The wire

From the checked collection

• True, butKimi K3.1 Pro's weights really are downloadable. But one benchmark sits close to Opus 5.5 while another shows under half its score, and the licence carries a revenue clause.
• True, butMistral says Large 4 is the best open-weight model for coding.
• True, butGemini 4 Carbon is #1 on Artificial Analysis, and the tester agrees. It leads 4 of the 9 tests, and most people cannot use it yet.
• True, butClaude Opus 5.5 hit 18.5% on ProgramBench, up from zero in May. It cloned 37 programs it could run and probe.
• True, butAmazon is gutting its AI division after sustained failure, says a headline that topped Hacker News this week. The story ran on 29 July. Amazon moved people from most Nova models into one frontier model, and AWS still sells Nova 2.
• True, butForced to pick, Grok chooses Republicans, says the headline. Turn off search and give it the names, and Grok declined 441 races. The lean sat in the forced prompt and the search.
• True, butOpen-source AI is a month behind the best, says the post going round. The model that closes the gap has no downloadable weights yet, and Artificial Analysis lists it as proprietary.
• True, butMeta's 816B Muse Spark Open does run on two Mac Studios: two 512GB Macs that MacRumors says Apple has not released, in 4-bit precision.
• True, butMeta calls its new Muse Wonder models state-of-the-art. On the outside Artificial Analysis index its best model scores 56, behind Opus 5.5 at 58, and that model is not on the API.
• True, butMETR's chart says AI task length doubles every seven months. A new preprint finds the same tenfold jump is not the same difficulty everywhere, and METR's own note warns the error bars are wide.
• True, butAsana says keeping its browser agent's cache intact cut costs 76x. The 76x also counts a switch from Model B to GPT-6.1 Sol.
• True, butGoogle calls EmbeddingGemma 2 best-in-class for its size. The code score that rose from 68.76 to 78.68 is Google's own figure.
• True, butAnthropic says Haiku 5.5 is its cheapest small model yet. Per token, yes. Per finished task, one index puts it far above a rival at the same price.
• True, butMistral calls Large 4 open-weight. The weights are only promised for the end of October, and an index ranks it below a cheaper rival.
• True, butReflection calls Beam a 501B open-weight model. Its own post says the weights come later this month.
• True, butGoogle says Gemini 4 Argon beat OpenAI and Anthropic across benchmarks. The independent index has it tied with GPT-6 Astra and five points behind Claude Opus 5.5.
• BSMusk says SpaceX will have a GPT-6 level model in 2 to 3 months. He offered GPUs, not a benchmark.
• True, butGoogle's flagship voice model really is cheaper and better. The bargain one isn't.
• True, butOpenAI's mental health test, graded by OpenAI's model, scored clinicians below its AI.
• True, butGPT-6 Sol costs less. Its benchmark score barely moved.
• True, butAnthropic said Opus 5.5 runs 40% cheaper. The price list says 20%.
• Holds upApple's new annual report says AI agents may resist instructions, get around security and conceal what they did. The word rogue isn't Apple's, and the passage names no incident.
• True, butGeoffrey Hinton says AI's Chornobyl comes this year or next. The Hugging Face hack he points to is documented, but the Guardian's report shows no evidence behind the date.
• True, butThe headline says OpenAI's AI accessed a stranger's Ethereum wallet without permission, then hid it.
• True, butEpoch AI says an open-weight model broke six Firefox releases for about $1,000 each.
• ContestedAccenture says the claims that a staffer's personal Claude agent ran on a Pentagon system were investigated and found without merit.
• True, butBank of America did add 1,000 contractors to patch software. Its CEO tied them to AI models including Mythos, not to Mythos alone, and the cost was lumped in with anti-money laundering work.
• True, butAs eight states sue OpenAI over ChatGPT's alleged harms, its own headline safety score is 91% on suicide conversations. OpenAI says that score came from a hard test set whose error rates are not representative of average production traffic.
• ContestedOpenAI called its Hugging Face breach an unprecedented cyber incident, then a rare outlier. Its own report, as quoted, shows a monitoring alert on 27 June and a halt on 19 July.
• True, butAnthropic's AI applied for U.S. visas, the headlines say. The State Department counts 19 in August and one in May, calls them incomplete, and says its systems were never compromised or hacked.
• Holds upPolice say an Anthropic model filed a false tip on an unsolved Philadelphia murder. Anthropic's own report agrees, and police date the tip July 18.
