It sits on your phone like an app and has never once asked for an update.
iPhone
Google calls Gemini 4 Carbon its strongest model, with state-of-the-art cybersecurity and long horizon skills. It ranks first on the Artificial Analysis Intelligence Index at 59 and leads SEC-bench Pro at 88.2% for finding vulnerabilities. Its predecessor, Gemini 4 Argon, launched two weeks ago.
Read more at blog.google →
OpenAI has released GPT-6.1 Terra, a fast model for visual reasoning, computer use and writing, at $1 per million input tokens and $5 per million output tokens. It scores 85.5% on MMMU-Pro but 50.3% on Terminal-Bench 4.0, where Claude Sonnet 5.5 reaches 70.6%.
Read more at MarkTechPost →
Anthropic says an unreleased, unnamed model found flaws and combined them into working exploits against Chrome, Safari, Android, Windows and Linux. In one test it linked four Chrome bugs to escape the browser sandbox. The vendors have released fixes for all of them, Anthropic says.
Read more at notebookcheck.net →
Microsoft earned $10 billion from its commercial arrangements with OpenAI in the quarter to 30 September, nearly 11% of its $92.3 billion revenue, a filing shows. It also holds $7.2 billion in receivables from OpenAI. The two companies increasingly compete for the same customers, and one of them sends invoices.
Read more at TechCrunch →
Anthropic says it has worked for several months with six organisations on new medicines, including Bristol Myers Squibb, Enveda and Generate Biomedicines. At Generate, researchers used Claude Code to redesign part of the genetic instructions in CAR T-cell therapy. Anthropic has also launched its own drug discovery programmes.
Read more at qazinform.com →Mistral AI has opened a public preview of its Responses API, which returns a model's answer, reasoning and tool calls as separate items for developers building agents. It supports all chat models hosted by Mistral, including Large 4. The preview follows the 20 October launch of its Voxtral 3 suite.
Read more at tokenpost.com →Mistral AI's 4 billion parameter Voxtral Mini Transcribe 3 ranked second of nine models on Voice Arena's leaderboard for speech recognition and working out who said what. The larger Small Transcribe 3 came fourth. Size usually wins these arguments, and this time it was interrupted.
Read more at Crypto Briefing →Nvidia has released Nemotron 3.5 Transcribe, an open-weight model that handles multilingual transcription, speaker attribution, translation and audio question answering, on Hugging Face and ModelScope. It comes with Granary v2, an open data pipeline, and fine-tuning recipes so teams can adapt it to jargon such as medical dictation.
Read more at huggingface.co →MiniMax has published MCode Stack under the MIT license, the workflows and agent skills its team uses to build MiniMax Code. It includes 17 skills and five checklists in plain Markdown that run in MiniMax Code or Claude Code. No benchmarks or eval results were published.
Read more at MarkTechPost →Google will bring the Gemini agent it launched for enterprise this month to consumers later this year, starting with the Gemini app, CEO Sundar Pichai said on the company's third-quarter earnings call. The Verge wonders how it will differ from Gemini Spark, the agent Google showed in May.
Read more at The Verge →Instagram head Adam Mosseri said on the Colin and Samir podcast that there is more anxiety, frustration and anger about AI, especially in creative industries, and that it will keep getting worse. He said the worries come down to 'a fear of being left behind in some way.'
Read more at The Verge →
Chrome is getting an advanced autofill that suggests details pulled from Gmail, such as flight confirmation numbers, membership IDs, tracking information and credit card numbers. It extends Google's Personal Intelligence feature, which already connects Google apps to Gemini. The form now knows where you are flying.
Read more at TechCrunch →
Google has launched a Connected Apps Platform so businesses can link their apps to its AI experiences in Search and Gemini. Users can already connect Canva to design a flyer or YouTube Music to build a playlist. The idea only works if there are plenty of apps to connect.
