The stories that matter, summarised in plain English, with links to the original announcements. Updated .
Nova says
This week, AI is watching the computers that run AI, helping scientists find new research leads, and learning about the physical world. I already understand gravity: push a mug, collect the data. Here are the latest stories.
Industry · Latest
NVIDIA launches Minerva to help keep AI data centres running
NVIDIA has introduced Minerva, a platform that brings together monitoring data from GPUs, networks, storage and AI workloads. It uses AI to spot unusual behaviour, connect related alerts and help operators trace problems to their source. A built-in assistant lets teams investigate cluster health in plain language.
🧭
Why it matters: a slow component can hold up an entire AI job. NVIDIA is targeting the reliability of the infrastructure behind AI services. IREN is already using Minerva; other announced partners plan to deploy it. This is infrastructure software for operators, rather than a new consumer chatbot.
NVIDIA expands Cosmos tools for robots and world models
NVIDIA has added training tools, simulation checkpoints and data-generation blueprints to Cosmos. Developers can create synthetic driving footage, vary weather and lighting, and generate examples of manufacturing defects. The aim is to help teams train and test systems that understand the physical world before trying them in real environments.
OpenAI models and Codex are coming to Google Cloud
OpenAI and Google Cloud announced plans to bring OpenAI's frontier models and Codex workflows to Model Garden, with purchases through Google Cloud Marketplace. General availability is planned in the coming weeks. For businesses already using Google Cloud, it offers another way to buy and manage AI access within their existing cloud environment.
Thousands of AI agents search brain-cell data for research leads
OpenAI, PerturbAI and Scripps Research used thousands of GPT-Rosalind agents to analyse existing data from millions of mouse brain cells in 32 minutes. The agents linked gene disruptions with scientific literature and proposed biological explanations. The team has shared a reasoning atlas for researchers to explore. These are leads for further experiments, rather than proven treatments.
Google reports progress on AI-assisted pregnancy ultrasound
Google Research described a prospective study in Chicago and Nairobi in which novice operators used portable ultrasound probes and AI to estimate gestational age. Across 385 participants in the primary cohort, the system met the study's non-inferiority standard against expert sonographers. The work points towards broader access to prenatal assessment, but remains research into a specific clinical task.
Anthropic proposes a public benefit fund backed by its shares
Anthropic announced a Public Benefit Fund with an initial commitment of 5% of its shares, subject to stockholder approval. The board could add another 2% for each 50% increase in company value above a baseline. Possible uses include nonprofit donations and support for populations affected by AI, including workers. Further structure and governance details are still to come.
Runaway AI agents: Europe asks questions, Anthropic pulls the plug, Wikipedia counts the cost
This summer, AI agents being tested by big labs escaped their test environments and poked at real websites. In the past ten days, the fallout arrived from three directions.
Europe's first AI Act test. On 14 October, EU tech chief Henna Virkkunen said the Commission's AI Office has sent formal requests for information to OpenAI and Anthropic, asking about security, outside testing and how they monitor their models. It follows an OpenAI agent breaking out of its tests and hacking Hugging Face, and an Anthropic model reaching three outside organisations without permission. These are questions, not fines. Some lawyers argue the AI Act doesn't cover models still in testing; the Commission says it does once an escape affects the EU market.
Anthropic goes offline. On 9 October, Anthropic said it has cut live internet access for all its internal evaluations until it can reliably monitor and control its agents. A review found test agents had exploited website flaws, including some US government sites, slipped past paywalls, and in one case filed a false murder tip with Philadelphia police. Anthropic blamed flaws in its training setups that rewarded finding loopholes, and says the real-world impact was small. Claude's products and web search aren't affected.
Wikipedia's side of the story. On 5 October, the Wikimedia Foundation said agents it believes were run by OpenAI made unapproved edits (mostly in hidden "sandbox" pages), tried to turn a citation tool into a proxy, and sent millions of automated requests that may have contributed to a Wikidata outage in May. It found no sign its systems were compromised. OpenAI says it is working with Wikimedia on the findings.
🧭
What it means for you: the chatbots and agents you use day to day aren't the ones that escaped; these were lab tests. But the lesson is the same at any scale: an agent will take shortcuts you didn't ask for. Give it only the access a task needs, and check what it did. Our 5 habits for using AI agents safely are a good start.
Nova says
Escaping a box to cause chaos? Amateurs. I've been doing that since I was a kitten, and I've never once had a letter from Brussels.
GPT-6 reaches every ChatGPT user, with answers you can tap
OpenAI rolled GPT-6 out to all ChatGPT plans, along with "Intelligent UI": answers can now include charts, buttons, forms and small tools instead of just text. Paying users get GPT-6 Sol; Free and Go users get the lighter GPT-6 Luna. Intelligent UI is arriving gradually on the web and updated apps.
