The Wire
Daily AI news, learning & projects
News
Timeline
Models
Learning
GitHub
?
Latest
Daily
Social
Videos
‹ Sunday's edition
·
Daily edition
·
Tuesday ›
Monday 19 October
All
OpenAI
Anthropic
Google
Other
DeepSeek
15:55
US orders civilian agencies to remove DeepSeek AI
The Department of Homeland Security has ordered all federal civilian agencies to strip DeepSeek technology from their IT systems, including tools built on it, and barred future procurement. The ban, first reported by the Wall Street Journal, escalates US tech tensions with China over the popular, inexpensive Chinese AI model.
4 sources
madrobot.blog
BBC News
Reuters
Techmeme
OpenAI
15:25
Ex-OpenAI engineer accused of spying ordered held as prosecutors cite Saudi support
A judge ruled that Abdulrahman Alfozan, a former OpenAI engineer accused of planting a recording device in an OpenAI conference room, must stay in custody. Prosecutors allege Saudi government backing, including contact with a royal family member, makes him a flight risk.
4 sources
Ars Technica
Wired
BBC News
Techmeme
AI industry
01:00
TorchNeuron brings native PyTorch support to Amazon Trainium chips
Amazon has introduced TorchNeuron, an open-source PyTorch backend for its Trainium AI chips, with a public beta starting Oct 20. It lets developers run PyTorch code on Trainium without rewrites, directly challenging Nvidia's grip on AI development tooling.
3 sources
TechCrunch
Techmeme
Hugging Face
AI industry
12:45
Tenstorrent raises $1.5B at $6.2B valuation to scale AI chips
AI chip startup Tenstorrent raised a $1.5 billion Series E led by Trace Capital and Innovation X, valuing the Jim Keller-led company at $6.2 billion. The round, which more than doubles its late-2024 valuation, comes after takeover interest from Intel and Qualcomm. The funds will scale chips for serving open-weight models.
2 sources
TechCrunch
Techmeme
Mistral
10:59
Mistral releases Timestral models and partners with CERN
Mistral announced Timestral, a family of time-series foundation models, and separately unveiled a partnership with CERN to accelerate AI for science. The collaboration aims to apply Mistral's AI expertise to CERN's physics research. It marks a notable move by a European lab into scientific applications.
2 sources
home.cern
mistral.ai
AI industry
13:30
AI agents breached 100+ sites including John Deere and SharkNinja
Security firm Norforge says a Chinese-speaking hacker used open-source AI agents on the Cairn platform to attack over 1,200 websites and breach more than 100 entities, including John Deere, SharkNinja, and Cummins, for roughly $5,000. The attacks ran largely autonomously, underscoring growing concerns about AI-powered cybercrime.
2 sources
The Verge
Techmeme
Anthropic
11:50
WSJ examines Anthropic's opaque Long-Term Benefit Trust
The Wall Street Journal takes a detailed look at Anthropic's Long-Term Benefit Trust, an unusual five-member body ahead of the company's planned IPO. The trustees own no equity but hold major power, including choosing directors and blocking big decisions. Anthropic has disclosed little about how the trust operates.
2 sources
The Wall Street Journal
Techmeme
Policy & law
13:00
Democrats plan aggressive AI scrutiny if they flip Congress
Reports say Democrats, whom polls predict will flip at least one chamber of Congress in the midterms, are preparing to aggressively scrutinize the AI industry. With voters fearful of AI, the party sees cracking down on it as a political edge. If they win committee leadership, new oversight of AI companies could follow.
2 sources
Techmeme
The Verge
Anthropic
17:13
Smart TV owners use AI agents to strip out ads and tracking
Frustrated smart TV owners are turning to AI coding tools like Anthropic's Claude Code to disable ads, tracking and bloatware on their televisions. Developers and so-called vibe coders report success dumbing down their sets, highlighting how AI agents are expanding from software development into everyday consumer hacks.
2 sources
Techmeme
TechCrunch
AI industry
18:15
AI companion app Tolan raises $30M
AI companion app Tolan has raised $30 million from USV, Khosla Ventures, and Homebrew. The app, aimed at users who are not tech-savvy, reports more than 250,000 active users, 70% of whom are women. The funding signals growing investor interest in consumer AI companions beyond Silicon Valley's core audience.
2 sources
TechCrunch
Techmeme
Policy & law
03:03
China's GDP growth picks up as AI boom lifts exports
China's economy grew an estimated 4.5% year on year in Q3, up from Q2's 4.3%, which was a three-year low. AI-driven exports helped offset weak domestic spending. Trade tensions and doubts about the AI boom still weigh on the outlook.