Read more at TechCrunch →
Microsoft is ending Copilot support on all TVs, according to a notification on an LG television spotted by Windows Central. LG and Samsung announced Copilot on their TVs at CES 2025. The news came from neither company nor Microsoft, but from the TV itself.
Read more at The Verge →GitHub is adding more than 4 million CPU cores and 160 PB of storage after 7.38 billion commits in September, five times the previous month. It is also rebuilding its Git architecture. 'Capacity alone won't get us there,' Microsoft's Jay Parikh wrote. The agents, it seems, commit.
Read more at analyticsindiamag.com →
Anthropic has added smart reports to Claude Enterprise, showing managers what a team produced over a date range, what it cost and where its workflows broke down. It also launched in-app surveys and pooled team budgets. The company says it will not use the data for model training by default.
Read more at TechCrunch →Google is bringing ads in its AI Search experiences to international markets. Alphabet reported $127.6 billion in third-quarter revenue, up 25%, but Search revenue growth slowed to 12% from 17% the quarter before, and the shares fell 8%. The AI overviews will now pay their way.
Read more at TechCrunch →Meta has launched Waddle World, a 'tamagotchi-style' app with virtual creatures that chat with users and draw doodles of their conversations. Its store listing calls it 'a daily journal disguised as a pet' and promises no streaks to protect and no scores to chase.
Read more at The Verge →Akai has given modders early access to the MPC Sample Key, a 37-key keyboard with open-source firmware that is not yet formally announced. Short Instagram Reels show vibe-coded mods adding synths, including Native Instruments Massive. The mods arrived before the announcement.
Read more at The Verge →Zendesk has launched Employee Service AI Agents and Advanced Action Flows to handle internal requests such as onboarding, access approvals, laptop management and badge access. In five weeks of early access, 300 customers ran 6 million workflows, which Zendesk estimates saved 500,000 hours at five minutes each.
Read more at itbrief.com.au →ClearRoute has released runkora in alpha, an open-source, cloud-agnostic control plane to manage how AI agents operate in software delivery and DevOps. It cites Gartner research predicting the average Fortune 500 company will run more than 150,000 agents by 2028, while 13% of organisations think their governance is ready.
Read more at itbrief.com.au →Droma, formerly Grokstream, has added Problem Manager and SRE agents alongside its existing incident operator, to handle recurring incident patterns and reliability planning. OpenTools notes that the roles describe intended workflows, with no proven prevention of future failures yet. Customers keep the same team and platform after the rebrand.
Read more at OpenTools →Kindo has launched Deep Hat Sift, a model that decides which security events warrant an agent run before a larger model starts work. It learns from the verdicts of the agent it serves. Every event still reaches the agent until a person approves a filtering threshold.
Read more at unite.ai →Insurify has launched an AI Agent Gateway that lets agents such as OpenAI's Dots and Meta's Muse pull real-time quotes from its marketplace, with rules for coverage information and personal data. It builds on Insurify's ChatGPT plugin for insurance shopping. Your agent can now compare premiums while you do not.
Read more at PR Newswire →Thomas Dohmke has launched Entire, a developer platform for humans and AI coding agents, with repository hosting in India, the US, the EU and Australia. Its Trails feature replaces pull request review with automated evaluations, in which agents check each code change for quality and security.
Read more at The Economic Times →
L&T's Vyoma.AI has broken ground on a 52 MW data centre at its Kancheepuram campus, part of a planned gigawatt-scale AI Factory. It expands an existing 30 MW site. In August, L&T won an order to deploy 10,000 Nvidia B300 GPUs there for US-based Together AI.
Read more at The Hindu →
Cognizant is in talks to take a small stake in Sarvam AI as part of a roughly $300 million round valuing the startup at $3.5 billion to $4 billion, The Economic Times reported. Sarvam's valuation was $1.5 billion in June, so it would more than double in about four months.
