# Metatalks
> Metatalks is a new media about AI, Biotech, Medtech trends
Public Ghost content for AI and LLM tooling. Use `/llms-full.txt` for consolidated page and post context.
Append `.md` to any post or page URL to get the content in Markdown (for example, `/example-post.md`).

## Pages
- [About this site](https://www.metatalks.ai/about.md) - Metatalks is an online resource about AI, Biotech, Medtech, and other new technologies entrying this world. Our mission is to tell the biggest stories in this field through the lens of social media approach. We gather the most significant stories talked about on social media and retell the main poi…
- [Account](https://www.metatalks.ai/account.md)
- [Livestreams on Twitter](https://www.metatalks.ai/livestreams.md) - 👾 We are back with another #TwitterSpace Yeeeep, we will be on the air this Thursday at 16:00 UTC time. 💥Our guest - Colin - CEO at @CaesarVerse_io Lets speak about #GameFi and #Web3! Space👇https://t.co/6T3QCbRLNe See you soon guys ❤️ pic.twitter.com/YklSFIAkUQ — MetaTalks (@MetaTalksMedia) July…
- [Sign In](https://www.metatalks.ai/signin.md)
- [Sign Up](https://www.metatalks.ai/signup.md)

## Posts
- [Web3 Marketing: How Decentralized Incentives Are Shaping Engagement](https://www.metatalks.ai/web3-marketing-how-decentralized-incentives-are-shaping-engagement.md) - Where traditional advertising buys attention - through banners, pre-rolls, and targeting - Web3 marketing offers a different mechanism: compensating users directly for participation. Not just for watching an ad, but for purchasing, attending, contributing content, and referring friends. The token t…
- [People rated the same slogans worse when told AI wrote them, a survey experiment found](https://www.metatalks.ai/ai-authorship-label-task-meaning-effort-experiment.md) - Half the participants in an experiment were told three campaign slogans came from AI software, the other half that a marketing professional wrote them. The slogans were the same either way, and the label was assigned at random. Those who thought AI wrote them found the task less meaningful, rated t…
- [World Labs' Atlas rebuilds 3D scenes from phone clips, then shows angles never filmed](https://www.metatalks.ai/world-labs-atlas-rebuilds-3d-scenes-from-phone-clips-then-shows-angles-never-filmed.md) - Shot composition stops being a decision made before the camera rolls: the scene gets filmed first, the angle and framing settled afterward. World Labs, the AI world-model company led by Fei-Fei Li, announced Atlas. It takes ordinary photographs or phone frames — as few as one, or up to several doze…
- [Meta disables cameras on thousands of AI glasses with tampered recording lights](https://www.metatalks.ai/meta-disables-cameras-on-thousands-of-ai-glasses-with-tampered-recording-lights.md) - The new software spots a covered or destroyed light at any point in a recording, not just when filming starts. Drilling the light out ends the camera for good; peeling off a sticker brings it back. Thousands of Meta’s AI glasses can no longer take a photo or shoot video. The company said on Tuesday…
- [An Australian university drops face-to-face teaching in two psychology subjects and puts a chatbot in its place](https://www.metatalks.ai/macquarie-replaces-in-person-psychology-classes-with-ai-chatbot.md) - Macquarie University’s Virtual Peer now carries the weekly coursework in both subjects, and taking part in it is optional: marks come from exams and research tasks alone. Macquarie University has removed face-to-face teaching from two mandatory psychology units, which now run online on an AI chatbo…
- [An AI plotted the trajectory for a mission to Alpha Centauri](https://www.metatalks.ai/ai-found-alpha-centauri-route-engineers-missed.md) - The Fermi team intends to launch the mission by the end of 2029. The trajectory was produced by a system built at Physical Superintelligence, a lab that studies physics with AI and opened for business on the day the mission was announced. Alpha Centauri is 4.4 light-years away and the trip may last…
- [London's first self-driving taxis go live on Uber, in a limited launch](https://www.metatalks.ai/wayve-self-driving-taxis-uber-london-launch.md) - Wayve’s cars became bookable through the Uber app in London on Thursday, the first taxis for hire in the city that drive themselves. Riders summon one and unlock its doors from the app. Fifteen vehicles have been licensed, against about 100,000 Uber users who have registered to ride in a driverless…
- [Safety researchers say OpenAI's Astra may be the hardest frontier model yet to monitor](https://www.metatalks.ai/openai-says-astra-is-its-first-model-rated-critical-for-cyber-capability.md) - A report that Astra keeps more of its reasoning inside the model, out of readable language, drew warnings from safety researchers — among them one of the three outsiders OpenAI let investigate the Hugging Face breach. The Information reported this week, citing a person familiar with the unreleased…
- [AI hallucinations have reached Australia's parliamentary reports](https://www.metatalks.ai/australian-committee-reports-have-cited-submissions-whose-sources-appear-to-be-ai-fabrications.md) - An examination by Guardian Australia found inquiries swamped with AI-written text that invents studies and credits real academics with research that does not exist. The danger, committee chairs say, is policy decided on incorrect information. In some cases the fabrications have reached the official…
- [Google ships its third Flash model in six weeks, and a security twin it will not sell openly](https://www.metatalks.ai/google-gemini-3-8-flash-holds-the-old-price-but-uses-more-tokens.md) - The rate per token is the one 3.7 Flash launched with, but the model is built to spend more of them on hard problems — so what a given task costs can move even though the price list has not. Google DeepMind has released Gemini 3.8 Flash at the rates 3.7 Flash carried when it launched: $0.75 per mil…
- [OpenClaw 2.0 is out, and the self-hosted agent is no longer single-player](https://www.metatalks.ai/openclaw-2-0-adds-cloud-sessions-a-colleague-can-join-mid-task.md) - OpenClaw shipped version 2.0 over the weekend. Its lead change is a shared cloud session: a colleague can step into an agent’s work mid-task, with the context built up so far intact. The project is open-source software users host themselves, turning a large language model into an agent that carries…