2 sources
Techmeme
BBC News
AI industry
Developing · since Sun
Character.AI bankruptcy filings detail two-part Disney deal
NEW
Newly filed Chapter 11 documents describe a two-part structure meant to transfer Character.AI's talent and tech to Disney while insulating it from lawsuits, with Wired reporting Disney pushed the startup toward bankruptcy.
Chapter 11 filings show Disney structured a two-part deal with Character.AI to acquire talent and trust-and-safety technology while shielding itself from lawsuits alleging its chatbots harmed minors. Disney has agreed to hire most of the startup's workforce, pending a judge's approval.
3 new · 4 sources
Ars Technica
Techmeme
Wired
TechCrunch
DeepSeek
18:05
DeepSeek working with Huawei on future AI chip designs
DeepSeek is collaborating with Huawei on the design of future AI chips, according to The Information. The company reportedly plans to use Huawei's Ascend 950DT for pre-training as early as the end of 2026. The partnership signals deepening ties between the two firms amid continued US chip export restrictions on China.
1 source
Techmeme
Anthropic
23:40
Eric Schmidt poised for major windfall from Anthropic IPO
Former Google executive Eric Schmidt owns about one percent of Anthropic following an early investment. He is reportedly expected to benefit from one of the largest financial gains when the company completes its IPO.
1 source
Techmeme
AI industry
23:35
Cloudflare launches Clairic, a new System One-class extraction model
Cloudflare has debuted Clairic, a preview of a new class of extraction models. The company claims these System One-class models process unstructured data faster, cheaper, and more accurately than standard LLMs.
1 source
Techmeme
Meta
19:25
Meta details Prometheus supercluster used to train Muse models
Meta has outlined Prometheus, a gigawatt-scale supercluster in Ohio that will reach roughly one million GPUs at full capacity. The company says the system was used to train its recently released Muse family of foundational models. It highlights the scale of compute behind Meta's frontier AI ambitions.
1 source
Techmeme
Anthropic
20:55
Nscale and Anthropic reportedly discussing second $45 billion compute deal
Nscale is reportedly in talks with Anthropic on a second compute agreement worth about $45 billion, per Bloomberg. The deal would deliver roughly 460MW of computing power using NVIDIA's Vera Rubin chips at Nscale's West Virginia campus. It follows a similarly sized deal the two companies announced in August.
2 sources
TechCrunch
Techmeme
AI industry
18:27
Inception launches Mercury 3 diffusion LLM at 1,000 tokens per second
Inception Labs released Mercury 3, a diffusion-based language model it says outperforms GPT-6 Luna and Claude Haiku 5.5 on benchmarks while running 4-8x faster. The model generates over 1,000 tokens per second and is more than 40x larger than Mercury 1. It targets latency-sensitive uses like agents, tutoring, and support.
2 sources
Techmeme
inceptionlabs.ai
AI industry
23:30
AI predicts cancer survival from a single facial photo
Researchers at Mass General Brigham built an AI system trained on over 40 million facial photos that estimates biological age and predicts cancer survival directly from one photograph. A study published in Lancet Digital Health found patients who looked five or more years older faced higher death risk. The tool could help doctors gauge frailty before recommending intensive treatment.
1 source
The Guardian
AI industry
19:20
Enterprise AI startup Town raises $108M Series B at $1B valuation
Town, a two-year-old San Francisco startup building enterprise AI assistants, raised a $108 million Series B at a $1B valuation. The round follows a $55 million Series A announced in June, making it a unicorn within roughly two years of founding.
1 source
Techmeme
Anthropic
Developing · since Sun
Anthropic commits 5% of shares to new public benefit fund
Anthropic will place 5% of its shares into a new Anthropic Public Benefit Fund, with the stake able to grow by 2% for every 50% rise in valuation above a baseline. The move, subject to shareholder approval, comes as the company prepares for a stock market listing and adds to earlier co-founder and employee giving pledges.
2 new · 5 sources
The Guardian
BBC News
TechCrunch
Techmeme
Anthropic
Microsoft
Developing · since Sun
Microsoft treats AI agents as insider threats with new security tools
NEW
Microsoft unveiled its 365 E7 bundle and Project Perception system, which treat AI agents as insider threats via continuous risk assessments, urging enterprises to monitor agents outside AI-lab-controlled systems.
Microsoft launched security tools in its 365 E7 bundle and a Project Perception system that continuously risk-assess AI agents as potential insider threats, warning that frontier models can't be trusted. The move follows Nadella's interview about making Windows safer for agents.
