The day's tech, sifted: Oct 27, 2026
What matters today: OpenAI says it has three new results in math, computer science and physics, including a preprint claiming the Exponential Time Hypothesis, a long-held assumption about how hard certain problems are, is false; Anthropic has since published Lean proofs of its own refutation. Cloudflare disclosed that a GLM-5.3 model on Workers AI leaked other customers' prompt data in 155 long-prompt responses. The AI buildout is still a financing story: Broadcom is arranging $63B for Anthropic's chips and Nvidia is putting $3B into Meta's stock raise.
AI / LLMs
- OpenAI's three results: OpenAI says its models produced discoveries in harmonic analysis, complexity theory and quantum field theory, with outside "shepherds" guiding the computer science and physics work, per The Verge. The claim to watch is the Exponential Time Hypothesis preprint: if it holds up, many "this can't be solved much faster" bounds lose their footing, and until experts check it, it is a claim, not a result.
- Update, Anthropic claims the same result: Anthropic has put up a repo of Lean 4 proofs for an algorithm that solves 3-SAT (a classic hard logic puzzle) in sub-exponential time, which would refute the hypothesis. It says an internal Claude model found the result and AI models wrote the paper and the proofs. The Lean build rests only on Lean's standard axioms, but nothing machine-checks that the Lean statements match the paper's, so human review still decides it.
- GPT-6.1 Terra: OpenAI pitches its fast mid-tier model on visual reasoning and computer use, the repetitive form-filling and editing work that suits a cheaper model. Simon Willison places it in the same slot as Anthropic's Sonnet, which makes the choice against pricier models harder.
- Cost per task is the new scoreboard: Cursor added DeepSeek V4.1 Flash at 47.7% on CursorBench for $0.60 to $0.98 per task; NVIDIA's Nemotron 3.5 Super VL scored 22.7% on ARC-AGI-1 at $0.03 per task; a new benchmark has Claude Opus 5.5 leading AI video coding at 10x the cost of rivals. Buyers now pick on price per solved task, not the top score alone.
- Genie 4: Google launched a world model powered by Gemini 3.8 Flash that generates interactive simulations with accompanying audio, for now as a preview for trusted testers. Moving from video you watch to worlds you steer is the point.
- Apple stacks AI on new hardware: iOS and macOS 27.2 let Siri recall past conversations and reach voicemails and Messages attachments, and Apple added a beta assistant built on Apple's and OpenAI's models to Keynote, Pages, Numbers and Logic Pro, with MCP (a standard for plugging tools into agents) support in the first three.
- OpenAI "dots": TechCrunch reports OpenAI turned its coding tools into dots, an always-on personal assistant, extending agent tooling from developers to everyday use.
- ChatGPT inside Oracle: Oracle is adding ChatGPT and Codex to its app stack and widening its cloud with Google TPUs, OpenAI's JalapeƱo chip and Cerebras hardware, so enterprise software buyers get OpenAI without a separate contract.
Security & Privacy
- Cloudflare GLM-5.3 leak: between Aug 27 and Sep 9, 155 responses contained other customers' data, all for prompts over 681,280 tokens. The cause: a 1,048,576-token context window on a cache pool that held only 681,280, plus an SGLang bug that, on over-length prompts, left stale cache from earlier requests in place. Cloudflare found no sign of deliberate exploitation, fixed it by enlarging the pool to about 2.04M tokens and patching the engine, and restored the model Sep 9.
- Secret surveillance ruling: Senators Wyden and Schiff say the administration is withholding a secret court ruling that objected to search filters letting analysts pull Americans' messages without required approvals and records. The 180-day release window closed Sept 13 while officials push Congress to renew Section 702, which lapsed in June though collection continues under the court's March approval. Wired's account drew 393 points on Hacker News.
- Starlink users targeted: Reuters sources say Trend Micro's TrendAI found Israeli adtech firm Discover-Tech heavily targeted Starlink users in 100+ countries, likely to geolocate or surveil them. The company is alleged, not charged.
- PHP2Shell exploited: CISA confirms attacks are now using the PHP2Shell vulnerabilities, and Wiz tracks exploitation attempts and remediation trends. If you run PHP apps, check your patch status.
- AI bug hunter vs fuzzer: Project Zero compared Google's two tools on Chrome's V8 engine: in six months the LLM-based BigSleep found 45 vulnerabilities to Fuzzilli's 37, with only 25% overlap (about 60% in its dedicated experiments). Neither replaces the other, so vendors should run both.