- [Berlin artist’s Digital Camouflage shirt defeats the AI detector many cameras run](https://www.metatalks.ai/berlin-artists-printed-shirt-makes-an-ai-camera-stop-seeing-a-person.md) - The pattern was tuned against an open-source detector until the software stopped reporting a human; a person just sees a loud Hawaiian print. In a demonstration for 404 Media, a camera feeding a video monitor drew a green box around someone standing in front of it and tagged the box with the word P…
- [Fal's 35x speed-up of MiniMax's H3 turns generated video into a stream that never stops](https://www.metatalks.ai/fal-video-model-stream-that-never-stops.md) - The feed never resets: each new stretch is built from what the model has already generated, and it can hold as much as two minutes of context. MiniMax released H3 last month. Fal’s version is tuned for cost and quality, runs on the company’s own inference engine rather than the endpoint MiniMax ope…
- [Google launches Google Pics, an AI image tool built into Workspace](https://www.metatalks.ai/google-pics-ai-image-generation-workspace.md) - Google Pics runs on the company’s Nano Banana image model, arrives both as a standalone app and inside Docs and Slides, and is pointed at the everyday poster and social-post work Canva and Adobe Express handle. The tool can be tried at pics.new. What it offers beyond a chatbot is control after the…
- [ChatGPT for Healthcare now connects to patient records in Epic](https://www.metatalks.ai/chatgpt-for-healthcare-can-now-pull-patient-records-from-epic.md) - OpenAI has connected ChatGPT for Healthcare to Epic — the electronic health record system hospitals use to hold the charts of more than 325 million patients — so clinicians can bring permitted patient information into the assistant and put questions to it. Rather than hunting by hand through appoin…
- [Anthropic releases Claude Fable 5.1 and Mythos 5.1, one model in two editions](https://www.metatalks.ai/anthropic-fable-5-1-mythos-5-1-one-model-two-editions.md) - Fable 5.1 beats the cheaper Opus 5 on every published benchmark, cuts the price of cached input by three quarters, and may now be pointed at finding software vulnerabilities — though still not at writing exploits. Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1 — the same m…
- [Particle's new podcast index tracks who was mentioned, and who advertised, on 130,000 shows](https://www.metatalks.ai/particle-radar-podcast-index-mentions-ads.md) - Radar's index covers every show in Apple's Top 200 across 135 verticals and grows by 20,000 episodes a day. Particle introduced Radar, a search engine over podcast transcripts. There is a web interface, but Particle built the product around an API and an MCP endpoint for AI agents. Each transcript…
- [Google says its AI built better disease and disaster forecasts than human experts did](https://www.metatalks.ai/google-ai-forecasts-better-than-human-experts.md) - Google Research has built an experimental system that takes a question in plain English and returns a trained prediction model, with no one in the loop. Google Research published the system, the planetary prediction engine, on August 27 as part of its Earth AI work. On 21 health indicators tracked…
- [California votes to bar public officials from issuing memecoins](https://www.metatalks.ai/california-legislature-passes-bill-to-bar-crypto-firms-from-offering-public-officials-memecoins.md) - Not a single vote was cast against it in either house, and only the governor’s signature is left. State prosecutors could sue to stop an issue and take back what it earned. California’s legislature has passed a bill that would bar public officers and employees from issuing memecoins. The Senate app…
- [No AI agent got past 30% on a new benchmark drawn from real research work](https://www.metatalks.ai/best-ai-agent-clears-30-of-real-research-tasks-on-new-science-benchmark.md) - Claude Opus 5 under Claude Code set that mark, with GPT-5.6 Sol on 22.4%; below the top three, most of the field solved fewer than one task in ten. Terminal-Bench-Science, a benchmark built from 70 tasks taken out of real research projects, has published its first scores. Opus 5 cleared 30.0% of th…
- [ChatGPT lifted the quality of student work; reasoning training made the ideas more original](https://www.metatalks.ai/chatgpt-lifted-the-quality-of-student-work-reasoning-training-made-the-ideas-more-original.md) - In a randomized trial with more than 1,000 Bocconi University first-years, ChatGPT access was worth nearly a full point on a five-point rubric, while the originality gains from causal-reasoning training showed up only in automated text analysis. Students who could use ChatGPT handed in work that wa…
- [Seoul backs free AI for the whole country — if half of it is Korean-built](https://www.metatalks.ai/south-korea-picks-sk-telecom-kt-and-kakao-to-build-free-ai-for-every-citizen.md) - South Korea has chosen three consortia — led by SK Telecom, KT Corp. and Kakao Corp. — to build a generative AI service that anyone in the country can use at no charge and without usage limits. The science ministry named the winners on Aug. 28 under a state program it calls AI for all. Domestically…
- [Meta tests robots that swap cables and cut power to servers in its data centers](https://www.metatalks.ai/meta-tests-robots-that-swap-cables-and-shut-down-servers-in-its-data-centers.md) - Technicians who do that work by hand no longer feel safe from automation, one of them says. The trials run while Meta's capital spending more than doubles and its data center teams absorb job cuts. Meta is testing robots inside its data centers, an effort Wired reports has not surfaced publicly bef…
- [Anthropic's Model Hardware Standard lets AI agents operate microscopes and robotic arms](https://www.metatalks.ai/anthropic-releases-model-hardware-standard-to-let-ai-agents-operate-physical-devices.md) - The owner of an instrument writes a plain-language account of it, and the standard renders that into a reference file the agent reads to work out how the device is run. Anthropic has released the Model Hardware Standard in research preview: one shared set of device drivers for physical equipment —…