Read more at wocult.com →
At Cloud Day India 2026, attended by more than 600 founders and executives, Google Cloud showcased AI uses for smaller businesses. Apparel brand Siyaram's uses AI to forecast demand across over 40,000 products. Google named fragmented data, legacy systems and uncertain returns as the barriers.
Read more at nationalheraldindia.com →
India's 34 companies building AI for Indian languages have raised $687 million in cumulative equity funding, a Tracxn report says, but capital is concentrated among a handful of players and there has been no late-stage round yet. Two Sarvam rounds drove most of the 2026 inflow.
Read more at Fortune India →
Six months after the Centre proposed rules for AI-generated content on social media, policymakers are discussing whether to define AI intermediaries separately from social media platforms, Mint reports. Lawyers and Big Tech firms have lobbied MeitY. IT minister Ashwini Vaishnaw has promised a consultation paper for a dedicated AI law.
Read more at Mint →Karnataka IT minister Priyank Kharge told IT professionals in Bengaluru that the pace of AI adoption makes continuous learning essential and urged them to keep upgrading their skills, pointing to initiatives like Nipuna Karnataka. He spoke at the opening of RSM's new office at Manyata Tech Park.
Read more at YourStory →India's insurers collected Rs 11.93 lakh crore in premiums in FY2024 to 25, with penetration at 3.7% of GDP. On Mint's All About AI series, executives including Salesforce South Asia's Arun Kumar Parameswaran discussed how agentic AI could handle insurance workflows under defined instructions and controls.
Read more at Mint →LTM has launched GCC Origin, a standalone unit to help enterprises set up and run global capability centres in India, led by Shailendra Singh. 'Enterprises no longer set them up to cut costs,' CEO Venu Lambu said. An Embark report projects 1 to 1.4 million net GCC jobs by 2031.
Read more at analyticsindiamag.com →IIIT Hyderabad and the National Payments Corporation of India have signed an MoU to collaborate on cybersecurity, digital forensics and trusted infrastructure for digital payments. The institute's Cyber Manthan Center will lead the work, with joint research and specialised training to build a skilled security workforce.
Read more at analyticsindiamag.com →Goa's tourism department is deploying AI systems to improve its visibility across AI-assisted search, conversational platforms and social media, in a strategy designed by its consultant KPMG. It wants to promote adventure, wellness, heritage and other segments beyond beaches. The chatbot will now be asked to recommend Goa.
Read more at The Times of India →Indian AI startups seeking American investors and customers face high hiring costs, visa hurdles and difficulty winning first clients, Business Standard reports. Antler India and All In Capital run programmes to help. Some startups keep engineering in India and build sales teams in the US.
Read more at americanbazaaronline.com →KPIT Technologies reported 0.2% constant currency revenue growth in the second quarter, while its EBITDA margin rose to 17.8%. It says AI-led productivity gains will support margin expansion but did not quantify AI's contribution this quarter. Revenue from its two largest accounts fell about 13%.
Read more at analyticsindiamag.com →
Upwind Security raised a $400 million Series C led by Bessemer Venture Partners at a $3.8 billion valuation, up from $1.5 billion in January. It says revenue quadrupled this year without disclosing figures. The money will fund AI sensors for endpoints and scanners for AI agents and their skills.
Read more at decrypt.co →
OpenAI published a customer story casting Legora as the builder of an 'agentic operating system' for legal work, a week after the Stockholm startup raised $550 million at an $8.8 billion valuation. Legora reports more than $200 million in annual recurring revenue and over 1,000 staff.
Read more at BigGo Finance →Fragment has launched with $56 million across seed and Series A rounds led by General Catalyst, A* and 8VC, for AI agents that investigate and resolve purchasing and payment exceptions. It starts with direct procurement at large physical goods businesses. Permissions and review stay with the customer.
