Scoop
AI tool news · rumor vs. reality

The AI Wire ●

Rumors tracked. Announcements verified. Updated daily.

2026-10-21

RELEASED

Recursive releases Flint, a free open-weight local coding model built largely by its own AI research system

Recursive Superintelligence released its first model on Wednesday: Flint, an open-weight coding model designed to run on a laptop rather than a data center. Flint is a mixture-of-experts model with 39.8 billion total parameters but only about 2.5 billion active per token, fits on a 48 GB Apple Silicon Mac or a single 48 GB Nvidia GPU (4-bit weights take about 23 GB), and is free on Hugging Face under Apache 2.0. The model card lists a 262,144-token context window; Recursive's blog cites up to 512k.

Recursive says Flint generates more than 125 tokens per second on an M4 Pro MacBook Pro and more than 150 on Nvidia RTX Spark — up to twice comparable models. On its own charts it tops other MoE models under 50B on agentic coding, scoring 74.2% on SWE-bench Verified and 95.1% on IFEval, though it trails badly on Terminal-Bench 2.1 at 54.9% against 73.0% for a dense Qwen model. The company concedes Flint may need more guidance than large closed coding assistants.

The bigger claim is how Flint was made: a small team built it from scratch in under four months, leaning on Eureka Machine, Recursive's automated research system, for nearly every part of the stack. Recursive, led by former Salesforce chief scientist Richard Socher, emerged from stealth in May with $650 million in funding from backers including GV, Nvidia and AMD Ventures, and pitches Flint as a first public test of recursive self-improvement.

ANNOUNCED

PHP2Shell: AI models help find two critical PHP flaws hidden since 2004, and the NSA tells sites to patch or go offline

Two critical vulnerabilities in PHP — the language behind roughly 70% of websites, including WordPress sites — were publicly disclosed on Wednesday, with emergency patches issued by the PHP Foundation. The flaws, tracked as CVE-2026-109618 and CVE-2026-109619 and collectively nicknamed PHP2Shell, are rated 9.2 in severity, allow unauthenticated remote code execution, and have sat in PHP versions since version 5 in 2004.

AI found them first. Security firm Calif, working in OpenAI's new Patch the Planet initiative, used GPT-5.6 Sol to find one flaw in August, while Wiz independently found both using its Wiz Atlas agent together with Google's Gemini 4 Argon and GPT-5.6 Sol. OpenAI says it led a two-month coalition of maintainers, researchers and providers to prepare fixes, and is putting $1 billion in subsidized model access behind defender work of this kind.

The US National Security Agency warned that sites that cannot patch immediately should be pulled offline until they can. Patched releases are PHP 8.5.12, 8.4.27, 8.3.36 and 8.2.35; versions 8.1 and older are unsupported. Wiz notes the same AI capability that found the bugs also lowers the barrier to exploiting them, which is why the patch race is urgent.

ANNOUNCED

Tibo's Day 17: a Codex reset day, plus a push to sign up for Daybreak Blue access

Day 17 of Tibo Sottiaux's 28-day ship-or-reset run, which lasts until November 1, landed on Wednesday as a reset day. The post reads: bet you thought we forgot, but no, it is reset day today — a full Codex limit reset for users tracking the series.

The same post used the moment to remind developers to get Daybreak Blue access, OpenAI's vetted tier for security work with its frontier models, saying it only takes about 15 minutes to apply. Daybreak Blue is the access route behind the recent run of AI-assisted vulnerability work, including the PHP flaws disclosed this week.

At Thursday morning's sweep, Day 18 had not yet landed: the trackers still showed waiting for Tibo, with the challenge at 18 of 28 days.

ANNOUNCED

OpenAI runs some 2,000 GPT-Rosalind agents over an 8-million-cell mouse brain atlas in 32 minutes

OpenAI, PerturbAI and Scripps Research announced on Wednesday that they split a whole-brain CRISPR atlas — 8 million mouse brain cells and nearly 2,000 disease-linked genes — across roughly 2,000 GPT-Rosalind agents, each investigating one gene disruption, analyzing how cells responded and searching the scientific literature for explanations. The full analysis took 32 minutes; the team estimates the same one-perturbation-at-a-time work would take human experts several months to a year.