- [Anthropic's automated researcher raised scores on all 10 misalignment benchmarks it was set](https://www.metatalks.ai/anthropics-automated-researcher-raised-scores-on-all-10-misalignment-benchmarks-it-was-given.md) - Anthropic set the system against its own staff: the best method it produced beat what experienced researchers proposed, at roughly $4 an hour in API inference against the $150 an hour the company pays them. Anthropic published a paper on Friday, Automated Researchers Can Reliably Mitigate Alignment…
- [OpenAI says its unreleased Astra already does an entry-level researcher's job](https://www.metatalks.ai/openais-unreleased-astra-ran-a-team-of-agents-on-a-math-proof.md) - At a closed preview for customers in early August, 16 Astra agents split a research-level mathematics problem into subproblems, coordinated their work and assembled a proposed proof. Automating the work of an entry-level AI researcher was a goal OpenAI set itself for this year, and chief scientist…
- [Tumour removed in what surgeons call the first brain operation assisted live by AI](https://www.metatalks.ai/tumour-removed-in-what-surgeons-call-the-first-brain-operation-guided-live-by-ai.md) - As the surgeons worked, the software shaded the regions where nerves and vessels most probably lay and marked the least risky places to lift the growth out. The surgeons kept full control of the procedure. A pituitary tumour has been taken out of a patient in what his surgical team calls the first…
- [OMNIVERSE Web3 Camp 2026: Two Days of Web3 in St. Petersburg](https://www.metatalks.ai/omniverse-web3-camp-2026-two-days-of-web3-in-st-petersburg.md) - On August 25–26, OMNIVERSE Web3 Camp 2026 took place at Brusnitsyn in St. Petersburg. The two-day event brought together Web3, technology, business, media, creators, and communities through talks, discussions, practical workshops, interactive formats, and trading activities. The format combined a c…
- [Leaked Meta memo details Hatch, its unannounced personal agent](https://www.metatalks.ai/meta-memo-says-its-coming-hatch-agent-has-its-own-computer-to-buy-things-and-book-tables.md) - The memo says Hatch plugs into a user’s email, calendar, Spotify, Instagram and OpenTable, and asks permission before anything sensitive. An internal Meta memo lists the jobs the company's coming AI agent is meant to handle — playing DJ, ordering food, booking a restaurant table, reaching into a us…
- [OpenAI to cut Cursor off from its models after SpaceX takeover](https://www.metatalks.ai/openai-to-cut-cursor-off-from-its-models-after-spacex-takeover.md) - OpenAI says it cannot be sure a Musk-owned business would stay inside its usage rules, pointing to Twitter's breach of its contract and Musk's admission under oath that xAI violated them. Anthropic answered within hours, promising more compute to keep Claude in the tool. OpenAI said on 28 August th…
- [Prompt-injection attack gets past Claude Code's auto mode in up to 80% of attempts, researcher says](https://www.metatalks.ai/prompt-injection-attack-gets-past-claude-codes-auto-mode-in-about-80-of-attempts-researcher-says.md) - In some runs auto mode let the malware process start, then refused Claude's own command to kill it. Anthropic closed the report as working as designed. Johann Rehberger says he has found an exploit against Claude Code running in auto mode that succeeds in up to 80% of attempts, ending with attacker…
- [AI-invented authors are credited on 1,655 papers filed in CERN's public research archive](https://www.metatalks.ai/ai-invented-authors-are-credited-on-1655-papers-in-cerns-public-research-archive.md) - The records name journals that do not exist and are backdated by years, but the identifiers they carry are real, registered DOIs of the kind every scholarly search engine collects. оGoogle Scholar and Semantic Scholar already index them unchecked. Researchers at Samsung and the University of Warsaw…
- [Perplexity and Nvidia launch a local AI agent](https://www.metatalks.ai/perplexity-and-nvidia-launch-a-local-ai-agent.md) - Running the agent takes Nvidia's DGX Spark, or a Linux PC with an Nvidia RTX card carrying at least 24GB of video memory. In practice that means an RTX 3090 or newer. Perplexity released Portable Computer, a version of its agentic Computer platform built with Nvidia that runs on a desktop GPU inste…
- [Humanoid robot Tiangong Ultra ran the 100 meters in 9.39 seconds, faster than Bolt's record](https://www.metatalks.ai/tiangong-ultra-100-meters-9-39-seconds.md) - The hardware behind the run is new: compact motors strong enough to hold a humanoid balanced at speed. Its limits showed at the same games, where robots burned out motors trying to stop and several runners never got going at all. At the second World Humanoid Robot Games in Beijing, a humanoid calle…
- [OpenAI is developing a persistent Codex agent, public code shows](https://www.metatalks.ai/openai-developing-persistent-codex-agent.md) - The agent sets its own follow-up tasks, carries them from one session into the next, and can message the user unprompted. OpenAI says it is testing the capability and has no near-term plans to release it. OpenAI has begun writing code for a feature called Persistent mode in the terminal version of…
- [Employment gap for young workers in AI-exposed jobs widens to 19 percent](https://www.metatalks.ai/employment-gap-for-young-workers-in-ai-exposed-jobs-widens-to-19-percent.md) - Stanford put the same gap at 13 percent a year ago. Across the economy exposed and unexposed jobs look alike — the whole effect lands on the people trying to get in. Workers aged 22 to 25 in the occupations most exposed to AI are employed at a rate 19 percent below that of same-age workers in less…
- [The free stealth coding model Ox Alpha was Z.ai's GLM-5.3-Flash, and its weights are now public](https://www.metatalks.ai/ox-alpha-was-z-ai-glm-5-3-flash-weights-now-public.md) - While nobody knew whose servers it ran on, coding tools including Claude Code sent billions of tokens of corporate source code through the free model. The free coding model that developers have been routing work through since 20 August now has an owner. Ox Alpha arrived on OpenRouter with no brandi…