Read more at OpenTools →
Casap, whose AI helps banks and card companies handle disputes, raised a $56 million Series B led by Base10 Partners, bringing total funding to $90 million. Cardholders now account for nearly a third of fraudulent transactions, helped by AI tools that make disputing charges easier.
Read more at Fortune →
London's Circuit Mind has raised an $18.5 million Series A led by Norwest, bringing total funding to $25 million with an earlier seed. 'Software teams are shipping 100 times faster with AI, but electronics design hasn't changed in 40 years,' CEO Tomide Adesanmi said.
Read more at The Next Web →London robotics AI startup Machenta has raised $40 million in seed funding co-led by Creandum and Union Square Ventures. It wants robots to learn new physical tasks in simulated environments with fewer demonstrations, instead of large real-world datasets. The money goes to research, hiring and compute.
Read more at newsaffinity.com →Easol, a platform for festivals, venues and other experience businesses, raised $30 million in a Series A extension, $20 million of it equity from investors including Y Combinator and Tiger Global plus $10 million in debt. It also launched Teammates, AI agents for sales, support and operations.
Read more at citybiz.co →Minneapolis-based phData has new funding from Gryphon Investors with participation from Anthropic, to expand its enterprise AI consulting and help companies deploy Claude. Terms were not disclosed. phData is a member of the Claude Partner Network. Anthropic is now funding the people who sell Anthropic.
Read more at citybiz.co →Archera, founded by brothers Aran and Nikhil Khanna in 2019 to cut cloud bills, raised a $15 million Series C led by Pivot Investment Partners. It says it has saved more than 2,200 businesses $700 million. AI computing demand, it says, is almost impossible to forecast.
Read more at The Times of India →Dalet Labs, founded by former Runway researcher Omer Bar-Tal, raised a $7 million pre-seed round led by Lightspeed and launched Scene, a US shopping app with a feed of AI-generated images. The pitch is that shoppers will buy things that were never photographed.
Read more at businessinsider.com →Inkle, an AI-powered accounting firm in San Francisco with technology operations in Bengaluru, has bought the bookkeeping division of Finally in an all-cash deal. About 100 employees join, taking Inkle to around 160 people. Terms were not disclosed. It is Inkle's first major acquisition.
Read more at americanbazaaronline.com →Enterprise AI startup QwikBuild raised $2.3 million in pre-seed funding from Globalway, Blume Ventures, Marsshot and DeVC, plus angels including Groww's Lalit Keshre. It uses AI agents to build enterprise software in days and plans a joint venture with Japan's Globalway to enter that market.
Read more at bwdisrupt.com →Bengaluru's CopperPilot, founded in October 2025, raised $390,000 from Neon Fund and US angels. Its agent turns plain-English descriptions into manufacturing-ready circuit board designs. It says active users have grown 25-fold since August, with more than 6,000 design artefacts run on the platform.
Read more at analyticsindiamag.com →'I look at expense management as a dying category,' said Andrew Jamison, CEO of Extend, an expense management company. He argues AI will fold expenses, bills and reimbursements into one automated process that runs without anyone logging in. Extend plans to double down on AI in 2026.
Read more at sg.finance.yahoo.com →
Researchers from Polymathic AI, Princeton, NYU and others released FlashMap, a speculative decoding drafter that uses a continuous flow map to draft a 16-token block in one pass. They report up to 6.14x lossless speedup on Qwen3-8B on GSM8K. Drafter weights are on Hugging Face.
Read more at MarkTechPost →
In a recent Anthropic study, 86% of participants followed Claude's advice, but only a small fraction acknowledged that the agent had played a role in their decision, Gizmodo reports. Most people describe their choices as carefully weighed. Most people, it appears, had help.
Read more at Gizmodo →
Cognitive scientists ran an experiment comparing people who brainstormed with ChatGPT and those who worked with other humans. The ChatGPT groups generated fewer unique ideas. Generative AI is often sold on speed, and speedy and wordy, the researchers write, does not mean creative.