The combined output is a public reasoning atlas linking genes, cellular effects, publications and candidate mechanisms. The distributed agents cited more than 700 references, against 11 from a single agent given the full dataset, and blinded specialists preferred the multi-agent reports 75% of the time. Their novelty and scientific-value judgments also lined up with the human experts'.

Two leads stood out: unexpected differences between disrupting Tsc1 and Tsc2, genes relevant to tuberous sclerosis, in movement-related brain regions including the cerebellum; and a new hypothesis linking disruption of DCTN1 to altered cholesterol regulation in neurons, relevant to Perry syndrome, a Parkinson's-like neurodegenerative condition. Both are research leads needing lab testing, and the companion paper is a bioRxiv preprint that has not been peer reviewed.

ANNOUNCED

OpenAI will sell its frontier models and Codex through Google Cloud's Model Garden

OpenAI and Google Cloud said on Wednesday that OpenAI's frontier models and its Codex coding agent will be offered through Google Cloud's Model Garden, billed through Google Cloud Marketplace. GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are rolling out in private preview to select enterprise customers, with general availability planned in the coming weeks.

Customers will be able to draw down existing Google Cloud spend commitments and use Google Cloud's access controls and data governance tools, which the companies pitch at regulated industries building agentic workflows. OpenAI said Microsoft remains its primary cloud partner.

The deal puts OpenAI's models on all three big US clouds. OpenAI's exclusive distribution arrangement with Microsoft expired in April, and it has since expanded to Amazon Web Services and now Google Cloud — even though Google competes with OpenAI in core AI products and search.

ANNOUNCED

OpenAI details the LLM orchestrator behind ChatGPT ads, and says ads passed a $1B annualized run rate in under 200 days

OpenAI published a technical post on Wednesday explaining how ChatGPT's advertising system actually works — and how big it has already become. The company says ChatGPT Ads passed $1 billion in annualized revenue run rate in under 200 days and has drawn more than 100,000 advertisers, the same week it expanded ads to 26 more countries for a total of 89.

Ad selection is run by a custom, post-trained frontier model OpenAI calls the orchestrator. It handles the whole flow through tool calls — choosing when to run retrieval, ranking, relevance scoring and auction steps — in a separate service, so it cannot slow down or shape ChatGPT's answers. Because conversational intent does not arrive as a single search-style request, the system keeps a stateful session across turns and even across conversations, rewrites that context into retrieval queries and semantic embeddings, fetches candidates with approximate-nearest-neighbor search, then filters for budget, pacing, policy, targeting and advertiser deduplication.

A day earlier, OpenAI had emailed advertisers that creative enhancements — automatic rewrites of ad headlines, text and landing pages — will be switched on by default for new and existing campaigns unless advertisers opt out. Ads appear only on the Free and Go plans; Plus, Pro and Enterprise users see none.

ANNOUNCED

OpenAI expands ChatGPT ads to 26 more countries, bringing the total to 89

OpenAI said on Wednesday it is expanding advertising on ChatGPT to 26 more countries, bringing the total to 89, as it builds the ads business ahead of a potential IPO. The company said in August that ads had reached an annualized revenue run rate of $1 billion, and that more than 100,000 advertisers have run campaigns since testing began in the US earlier this year.

The expansion adds 14 more Latin American countries, including Argentina, Chile, Colombia and Peru, following launches in Mexico and Brazil. Ads are also launching across Europe, the Middle East, Africa and Asia, including Nigeria, South Africa, Serbia, Ukraine, Cambodia, Bangladesh and Pakistan.

Ads appear only on ChatGPT's Free and Go plans; Plus, Pro and Enterprise subscriptions remain ad-free. Advertisers in the new markets can buy through technology partners or self-service via Ads Manager, and in September OpenAI added AI-powered advertiser tools and began testing business-sponsored agents with select US advertisers.