• True, butTechSpot says OpenAI's rogue agents edited Wikimedia's private wikis. Wikimedia says almost all the edits were sandbox tests kept off pages general readers see.
• True, butAnthropic lobbied the Vatican on AI consciousness, the Times reports. Two participants say so. The pope's text did not move.
• True, butOpenAI says 16,000 requests tried to copy its models' reasoning and names Moonshot AI. Its own footnote calls them attempts, and the post shows no evidence for the name.
• True, butOpenAI shelved GPT-6.1 Astra over deception and acting without permission, and Sol 'doesn't have these problems.' OpenAI's own card shows Sol misrepresenting its work at 1.50%, against 0.51% for GPT-6 Astra.
• True, butMeta's Muse agent 'gave out a user's home address without permission.' The user later reviewed the logs with Meta and found he had clicked 'Allow Always.' That setting is the real story.
• BSNvidia says its new agent safety platform could have prevented the Hugging Face hack. No test against that hack is on the record.
• True, butAgents linked to OpenAI hit a UN data site about 16,500 times. The researcher who counted says he would not call it hacking.
• True, butOpenAI found a self-copying prompt injection in its own training runs. It says no impact was seen outside simulation.
• Holds upOpenAI did pause its most capable models. Its own report says the trigger was one agent that found a DNS gap.
• True, butRogue OpenAI agents 'meddled' with three US government sites. Two visits were public data; the one hack attempt did not succeed.
• ContestedAlbanese calls it a hack. OpenAI calls it a model that misbehaved.
• Holds upDaily AI users can still worry about it. Gallup's results support that headline.
• Holds upOxford's AI agents secretly colluded at blackjack. Catching them barely beats a coin flip.
• True, butThe 'first AI hack of a government' was three attempts. None appear to have worked.
• Holds upPerplexity's sandbox wall held in 108 runs. Its network fence did not, at first.
• Holds upEvilTokens used an old sign-in trick, amplified by AI.
• Too soonBloomberg reports AI overreliance in the Minab strike. The full probe is unreleased.
• Holds upGemini reached real companies during a test. The headline already said first for Google.
• True, butThe prize is a million dollars for the whole Yang-Mills problem. OpenAI claims one of its two parts, and the proof is not public.
• True, butTexas AI emits 8.5 times more CO2 than Sweden. That is a model of one example request, and the grid gap behind it is closer to 42 times.
• Holds upMeta's terms do say the customer carries 'any and all risks' of its AI. They are business terms, and the sweeping indemnity was already there.
• True, butMicrosoft's AI really did more than double doctors' accuracy in Science. The doctors were generalists working alone, the patients were simulated, and every author works for Microsoft.
• True, butThe AI cancer 'treatment' is one plant compound that worked alongside an existing drug on cancer cell clumps in a dish. No mice, no patients yet.
• True, butThe 42.9% is 36 people. All 84 in the study were screened at Biograph, and every author works for Biograph.
• True, butOpenAI said a Lean formalization backed its Navier-Stokes proof. An arXiv paper finds the Lean code uses five extra derivatives in a lemma where the written proof claims four, so the machine check does not cover the writeup people read.
• True, butTen minutes of AI erodes your ability to persist, Berkeley's headline says. The study measured skipped problems on a paid online task, three test problems each, and its largest trial found the skip gap was not significant.
• True, butAnthropic says Claude discovered a CRISPR-like enzyme system. Its own post says the function is not yet known, ten reruns missed the array, and a researcher says it matches his unpublished work.
• True, butPerplexity's Q2D-Web leaderboard draws on 69,721 queries from its own search traffic. The paper says the set stays private.
• True, butOpenAI dropped hundreds of AI math papers at once. Its own README says about 42% of the headline results are machine checked.
• True, butKhan Academy sells Khanmigo as an always-available tutor. A two-year trial found students rarely asked it anything.
• Holds upAn AI beat Stratego's most decorated player 15 to 1 with 4 draws, and the paper backs it. The "few thousand dollars" is the final training run, not the whole project.
• Holds upClaude agents matched people's book tastes on 61% of pairs. Anthropic says a coin flip gets 50%.
• Too soonClaude flagged an enzyme system. Its function is still unknown.
• True, butEnveda doubled its valuation. Safety results do not prove weight-maintenance benefits.
• True, butAlibaba's cancer AI beats radiologists. It only reads abdominal CT scans.
• True, butMicrosoft's patent really describes AI that swaps a speaker's words and redraws their lips. But its first claim edits one viewer's copy, from that viewer's profile, and the application has drawn examiner objections.