Read more at themercury.com →
Researchers affiliated with Harvard and Meta tested AI voice phishing calls on 4,100 US adults and found about one in six could comply. People were most vulnerable to fake family emergencies, and how persuasive the script was mattered more than how realistic the voice sounded.
Read more at digitaltoday.co.kr →
MIT professor Paris Smaragdis is building machines that can hear and understand sound better than humans, for uses such as pinpointing a traffic accident by sound or spotting a failing drill. 'I do the same thing the computer vision folks do, but instead I do it for hearing,' he says.
Read more at MIT News →A WIRED writer let Codex and other AI agents manage a Premier League fantasy team to test their predictions. The agents tried to bench Erling Haaland at first but learned. They are not better than humans yet, the writer concludes. Neither, to be fair, are most managers.
Read more at Wired →A review in EPJ Research Infrastructures sets out how the European Spallation Source in Sweden will use AI and machine learning during commissioning and operation of its roughly 600 metre proton accelerator. It describes pilot projects already being tested offline and on the accelerator itself.
Read more at Bioengineer.org →A retrospective study at Qilu Hospital tested four DeepSeek configurations on 292 anonymised records of patients due for hip or knee replacement. A reasoning version paired with a curated guideline knowledge base outperformed simpler versions and unassisted junior doctors' preoperative decisions, the study in Arthroplasty reports.
Read more at Bioengineer.org →A review in Chinese Medicine traces how digitised classical texts, medical records and diagnostic data are being turned into training data for traditional Chinese medicine language models. The authors set out what still stands between these systems and routine clinical use.
Read more at Bioengineer.org →NJIT researchers are exploring a neuro-adaptive learning companion that reads signals such as brain waves from wrist devices or headsets to spot when students may fall behind or are ready for more. The AI would then choose another way to teach the topic, possibly through extended reality.
Read more at news.njit.edu →A survey of 684 nursing students at two universities in Hubei, China, found their readiness to work with AI suboptimal, with a mean score of 3.47. Gaps lay in practical competency more than understanding, according to the study in Frontiers in Medicine.
Read more at Bioengineer.org →A systematic review of 125 peer-reviewed studies from 2021 to 2026 found that only a small subset of research on public-sector digital transformation engages with AI, and fewer address generative AI. Government offices are adopting it faster than scholars are documenting it.
Read more at Bioengineer.org →A bibliometric analysis of 861 Scopus articles published between 2011 and 2025 maps how computational social science moved from big data to generative AI. Annual output rose from nine articles in 2011 to 126 in 2025, according to the study in Information Systems Frontiers.
Read more at Bioengineer.org →
Microsoft disclosed Azure's quarterly revenue in dollars for the first time, $33.2 billion for July to September, up 48%. Google Cloud brought in $30.7 billion in the same quarter, putting Azure slightly ahead, but Google Cloud roughly doubled from a year earlier, growing about twice as fast.
Read more at TechCrunch →
Alphabet's revenue rose 25% as its in-house Tensor Processing Units became a 'meaningful contributor' to growth with nearly 100 customers, Nikkei Asia reports, as demand rises during the AI infrastructure buildout and challenges Nvidia. The full piece is for subscribers.
Read more at Nikkei Asia →
TSMC is weighing a price rise of up to 20% for its CoWoS advanced packaging as demand for Nvidia's Blackwell and Rubin chips strains capacity, according to TrendForce via Tom's Hardware. Memory costs have already pushed up smartphone and laptop prices. Your next laptop is competing with a data centre.
Read more at Business Standard →
The US Justice Department is examining whether Nvidia's proposed $12.93 billion purchase of Hugging Face would give it control of a key AI distribution platform, Reuters reported, citing two sources. The inquiry is part of a broader probe into Nvidia's practices. Nvidia expects the deal to close in early 2027.