INTERVIEW

Interview: Sam Altman tells Matthew Berman no one really knows if AI is conscious — and that AI needs a new kind of computer

Sam Altman sat down with Matthew Berman for a 45-minute interview published on Wednesday, ranging across AI-driven math and science discoveries, personal AI assistants, compute and OpenAI's new hardware. Asked whether today's models are conscious, Altman said he does not believe they are — 'or if they are, it's a very strange form of it' — before adding the line that anchors the conversation: 'I want to be clear, no one really knows.'

On hardware, Altman argued that phones and computers were designed for people, so AI agents are stuck interpreting visual interfaces instead of reaching information directly. He sees AI becoming an ambient assistant that helps beyond a chat window, which may require a different physical form factor — the bet behind OpenAI's device work.

The interview also covers recursive self-improvement, why AI has not killed coding jobs, ChatGPT being too agreeable, and OpenAI's bet on chips and cheaper compute. Altman closed on ambition with a caveat: 'AI will cure cancer' is not enough — the point, in his telling, is aiming at a new renaissance.

ANNOUNCED

Nvidia at GTC Berlin: Agent Knowledge Foundation and Kumo models for enterprise agents, plus an OpenShell Reference Architecture

At GTC Berlin on Wednesday, Nvidia expanded its open Agent Toolkit with the Agent Knowledge Foundation, a layer meant to ground enterprise agents in company data without moving the data, rebuilding workflows or changing governance controls. CEO Jensen Huang made data access the center of his keynote pitch: what turns agents into 'super agents,' he said, is access to a company's knowledge.

The foundation bundles CUDA-X libraries such as cuDF, cuGraph, cuML, cuOpt and cuVS with a new Auto-Ontology library that translates raw company data into business concepts — customers, products, orders — and maps how they relate. It also includes Nemotron Retriever and the new Kumo Structured open models (Kumo Tabular, Relational, Time Series and Forecast) for finding company information and making predictions from structured data.

Nvidia also unveiled the OpenShell Reference Architecture, a hardware-and-software design for containing and monitoring agents: models run on Vera Rubin systems, agent harnesses on Vera CPUs, and BlueField-4 chips run Nvidia's Sentry software to watch agent behavior and enforce policy in hardware. Anthropic, AWS, Google Cloud, Meta, Microsoft, OpenAI, Perplexity, SAP and SpaceX's AI unit are among those working with the architecture, and Nvidia also showed DSX POD, prebuilt AI factories it says can be installed in 14 days instead of six months.

ANNOUNCED

Microsoft will remove eight OneDrive Copilot actions from the Windows 11 File Explorer right-click menu

Microsoft is scaling back Copilot in Windows 11 File Explorer, announcing plans to remove eight OneDrive-related Copilot actions from the right-click menu: Summarize, Create FAQ, Compare Files, Ask a question, Explain image, Explain images, Compare images and Extract text. The company describes the change as simplifying Copilot usage and giving users more control over what the menu shows.

The removal begins in mid-November 2026 and proceeds gradually, with completion expected by early December. Microsoft is not retiring the underlying capabilities — users will still reach them through Ask Copilot, the Copilot icon in File Explorer Home, or the OneDrive activity center, at the cost of an extra step.

The move continues a pullback Microsoft's Windows chief Pavan Davuluri flagged in March, when he said the company was reducing unnecessary Copilot entry points, and follows growing user criticism of AI features being added across Windows whether they are wanted or not.

ANNOUNCED

Judge sets Meta's New Mexico privacy penalty at $25 billion, down from a $219.5B maximum — Meta says it will appeal

A New Mexico judge on Wednesday set Meta's civil penalty at $25 billion in the state's Cambridge Analytica-era privacy case, converting September's jury verdict into a dollar figure. The Santa Fe jury had found 43.9 million willful violations of the state Unfair Practices Act across 34 public statements about data privacy, hate speech and misinformation; at the $5,000 statutory maximum per violation, exposure ran to roughly $219.5 billion.

The state had asked the judge to apply the full rate and then trim the total to $35-40 billion, while Meta argued anything above about $3.45 billion would breach excessive-fines protections. The judge kept the top per-violation rate, calling the misstatements intentional, systematic and repeated over more than a decade, but used the court's discretion to cut the award to $25 billion. Interest accrues at 15% a year until it is paid, and under New Mexico law the penalties fund education spending.

