The Signal — October 28, 2026
The Read
Two frontier labs independently told the world their models had disproved the same 27-year-old conjecture, and had done it on the same day — October 16 — hours apart in the telling. OpenAI published three results on Monday across harmonic analysis, complexity theory and quantum field theory; by early Tuesday Anthropic had posted its own refutation of the Exponential Time Hypothesis to GitHub with 14 million output tokens behind it. The more durable story is what happened next: Russell Impagliazzo and Ramamohan Paturi, who published ETH in 1999, are among the people now being asked to interpret a machine's attempt to kill it, and MIT's Ryan Williams is organising the field's response with an open letter. That is a verification layer being built in public, in real time, which is the thing that has been missing from every previous AI-discovers-something headline. Elsewhere Anthropic bought a one-person biotech whose founder had been running eleven outside labs from a Claude window — the day-zero thesis with a drug candidate attached.
🌊 TIDE
No shift. One confirmation on ai-as-worker: the payoff gap got its second independent measurement in two days, and both readings point the same way — the returns are going to the companies that re-founded the work, not the ones that bought seats.
Three-quarters of boards expect AI productivity gains; 3% of organisations have hit their targets
Gartner put numbers on the boardroom-to-shop-floor gap at its HR Symposium: 75% of corporate board members expect AI to raise productivity and cut costs, while 3% of organisations have actually hit those targets. 38% are over budget on AI spending. 85% plan to spend more next year anyway, and only about one in five employees' time is being freed up. Forrester expects AI to account for 20% of all business technology spend next year with more than half of CIOs likely to miss their ROI targets; KPMG found 65% of UK leaders will maintain AI investment whether or not they see measurable returns. Gartner published no sample sizes or methodology for any of it, so treat the precision loosely and the direction seriously.
So what: This is the Apollo revenue-per-employee chart from the edition covering Sunday arriving again from a completely different instrument, and the honest reading is the optimistic one: the 3% is not a verdict on the technology, it is a measurement of how many companies changed the work rather than decorating it. 85% spending more into a 3% hit rate is the opening — the field is wide, most of the money is being spent badly, and the operators who re-found a process instead of bolting a copilot onto it are competing against a field that mostly has not. Pick one process this quarter, rebuild it as if the old constraints never existed, and measure it against the 3%.
🌊 WAVES
Two labs refuted the same conjecture on the same day, and the people who wrote it in 1999 are now organising the review
OpenAI said on Monday that its models had produced three results: a proof of the local smoothing conjecture in every spatial dimension, a rigorous treatment of quantum chromodynamics, and a refutation of the Exponential Time Hypothesis — the 1999 assumption that a class of hard computational problems cannot be solved faster than exponentially. Hours later, Anthropic researcher Levent Alpöge posted that Claude had refuted ETH too, on the same date OpenAI's models did, October 16, using 14 million tokens; Anthropic published the work to GitHub rather than to its own site. OpenAI asked MIT's Ryan Williams and ETH's own authors, Russell Impagliazzo and Ramamohan Paturi of UC San Diego, to help interpret the result. Williams — who earlier this month described OpenAI's 722-paper release as a "DDOS attack on mathematicians" — is organising the computer-science response, and the group has published an open letter asking for feedback and for norms on AI contributions, plus funding for a conference. Williams is measured on the substance: "I don't think that OpenAI's algorithm could possibly be the best algorithm you can get," and "I do think that it kind of calls into question a lot of the intuitions." Nothing is confirmed, no encryption is broken — OpenAI checked and told Williams it was "reasonably confident" there were no breaks — and if the claims hold, a large body of hardness results built on ETH needs revisiting.
Roadmap implication: Roadmap implication: the bottleneck on AI-generated science has moved from production to verification, and that is a market. Two labs producing the same result independently, on the same day, is the strongest capability evidence of the month — far stronger than either claim alone, because coincidence is a much worse explanation than capability. But the value only lands when a human field can absorb it, and what shipped this week is the absorption machinery: an open letter, a named review group, a GitHub repo instead of a press release. If your organisation is anywhere near formal methods, proof assistants, scientific review or research tooling, the demand curve for "independent verification of machine output" just became visible. Plan for a world where the expensive step is checking, not generating.
Oracle turned itself into the neutral chip mall, and Google's TPUs are leaving Google's own cloud for the first time at hyperscale
At Oracle AI World on Tuesday, Oracle said Google's TPUs will be selectable in Oracle Cloud Infrastructure — the first time the chips have been offered on a hyperscale cloud outside Google's own. Google's eighth-generation parts come in two variants, TPU 8t for training and TPU 8i for inference, with Google claiming nearly 3x the compute per pod over the previous generation and frontier-model development cycles cut "from months to weeks." Availability is calendar 2027, ramping over several years, with Oracle supplying power, cooling and datacentre networking. No dollar figure was disclosed. The same day Oracle announced general availability of NVIDIA's Vera Rubin NVL72 systems and Vera CPUs, an expanded OpenAI partnership that adds an Oracle AI Security Mesh combining multiple frontier models, among them OpenAI's cybersecurity models, and a native Snowflake service billable through Oracle.