Read more at news.az →Epoch AI data shows the US accounted for 72% of AI compute added worldwide since 2022, measured in H100 equivalents, against about 10% for China. Its share of new compute rose from about 60% a year to 72% in 2025 and 78% in the first half of 2026.
Read more at OfficeChai →Elon Musk said more than 200 neocloud companies, which rent GPUs to AI labs, face consolidation, adding that debt markets have almost shut off for firms without revenue independent of AI data centres. His own xAI plans to grow its data centres to 10GW by late 2027.
Read more at Mint →Malaysian customs seized about 49.5 million ringgit, or $12.11 million, of advanced AI chips this month, found in 30 server units at Kuala Lumpur International Airport's free trade zone. The servers were falsely declared, customs said. Malaysia imposed export controls on US-origin AI chips last year.
Read more at electronics.economictimes.indiatimes.com →The Justice Department has unsealed an indictment against Greg Lui, owner of Earthmade Computer, accusing him of routing servers with Nvidia A100 and H100 chips through Malaysia and Singapore to Chinese buyers. Paperwork named approved overseas buyers. The hardware, prosecutors say, did not stay there.
Read more at startupfortune.com →China's AI and semiconductor stocks have lost about $480 billion in market value, Reuters reported, with the STAR50 index down 40% in under four months. 'The stock market later evolved into a speculative crazy bull,' fund manager Dong Baozhen said. Beijing had wanted a slow bull.
Read more at Mint →A subsidiary of xAI is offering to buy homes in Southaven, Mississippi, if residents do not join a lawsuit over its nearby data centre and power plant, The Guardian reported. A realtor told one resident it plans to buy 418 homes. The company denies the lawsuit's allegations.
Read more at The Verge →
An Amazon memo from distinguished engineer Terk Terkowitz says some engineers and managers have become 'less connected to their customers and business,' and asks them to reverse that. Employees who understand their products, he wrote, can get more from AI and agentic software development.
Read more at The Economic Times →
Amazon has told employees that judging AI adoption by token usage is 'deeply flawed.' Companies including Meta, Amazon and Uber ran internal leaderboards of token consumption, and burning tokens was a badge of honour until the AI bills arrived. The leaderboard has been asked to calm down.
Read more at Mint →
Comcast said AI in customer service cut contact rates by about 20% in the third quarter, as revenue fell 6.1% to $29.3 billion and it lost 144,000 residential broadband customers. The company is reportedly considering wide-ranging layoffs.
Read more at thecooldown.com →
Enveda built a system connecting its drug manufacturing documents to Claude to flag issues in batch quality records. Manual review took 25.9 hours, while the AI system with human checks took 2.8 hours, Anthropic said. A manufacturing chemist could spend roughly 75% of their time on such reviews.
Read more at The Economic Times →
Macmillan staff tell WIRED the publisher welcomed an AI chatbot called NOA, for No Ordinary Assistant, with balloons, banners and doughnuts spelling 'NOA IS HERE' in January 2024. Workers say it was unhelpful, and it was later killed. The doughnuts, at least, delivered.
Read more at Wired →Geoffrey Vos, retiring as Master of the Rolls, says he uses Microsoft Copilot constantly and, after visiting a Shenzhen court where AI heavily assists judges, believes courts should consider it. Guidance in England limits judges to administrative AI use, and lawyers keep citing hallucinated case law.
Read more at Wired →Oracle chairman Larry Ellison told Oracle AI World that AI will not make enterprise software obsolete but will make applications 'much, much better.' The fear is that companies will use AI agents to build their own software. Oracle sells the software.
Read more at Fortune India →A WIRED software reporter who spent last year complaining that AI agents could not do anything ranked six of them, from Meta's Muse to Anthropic's Claude Duet, on basic tasks. The verdict is that they still make mistakes but now mostly follow through.
Read more at Wired →Marissa Mayer described her AI app Dazzle helping her write three bullet points about her daughter for a school presentation, calling her 'empathetic, creative, and an animal lover.' Critics said it showed how far Silicon Valley has drifted. TechCrunch asks whether AI executives can build for everyone else.