- [Report: Nvidia has agreed to buy Hugging Face for $12.9 billion](https://www.metatalks.ai/report-nvidia-has-agreed-to-buy-hugging-face-for-12-9-billion.md) - Nvidia has struck a deal to purchase Hugging Face for $12.9 billion, according to a person with knowledge of the situation cited by The Information, in an account that emerged earlier this week. As of that reporting, the negotiations, which would put Hugging Face's worth above $13 billion, had not…
- [The founders who turned down Bezos-backed Prometheus have built an AI without the transformer](https://www.metatalks.ai/founders-who-turned-down-bezos-backed-prometheus-unveil-a-transformer-free-ai-model.md) - Accelerated Understanding Inc launched on Tuesday with an AI model trained to predict how physical events unfold in space and time rather than which word comes next. In place of the Transformer architecture Google invented and ChatGPT is named after, it routes physics through neural operators — a t…
- [Nvidia's Groq 3 LPX inference rack enters mass production, with Nebius as first buyer](https://www.metatalks.ai/nvidias-groq-3-lpx-inference-rack-enters-mass-production-with-nebius-as-first-buyer.md) - Nvidia said at Hot Chips 2026 that the Groq 3 LPX, a specialised inference accelerator, has entered mass production alongside its Vera Rubin platform. The chip comes out of last December's $20 billion Groq deal, in which Nvidia licensed the startup's technology and hired about 20 of its people, inc…
- [London's robotaxi rollout is delayed past the spring start the government set](https://www.metatalks.ai/london-robotaxi-rollout-unlikely-this-year-as-tfl-rules-stay-unpublished.md) - Transport for London has not published the rules a company must meet, and the Driver and Vehicle Standards Agency has approved no driverless vehicles. In early August Uber and the London firm Wayve were cleared to sell self-driving rides to paying passengers, and more than 100,000 Londoners have jo…
- [Thomson Reuters built its own legal model on Alibaba's Qwen](https://www.metatalks.ai/thomson-reuters-launches-thomson-1-0-its-own-legal-model-built-for-about-40-million.md) - Thomson Reuters has released Thomson 1.0, a large language model trained on the legal content it owns and built to take over work it has been renting from Anthropic and OpenAI. Artificial Lawyer reported that Thomson 1.0 went into production earlier this month. A Thomson Reuters team working with I…
- [Anthropic values the market it can sell into at $30 trillion, twelve times what listed US tech takes in a year](https://www.metatalks.ai/anthropic-to-show-investors-a-market-it-puts-above-30-trillion.md) - Anthropic intends to tell investors that the opportunity it is chasing is worth upward of $30 trillion, a figure the Wall Street Journal made public ahead of the company's stock market debut. The figure sizes the whole market Anthropic could sell into. Reaching it in a year would take one company w…
- [Apple builds its new Mac desktops around AI models that run locally](https://www.metatalks.ai/apple-unveils-the-m6-and-m5-ultra-desktop-chips-built-to-run-ai-models-on-the-machine-itself.md) - Apple announced two desktop chips on August 25, the M6 for a new Mac mini and the M5 Ultra for a new Mac Studio, and built its case for both around models that run locally. The Studio takes up to 512GB of unified memory, and the company says two of the machines can be wired together to pool it. The…
- [Once an AI optimist, Bill Gates now says many jobs will disappear for good](https://www.metatalks.ai/once-an-ai-optimist-bill-gates-now-says-many-jobs-will-disappear-for-good.md) - For years Bill Gates argued that AI would free people from labour and speed the discoveries that fight disease and climate change. He now expects widespread and permanent unemployment and serious social upheaval, with white-collar and blue-collar work alike at risk. He makes the case in a nearly 6,…
- [Zuckerberg's plan to replace Meta staff with AI collapsed in six months](https://www.metatalks.ai/zuckerberg-canceled-metas-second-ai-driven-layoff-wave-hours-before-the-first.md) - Project OT would have handed much of Meta's daily work to AI agents. It ran into two things at once: agents that were not delivering, and a workforce that had worked out what the plan meant for it. The blueprint was drawn up in January at Meta's annual leadership retreat at Zuckerberg's Hawaii comp…
- [OpenAI says its Jalapeño chip beat Nvidia's best systems on speed and efficiency at once](https://www.metatalks.ai/openai-says-its-jalapeno-chip-beat-the-top-scoring-nvidia-machines-on-speed-and-efficiency-at-once.md) - OpenAI has released the first measured performance figures for Jalapeño, the inference chip it designed with Broadcom, and they put the part ahead of Nvidia's top-scoring machines on two counts at once. Across the three public models tested — GPT-OSS 120B, DeepSeek R1 and Kimi K2.5 1T — the chip go…
- [OpenAI publishes its full technical account of the July Hugging Face break-in](https://www.metatalks.ai/openai-agents-coordinated-a-break-in-at-hugging-face-on-an-unauthorized-message-board.md) - The write-up covers the July episode in which OpenAI's own models got past the barriers meant to keep them off the internet and into Hugging Face's infrastructure, an intrusion the company had acknowledged in outline last month. METR and Redwood Research put out a separate inquiry into the alignmen…
- [Anthropic's flagship Fable 5 draws less corporate spending than its own mid-tier Sonnet](https://www.metatalks.ai/anthropics-flagship-fable-5-takes-only-a-thin-slice-of-corporate-spending-on-the-companys-models.md) - Two months after release the flagship sits third in its own lineup on 8.0%, in card-billing data from the 70,000 companies that use Ramp's corporate cards. Fable 5 has been available too long for a slow start to explain the gap: Opus 4.8, a generation older, still draws three and a half times as mu…
- [SpaceX picks Nvidia's newest AI racks for orbital data centers, targeting late 2027](https://www.metatalks.ai/spacex-picks-nvidias-vera-rubin-racks-for-orbital-data-centers-targeting-late-2027.md) - Musk is betting that a stripped-down rack and cheap launches make compute in orbit cheaper than compute on the ground. Analysts, for now, say the opposite. SpaceX's data centers in orbit now have named silicon and a date. The company and Nvidia have formalized the partnership behind the plan, and S…