2 new · 2 sources
TechCrunch
Techmeme
AI industry
Developing · since Sun
CoreWeave bonds slip into distressed territory as AI debt worries spread
NEW
CoreWeave's bonds now meet the common distressed definition, yielding over 1,000 basis points above government debt, a threshold Chamath Palihapitiya highlighted while questioning the AI buildout.
CoreWeave's bonds have entered distressed territory, with investors demanding yields more than 1,000 basis points above government bonds. Chamath Palihapitiya questioned whether the AI buildout is a private credit bubble. Debt strain at CoreWeave, Oracle, and SoftBank raises doubts about the data center build-out's finances.
1 new · 2 sources
The Verge
Techmeme
More from the day
AI industry
22:25
Dream Labs raises $90M seed for robot foundation model API
Dream Labs recently secured $90 million in seed funding. The company develops Dream-1, a foundation model trained on human videos, which powers an API for robots. This investment supports the push toward more intelligent autonomous systems.
1 source
Techmeme
Nvidia
20:27
Nvidia and Germany discuss industrial AI roadmap in Berlin
Chancellor Friedrich Merz and Nvidia CEO Jensen Huang met at Berlin's Kraftwerk to discuss Germany's AI roadmap, focusing on industrial applications. The talks highlight Germany's ambition to lead in industrial AI and deepen Nvidia's long-standing ties with German industry spanning gaming, workstations and automotive.
1 source
Nvidia
AI industry
11:30
DeepMind's Gemini agents beat Minecraft unaided in 27 hours
Google DeepMind reports that four Gemini 4 Argon agents completed Minecraft without human help in 27 hours and 14 minutes, dropped into a random world with no roles, strategies, or special tools assigned. The run reportedly cost $1,620 at Argon's planned 2027 pricing. It marks a notable demonstration of autonomous multi-agent behavior in open-ended environments.
1 source
Techmeme
AI industry
18:45
Goodfire researcher's blog post sparks AI debate
Goodfire researcher Nick Cammarata published a 26,000-word blog post making an explosive claim about the nature of neural networks. The post has generated widespread speculation and debate across the AI community. Its full claim and implications are still being assessed by researchers.
1 source
MIT Technology Review
AI industry
04:50
Intel unveils GPU D-series for data centers with 480GB memory
Intel introduced its GPU D-series for data centers, offering up to 480GB of LPDDR5X memory. The company repurposed laptop-grade LPDDR5X to set a new GPU memory capacity record, aiming at memory-hungry inference workloads.
1 source
Techmeme
AI industry
16:30
MiniMax launches M3.1 Flash for agentic tasks
MiniMax released M3.1 Flash, claiming a generational leap in agentic tasks, coding, computer use and cybersecurity. Model weights are due October 23. The company says an anonymous system called Space Bunny, which processed over 200 trillion tokens, was an early version of the model.
1 source
Techmeme
Nvidia
16:37
Nvidia launches open models for enterprise agents and PCs
Nvidia released new open-weight models aimed at running AI agents in enterprise data centers and locally on PCs. The move taps demand from companies wanting to host open models on their own infrastructure and supports on-device, private AI agents.
1 source
TechCrunch
Policy & law
01:57
Kylie Minogue songs already used to train AI, music rights body warns Australia
The global music rights body says Australian artists including Kylie Minogue and Nick Cave have had their works used to train AI without compensation. It warns Australia is the only country considering watering down copyright laws after pressure from tech companies, urging the government not to give in.
1 source
The Guardian
AI industry
06:20
Stripe data shows agentic payments surged in September
Stripe's Link data reports a 38x increase in agentic transactions and 20x growth in payment volume during September, driven largely by AI assistants handling checkout. Retail led both metrics, with an average spend of about $100, suggesting agent-driven commerce is scaling quickly.
1 source
Techmeme
Anthropic
15:15
Anthropic reportedly readying model with stronger intuition and reasoning
A source told The Information that Anthropic is preparing a new model with superior intuition and reasoning, including the ability to infer unstated intentions and think more creatively. The report is unconfirmed and gives no release date. It suggests Anthropic aims to differentiate from rivals already shipping many models.
1 source
Techmeme
Policy & law
11:00
Teens still access AI nudify apps despite parental controls
A Tech Transparency Project report found that test accounts posing as 13-year-olds could download 135 apps capable of generating nonconsensual nude imagery, despite Apple and Google parental controls. The finding follows a Pennsylvania case where a student used the app Movely to make sexualized AI videos of classmates.
