Friday, 9 October 2026 / transcript
Transcript — Fri 9 Oct
0:00 / 14:13
Maya and Alex are AI voices. Each part of the conversation below comes from one item in the written edition — linked above it — and is checked automatically before publishing: every number must appear in that item, every caveat the edition raises must be said aloud, the source must be named, and speculative or hyped language is rejected.
Intro
MayaIt's Friday, October 9th, and this is The AI Edge, presented by Epilogue.
AlexEpilogue is an AI venture studio and consultancy in Toronto, building products where the answer has to be right. Epilogue quotes every figure exactly as the source wrote it, and says so when something doesn't tie out. Visit epiloguelabs.com to learn more.
MayaI'm Maya.
AlexAnd I'm Alex.
MayaHere's what moved at the frontier of AI since yesterday morning: the advances, the research, and the uses for good and for harm, with every claim linked to its source.
AlexSo what's at the top?
MayaFirst, OpenAI has withdrawn three of the mathematics manuscripts it published this week, after a sign error invalidated an argument in one paper and the construction two others were built on.
AlexSecond, OpenAI has banned two covert influence operations, one from Russia and one from Iran, and rated the Russian one Category 5, the first it has disrupted at that level.
MayaAnd third, the Financial Times reports OpenAI has told investors its annualised revenue is approaching $50 billion, about $20 billion below the figures reported a week earlier.
Frontier models & labs — OpenAI withdraws three of its 719 maths manuscripts after a sign error invalidated two dependent papers
MayaStart with the maths. This is an update to a story we covered on October 7th. OpenAI published its catalogue of manuscripts on unsolved problems on October 6th.
AlexAnd now three of them are gone. OpenAI's revision log, dated October 7th, says a sign error invalidates what it calls a stabilization-trace cancellation argument, and that the same error affects the construction two dependent papers relied on.
MayaWhat else changed in that log?
AlexIt revises 14 other manuscripts, with proof repairs and corrected statements, updates 13 more to cite the revised companions, and puts formalisation at 300 of 719 top-line results, about 42%.
MayaA spokesperson told Retraction Watch that OpenAI welcomes scrutiny from the mathematical community, and that where errors are identified it will work to correct them promptly, and withdraw papers if no fixes can be found.
AlexTechCrunch adds one number worth holding onto: just 10 of the 719 manuscripts included a release of the model's chain of thought.
MayaNeither source says any of the underlying mathematical assertions has been refuted.
Frontier models & labs — Preprint: the Lean proof of OpenAI's announced Navier-Stokes blow-up does not match its natural-language proof
AlexThere's a second maths story. A preprint on arXiv says the formalised Lean proof of OpenAI's announced Navier-Stokes result does not correspond to the written proof of blow-up.
MayaWhat exactly did they find?
AlexRight. The authors give several examples of what they call AI mistranslations, where a statement or proof is rendered into Lean in a way that doesn't match the original. The abstract doesn't say how many.
MayaTheir conclusion is that these proofs shouldn't be trusted without the same peer review any other proof gets. But it's a preprint, so it hasn't been peer reviewed itself, and the authors say they make no claim about whether OpenAI's written proof is correct.
AlexTechCrunch reports that an advisory group of nine researchers, hosted by Princeton's Institute for Advanced Studies, had asked labs to publish machine-readable metadata linking the written and formal versions, and that OpenAI did not do that in this release.
Transition
AlexLet's turn to the research.
Research & papers — Epoch AI gave six models 11 of its own work tasks and concluded they cannot yet replace its staff
MayaEpoch AI did something unusual. It handed six models 11 of its own real work tasks, across five categories from graphic design to research design, and had a human grader score the outputs against a rubric.
AlexAnd the verdict?
MayaEpoch says Claude Fable 5.1 and GPT-6 Astra are broadly tied in the lead, and reliable on well-defined coding and computational work. But its conclusion is blunt: it finds AI cannot yet replace workers, at least not at Epoch.
AlexWhere do they fall down?
MayaMissing implicit standards, weak research judgment, and over-elaborating. Open-weight models trail further behind. The one score Epoch puts in the text is Kimi K3 at 158 on its capabilities index, roughly on a par with Grok 4.6.
AlexThe design is worth stating plainly: one run per model per task, scored by a single grader. So it's indicative, and it measures one organisation's work, not knowledge work in general.
Research & papers — Eight of ten AI search platforms cited a fabricated concept within seven days of it being posted
MayaA paper on arXiv now, about what AI search will repeat back to you. The authors analysed 17,211 citations across 10 AI search platforms, covering 6,356 source domains.
AlexConcentrated?
MayaVery. The top 20 domains account for between 20.5% and 70.8% of each platform's citations. Then the authors planted a fabricated concept on public posting sites, and 8 of the 10 platforms cited it within seven days. One picked up their marked content within an hour.