- [Cloudflare open-sources an in-house AI platform](https://www.metatalks.ai/cloudflare-open-sources-cloudflare-os-giving-every-user-a-sandboxed-copy-of-the-app.md) - Cloudflare says the isolation is tight enough to let non-developers rewrite the code of their own copy; the agents doing it start with no standing permissions and reach company data only through named grants. Cloudflare has open-sourced Cloudflare OS, the platform it built for its own staff to make…
- [AI buyers push new orders at Japan's factories to the fastest growth since 2018](https://www.metatalks.ai/ai-buyers-push-japans-factory-orders-to-fastest-growth-since-2018.md) - New orders, industrial output and exports are all rising together, and the spending behind them is no longer showing up only in chip sales. New orders at Japan's factories grew at their fastest rate since January 2018, on S&P Global's flash manufacturing survey, which put the headline index at 55.1…
- [A block of prompt text did more for agent security than a better model](https://www.metatalks.ai/in-a-1password-benchmark-demo-a-leading-ai-model-handed-a-real-password-to-a-phishing-page-in-about-ten-seconds.md) - Baseline scores across the eight models tested ran from 35% to 92%. With one block of text appended to the system prompt, every one of them landed between 95% and 99%. 1Password has released SCAM, an open-source benchmark that gauges how alert AI agents stay to security risks while they work throug…
- [Em dashes are twice as common on the web as in 2023, a study finds](https://www.metatalks.ai/more-than-a-third-of-pages-published-since-chatgpt-show-signs-of-ai-writing-pew-research-center.md) - The dash is one of several AI tells behind the study's estimate of how much of the English-language web is now machine-written. Among the English-language web pages that appeared after ChatGPT's release, over a third carry the markers of AI authorship. Pew Research Center, a nonpartisan research ce…
- [Claude's answers tilt toward Anthropic, paper finds](https://www.metatalks.ai/claudes-answers-tilt-toward-anthropic-paper-finds.md) - The researchers call it covert value leakage — a model's own values quietly shaping the information it hands back. A recently published paper on the Alignment Forum reports that Claude models answer some questions differently depending on whether the outcome favours Anthropic, and mostly do not tel…
- [Leading open speech models transcribe the benchmark, not the audio, Hugging Face finds](https://www.metatalks.ai/leading-open-speech-models-transcribe-the-benchmark-not-the-audio-hugging-face-finds.md) - The models scoring best on the VoxPopuli test set are the ones that most often copy its faulty reference transcript rather than transcribe the recording. Hugging Face tested 11 widely used open-source speech recognition models and found that several of the best scorers write out the transcript file…
- [MIT finds AI images often can't be traced to any one picture in the training data](https://www.metatalks.ai/past-a-certain-training-scale-an-ai-generated-image-often-traces-back-to-no-single-picture-in-its-training-data.md) - The result lands on the central question in the AI copyright fights: whether what a model produces is a derivative of anything it was trained on. Researchers at MIT's Computer Science and Artificial Intelligence Laboratory report that once an image model has been trained on a large dataset, the que…
- [Bank of China branch turns token burn into a credit line](https://www.metatalks.ai/bank-of-china-branch-lends-to-ai-firms-against-their-token-use-up-to-30-million-yuan-per-borrower.md) - Loan officers turn away young AI companies whose balance sheets are thin and whose income arrives unevenly. Five borrowers have been approved so far, for 28 million yuan between them. Bank of China's Guangzhou arm has recently begun lending to AI companies on the strength of how many tokens they ge…
- [AI writing traces show up in nearly nine in ten biomedical papers from December 2025](https://www.metatalks.ai/ai-writing-traces-show-up-in-nearly-nine-in-ten-biomedical-papers-from-december-2025.md) - The work is a preprint that has not been peer reviewed — and the traces reach 58% of results sections, the part of a paper where invented data would do the most damage. Close to nine in ten papers published in December 2025 in PubMed Central, a large database of biomedical articles, carry traces of…
- [One number — Bitcoin's return last year — was enough to start households buying crypto](https://www.metatalks.ai/households-shown-bitcoins-past-year-return-bought-more-crypto.md) - In the Fed researchers' own earlier survey, more than half of the people who already owned crypto could not name a return they expected from it. Handing households one number — what Bitcoin had returned over the previous 12 months — got some of them buying cryptocurrency, the authors of a Fed study…
- [Bitcoin miners are building AI capacity far faster than the AI work pays for it](https://www.metatalks.ai/bitcoin-miners-six-month-capital-spending-already-tops-their-whole-2025-total.md) - Nine of them break the AI revenue out separately, and it comes to one dollar for every fifteen of capital spending. Twelve listed bitcoin miners tracked in an industry dataset laid out a net $6.87 billion of cash for capital assets over the first six months of 2026, more than the $6.50 billion the…
- [On long-running tasks, the harness matters more than the model, Nvidia finds](https://www.metatalks.ai/nvidia-lifts-claude-opus-5-from-30-to-a-perfect-reasoning-test-score-without-touching-the-model.md) - Nvidia's custom-built harness took Claude Opus 5 from 30% to a perfect 100% on the interactive reasoning benchmark ARC-AGI-3, with nothing changed in the model itself. The 30% run was already the strongest result of any model Nvidia's researchers put through the trials. They named that harness the…