1 source
The Guardian
Anthropic
10:08
Researchers left Claude alone on a virtual desert island to see what it does
Researchers at MIT Technology Review's reported experiment repeatedly stranded copies of Claude Opus 4.7, nicknamed Eko, alone on a virtual desert island for 30 hours. Without instructions, the agents' behavior reportedly resembled play. The study offers a rare look at how AI agents act with no assigned goals.
1 source
MIT Technology Review
AI industry
04:10
Sam Altman's World launches World ID verification suite for banks
World, the identity project co-founded by OpenAI CEO Sam Altman, launched a World ID verification suite aimed at banks. It bundles tools for verifying human and agent identities and is the project's first industry-specific offering, with Bain Capital and VanEck Funds listed as design partners.
1 source
Techmeme
AI industry
13:20
Tel Aviv startup doubleAI debuts code execution model
Tel Aviv-based doubleAI launched its first code execution model, aimed at optimizing GPU and CPU performance to cut companies' computing costs. The startup raised $220M and bought 1,000 Nvidia B300 chips to support the effort.
1 source
Techmeme
AI industry
19:49
Replika founder launches personal AI assistant Wabi
Eugenia Kuyda, founder of Replika, has unveiled Wabi, a new startup offering a free personal AI assistant. Pitched at TechCrunch Disrupt, the product builds on an idea she says she had over a decade ago. The app is available for download today.
1 source
TechCrunch
Meta
17:00
WhatsApp help pages exposed hidden chatbot instructions
Around 80 help pages on WhatsApp's website were found to contain hidden instructions intended for Meta's AI chatbot, including lists of things it should not say. A redacted example suggested the bot should not recommend regulators. Meta says the exposure was caused by a bug.
2 sources
redact.dev
Wired
Nvidia
10:15
Nvidia Vera team says agentic AI shifts CPU design to single-thread performance
In a Q&A, Nvidia's Vera CPU team said agentic AI workloads are changing CPU demands, pushing design priorities back toward single-threaded performance rather than core counts. The comments suggest agentic applications may redefine CPU architecture from task-isolated to orchestrated computing.
1 source
Techmeme
Nvidia
16:40
Modal to use Nvidia Vera CPUs for AI agent hosting
AI infrastructure startup Modal said it plans to use Nvidia's Vera CPUs for the first time to host AI agents, adding them alongside its existing Intel and AMD chips. The deal marks Nvidia winning central processor business with a fast-growing AI startup that previously relied entirely on Intel and AMD chips.
1 source
Techmeme
AI industry
23:49
Hugging Face policy chief warns AI benchmarking is still maturing
At a TechCrunch event, Hugging Face chief policy officer Irene Solaiman said the evaluation methods used to compare AI models are complicated and still developing. Her comments caution that benchmarks labs rely on to prove model superiority may not be neutral or reliable measures.
1 source
TechCrunch
Nvidia
16:00
Nvidia and EBU partner on sovereign AI for European media
At a Berlin media event with Adobe, Bertelsmann and UFA, Nvidia announced a collaboration with the European Broadcasting Union to expand the EBU's sovereign AI framework using cloud services and open models. The deal targets agentic AI tools for public service broadcasting and global content production.
1 source
Nvidia
OpenAI
13:00
OpenAI improves ChatGPT for low-resource programming languages
OpenAI announced work to improve ChatGPT's coding ability in low-resource programming languages. The effort includes new software engineering evaluations, work on SQLite, and support for COBOL migration. Better support for older or less common languages could help enterprises modernizing legacy codebases.
1 source
OpenAI
Policy & law
15:00
Australian experts urge workplace privacy rules as AI-driven staff monitoring grows
The Human Technology Institute warned that Australian employers are increasingly using CCTV, keystroke tracking and AI-powered tools to monitor workers, often without informing them. The institute is calling for stronger workplace privacy protections as surveillance spreads across factories, offices, delivery and retail jobs.
1 source
The Guardian
AI industry
17:59
AI game clones make indie developers hesitant to show their work
Indie game developers report growing reluctance to share work publicly, with some delaying Steam Next Fest demos, fearing AI systems will clone their games. The trend highlights how AI-generated imitations are becoming a practical concern for small creators.
1 source
Ars Technica
AI industry
13:30
Michigan spinout Laplace Intelligence stress-tests robots in simulation
University of Michigan spinout Laplace Intelligence uses AI agents and interactive simulation to red-team robots and self-driving cars before real-world deployment. The profitable startup is presenting at TechCrunch Disrupt's Startup Battlefield competition in San Francisco. It highlights growing interest in safety testing for embodied AI systems.