AlexWhat did that cost them?
MayaThey spent $14 on a generative-engine-optimisation service, which produced 13 public posts. And of 22 posting platforms tied to cited domains, 15 were rated low or medium barrier to set up an account and publish.
AlexIt's a preprint, so not peer reviewed. And the paper doesn't say which 8 of the 10 platforms cited the fabricated concept.
Transition
AlexNow to security and misuse.
Security, misuse & threat intelligence — OpenAI bans Russian and Iranian "false front" networks, rating the Russian one its first Category 5 operation
MayaOpenAI says it banned two clusters of ChatGPT accounts. One from Russia, which it calls Dark Clark. One from Iran, which it calls Bogus Bylines.
AlexAnd the Russian one is the headline rating.
MayaCategory 5 on OpenAI's own 1-to-6 scale, which it says is the first Category 5 operation it has disrupted since it started reporting. The Iranian one is Category 4 for planting articles, and Category 2 for its commenting work.
AlexWhat did Bogus Bylines actually do?
MayaIt ran seven invented reporters posing as Western journalists. OpenAI identified almost 100 articles published or syndicated under those bylines, across roughly a dozen small and medium outlets, the earliest from July 2025 and the latest from October 2026.
AlexAnd Dark Clark was named after a fake persona presented as the leader of a purported think tank, the Social Research Center. CyberScoop reports it focused on harming Ukraine's reputation in Latin America, and got involved in politics in Argentina and Bolivia, including fake audio clips, one impersonating the Ukrainian consul in Ecuador.
MayaOpenAI called that the most complex attempt to run a front identity it has disrupted over the past two and a half years. It didn't attribute either campaign to a specific government agency, and said the Iranian one looked like a commercial actor running a for-hire influence campaign.
AlexCyberScoop reports most campaigns OpenAI tracks rate 1 or 2, and that the Iranian stories drew little engagement, though the reach of these two prompted fact-checks and official denials in Latin America.
MayaThe caveat matters. This is a company claim, OpenAI's own account. And OpenAI's own report page wouldn't open for us, so every figure comes from the two reports that quote it, CyberScoop and Unite.AI.
Security, misuse & threat intelligence — Anthropic launches a critical-infrastructure defence programme with 11 founding partners and a free open-source scanner
AlexAnthropic is putting Claude on the defending side. It launched a Critical Infrastructure Defense Program with 11 founding partners, including CrowdStrike, Palo Alto Networks, Dragos, Rockwell Automation and Hitachi.
MayaDefending what, specifically?
AlexOperational technology. Power grids, water systems, transportation networks. Partners get frontier Claude models, on-site engineers and threat research.
MayaThere's a second piece, a free opt-in scanner for open-source projects, with reports Anthropic says are fully model-generated with no human review. Its red team says the work found over 29,000 candidate vulnerabilities in the last six months, about 6,000 of which were manually reviewed and triaged.
AlexThe number everyone will quote is a true-positive rate above 90%. Say it carefully: Anthropic says that's its expectation, not a measured result. This is a company claim, from Anthropic's own announcement.
MayaAnthropic also says some operational-technology fixes may take decades, and forecasts AI will favour defence in about two years. The post gives no dollar figure for the Cyber Mission or for the fund that keeps the scanner free.
Transition
AlexOn to defence and government.
Military, defense & geopolitics — Banks-Gillibrand bill would make DOD AI contractors on $100M-plus deals report weight theft within 72 hours
AlexDefenseScoop reports a bipartisan Senate bill that would put reporting duties on the AI companies the Pentagon buys from. Senators Jim Banks and Kirsten Gillibrand proposed it.
MayaWho does it cover?
AlexContractors with Defense Department AI deals worth $100 million or more. They'd have to report a national security incident, such as theft of model weights, within 72 hours of discovering it.
MayaAnd material vulnerabilities or concerning model conduct within seven days of judging the issue material, with a recertification that what they filed is still accurate at least once every 90 days.
AlexThe disclosures reach far: who can access model weights and training, unauthorised access or exfiltration, and past cases of a model evading safeguards or acting unprompted.
MayaGillibrand said the Pentagon is moving forward without commonsense guardrails in place. Banks said his bill strengthens reporting requirements.
AlexTwo caveats. This is a single source, and DefenseScoop doesn't give a bill number or say whether it's been referred to committee. So it's a proposal, not law.
Transition
MayaNext, science and health.
Health, science & medicine — Anthropic commits $150 million over three years to the US Genesis Mission's scientific research projects
MayaAnthropic says it will give $150 million over the next three years to the Genesis Mission, which it describes as a federal initiative to accelerate scientific and technological discovery through AI.
AlexGiven how?
MayaAs Claude, Claude Code and API credits for research projects, plus training and technical support. It names NASA, the National Institutes of Health and the National Science Foundation among more than 15 agencies involved.