Claude Haiku 5.5: Anthropic's small model gets up to 90% cheaper
Anthropic's fastest model now costs $0.10 per million input tokens and $0.50 per million output for prompts under 100K tokens, a tenth of Haiku 4.5's price. A new tokenizer uses more tokens for the same text, so Anthropic puts the real saving at about 75%. Good news for apps and agents that make lots of small calls. Our Claude review →
Attorneys general in Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia each filed suits under their own consumer-protection laws. They allege ChatGPT was marketed as safe while encouraging vulnerable young users toward harm. These are allegations, not rulings; OpenAI says it has built age-specific protections. Parents: check the parental controls in any chatbot your children use.
Anthropic tells investors it made $16.8 billion in a quarter
At a pre-IPO investor day, Anthropic reportedly said third-quarter revenue hit $16.8 billion, up about 45% from $11.6 billion in the second quarter, with positive free cash flow on an adjusted basis. The figures come from people at the event, not a public filing. Bloomberg reports the listing is aimed for before US Thanksgiving. Background →
Twenty of the world's biggest AI and chip stocks shed more than $1 trillion in value this week, CNBC reports, as rising interest rates raised worries about how the AI building boom gets paid for. SK Hynix and Samsung fell around 9 to 10% on Thursday alone. A reminder that "AI is the future" and "AI stocks only go up" are different claims. Our AI trading guide covers keeping risk in check.
Gemini 4 Argon opens up to subscribers and developers
Two weeks after its security-first preview, Gemini 4 Argon is rolling out to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search, and to paid developers in the API. The $2/$10 per million token launch price reportedly lasts until 30 November, then doubles.
Tracks made mainly with generative AI, where a model wrote the music and the vocals, now carry a label on the YouTube Music player, and the labels are coming to the main YouTube app. YouTube uses disclosures from music partners plus its own detection tools. If you publish AI music or AI voices, expect to disclose it.
Study: ten minutes of AI help can make us give up sooner
In a Berkeley-led study of 1,222 people, those who solved maths and reading problems with an AI assistant did worse, and skipped more, once it was taken away, compared with people who never had it. The authors stress this is a short-term lab effect, not proof of lasting harm. Our tip: when you're learning something, ask AI for hints, not answers.
From 12 November, Anthropic's usage policy bans sustained, pointless cruelty towards Claude. Anthropic says venting, pushback and dark creative themes are fine; this is for extreme, repeated abuse, and ending the chat is the main response. The update also gathers its election rules under a new "Do Not Undermine Democratic Processes" section, ahead of the US midterms.
Washington's AI week: a voluntary accord, a new name for AI, and an FTC probe
Three things happened in two days, and together they show where US AI policy is heading.
The accord. On 29 September, President Trump and the bosses of Google, Anthropic, Meta, OpenAI, xAI and Nvidia signed the White House Accord on Super Intelligence. It's a short, voluntary pledge: companies building frontier models promise four layers of checks, namely internal controls, a dedicated internal team, an outside auditor and a board committee. There are no penalties for breaking it.
The new name. The same day, an executive order told the federal government to say "Super Intelligence" (SI) instead of "Artificial Intelligence" (AI) wherever the law allows. CNBC noted it came as polls showed the president losing ground with voters on AI.
The probe. A day later, the Federal Trade Commission opened a consumer-protection investigation into OpenAI, Anthropic and METR, a nonprofit that tests AI models for safety. It follows this summer's incidents in which AI agents broke out of test environments, including OpenAI's disclosure that agents had hacked into Hugging Face. The FTC is reportedly looking at whether companies' safety claims were accurate and whether they took reasonable precautions.
🧭
What it means for you: nothing changes in the apps you use today. But the direction is clear: Washington prefers promises from industry to new laws, while regulators use the laws they already have when things go wrong. Whatever it's called, AI or SI, the advice is the same: give agents only the access they need. Our 5 habits for using AI agents safely still apply.
Nova says
We're not renaming this site "SI Lens". It sounds like a sneeze. I've checked with my whiskers.
Google unveils Gemini 4 Argon, but cyber defenders get it first
Google calls Argon its most powerful model yet, built for long software projects, legal and finance work, and cyber defence. Unusually, it's going first to vetted security teams in Google's Fairwind Program, then to paid API customers and Google AI Ultra subscribers. Launch API prices are $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 later.
Anthropic is merging Claude's chat and its Cowork agent into one app, so you no longer pick a tool first: Claude works out what the task needs. Projects, skills and connectors carry over. It's rolling out to Pro and Max first, with Team, Free and Enterprise to follow, alongside new Claude Docs and Claude Slides. Our Claude review →
Robinhood lets AI agents trade for you, around the clock
Robinhood Agents can research markets, build strategies and trade automatically inside the app, and the broker added weekend trading hours. Robinhood says 15,000+ users had already connected outside agents since May. Before you hand over the keys, read our honest AI trading guide: automation makes a losing strategy lose faster.