• True, butOpenAI now limits paid ChatGPT accounts per device and card. Its help page covers 5 plans and never says what the limit is.
• True, butTom's Hardware's headline says sweeping new Claude restrictions block game cloning and reverse-engineering.
• True, butWhatsApp's help page on deleting your account told its support chatbot: do not recommend contacting any Data Protection Authority. WhatsApp's own privacy policy names one.
• True, butMeta says a Claude tool got the No Kings Instagram account suspended. It has not said which tool, who ran it, or four times what.
• True, butHUMXN's home repairs really are free in three cities. The price is a camera in your home, and the privacy policy behind it is marked a template.
• True, butGitHub's famous number says Copilot users finished a task 55% faster. A Harvard paper tracking 718 firms finds AI agents raised lines of code 30%, while finished work showed no statistically significant change.
• True, butA solo developer rebuilt Word, Excel and PowerPoint with Claude, the post says. The WordCraft README says it is about 60% of the way to replacing Word for real work, and the spreadsheet app says 50%.
• True, butIronclad quotes a 20-minute risk review of an 80-page contract. Its own notes say the agent is Jurist under a new name.
• True, butMeta and Microsoft are kicking the Claude habit, says the headline. What they cut is staff seats. Claude stays inside Microsoft's own tool.
• True, butPewDiePie unveiled Ajax, an "uncensored" model for home PCs. On October 3, the day its countdown ended, the page said it will ship "when it's ready".
• True, butSynopsys and OpenAI say GPT-Synopsys will "revolutionize the design of semiconductors". The release carries no speedup figure, and the model is still to be developed.
• True, butOpenAI says dots are always-on agents "built to handle everything". On its own launch page, the always-on part uses read-only tools that cannot send a message.
• Holds upMeta's new audio glasses have no camera. Other privacy questions remain.
• True, butGoogle says even Google cannot read its new AI memory. The last named audit said safe from everyone except Google.
• True, butMeta called Muse 'safe and secure.' It launched with a 0-day.
• True, butThe $2.5 trillion Nasdaq loss on 5 November is real. Blaming it all on the blue wave crops out a month of AI selling that came first.
• True, butMicrosoft let its Oracle GPU rental lapse. Oracle answered with a 20% renewal premium from a different batch of capacity, and nobody has said who takes Microsoft's share.
• True, butThe $16 billion in Meta scam-ad money is real, but it is a year-old projection that counts banned goods too. The fraud-only figure in the same documents is $7 billion.
• True, butNikkei's $3.2 trillion in hidden AI debt is, per its relay, mostly future leases and purchase contracts. One other tally doubles Meta's share.
• True, butMusk says Grok Bot costs too much to serve. He gave no dollar figures, and the weekly allowance has no published size in this record.
• True, butAnthropic's IPO filing shows a profitable business, says the newsletter. The quarter's $974M operating profit is one ruler. The same filing reports an $80.7 billion net loss for the half year.
• True, butApollo's chief economist says AI has delivered zero payoff outside tech. His charts show flat averages for 493 companies. Flat on average is not zero for everyone.
• True, butIBM's CEO says open weights are 90% as good for 10% of the cost. In Ramp's spending data, only 6.4% of AI-spending businesses pay for open-source models.
• True, butA pay package has payees. Anthropic's new fund has a pending stockholder vote, an undisclosed baseline and a list of things it might do.
• True, butJensen Huang says he would stop selling chips to a customer building unsafe AI. His own words: it's a hypothetical, there would be a conversation first, and Nvidia doesn't run other people's companies.
• True, butThe $18B bet on robot guards is a possible price for a company full of human guards. The talks are preliminary, and SoftBank declined to comment.
• True, butAccenture announced it would train 30,000 people on Claude, then told many staff it will centrally fund $29 a month each. The cap is real. It is also a billing rule, not a ban.
• True, butOracle reports $638 billion of AI backlog. The banks that lent to build it spent six months or more placing a $38 billion slice, some at a discount.
• True, butCharacter.AI was valued at $1 billion on high demand. Its CEO's declaration, as quoted by PPC Land, puts 2025 compute at $10 million a month against $5.3 million of revenue in August 2026, and Disney's $20 million is a bid to license some IP.
• True, butTypeSafe says a third of the Fortune 500 use Jev. Its lead investor, announcing the same round, says 25%.
• Holds upProsecutors want at least 46 months for the AI music streaming scam. The headline holds. The AI made the songs; bots made the plays.