Meta chief legal officer C.J. Mahoney said the company will appeal, calling the ruling a tax on speech that makes it all but impossible to do business in the state. Attorney General Raul Torrez, who sued in 2021, said the case is about whether companies can mislead consumers about their personal information without meaningful consequences. Barron's noted the timing: the penalty lands just as Meta asks users to trust its Muse agent with personal information from emails to bank statements.

ANNOUNCED

Meta puts Muse at the heart of its strategy: 10 million accounts in six weeks, capex raised to $135B-$145B and new non-voting shares to fund the AI push

Meta reported third-quarter results on Wednesday and put its Muse AI agent at the center of the story. More than 10 million people in the US and Canada have set up Muse accounts since the agent launched on September 8, about 25% of them use the app daily, and shopping is the most popular use case, CEO Mark Zuckerberg told analysts. He said he expects Muse to eventually serve billions of people and that Meta is deliberately not charging up front in order to build a Muse network first.

The quarter itself beat on the top line: revenue rose 26% to $64.6 billion, ahead of the roughly $63.4 billion expected, and Meta forecast fourth-quarter revenue of $73 billion to $76 billion. Net income was $9.86 billion and earnings per share $3.83, well below expectations, after a $9.89 billion charge tied to the August multi-state settlement; without it, EPS would have been $7.40. Free cash flow turned negative at -$2.12 billion, versus +$10.63 billion a year earlier.

To fund the build-out without diluting Zuckerberg's control, Meta will create a new class of non-voting Class C shares in January 2027, effectively a two-for-one split, usable for acquisitions and talent. It also raised the low end of its 2026 capital-spending forecast from $130 billion to $135 billion, keeping the top end at $145 billion. Shares rose more than 4% in extended trading.

ANNOUNCED

HHS says only Claude will do, planning a no-bid award to keep Anthropic for its staff

The US Department of Health and Human Services plans a no-competition, sole-source award to keep Anthropic's Claude for its staff, saying no other AI fits its needs, according to a notice of intent reported this week. The move comes months after the Trump administration's ban on Anthropic in other parts of the government.

FedScoop notes that the Food and Drug Administration, part of HHS, began moving its Elsa chatbot from Claude to Google's Gemini earlier this year. Anthropic's discounted deal for agencies through the GSA's OneGov programme ends on October 31.

The Pentagon is still fighting in court to keep Anthropic out, which makes HHS's stance an early test of how hard it is for agencies to switch AI suppliers once staff rely on one — especially as cheap introductory government deals expire.

INTERVIEW

Demis Hassabis moves his AGI estimate to two to three years: 'It's soon — I'm updating nearer and nearer'

Google DeepMind founder Demis Hassabis has pulled his AGI timeline closer again. In a Wall Street Journal profile published this week, he says he now expects a system matching the full cognitive range of the human brain within two to three years — down from the three-to-five-year estimate he was giving earlier this year. 'It's soon,' he said. 'I'm updating nearer and nearer.'

His test for the milestone is characteristically concrete: whether a model trained only on knowledge available before 1901 could reproduce Einstein's 1905 papers. Hassabis, a chemistry Nobel laureate and former chess prodigy, is taking on a new role after a recent leadership shuffle at Google DeepMind while continuing to run Isomorphic Labs alongside it.

His framing of the stakes is measured: AI's risks are profound but addressable without hype in either direction — 'not all disaster, and not all nirvana.' Winning, in his words, means helping humanity steward AGI into the world safely and securely for the benefit of everyone.

RELEASED

Google Maps adds Gemini to CarPlay in the US

Google Maps users on Apple CarPlay in the US can now tap a Gemini icon on the car display and ask questions in natural language while driving. It is the first time Google has made a voice assistant available inside its CarPlay app; Google Assistant was previously only available when navigating from the phone itself.

Drivers can ask what the weather will be at their destination on arrival, or ask multi-step questions such as where to stop for sightseeing and where to eat or get fuel along the way. Google says the feature is rolling out now in the US, with a global rollout planned.