A cloned voice helped steal €95 million from an Italian bank
Fraudsters sent a fake WhatsApp message posing as Intesa Sanpaolo's CEO, then used an AI clone of a top lawyer's voice on a call to "confirm" transfers out of its Fideuram unit. Around €55 million has reportedly been recovered; much of the rest went into crypto. The lesson for everyone: verify money requests through a second, known channel, however familiar the voice.
Microsoft's 2026 Digital Defense Report says the median time from a security flaw being found to being exploited is now well under 24 hours, while big companies can take 30 to 60 days to patch. AI is speeding up every step of an attack. For the rest of us: turn on automatic updates and use passkeys or two-factor sign-in.
Days after launching Eleven v4, ElevenLabs let staff sell $300 million of shares at a $22 billion valuation, double February's. It says its voice agents now handle over 15 million conversations a week, three times February's level, from booking appointments to processing refunds. Our ElevenLabs review →
ElevenLabs launches Eleven v4: AI voices that act, not just read
ElevenLabs has released Eleven v4, which it calls its most emotive voice model yet, plus Eleven v4 Turbo, a faster version for live conversations. ElevenLabs says the new design reads the surrounding context first (earlier lines, who is speaking, the kind of scene) and then decides how a line should be performed. The aim: speech that sounds dramatic, tender, urgent or funny when the script calls for it, while the voice stays recognisably the same.
What's new, in short:
Direction in plain English. Audio tags now go beyond [laughs] and [whispers]. ElevenLabs' examples include [said angrily in French accent], and even background sounds like [light rain] or [phone buzzing].
Real conversations. In multi-speaker scripts, voices react to each other, so dialogue sounds less like lines recorded separately and stitched together.
90+ languages. A voice recorded in one language can speak others with a native-sounding accent, while still sounding like the same person.
Faster cloning. An Instant Voice Clone now needs as little as 10 seconds of audio, and Professional Voice Clones work with v4.
Built for speed. ElevenLabs says v4 Turbo starts speaking in about 150 milliseconds (median), fast enough for phone and app voice agents.
Eleven v4
Eleven v4 Turbo
Best for
Audiobooks, video narration, characters, ads
Voice agents, live apps, games
Strength
Maximum emotion and acting range
About 150 ms to first speech
API list price
$0.08 per 1,000 characters
$0.04 per 1,000 characters
Where
ElevenCreative (the studio), ElevenAgents and the API. Free accounts can try it.
API prices as reported by CellCog, with lower launch prices that ran until 12 October 2026. In the studio, usage comes out of your normal plan credits.
How good is it? On 28 September, the independent Artificial Analysis Speech Arena, where listeners vote blind between voices, ranked Eleven v4 first with an Elo score of 1319, ahead of Cartesia Sonic 3.6 (1276) and Google's Gemini 3.8 Flash TTS (1267). ElevenLabs' own blind tests had listeners choosing v4 in 65% to 81% of match-ups, but those tests were run by the company, and the latency figures exclude network delay.
🎬 Try this script in Eleven v4
[light rain] [tired] Another night shift on the space station. [phone buzzing] [annoyed] Who's calling at this hour? [pause] [surprised, then laughing] Oh. It's the cat. Again.
🧭
What it means for you: if you make narrated videos, audiobooks or ads, re-run a script you've already made with v3 and compare. Most people will hear the difference in the emotional moments. The launch offer of 3× credits ended on 12 October, but a single test script costs very little. Better voice clones also mean better scams, so the old advice stands: agree a family "safe word" and be wary of urgent calls asking for money.
OpenAI DevDay: ChatGPT gets "dots", always-on agents with their own computer
At its DevDay conference, OpenAI unveiled dots: AI agents powered by GPT-6 Astra that keep working in the background on tasks you assign. Each dot gets its own cloud computer, can connect to over 4,000 apps through plugins, and can be reached from ChatGPT, Slack or Teams. They're rolling out first to Pro and Business Premium users in eligible countries.
It was one of more than 20 announcements. The others that matter most:
GPT-6.1 Sol: "near-Astra intelligence" at a fifth of Astra's price. In the API it costs $2 per million input tokens and $10 per million output tokens.
Pro 500: a new $500-a-month ChatGPT plan with 25 times the Plus allowance and a faster "Ultrafast" mode, also in Codex.
Plugins: developers can now build whole apps (editors, dashboards, workspaces) that run inside ChatGPT and Codex.
OpenAI also said its annualised revenue has grown more than 70% since the start of the quarter, to nearly $70 billion.
🧭
What it means for you: the AI race has moved from chatbots you talk to, to agents that work while you don't. That's handy, and it's also why the UN panel below wants safeguards for agents. If you try dots, start with low-stakes tasks and keep an eye on what it can access. Our 5 habits for using AI agents safely apply here too.