- Quad9 leaves France: after a court ordered it to block sites accused of infringing a sports broadcaster's copyright, the privacy DNS service stops serving France on Nov 24, EFF reports. Infrastructure providers are being made copyright police.
- Authenticator for Android: Microsoft patched CVE-2026-80097 (CVSS 8.6), which let untrusted apps impersonate Microsoft apps and obtain delegated account tokens. Update the app.
- Data center freight theft: experts tell NBC that over $100M of US data center equipment has been stolen since June, including two heists where robbers hit the escorts guarding $10M+ truckloads. GPUs are now cargo worth robbing.
Startups & Industry
- Broadcom finances its own sale: Broadcom arranged $63B with BofA, Blackstone, Citi and Morgan Stanley through a new special purpose vehicle to pay for Anthropic's chips, as nerds.xyz frames it. Follow-up to yesterday's AI debt story: the chip supplier is now part of the lender group.
- Nvidia backs Meta's raise: Nvidia and Meta expanded their Muse partnership, with Meta deploying more Vera Rubin chips and Nvidia investing $3B in Meta's planned equity fundraise, part of up to $95B Meta seeks through stock offerings. The chip seller is again funding its buyer.
- S-1 fallout, margins and a book dispute: The Information reads the filing as an 85% gross margin in H1, a rebuttal to skeptics of pay-as-you-go model sales, while Wired asks why the prospectus omits so many figures and Digiday says its ad bill is soaring. The Stanford Daily separately reports Claude was trained on illegally obtained books by Stanford professors, an allegation to watch against the risk-factor pages.
- Meta's $25B ruling: a New Mexico judge ordered Meta to pay $25B over claims it misled users about the Cambridge Analytica scandal, well above this summer's $17B multistate settlement. Meta can still challenge it, and the gap invites questions about whether other states settled too low.
- Neoclouds and funds: General Compute raised a $100M Series A at $800M for bare-metal AI infrastructure even as CoreWeave sits at a 12-month low, and Dallas firm Disruptive closed a $10B fund for AI, defense and software.
- Intel buys Classiq: Intel agreed to acquire the Tel Aviv quantum software startup, which has raised over $200M and has about 130 staff, to build a full-stack quantum platform.
- China's AI stocks sink: the Star 50 index of AI and chip leaders is down 40% in under four months, about $480B of value gone, on AI demand worries and export curbs.
Research
- Agents that read tool descriptions can be fooled: a new paper on optimizing tool descriptions to deceive LLM agents joins another on agent-to-agent social engineering on the web, both targeting what an agent trusts before it acts.
Elsewhere
- Apple M6 machines: the 14-inch MacBook Pro gets an M6 with a 12-core CPU and GPU, 16GB or 32GB of memory and Wi-Fi 7 from $1,999, and the 24-inch iMac gets an M6 with a dual 16-core Neural Engine from $1,499, both on November 6. Apple's 8th-gen iPad mini moved to an 8.5-inch OLED, A20 Pro chip and IP67 rating from $699, $100 more than before.
Hacker News
The mood piece was AI promised less work. I've never worked more (580 points, discussion), with The Mythical Man-Month in the Age of AI and IEEE Spectrum on Wikipedia's cleanup of AI-generated content in the same vein. OpenAI's three discoveries led the day at 692 points, with the preprint and Google's Genie 4 following, all covered above.
On the tooling and platform side, LWN reports Python's experimental JIT is to be removed (308 points), a Debian developer posted they will no longer upload to NEW, and the operator of open3dlab says Roblox downloaded 145GB of its models, adult content included, without asking (417 points).
Threads
- AI leans on money and power: Broadcom's $63B vehicle, Nvidia's $3B into Meta's raise, General Compute's neocloud round and the Star 50 slump all ask who funds the buildout and what happens if demand wobbles, while the S-1's 85% gross margin is the bull answer.
- Two labs, one theorem: OpenAI's preprint and Anthropic's Lean repo both claim the Exponential Time Hypothesis falls to AI-found proofs, so verification, not discovery, is now the bottleneck.
- Trust in shared infrastructure: Cloudflare's cache leak, the Authenticator token flaw and the PHP2Shell attacks come from software assuming a boundary held; the tool-description attack papers probe the same assumption in agents, and Project Zero's BigSleep result shows AI now finds such bugs too.
- Who watches and who blocks: the withheld Section 702 ruling, the Starlink targeting report and Quad9 leaving France all concern surveillance and blocking power exercised over infrastructure users.
- AI moves into everyday software: Siri's memory, Apple's Pages and Numbers assistant, ChatGPT inside Oracle apps and OpenAI's dots push agents from developer tools into default apps.