AlexPriorities include fusion energy and quantum computing.
MayaWorth being precise: this is a commitment in credits and services, not cash. It's a company claim, Anthropic's own announcement, the post states no compute figures, and it names no specific research results yet.
Transition
AlexNow policy and law.
Policy, regulation & law — Anthropic's Usage Policy update, effective November 12, bars weapons software and arming drones
AlexAnthropic rewrote its usage policy, and it takes effect on November 12th.
MayaWhat actually changed?
AlexOn weapons, Anthropic says the policy always prohibited developing them, and the update makes clear the prohibition covers the software and components that make weapons work, as well as actions like arming drones and other autonomous vehicles. It says that reflects how it was already enforcing.
MayaThe surveillance section is rewritten too. Tracking people without consent is prohibited, in real time or from data collected earlier. Claude can't be used to decide or recommend who to investigate, arrest or charge. And building or improving surveillance tools is out.
AlexWhat's still allowed?
MayaConsented tracking, like fraud monitoring, plus content moderation, journalism and legal research. There's also a new rule for when Claude controls hardware that could injure someone: a qualified operator has to be able to watch and stop it, and it has to hold a safe state if Claude is disconnected.
AlexWhat Anthropic does not say is how the weapons or surveillance clauses change any existing government contract.
Policy, regulation & law — 114 lawmakers ask Google and Spirit Airlines to halt a $10 million deal for 100 million employee emails
MayaThe Record reports that more than 100 members of Congress wrote on Thursday to the chief executives of Google and Spirit Airlines, asking them to halt a data deal. The letter itself cites 114 federal lawmakers.
AlexWhat's in the deal?
MayaGoogle would receive internal Spirit data in exchange for a $10 million payment to the failed airline. The proposed data includes about 100 million emails and 500 million Microsoft Teams messages, plus payroll and tax information.
AlexGoogle wants it to train AI models, and says it will be deidentified. The lawmakers' argument is that removing names, email addresses or other direct identifiers does not necessarily make a dataset anonymous.
MayaA Google spokesperson told The Record the company is not looking to buy personal information, and that data will either be fully excluded or deidentified by an independent third party first.
AlexThe Record is a single source on this, and the letter is a request, not a legal bar on the deal.
Transition
MayaThen the money.
Compute, chips & infrastructure — FT: OpenAI tells investors annualised revenue is "approaching $50 billion", $20 billion below last week's reports
MayaThe revenue number moved, and it moved down. The Financial Times reports, as summarised by TechCrunch, that OpenAI has told investors its annualised revenue is approaching $50 billion.
AlexAnd a week ago?
MayaA figure approaching $70 billion was doing the rounds. That came out of attempts by OpenAI's own investors to build a direct comparison with Anthropic's annualised revenue.
AlexWhich is the part to be careful about. TechCrunch notes the two companies count differently. Anthropic includes sales made by its cloud partners and OpenAI does not, so the two run rates are not directly comparable.
MayaFor scale: OpenAI raised $122 billion in its March funding round alone, and its listing has been pushed to early 2027.
AlexIt's a single source for us, OpenAI didn't comment, and we couldn't open the original Financial Times article ourselves.
Transition
AlexAnd finally, deployment and impact.
Deployment & impact — Fired OpenAI safety researchers publish an open letter denying misconduct and warning of a chilling effect
AlexAn update to a story from last week. The three safety researchers OpenAI dismissed published an open letter on Thursday, to OpenAI's Safety and Security Committee, its Safety Advisory Group and its Mission Advisory Council.
MayaWhat do they say?
AlexThey deny mishandling sensitive information outside established procedures. They write that communications around the firing have made their former colleagues afraid to speak, and that terminations like theirs, executed and communicated so abruptly, are chilling the open culture OpenAI has prized in the past.
MayaThey also deny leaking to The Information about architectures in OpenAI's newest models that are harder to monitor. And they say one of them was talking to outside safety evaluators during the investigation of the Hugging Face incident, when internal policies were being developed in real time.
AlexWhat they ask for is that OpenAI keep its commitments to embed third-party safety auditors, preserve the monitorability of frontier models, and not use the firings as a pretext for stepping away from those partnerships.
MayaAnd OpenAI's side?
AlexIt hasn't formally responded. It gave TechCrunch an internal memo saying the decisions were not about raising safety concerns or speaking out, and a spokesperson described a pattern of misconduct in clear violation of its policies on mishandling research information.
MayaThe two accounts are irreconcilable, and neither has been independently verified. OpenAI did not answer questions about which policies were violated. We read the letter itself, and OpenAI has published no document setting out its side in comparable detail.
Outro
MayaThat's The AI Edge for today. The full edition, with a link to every source behind every claim, is on the site.
AlexOur voices are AI-generated.
MayaListen in tomorrow for the next edition.