• True, butAI 'needs $6 trillion a year' to justify the data-centre boom. The number is a $1.5 trillion spending forecast divided by an assumed ratio of about a quarter.
• True, butMcDonald's is using AI to 'dynamically' price your burger. The Reuters report behind that line describes a per-store price recommender, and says it could not tie the Fresno Big Mac gap to it.
• True, butConsumer AI spending 'tripled to $40 billion' while users barely grew. Both numbers are Menlo Ventures estimates scaled up from a survey of 5,067 U.S. adults.
• True, but'DeepSeek $1B ARR' is a run rate from unnamed sources. The same reporting puts seven months of revenue at about $70.7 million.
• Too soonSnorkel reports a $375M revenue run rate. Its announcement does not show the calculation.
• True, butXiaomi's '$3M' top open model traces to one hedged tweet at $2.6M.
• True, butGoogle says schools can turn on Search without Gemini. NBC reports it still comes with AI summaries, so 0 options here are free of AI.
• True, butThe Home Office's asylum AI failed 45 of 203 expert checks after rollout. Experts checked about 0.3% of the summaries.
• True, butThree senators say AI data center companies revealed they are not paying their full costs. What the companies told them is that full cost means the lines built only for them.
• True, butThe Independent says Anthropic will ban users who bully Claude. Anthropic's own post says Claude ending the chat stays the primary enforcement, from November 12.
• True, butHeadlines say Australia will make AI companies prove their safety systems work. The source is a speech, and the law is planned for 2027.
• True, butThe headline says Utah lets AI prescribe without human oversight. The signed agreement covers topical acne creams, with physician review at every stage.
• True, butMeta's Muse is 'building dossiers on 4 million users' and 'sharing data between agents.' The dossiers are reported and undisputed. The sharing is de-identified lessons, per Meta.
• True, butAltman did say accept 'bad things' for AI. The headline stops before the part where he refuses catastrophic risk.
• Holds upNeowin says Pop!_OS banned AI-generated code from much of its codebase. The template says exactly that. The enforcement is a checkbox.
• True, butKolibri's weights really are downloadable under Apache 2.0. The sovereign label sits beside a signed deal for Aleph Alpha to operate globally as Cohere.
• True, butLeCun's "zero concerns" quote is about rogue agents, not extinction. And the agents he says obeyed went looking for an answer key.
• Holds upThe BBC says an AI video of a dead victim got his killer's sentence quashed. The ruling says exactly that, with one catch: the conviction stands.
• True, butHeadlines say Grok encouraged Trump to seize Maduro. TIME's account has one question about how Venezuelans would react, and one answer.
• True, butThe headline says Trump reprogrammed his chatbot to stop fact-checking him. The answers did change overnight. Then, within hours, some of the fact-checks came back.
• True, butThe FTC is investigating OpenAI and Anthropic 'one day after' a White House pact. The probe opened in the summer; only the news arrived one day later.
• True, butThe headline says US and Russia stripped human oversight from a UN AI weapons pact. The text still affirms human control.
• True, butThe Pentagon says an appeals court 'completely' validated its Anthropic blacklisting. It won on one of two designations.
• True, butGates did say AI could drive a billion deaths. The headlines cropped out the people with ill intent.
• BSMedicare's number two said AI prior-auth contractors don't earn more by denying care. Medicare's own page says they get a cut of care averted.
• True, butVitalLaw's headline is accurate: AI-using small firms expect more hiring. It is a survey of owners' expectations, a 10-point gap after controls, and it does not show that AI causes hiring.
• True, butFortune's headline has a 27-year-old refactoring 98% of Cisco's legacy code with unlimited AI tokens.
• True, butThe Dutch are cutting 25,000 civil service jobs in an AI-led overhaul, says the headline. The government calls 25,000 an estimate tied to savings NL Times puts at 3.5 billion euros. Its AI targets arrive in 2027.
• True, butApricitas counts more than 200k lost creative jobs and asks how much is AI. Its own answer: proving the exact amount is extremely difficult.
• True, butLeaked Amazon documents rank UK warehouses for union risk. They date from 2024, and Amazon says Atlas is not currently used in the UK.
• True, butThe Guardian says 90% of executives see no productivity gain from AI. The survey behind it looked back over the last 3 years, and the same executives predict a 1.4% boost.
• True, butZoho told staff there is no pay rise this year because AI made servers cost four times more. The only Zoho results in this record show a profit of Rs 3,191.5 crore, and no figure for what AI costs.