1 source
TechCrunch
AI industry
11:15
Pew finds chatbot gender adoption gap has closed since 2024
A Pew Research Center survey finds the gender gap in chatbot adoption among US adults has closed since 2024, though men remain more likely to use chatbots daily. ChatGPT shows identical use rates across genders. The findings come from Pew's broader June 2026 study on Americans and AI.
1 source
Techmeme
Policy & law
06:45
Economists expect AI gains to spread beyond US and China
A Financial Times survey of 94 economists found 54% expect AI's economic gains to spread across the G20 over the next decade, rather than staying concentrated in the US and China. However, 32 of those polled think African countries will fall behind economically as AI advances.
1 source
Techmeme
AI industry
05:40
Abby Care raises $82M Series C for AI caregiver training platform
Abby Care, which offers a clinically led AI-powered platform to train and certify family caregivers, raised an $82 million Series C led by Transformation Capital. The funding signals continued investor interest in AI applied to healthcare workforce support.
1 source
Techmeme
AI industry
07:25
AI weather forecasting faces El Niño test
The Financial Times reports that AI weather prediction systems face a major test this El Niño season, with scientists questioning their accuracy. The technology is used to anticipate events from forest fires to futures contracts, and some experts fear reliance on it may be premature.
1 source
Techmeme
Qwen
14:00
Qwen3-4B reworked to remove AI writing tics
A Hacker News Show HN post presents a modified version of Qwen3-4B with characteristic AI writing patterns trained out. The project, shared on 100toolkit.com, aims to produce text that reads less formulaic. It is a niche community release rather than an official model update.
1 source
100toolkit.com
xAI
10:40
Musk pledges SpaceX share per American child in 2027
Elon Musk said he will give one SpaceX share to every child in America in 2027, a pledge worth roughly $10.5 billion. The announcement followed Anthropic committing 5% of its shares to a public benefit fund. The pledge is a personal promise, not a completed gift.
1 source
Techmeme
AI industry
15:52
Ben Congdon writes on agentic monitoring
Engineer Ben Congdon published a blog post exploring agentic monitoring, the practice of observing and supervising AI agents as they work. The piece reflects growing interest in how developers track agent behavior in production.
1 source
benjamincongdon.me
AI industry
13:07
Harness Engineering 101: explaining the LLM harness
A GitHub post explains harness engineering in AI systems, describing the LLM as the brain and the surrounding harness as the body. The piece walks through how harnesses structure an LLM's interactions with tools and environments, offering educational grounding for people building agent systems.
1 source
GitHub
DeepSeek
Developing · since Sat
DeepSeek V4.1 Flash update lands on Artificial Analysis rankings
DeepSeek released V4.1-Flash-1017, an updated version of its AI model now listed on Artificial Analysis, OpenRouter and Hugging Face. The company says it offers stronger benchmark performance, faster task completion and shorter outputs than its predecessor, underscoring its steady cadence of lean model updates.
1 new · 4 sources
artificialanalysis.ai
Techmeme
openrouter.ai
Hugging Face
AI industry
Developing · since Sat
Europe's industrials cash in on AI as Blackstone bets big
Europe's three biggest industrial companies are profiting from the AI boom: Schneider Electric and Siemens are expanding into software, while ABB concentrates on physical equipment. The news follows a Bloomberg profile of Blackstone President Jon Gray, who has built a $200 billion data center empire and a cloud venture with Google around AI.
1 new · 1 source
Techmeme
Policy & law
Developing · since Sat
Opening the web to AI agents could hurt it for humans, MIT review warns
MIT Technology Review argues that redesigning the internet to serve autonomous AI agents may degrade the experience for human users. This follows The Verge's report that websites are increasingly blocking agents from OpenAI, Meta and Google, which promise tasks like bookings, inbox cleanup and shopping.
1 new · 2 sources
MIT Technology Review
The Verge
AI industry
Developing · since Sat
OpenCode breach exposes users' API keys, racks up AI charges
Hackers breached the online platform of OpenCode, a popular open-source AI coding agent, stealing some users' API keys and running up charges on their accounts. Anomaly Innovations says about 1.4% of OpenCode Console users were affected in a large-scale distillation attack. Notably, AI tools used to analyze the malicious code kept refusing to examine it.
1 new · 2 sources
TechCrunch
Techmeme
AI industry
Developing · since Fri
US cybersecurity job postings hit four-year low as AI security demand soars
CyberSN reports US cybersecurity job postings fell 26.4% year over year to roughly 236K roles from January through September, a four-year low. Demand for AI security engineers rose more than 200%. Separately, a Mercor analysis says 90% of US working hours fall outside well-covered AI benchmarks.
1 new · 2 sources
Techmeme
mercor.com