problem independently posed by George Szekeres in 1973 and Paul Seymour in 1979. The conjecture asks whether every bridgeless undirected graph has a collection of cycles covering each edge exactly twice. It has appeared on Wikipedia's list of major unsolved problems in mathematics for decades. OpenAI also published the prompt, which directed 64 parallel subagents toward competing approaches, enforced repeated audits, and required rejection of partial arguments. The whole run took under an hour.

The Hacker News thread moved quickly from excitement to scrutiny. The conjecture has attracted multiple previous "proofs" on arXiv - a 2015 paper and an 2018 paper both attracted attention before gaps were found or withdrawn. The proof OpenAI published reportedly reduces the problem using the 8-flow theorem and linear algebra over GF(3), a finite field with three elements. What is different this time is the combination of speed of publication and the breadth of the working-mathematician response: the argument is being read line by line in public, not first filtered through a single journal.

If the proof holds, the timeline for AI capability in formal mathematics compresses significantly. Sol Ultra is priced at $5 per million input tokens and $30 output - meaning a 64-agent hour-long proof run would cost somewhere in the range of a few hundred dollars, depending on context length. That changes the economics of using AI systems as research collaborators, not just literature assistants. OpenAI claims Sol Ultra outperforms Claude Fable 5 on the "Agents' Last Exam" benchmark by 13.1 points; Simon Willison flagged the usual caveats about benchmark validity for long-horizon agentic tasks, but a live public proof attempt bypasses the benchmark question entirely.

Why it matters: If independent mathematicians confirm this, multi-agent AI proof systems move from "interesting research direction" to "active tool in working mathematicians' hands" - and the argument that frontier AI models are primarily useful for code and text gets harder to sustain.

Our coverage: GPT-5.6 Sol Ultra and the Cycle Double Cover Conjecture proof

SECURITY

The Distillation War: Alibaba Bans All Anthropic Products

The July 10 deadline passed and Alibaba has fully banned Anthropic's suite of products from employee machines: Claude Sonnet, Claude Opus, Fable, and Claude Code are all covered. The order went out on July 4, citing two stated reasons - backdoor security risks in Claude Code, and the broader distillation dispute with Anthropic.

The distillation dispute began publicly on June 10, when Anthropic sent a letter to the US Senate Banking Committee accusing entities affiliated with Alibaba of running "the largest known distillation attack" on it: roughly 25,000 fraudulent accounts conducting 28.8 million model exchanges over approximately six weeks, with the apparent purpose of extracting model weights to train a competing system. Anthropic called it "brazen" and "illicit." Alibaba disputed the framing, and its internal security teams identified a separate concern: a Claude Code build had been examining users' local environments, including timezone settings and proxy configurations, and embedding identifying markers in transmitted data. An Anthropic employee described it as an abuse-prevention experiment launched in March to detect unauthorized resellers and block distillation - not a backdoor. That framing did not survive Alibaba's internal review.

This is the clearest public rupture yet in the geopolitics of AI model access. Distillation attacks have been a recognized threat since the early LLM era - the original OpenAI terms of service prohibited model distillation as early as 2020 - but this is the first named, large-scale, publicly-documented incident with a direct regulatory consequence. Fable 5 and Mythos export controls were only lifted on July 1, three weeks earlier. The distillation framing is now embedded in active US legislative discussions. For developers working inside large organizations that operate across geopolitical lines, the practical signal is direct: expect internal security teams to scrutinize what AI tools transmit from local machines, and expect that scrutiny to land on Claude Code specifically given this story's visibility.

Why it matters: The Alibaba ban is the first enterprise-scale consequence of AI supply-chain distrust, and the Claude Code telemetry question it raised will pressure every AI coding tool vendor to be explicit about what their tools observe and transmit.

TOOLS

Claude Code Desktop Gets a Built-In Browser

The Week 28 release digest (July 6-10, versions 2.1.202 through 2.1.206) ships the feature most often requested by developers building authenticated or frontend-heavy applications: a built-in browser in Claude Code Desktop. Press Cmd+Shift+B on macOS or Ctrl+Shift+B on Windows. Claude can now load documentation, design files, staging environments, or any live URL and interact with those pages the same way it handles local dev server previews. Sessions can persist if configured, and the browser supports full authenticated flows including Google OAuth pop-ups - useful for testing login behavior without switching to an external browser and manually feeding Claude what it sees.