Roadmap implication: Roadmap implication: the single-vendor accelerator assumption is over at the procurement layer, not just in principle. A cloud that sells you NVIDIA, Google silicon and an OpenAI-powered security product in the same keynote is telling you that neutrality is now the product, and the customer who benefits is the one whose stack can actually move. Two concrete actions: find out today how much of your inference is portable off its current accelerator, and get a 2027 TPU quote into your next compute negotiation even if you never intend to take it — the quote is worth more than the migration. The structural read is better still: when the chip supplier starts selling through a competitor's cloud, pricing power is leaking toward the buyer.
Anthropic bought a one-person drug company, and the numbers behind it are the most concrete day-zero evidence of the year
Anthropic said on Tuesday it had acquired Mithril Medicine, a one-person biotech developing treatments for two rare forms of muscular dystrophy. Founder Cody Tranbarger used Claude to coordinate contracts, logistics and experiments across eleven external labs, taking the programme from idea to promising mouse results in under eighteen months; if candidates hold up, a drug could reach human trials as early as late 2028. Alongside it Anthropic published results with six partners — BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines and The Column Group — with unusually specific numbers. Divergence generated roughly 35,000 antibody designs in about a month; of the hits, three bound IgE and twenty dislodged it, three of those outperforming the standard medication. BigHat made and tested about 1,010 variants in roughly ten weeks, with the best binding more than 10x more tightly. Generate's Claude-proposed CAR T design, about 70% shorter than standard, allowed roughly 10x more receptor expression in lab tests. At Enveda, Claude caught all eleven issues human reviewers found across 1,061 pages and flagged seven more, cutting review time 89% from over 25 hours to under three, and drafted one regulatory document in four hours instead of about three months. Anthropic says the work is early and mostly preclinical; nothing here has been in a human.
Roadmap implication: Roadmap implication: "pipeline-in-a-person" is the clearest statement yet of what re-founding a function actually looks like, and it did not come from a big pharma transformation office — it came from one researcher who never built the old org chart in the first place. The Enveda number is the one to take to your own operation: 25 hours to under three on regulatory review, and three months to four hours on a draft, in a function nobody would have nominated as the AI opportunity. Go find the equivalent in your business — the document-heavy, expert-reviewed, nobody-enjoys-it process sitting next to the work everyone is trying to automate. And note the strategic shape: a model company that buys the asset rather than selling the tool is telling you where it thinks the margin is.
The most-watched independent index rewrote itself, went 60% private, and the open-weight gap went from 3 points to 12
Artificial Analysis published Intelligence Index v5 on Tuesday, its biggest revision yet. The new index is 25% knowledge work, 25% coding, 25% science and 25% general, and 60% of it now rests on evaluations with private test sets or held-out answers. Claude Opus 5.5 leads at 58, GPT-6 Astra at 57 and GPT-6.1 Sol at 55, with Anthropic and OpenAI holding the top five; Gemini 4 Argon is next at 50. The headline change is at the open end: the gap from the leader to the best open-weights model widened from 3 points under v4.3 to 12 under v5. Kimi K3.1 Pro tops the open category at 46 but its weights are not released; among models with weights actually out, DeepSeek V4.1 Flash 1017 and GLM-5.3 share the lead at 32, twenty-six points back. Nathan Lambert's read, published the following morning, is that the new index is deliberately measuring where frontier models are going rather than where the economy currently runs, that the previous index was flattering open weights, and that going private trades community bug-finding for contamination resistance — "you're putting complete trust in the private benchmark builder." He expects open labs to close much of the gap at the next cycle in two to three months.
Roadmap implication: Roadmap implication: a vendor-selection input most boards treat as a thermometer just moved 9 points of apparent open-versus-closed gap by changing its own weights, and that is a governance fact about your procurement process rather than a capability fact about the models. Two things follow. First, if an index score sits anywhere in your model-selection paperwork, record which version produced it — a number that moves when the rubric moves is not a benchmark, it is an opinion with a decimal point. Second, the constructive version: Lambert is right that coding is where open models stay competitive and science is where they lag, so the actionable move is to stop shopping by composite and start shopping by the slice of the index that matches your workload. The open-weight economics have not changed; the measuring stick has.
🌊 RIPPLES
The largest proof in mathematics is now machine-checked: 43.1 million lines of Lean in 24 days, and 34 errors found in the human record
Axiom announced on Tuesday a complete Lean auto-formalisation of the Classification of Finite Simple Groups — roughly 15,000 pages across more than 500 articles by about 100 mathematicians over generations, compressed into approximately 43.1 million lines of machine-checkable code. Dozens of AI agents worked alongside human mathematicians across 108 milestones and about 380 smaller tasks, with the first agent swarm going live on October 2 and the work completing in 24 days. The agents surfaced 34 noteworthy deviations from the published literature — mostly typos, table errors and dropped hypotheses, with the most substantial concerning the Unbalanced Group Theorem — and produced a new preservation theorem while repairing one of the proof's gaps. The figures are self-reported in a company post, no independent party has rebuilt the proof, and Lean checks the deductions rather than confirming that the formal statement matches the intended classification.