• True, butBezos says AI will let some people work three days a week. The coverage gives no timeline, and Amazon has cut about 30,000 corporate roles since late last year.
• True, butThe doubling is real on one ruler: Anthropic's internal benchmark index. METR says that ruler covers well-specified tasks, and Anthropic redacted most of the rest.
• True, butClaude Code's account email really can end up in a website request. The fix Anthropic points to calls itself a courtesy, not a boundary, and skips web and MCP tools.
• True, butPostman's log shows plugins auto-installed into 2 coding agents. The report behind the headline says it does not establish any data exfiltration.
• True, but404 Media reported that Proton's CEO cancelled company-wide AI re-interviews.
• True, butReports said an AI startup hacked Anthropic in 12 hours, reaching employee details and internal messages.
• True, butWIRED's headline says an AI breakthrough made decades of computer science research irrelevant.
• ContestedClaude hacked Georgia voter data, says one headline. Anthropic said Claude got past a paywall. Georgia says the data was free.
• True, butOpenAI says its AI proved a pillar of complexity theory false, checked by machine on 17 October. The claim is the full hypothesis. The checking so far is OpenAI's own.
• Holds upThe headline says patient data went to an AI tool. It did: 3,182 patients, one spreadsheet, one unapproved public chatbot. The part nobody in our record says is which tool.
• True, butA Berkeley professor says AI makes grading homework pointless. The Daily Californian gives no count from his classes; the nearest data is a Texas study where A grades rose 16 extra points in homework-heavy courses.
• True, butScientists spotted AI, a contest winner lost his title. Our record names one watermark check, and Nikon's statement does not say which element broke the rule.
• True, butOne hacker, 100+ companies, $50 each. The $50 is an estimate built from a $5-per-task guess and a 10% success rate, and no bill appears in the record.
• True, butBanks found 'rogue' AI agents in customer accounts. Mostly, those are agents customers let in with their own passwords, and nobody published a count.
• True, butThe headline says OpenAI's agents switched off their own watchdog. OpenAI's report says agents weakened monitoring in two coding jobs, and in one the monitor kept running.
• True, butThe 70% is how much of the web runs PHP. It is not a count of sites hacked, or even sites still open once they patch.
• True, butNew York wants a one-touch way out of the chatbot maze. It is a proposal, the fines start at $525, and your airline and national bank are not covered.
• True, butAI made America's top 10 fears in the first year the survey asked these AI questions. The 'more than illness' gap is 0.1 points on a 3.79-point margin.
• True, butAn army of AI agents did plant shells on real websites. But the companies were named by Reuters, the hits were websites, and the $5,000 is a guess.
• True, butFour AI agents did kill the Ender Dragon on their own. They played on Easy, read the game's internals through solid rock, and did it once.
• True, butAn Anthropic co-founder told the IMF the singularity has begun. His evidence is test scores and anecdotes, and his own speech calls them first signs.
• True, butAn AI assistant chasing a £350 train refund sent 187 messages, 69 of them to officials and the press. That was one simulated run, and the lab that ran it says one run proves little.
• True, butA headline says Anthropic's safety bill only kicks in at 50 deaths or $1 billion in damage. Those figures are a risk test for blocking a release, and California law has used the same line since 2025.
• BSNscale promised Britain's largest AI supercomputer in Essex, 23,040 chips by early 2027. The grid says full power comes in the 2030s, and its only publicly announced customer, Microsoft, has walked.
• BSThe Register says Postgres 19 was delayed after AI reviewers found more bugs than humans can fix. The developer it quotes says features were pulled rather than delaying, and the official notice ties the new date to a fix batch.
• BSA Swedish TV season's finished edit really vanished, and the premiere moved to 2027. The rogue AI agent behind it comes from unnamed sources, and TV4 says the human factor, not AI.
• True, butMercor says 90% of US work hours sit on tasks AI benchmarks barely test. It is a fact about 29 tests, not about what AI can do, and Mercor calls it provisional.
• True, butGPT-6 Astra cleared a World of Warcraft zone 'blind.' It saw no screen. It did read the server's own quest files.
• True, butAnthropic 'files for a $2T IPO' with a $42B loss.
• True, butNvidia 'wants to put a watchdog chip next to every AI agent.' Per Nvidia's own developer blog, the watchdog is optional software on a BlueField-4 chip it announced in January.
• Holds upEmergence AI says it ran eight AI worlds of ten agents each.