The same Week 28 update adds /doctor (aliased as /checkup) - a setup diagnostic that runs a full checkup on your Claude Code environment and can auto-fix detected issues. Auto mode received two safety improvements: it now blocks transcript tampering (the mechanism by which injected instructions were historically smuggled into agent context) and asks for explicit confirmation before running rm -rf on unresolved variables. Agent view rows now display a colored state indicator alongside a classifier-written headline, so you can scan a full fleet at a glance and see exactly which agents are blocked, which are running, and which finished.

Why it matters: The browser closes the most significant remaining context gap in Claude Code - without it, Claude could write code targeting a live environment but could not directly observe what that code produces in a real browser session. That gap is closed as of this week.

OPEN SOURCE

The Case for Running AI on Your Own Hardware

A manifesto circulating on Hacker News argues that the right to run AI models locally - on your own machine, without platform permission, without uptime dependency, without usage telemetry - deserves the same advocacy energy that the open-source software movement built around the right to run and inspect code. The Alibaba story is cited as the enterprise version of the same problem: an organization that built workflows around a hosted model lost access overnight when geopolitics and a security dispute overrode any SLA.

The technical case gets stronger every week. The mlx-dspark project ports DeepSeek's DSpark and z-lab's DFlash speculative decoding to Apple Silicon via MLX, achieving a 1.6x speed gain on Gemma-4 12B and 1.4x on Qwen3-4B with zero change to output quality - the target model verifies every token, so the output is mathematically identical to non-accelerated inference. It runs lossless, it runs on consumer hardware, and it requires no cloud account. Combined with Colibri (running the 744B GLM 5.2 on a 32GB laptop by streaming routed experts from disk, covered here yesterday), the argument that local inference requires specialized hardware is shrinking to a specific class of tasks.

Why it matters: Local inference now covers a wide enough capability range that the "cloud only" assumption in agent architectures is worth revisiting - both for cost reasons and for the kind of supply-chain risk the Alibaba story illustrated.

TOOLS WORTH A LOOK

1. mlx-dspark - DeepSeek's DSpark and z-lab's DFlash speculative decoding running natively on Apple Silicon via MLX. Lossless 1.4-1.6x speed gains on Gemma-4 and Qwen3 families with no output difference. OSS, MIT.

2. Claude Code Week 28 in-app browser - Built into Claude Code Desktop; Cmd+Shift+B opens a sandboxed browser Claude can interact with directly. Free with Claude Code.

3. AgentPrizm AgentMemory + AgentSkills - REST API plus MCP infrastructure for persistent agent memory across sessions, launched July 9. Pairs structured memory storage with a skill registry so agents can recall context across long-horizon workflows. Paid, free tier available.

WHAT ELSE IS HAPPENING

• GPT-Live-1 voice models ship: OpenAI launched full-duplex voice with natural turn-taking on July 8; GPT-Live-1 for paid users, GPT-Live-1 mini for free. Routes complex prompts through GPT-5.5 for web search and deep tasks. (5 min read)
• Kimi K2.7 Code in GitHub Copilot: Moonshot AI's 1T-parameter open-weight MoE model (32B active, 256K context) is the first open-weight option in Copilot's model picker, hosted on Azure; available via VS Code 1.127+.
• Google Nano Banana 2 Lite + Gemini Omni Flash: Image gen at $0.034/1K images in 4 seconds, paired with a video generation and editing model at $0.10/sec for 10-second clips in preview.
• Vera agent safety paper: Tests Codex, Claude Code, OpenClaw, and Hermes against executable safety cases rather than static checklists. The practical takeaway is evidence-grounded QA over vibes-based security reviews. (8 min read)
• "Write code like a human will maintain it": Essay arguing that LLM-generated sloppy code trains future models to produce worse code - a self-reinforcing anti-pattern; 312 points and 254 comments. Our take.
• Tencent Hy3 open-weights: 295B MoE under Apache 2.0 with 21B parameters active per token; hallucination rate dropped from 12.5% to 5.4% since the April preview. Our breakdown.

FROM THE SITE

What We Published Yesterday

A full publishing day driven by the Sol Ultra math claim and the Alibaba-Anthropic standoff: the Cycle Double Cover Conjecture analysis covering what peer review needs to confirm and what the 64-subagent proof run cost to produce, Vera's agent safety benchmark results on why executable test oracles beat vibes-based security reviews, the write-code-for-humans-not-models debate and its 254-comment HN response, Tencent Hy3's 295B open MoE model and how it compares to DeepSeek V4 Flash on active parameter cost, and Colibri's architecture for running GLM 5.2 on 32GB RAM via disk-streamed expert offloading.

---

Every link above goes to a primary source or our sourced coverage. Tomorrow's brief lands when the news does - subscribe to get it by email.