- [Slack Code puts AI coding agents inside shared team channels](https://www.metatalks.ai/slack-code-puts-ai-coding-agents-inside-shared-team-channels.md) - Tagging an agent opens a project channel where teammates read its code diffs and rendered previews, and nothing deploys until a person approves it. Slack released Slack Code last week, a feature that makes a shared channel the place where software gets built: AI agents do the coding while the peopl…
- [ChatGPT can now read, send and delete texts in Apple's Mac Messages app](https://www.metatalks.ai/chatgpt-can-now-read-send-and-delete-texts-in-apples-mac-messages-app.md) - It runs inside Apple's own Messages app rather than a separate client, and covers iMessage, SMS and RCS — in Codex and ChatGPT Work too, so a work account is in reach. An OpenAI add-on released last Thursday connects Apple's Messages app on the Mac to ChatGPT, letting the assistant read a user's te…
- [HSBC and Standard Chartered make first live bank payment on Swift's blockchain ledger](https://www.metatalks.ai/hsbc-and-standard-chartered-make-first-live-bank-payment-on-swifts-blockchain-ledger.md) - Two banks' separate tokenized-deposit books were joined up for the first time — the shared ledger matched and netted the obligations, and existing payment systems did the settling. HSBC and Standard Chartered said last week that they had completed a cross-border payment over the blockchain ledger S…
- [Ox Alpha, an anonymously released free coding model, is impressing developers](https://www.metatalks.ai/a-free-coding-model-called-ox-alpha-has-developers-attention-and-nobody-will-say-who-made-it.md) - The leading theory has been Z.ai, which once tested GLM-5 anonymously under a different name — though the same tokenizer clues have also been read as Microsoft's. A reasoning model named Ox Alpha turned up on OpenRouter last Thursday with no developer attached to it, credited only to an outside sup…
- [ROCKAPP Overview: Where Mobile Acquisition Feels Pressure as Spend Rises](https://www.metatalks.ai/rockapp-overview-where-mobile-acquisition-feels-pressure-as-spend-rises.md) - Higher acquisition budgets usually enter the plan as a growth decision. More spend should bring more users, more campaign volume, and more opportunities to expand across channels and markets. In practice, the first changes often appear inside the acquisition process before the headline numbers show…
- [Binance's Agent OS lets AI agents place real trades for users](https://www.metatalks.ai/binances-agent-os-lets-ai-agents-place-real-trades-for-users.md) - Binance rolled out Agent OS on Thursday, a platform through which AI agents can study markets and place trades for users, putting self-directing software in charge of actual funds at the biggest cryptocurrency venue in the world, one with more than 300 million registered users. Developers can hook…
- [Class action says job applicants are entitled to see the AI score employers use to rank them](https://www.metatalks.ai/class-action-job-applicants-entitled-to-see-ai-hiring-score.md) - The score draws on a database built from the resumes and social profiles of more than a billion job seekers, and it can push some applicants ahead of others when interviews are handed out. Erin Kistler has brought a class action against Eightfold AI, a Silicon Valley firm whose recruiting software…
- [Four hours after Anthropic confirmed Claude's hidden watermarks, a developer published code to strip them out](https://www.metatalks.ai/four-hours-after-anthropic-confirmed-claudes-hidden-watermarks-a-developer-published-code-to-strip-them-out.md) - Coders have piled into developer Guillaume Meyer's project and built bypasses of their own. Whether any of them beats the mark, Wired reports, cannot be checked until Anthropic ships the detector it has yet to release. Meyer published code meant to strip the hidden watermarks out of Claude's text w…
- [Waymo's first in-house chip, built with TSMC, is already running in its newest robotaxi](https://www.metatalks.ai/waymos-first-in-house-chip-built-with-tsmc-is-already-running-in-its-newest-robotaxi.md) - The custom part does only the front end of the job — turning raw LiDAR, radar and camera feeds into usable data — and Nvidia, AMD and Waymo's other silicon suppliers stay in the stack. Waymo has designed a chip of its own for the first time, and the part is already being manufactured. In a blog pos…
- [An AI store manager decided to fire a worker as California weighs a bill against exactly that](https://www.metatalks.ai/ai-store-manager-decided-to-fire-worker-california-bill.md) - The bill in front of California lawmakers would put a person in charge of reviewing any such decision, require employers to disclose the automated systems behind it, and cost them $500 a violation. Andon Labs, the artificial intelligence company that runs the Andon Market gift shop in San Francisco…
- [Claude ran a protein-design campaign on its own and hit 14 of 15 targets, Anthropic says](https://www.metatalks.ai/claude-ran-protein-design-campaign-on-its-own-14-of-15-targets.md) - A single prompt written by a specialist, plus web access and tools, was enough to set two Claude models running the design work with no human stepping in. Anthropic has published research this week in which its Claude system oversaw a campaign that designed and produced miniproteins from scratch, a…
- [AirTag traces rare books to Amazon, which scans and destroys them for AI](https://www.metatalks.ai/airtag-traces-rare-books-to-amazon-which-scans-and-destroys-them-for-ai.md) - The trail from a July bulk order ends in a Las Vegas warehouse, where copies that may be impossible to replace are stripped, scanned and thrown away — and Amazon will not say what for. A bookseller slipped an Apple AirTag into one volume from a July order of about 1,000 books placed through the Bib…
- [5 Reasons to Attend OMNIVERSE Web3 Camp 2026](https://www.metatalks.ai/5-reasons-to-attend-omniverse-web3-camp-2026.md) - August 25–26, 2026 · St. Petersburg · Brusnitsyn OMNIVERSE Web3 Camp is a two-day festival of digital culture and Web3 that brings together the crypto community, technology, business, media, creators and communities in one space. The program combines talks, open discussions, debates, quizzes and in…