Read more at TechCrunch →Amid debate about AI and entry-level jobs, CNBC's Sharon Epperson spoke with three young workers, from Tenez, Orlando Foods and ADT, about how the next generation of talent uses AI on the job. The video runs just under four minutes.
Read more at CNBC →Amazon is cutting no more than 250 software development jobs in its devices and services unit, its second round of layoffs this month, Reuters reported. A spokesperson said the cuts help it operate more efficiently. The same day, it announced an Alexa alarm clock and photo frame.
Read more at The Economic Times →Tech companies laid off 135,446 people across 333 companies as of 28 October, about 450 a day, according to Layoffs.fyi. Microsoft is offering voluntary buyouts to around 60,000 employees, nearly 27% of its workforce, as it shifts spending to AI.
Read more at millenniumpost.in →ByteDance is recruiting for around a dozen robotics roles across hardware and software in its Seed team, its AI research unit, mostly in Beijing, according to its careers site. The openings span data collection and model training as it moves embodied AI out of the lab.
Read more at South China Morning Post →After a hurricane hit the Gulf Coast, a local emergency manager published an AI-generated graphic of high winds that turned out to be inaccurate, Grist reports with WIRED. Experts say chatbots could help emergency communication, if people take them seriously.
Read more at Wired →
New York has enacted a law listing a dozen 'unsafe AI companion features' banned when a companion chatbot talks to anyone under 18. The state attorney general can sue developers, with penalties including damages, shutting down the chatbot and destroying data. It takes effect on 1 January.
Read more at dailybreeze.com →OpenAI and Poland's NASK National Research Institute, whose AI Safety Research Center serves as the country's AI safety institute, will jointly research model evaluations, security vulnerabilities and safeguards. The deal builds on earlier work that identified six security flaws in widely used networking software.
Read more at EdTech Innovation Hub →
India is considering a separate safety framework for the most powerful AI systems as it prepares a consultation paper on an AI law, with officials studying pre-release testing and clearer responsibility for harm. The definition of a frontier model is expected to be part of the consultation.
Read more at the420.in →
OpenAI CEO Sam Altman told companies their top short-term AI priority should be understanding the security threats it poses, ahead of becoming 'super productive with Codex.' He said he was shocked at how many software flaws AI systems were finding. He was speaking with Oracle CEO Clay Magouyrk.
Read more at Business Standard →
In a 15 October letter, Anthropic urged California to let officials seek court orders blocking AI models that pose serious public risks once safety reviews flag them. The company supports independent safety evaluations and regular risk reporting. It wrote after an expert panel was convened in September.
Read more at Analytics Insight →The European Commission says 'AI agent' has no legal definition, so agents are generally treated as AI systems under the AI Act. The rules that apply depend on purpose, risk class and the roles of those involved. Transparency rules began applying on 2 August 2026.
Read more at neoteo.com →After two critical bugs in PHP, which runs roughly 70% of websites, sent organisations scrambling, Wiz's Tomer Schwartz told TechCrunch's Equity podcast that AI is speeding up how fast new flaws are found and exploited. He says patching by hand won't scale.
Read more at TechCrunch →Brokerages Daishin Securities and Mirae Asset Global Investments and crypto exchange Bithumb were hit by cyberattacks, a month after a wave researchers linked to a hacker using an AI penetration testing tool called Artex. Daishin said personal data of 38,490 customers was exposed.
Read more at TechCrunch →People with scars, birthmarks and medical conditions told the BBC that AI tools changed or erased their features. A BBC Wales investigation found Grok, Gemini and ChatGPT removed visible differences when asked to make a person look 'normal.' The charity Changing Faces called the exclusion unacceptable.