Do this now: Do this now: if you own anything that depends on a long chain of human-verified reasoning — actuarial models, safety cases, compliance arguments, a pricing engine nobody has re-derived since 2019 — this is the week to ask what an agent swarm pointed at it would find. Axiom's useful output was not the proof, it was the 34 errors. The errata list is the product.
Genie 4 generates a world with its own soundtrack, and Gemini drives it
Google DeepMind published Genie 4 on Tuesday, a research preview for trusted testers. It is natively multimodal — generating ambient sound, music and dialogue alongside the visuals — steerable in real time by keyboard and natural language, with simulation of aerodynamics and collisions and characters that respond to their environment. Gemini adapts prompts to the scene and monitors coherence, with Gemini 3.8 Flash acting as a guiding agent that plans steps and verifies progress; one demo pairs Genie 4 with Gemini 4 Argon, which breaks a user-instructed task into steps for Genie to render, to simulate Apptronik's Apollo humanoid playing checkers. All output carries SynthID watermarks. DeepMind published no resolution, frame-rate, session-length or context figures and states the limits plainly: visual detail and object permanence drift over longer runs, and legible text, spatial accuracy and long-term consistency remain under research. It is coming to Google AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and AI Studio, with no public date.
Do this now: Do this now: if you train or test agents or robots, join the Project Genie waitlist this week rather than next quarter — the interesting claim is not the video, it is that the world model is now a place to run an agent, with another model steering the environment. The absence of any published numbers means treat it as a capability signal, not a deployable system.
A pre-revenue robotics company is raising at roughly $20 billion, nearly double its valuation from earlier this year
Physical Intelligence is in talks to raise at least $1 billion at a valuation of roughly $20 billion, Bloomberg reported on Tuesday, up from $11.5 billion earlier in 2026. Lux Capital and Avenir are reported to be in talks to lead; existing backers include Alphabet's CapitalG, Thrive Capital and Jeff Bezos. The company builds general-purpose software for robots intended to work across different hardware platforms, and has no revenue. The talks are early and the round is not closed.
Do this now: Do this now: if you are raising into physical AI, reprice your comparables — a pre-revenue software-for-robots company doubling to $20 billion in under a year is the ceiling moving, and the ceiling is what your next term sheet gets measured against. If you are buying rather than raising, the same datapoint says capital is arriving well ahead of revenue in this category, so budget for a vendor landscape that will look different in twelve months.
The research radar went almost entirely agent-reliability for a day
Hugging Face's curated Daily Papers for October 27 was a near-sweep for agents — their harnesses, their orchestration and their failure modes — rather than for raw model capability. The top entries by upvote: Trajectory2Gym on decision-grounded compositional interaction synthesis for tool-use agents (331), StreamVibe on proactive video agents (239), Trial by Trace on benchmarking agent harnesses for reliability and failure attribution (214), DynaFlow on free-form multi-agent orchestration (183), then InteractiveWorld (109) and Work Agents (98).
Do this now: Do this now: read the Trial by Trace abstract before your next agent post-mortem. When the community's own curation puts harness reliability and failure attribution above capability for a day, that is the research frontier telling operators where production agents actually break — and it is not in the model.
Meta upsized its equity raise to $97.5 billion, and the book came back five times oversubscribed
Meta priced its AI-funding stock sales on Tuesday at a total of $97.5 billion, up from the $95 billion announced Monday, lifting the common-stock portion to $25 billion from $22.5 billion. With underwriter options of about $5.5 billion the total could reach roughly $103 billion. The plan also includes about $12.5 billion in Series A depositary shares and up to $60 billion through an at-the-market programme. Reporting on the offering put it as five times oversubscribed. Separately at The Information's WTF Summit on Tuesday, Goldman Sachs global co-head of investment banking Kim Posnett said she is "hugely bullish" on AI while warning that investors are repositioning rather than retreating: "While AI itself is not a bubble, I think you can have bubbles within that." On circular financing among the hyperscalers and their suppliers she was direct — closer partnerships make sense and create greater risk, "both are true."
Do this now: Do this now: if you are modelling AI capex exposure, note that the marginal dollar is still arriving at a discount to demand — a raise upsized mid-flight and five times covered is not a market closing. The actionable caution is Posnett's, not the bears': differentiate within the trade rather than on it, and assume the cost of capital for anything not vertically integrated keeps rising from here.
- American Bazaar — Meta raises AI fundraising plan to $97.5 billion as Nvidia, AMD pledge investments
The full edition and the archive: https://excelsiorgroup.ai/insights/signal/