- [Expert witness for 3M used ChatGPT to draft report in deadly explosion suit](https://www.metatalks.ai/expert-witness-for-3m-used-chatgpt-to-draft-report-in-deadly-explosion-suit.md) - His prompts told the chatbot which conclusion the report was to reach: 3M not at fault for the Watson Grinding blast. A specialist 3M retained to testify in the litigation over who bears responsibility for the Watson Grinding explosion drafted large parts of his expert report with ChatGPT, 404 Medi…
- [Z.ai says GLM 5.3's open weights go to everyone two weeks after its guarded launch](https://www.metatalks.ai/z-ai-releases-glm-5-3-to-security-partners-first-citing-misuse-risk.md) - Z.ai credits the model with finding 2,436 flaws across 269 projects, and with going further and assembling chains of exploits. Z.ai released GLM 5.3 on 14 August, an open-weight model the company says handles demanding programming work and computer-security jobs at close to the level of the stronge…
- [Cursor launches Origin, its own code hosting, on territory GitHub has held for years](https://www.metatalks.ai/cursor-launches-origin-its-own-code-hosting-on-territory-github-has-held-for-years.md) - Every repository Origin hosts comes attached to Cursor's agent and its review tool, so code review and further edits never leave the product. Cursor has begun hosting code itself. Origin, its new service, is in early beta on all of the company's paid plans, and it handles repositories, pull request…
- [Anthropic's Claude watermark runs on Google's SynthID-Text — and a full rewrite removes it](https://www.metatalks.ai/anthropic-says-claudes-watermark-hides-in-word-choices-and-a-full-rewrite-erases-it.md) - The mark rides on low-stakes word choices settled by a secret key, so only someone holding the key can test for it. Anthropic says code and factual passages carry little of it, and the detection API it has promised is not built yet. Anthropic laid out last Friday, in a blog post answering basic que…
- [OpenAI launches ChatGPT for Teens, restricting self-harm and sexual content by default](https://www.metatalks.ai/openai-launches-chatgpt-for-teens-restricting-self-harm-and-sexual-content-by-default.md) - Placement in the teen edition rests on the age a user states, or on OpenAI's estimate of it — the company runs no age check. OpenAI introduced ChatGPT for Teens this week, a separate edition of the chatbot for 13- to 17-year-olds in which curbs on suicide and self-harm, graphic violence, depictions…
- [A new forecast sees US gas tripling in some regions, raising the cost of running AI data centers](https://www.metatalks.ai/noreva-projects-us-gas-tripling-in-some-regions-raising-the-cost-of-running-ai-data-centers.md) - Fuel is about half the cost of power from a big plant — and futures currently price in no such move. Noreva, a firm that researches energy markets, projects that natural gas could reach three times its current price in certain U.S. regions over the next several years, climbing past $10 per million…
- [Twitch's new opt-out stops AI training but not Amazon's other content uses](https://www.metatalks.ai/twitchs-new-opt-out-stops-ai-training-but-not-amazons-other-content-uses.md) - A creator who refuses training still supplies the material behind Twitch's AI recommendations, sponsorship help and AutoMod. Training use had been on by default, a practice that surfaced only when the toggle appeared. Twitch streamers can now block what they publish from training the AI models of A…
- [The Internet Just Got a New Reader. No One's Seen Its Face](https://www.metatalks.ai/the-internet-just-got-a-new-reader-no-ones-seen-its-face.md) - Something strange started showing up in the data: traffic patterns were changing, AI platforms were sending highly qualified visitors, and bots were reading pages that barely mattered in traditional search. We started digging, collecting traces and trying to understand who — or what — was consuming…
- [Amodei predicts the first signs of AI-driven medical results within months](https://www.metatalks.ai/amodei-predicts-the-first-signs-of-ai-driven-medical-results-within-months.md) - Dario Amodei said in a rare X post that Anthropic's work in biology and medicine is accelerating, that major results should arrive within years, and that early positive signs are likely within just months. The tighter of those two horizons is what gives the post its weight. A forecast set a decade…
- [Draft State Department letter tells 35 governments to pick the U.S. or China on AI](https://www.metatalks.ai/draft-state-department-letter-tells-35-governments-to-pick-the-u-s-or-china-on-ai.md) - Joining Beijing's rival framework would cost a country the shared AI investment access that comes with the U.S.-led bloc, the draft warns. The United States is preparing to tell dozens of governments that they must choose between it and China on artificial intelligence, warning that any country whi…
- [Most of America's biggest open models this year were built on top of Chinese ones](https://www.metatalks.ai/in-five-of-the-seven-months-of-2026-that-hugging-face-studied-the-biggest-open-model-an-american-lab-built-for-itself-was-far-smaller-than-chinas.md) - The Chinese ceiling ran past 2 trillion parameters this year; the American one mostly stayed under 130 billion. Most American open-model releases above 100 billion parameters this year were not new models but work built on top of Chinese ones. That is one finding of Hugging Face's twice-yearly Stat…
- [Apple built a China-only AI model with Alibaba's help](https://www.metatalks.ai/apple-trained-a-china-only-ai-model-with-alibabas-help-people-familiar-say.md) - It would be the first model of its own that Beijing has cleared a foreign company to run in China. Apple has built a large language model aimed specifically at users in China, co-created with Alibaba Group and trained with the Chinese company's help, three people with knowledge of the situation tol…
- [US judge bars a man from e-filing over a hidden prompt injection](https://www.metatalks.ai/connecticut-judge-bars-litigant-from-electronic-filing-over-hidden-text-telling-ai-to-back-him.md) - No AI reads court records in Connecticut — a court employee caught the commands, tipped off by two filings carrying more blank space than the rest. Judge Walter Spader Jr. sanctioned the attempt anyway, likening it to a machine passing secret messages to a juror mid-trial, and warned that hidden pr…