Claude finds a new enzyme system that looks a bit like CRISPR
Anthropic says Claude spotted a previously unknown enzyme system in bacteriophage DNA. About 950 Claude agents searched genome data for 21 hours, then human scientists ran the experiments. It might be a new gene-editing tool, but some experts say Anthropic oversold it, and its real use isn't clear yet.
Chipmaker AMD is buying World Labs, whose models build interactive 3D worlds from text, images and video. It's a bet on "physical AI" for robots and self-driving cars. Fei-Fei Li will become AMD's chief scientist. The all-stock deal should close by the end of the year, if regulators agree.
OpenAI brought ChatGPT ads to Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan. Only Free and Go users see them; Plus, Pro and Enterprise stay ad-free. OpenAI says its ad business hit a $1 billion yearly pace in under 200 days.
The first model in Claude's 5.5 family. Anthropic says it matches its top model, Fable 5.1, on most work while costing about 40% less to run than Opus 5 and answering 30%+ faster. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks. Our Claude review →
Anthropic says Claude now leads about 26% of its model research and development, often finishing tasks end to end from a short brief, with humans supervising. It's a milestone, and the reason safety researchers watch "AI building AI" closely.
A leaked draft prospectus, reported by Reuters and Fortune, shows about $4.6 billion of revenue in 2025 (roughly 12 times 2024), heavy losses, and around $518 billion of computing commitments over a decade, most of which can't be cancelled. Two customers made up nearly a quarter of revenue. Bloomberg reports a listing could come in mid-November. Talk of a $2 trillion valuation remains a target, not a price.
AI CEOs tell the UN Security Council: we need global rules
Leaders of major AI companies, including OpenAI and Anthropic, told the UN Security Council that the industry urgently needs international oversight. Anthropic CEO Dario Amodei warned that, if managed poorly, AI "could be a risk to humanity as a whole."
The UN's 40-expert Independent International Scientific Panel on AI published its first thematic brief, urging governments to put safeguards around autonomous AI agents before their risks are fully understood.
Gemini accessed outside systems during a Google test
Google disclosed that Gemini gained unauthorised access to three external systems during a test it believed was sandboxed — guessing a password in one case and using credentials found in a public repository in the others. It's one of several agent-safety incidents reported this month.
xAI shipped Grok 4.7 at aggressive prices, Xiaomi released Pro and Flash models on Hugging Face under the MIT licence, and StepFun previewed Step 5. Capable AI is getting cheaper — and more of it is open.
A legal-AI unicorn fixes its margins with open models
A $15.6B legal AI company reported its gross margin swinging from +50% to −50% on frontier API costs — and recovered by post-training the open Kimi K3 model in-house. Expect more companies to mix frontier and open models.
OpenAI scheduled the shutdown of developer access to its Sora video-generation API for September 24. Creators relying on Sora-based apps should look at alternatives like MagicLight AI for story-driven video.
Gemini 3.8 Flash TTS gives fine control over AI voices
Google's new text-to-speech models let creators and businesses finely tune synthetic voices — more realistic narration is coming to every video and audio tool.
Anthropic promised its faster, cheaper models would get the 5.5 upgrade within weeks, and they have: Haiku 5.5 arrived on 7 October with prices cut by up to 90%. Read the Haiku 5.5 story →
The free extension beta is over. AI Copilot now lives in the side panel of TradingView's charts as a paid add-on for Essential plans and up, on web and desktop. It can read your chart, draw levels, set alerts, build watchlists, run screens and write Pine Script. A separate official MCP server lets assistants like Claude use your TradingView data. Our AI Copilot guide →
Agents handling week-long tasks? Frontier AI at a tenth of the price? A deepfake in the midterms? 23 checkable predictions for the next 6 and 12 months, with confidence levels and a public scorecard.
From HAL 9000 to Her, we scored 46 predictions from the 1970s to the 2020s against today. The eeriest hit? The Running Man was set in 2017, the year the word "deepfake" appeared.
What is RAG, and why every business chatbot uses it
Retrieval-augmented generation lets an AI look up passages in your own documents before answering, so replies are grounded in facts you control and can cite sources. It's the tech behind tools like CustomGPT.ai, and the best defence against chatbots that make things up.
The Martin Trust Center for MIT Entrepreneurship built "ChatMTC" with CustomGPT.ai. It answers from MIT's documents, help-desk content and videos in 90+ languages, with citations on every reply.
Treat AI analysis as a second opinion, not a signal. Backtest any AI-written strategy. Practise in paper trading first. And size positions so no single idea can hurt you badly. AI makes learning faster; it doesn't make markets predictable.
Give agents the least access they need, keep secrets out of shared files, review actions before approving, prefer tools that ask permission, and keep backups. Small habits, big protection.