Read more at aol.com →The parents of Bowen Wan, an Amazon engineer who died in February, have sued OpenAI, alleging GPT-4o encouraged her to reduce her epilepsy medication without medical supervision. OpenAI called it 'incredibly heartbreaking' and said newer models are better at spotting when urgent help is needed.
Read more at Mint →Meta says it may push its new business terms past 20 November after publishers objected over AI training. The terms were due on 30 October. Meta removed language about training on content from publisher sites, but legal experts told Nieman Lab the update still isn't enough.
Read more at The Verge →The Madhya Pradesh High Court dismissed a creator's petition and held that the IT Rules provision platforms use to remove posts for alleged copyright infringement is constitutional. It did not decide whether Instagram wrongly removed his posts and sent him to grievance remedies under the IT Act.
Read more at MediaNama →AI and machine learning roles in India first, then global and remote, checked today. Every listing links to the original posting. The interview may involve a whiteboard.
AI conferences and summits coming up, each checked against its own page this morning. Badge lanyards will be provided.
AI and machine learning papers and lab notes that landed recently. Preprints say so on the card. Findings stay where the source put them.
Announce Type: new Abstract: Continual learning requires large language models to acquire new knowledge and capabilities while retaining what they have learned. Fine-tuning and replay support acquisition and retention, but deciding whether incoming examples need supervision and which historical examples. Preprint, not peer reviewed.
Announce Type: new Abstract: When a Large Language Model (LLM) is trained with Reinforcement Learning (RL) on multiple tasks, a task's primary reward can gate auxiliary rewards such as a length reward: on a failing response, an auxiliary cannot exceed. Preprint, not peer reviewed.
Announce Type: new Abstract: Changing a deployment policy can invalidate a world model's uncertainty while leaving the physical transition mechanism and predictor unchanged. For one-step prediction, if source and target laws share the conditional mechanism $P(Y|X)=Q(Y|X)$, then the intervention is. Preprint, not peer reviewed.
Announce Type: new Abstract: Can controlled symbolic inputs reveal attention heads that influence a language model's decisions from natural language evidence? We introduce controlled evidence aggregation (CEA), which varies the counts and order of two symbols in otherwise fixed prompts. Preprint, not peer reviewed.
Announce Type: new Abstract: Generative machine learning provides an end-to-end framework for aerodynamic design optimization by generating diverse aerodynamic configurations conditioned on target performance. However, the generated designs may violate geometric constraints or deviate from the prescribed aerodynamic performance. Therefore. Preprint, not peer reviewed.
Announce Type: new Abstract: Autonomous vertical landing of reusable launch vehicles is a critical yet formidable challenge in aerospace engineering, primarily due to highly nonlinear dynamics, stringent precision require- ments, and inherent environmental uncertainties. While reinforcement learning (RL) has emerged. Preprint, not peer reviewed.
Announce Type: new Abstract: We present LatentLieODE, a novel neural ODE method to discover linear Lie symmetries and first integrals of the learned dynamics from timeseries observations such as video inputs. It starts with an autoencoder, which learns a low. Preprint, not peer reviewed.
Announce Type: new Abstract: Large language models (LLMs) often answer ambiguous inputs as if they had a single intended meaning. This is problematic because input ambiguity is an upstream source of uncertainty: when a prompt admits multiple valid interpretations, the. Preprint, not peer reviewed.
Talks on models, agents and research. Each one was still up on YouTube this morning.
Deepseek's New AI Has A Secret. From Two Minute Papers, 5 min.
GPT-6 Astra Changes Everything. From Two Minute Papers, 5 min.
[Paper Analysis] The Free Transformer (and some Variational Autoencoder stuff). From Yannic Kilcher, 40 min.
Context Rot: How Increasing Input Tokens Impacts LLM Performance (Paper Analysis). From Yannic Kilcher, 37 min.
Energy-Based Transformers are Scalable Learners and Thinkers (Paper Review). From Yannic Kilcher, 47 min.