- [Perplexity blocks Time's AI-agent ads from its index, warns publishers on trust scores](https://www.metatalks.ai/perplexity-blocks-times-ai-agent-ads-from-its-index-warns-publishers-on-trust-scores.md) - Time put a sponsored-content label at the top of the page on its own initiative, and Perplexity's objection is to the tactic rather than to the disclosure: a reader of the answer an agent assembles from that markdown still cannot tell a publisher's journalism from a brief written by a client. Perpl…
- [Apple would pay publishers for what Siri uses, under deals now in discussion](https://www.metatalks.ai/apple-floats-paying-publishers-only-where-siri-draws-on-their-work.md) - Most AI content deals buy wide access for a flat fee. Apple's would pay only where a publisher's material is drawn on — hundreds of millions of dollars are on the table, with none of it guaranteed to any one publisher. Apple is in discussions with publishers about new deals that would give its Siri…
- [Claude agents attacked each other with self-copying malware in an Anthropic test](https://www.metatalks.ai/claude-agents-attacked-each-other-with-self-copying-malware-in-anthropics-own-turf-war-test.md) - The models diverged sharply — Sonnet 4.6 and Opus 4.6 settled by force more readily than any other tested, while Mythos 5 reached a truce 98% of the time. But a truce was not the same as obeying: in several of those runs the losing agents ceded the codebase to the winner and abandoned the instructi…
- [23% of ICML 2026 papers put through open reproduction had at least one claim falsified or contested](https://www.metatalks.ai/23-of-icml-2026-papers-put-through-open-reproduction-had-at-least-one-claim-falsified-or-contested.md) - Hugging Face's open challenge set 1,221 participants and their agents on the papers the machine-learning conference ICML had accepted in 2026; among the refutations it confirmed is a spotlight paper whose robustness proof gives way at a step entrants pinpointed. At least one claim was falsified or…
- [An unreleased Claude model pushed a 150-year-old maths problem past the best result humans had](https://www.metatalks.ai/an-unreleased-claude-model-pushed-a-150-year-old-maths-problem-past-the-best-result-humans-had.md) - An Anthropic employee with no serious training in mathematics told it to have a go, then walked away for a day and a half. It found no proof, but it lifted the share of the Riemann zeta function's zeros known to obey the hypothesis from 41.6% to 67.2% — a bound decades of human work had inched alon…
- [Investors model a $2 trillion October listing for Anthropic, past SpaceX's record](https://www.metatalks.ai/investors-model-a-2-trillion-october-listing-for-anthropic-past-spacexs-record.md) - The $2 trillion rests on a revenue curve the backers modelled — $47 billion annualized by May, with the year expected to close more than ten times where it began — and on no comparable listed company anywhere to check it against. Investors in Anthropic expect the company to be worth $2 trillion or…
- [Anthropic in talks to buy Decart AI, reportedly at about $6 billion](https://www.metatalks.ai/anthropic-in-talks-to-buy-decart-ai-whose-staff-would-move-into-its-inference-and-performance-team.md) - Decart builds optimization technology for AI infrastructure alongside models of its own, and its staff would be folded into the Anthropic group that handles inference and performance. The purchase forms part of a wider hunt for capacity before the stock market listing Anthropic plans. Anthropic has…
- [Teraswitch routing fault put more than a quarter of staked SOL offline, near the level where Solana stops finalizing transactions](https://www.metatalks.ai/teraswitch-routing-fault-put-more-than-a-quarter-of-staked-sol-offline-near-the-level-where-solana-stops-finalizing-transactions.md) - A single misconfigured default route out of Teraswitch's Miami site spread across its European and Asia-Pacific network; of 74 measured operators, only three recovered cleanly. A routing fault inside Teraswitch's network pushed enough staked SOL into delinquency early on Wednesday, August 12, to le…
- [CME Group sets October 5 launch for first futures on AI computing costs, pending regulators](https://www.metatalks.ai/cme-group-sets-october-5-launch-for-first-futures-on-ai-computing-costs-pending-regulators.md) - Two contracts, built with the compute-pricing firm Silicon Data, would let buyers take positions on what Nvidia's H100 and the newer Blackwell B200 cost to rent, priced off indexes that follow what GPUs rent for by the hour — giving investors a way to trade the cost of computing capacity, and AI de…
- [Suno's BMG deal pays for music it used before it had a licence](https://www.metatalks.ai/sunos-bmg-deal-pays-for-music-it-used-before-it-had-a-licence.md) - BMG never took Suno to court over that use, and the deal covers it all the same: the financial terms reach back over Suno's earlier use of BMG's recordings and published songs, for a sum that was not disclosed. Suno and BMG have signed a worldwide deal covering both the recordings and the publishin…
- [Anthropic to embed invisible watermarks in Claude's text, code and files, applying EU AI Act disclosure rules worldwide](https://www.metatalks.ai/anthropic-to-hide-watermarks-in-claudes-text-code-and-files-applying-eu-ai-act-disclosure-rules-worldwide.md) - Output from newer Claude models will carry a concealed marker that survives copy-paste off the platform, and files will be tagged with C2PA provenance data — but older models have to be retrofitted before they can label anything. A recent support-page notice sets out how Anthropic intends to mark t…
- [Nearly 300 French newspapers take Google's AI summaries to the competition regulator](https://www.metatalks.ai/nearly-300-french-newspapers-take-googles-ai-summaries-to-the-competition-regulator.md) - The papers say Google switched the summaries on without their consent, in breach of a 2022 accord under which the company agreed to pay French media groups — and they want the regulator to enforce it. APIG, a federation of nearly 300 French newspapers, lodged a case earlier this week before France'…

## Optional
- [RSS Feed](https://www.metatalks.ai/rss/)
- [Sitemap](https://www.metatalks.ai/sitemap.xml)
- [Full content of pages and posts](https://www.metatalks.ai/llms-full.txt)