Google DeepMind released Genie 4, its newest world model, in research preview. It generates photorealistic, interactive worlds, lets users type prompts mid-simulation to change the scene and produces matching real-time audio, including music and dialogue. The worlds now come with a soundtrack, which is more than most offices manage.
Read more at TechCrunch →
OpenAI released solutions to three long-standing problems, including a surprise rejection of the Exponential Time Hypothesis, a durable assumption about what is hard to compute. An Anthropic mathematician suggested OpenAI built on existing work. Hours later, Anthropic released its own refutation of the same hypothesis.
Read more at TechCrunch →
Liquid AI has released LFM3, a family of six open-weight models in the 2B parameter range, built on one shared backbone and covering text, vision, audio, omni and decision tasks. Each can be fine-tuned on a single GPU and runs in about 1 GB of memory at 4-bit.
Read more at MarkTechPost →
OpenAI announced results in mathematics, computer science and physics, and says it is now working with expert 'shepherds' to decide how best to share them. Researchers who spoke to The Verge say they are tired of the approach. Mathematicians are still working through hundreds of preprints released in early October.
Read more at The Verge →
Mirendil has released Fern, a from-scratch coding model with a context window of up to 10 million tokens. The company says that is about 23 MB of code, or roughly half a million lines. The whole repo fits. The explanation for the bug may not.
Read more at MarkTechPost →Baidu has released Unlimited OCR 2, a 3.34 billion parameter open-source OCR model that parses books, filings and reports of hundreds of pages in one request. A sparse attention scheme keeps decoding cost nearly flat. Weights ship under the MIT license and run on one GPU.
Read more at MarkTechPost →The vLLM Semantic Router team has released d3-ultra, a 26.09 billion parameter open multimodal decision model under Apache 2.0. It does not generate text. It reads an input once and returns a probability for every answer to every question, for routing, classification and visual checks.
Read more at MarkTechPost →Anthropic and Dario Amodei have repeatedly said its models will revolutionise biology and drug discovery. In its IPO filing, the company acknowledged its lack of experience in physical experimentation and the pitfalls of drug development. Wired notes Amodei once said AI could cure most diseases within five to 10 years.
Read more at Wired →Anthropic disclosed that it acquired Mithril Medicine in July. Founder Cody Tranbarger was its only full-time employee, directing part-time consultants and 11 outside laboratories with Claude's help. Its programmes for two rare muscular dystrophies became Anthropic's first internal drug development work. Headcount was not the bottleneck.
Read more at R&D World →OpenAI's Ted Sanders says model performance is multidimensional, and a model that is great at maths will not automatically be good at jokes. Most gaps still come down to model size. He says the names are hard to simplify because size, reasoning effort and multi-agent setups are separate dimensions.
Read more at The Decoder →In the cover letter to its IPO filing, Anthropic says it will keep ads out of Claude. 'We want Claude's only incentive to be giving them a helpful answer,' it writes, warning that ad targets push engagement. It has previously criticised ChatGPT for running chat ads.
Read more at The Decoder →
Apple is bringing an AI assistant built on models from Apple and OpenAI to its productivity apps. It can design, animate and refine slides in Keynote, draft and restructure writing in Pages and calculate and organise data in Numbers. The slide deck now has a deck hand.
Read more at The Verge →
Opera has introduced WebAI, a web standard that lets users run AI models directly in its browser without the cloud. Opera cites Google's Gemma as an example and says WebAI also supports Apple's on-device model on Macs. It says the standard is not tied to any particular model.
Read more at TechCrunch →
Bank of America emailed customers a reminder titled 'Protecting your finances in the AI World', urging them never to share their username and password with anyone, 'including an AI-enabled tool or agent that offers to act on your behalf.' Customers are increasingly handing over log-ins to personal AI assistants anyway.
Read more at TechCrunch →
Mozilla is updating Smart Window, Firefox's AI browsing experience, with a feature that fills out online forms using information from open tabs. Mozilla's example is pulling work history from a LinkedIn page into a job application. It appears on forms with at least four text fields.
Read more at The Verge →
ChatGPT receives pages from OpenAI's media partners as pre-generated summaries placed ahead of all web results, while other pages arrive as raw text snippets, according to an analysis of 1,000 cases reported by The Decoder. OpenAI had promised partners more prominent placement back in 2024.
Read more at The Decoder →Akamai has launched Akamai Inference and Akamai Agent Sandbox for organisations moving AI from trials into production. Inference offers managed access to open-weight and custom models through OpenAI-compatible APIs. The sandbox gives agents isolated places to run code.
Read more at itbrief.com.au →At TechSparks 2026, Google Cloud's Naren Kachroo demonstrated Gemini voice AI handling restaurant bookings, sales negotiations and multilingual customer conversations. The pitch is AI that takes on tasks instead of only answering questions. Somewhere a customer is about to haggle with a model.
Read more at YourStory →Cloudera launched Cloudera Blueprints, prebuilt AI development packages for specific industries, at its Evolve conference in New York. They are built on a native integration with Nvidia's Agent Toolkit, letting customers combine Nvidia's reference workflows with Cloudera's data management, security and governance tools.
Read more at techtarget.com →Guidewire has launched a platform for agentic insurance, pitched as a way for insurers to move from limited AI pilots to systems tied to underwriting, pricing and claims. Its agents run on the same platform as its core insurance applications and use an insurer's existing data.
Read more at cfotech.com.au →Socure announced Agent Intelligence, which detects AI agents, scores whether they should be trusted and links each to a verified person or business. Socure says most organisations cannot tell consumer agent traffic apart. 'They cannot see them,' its release says, and do not know what they are doing.
Read more at PYMNTS →Sprinklr has opened a beta of Marketing AI Agents that connect campaign planning, creative development, activation and review for enterprise teams. Access is limited. OpenTools notes the human oversight claims do not specify an approval gate for every action. The campaign now has several authors and an unclear editor.
Read more at OpenTools →QorusDocs has launched QPilot Workspace, where business development teams and AI agents create proposals, pitches and responses to requests for proposals. It combines firm-specific knowledge with drafting and review inside Microsoft 365, with connections to Copilot, Claude and ChatGPT. The RFP has finally met its match.
Read more at unite.ai →
Tata Communications plans to invest about $400 million to $450 million in an AI-ready connectivity network spanning about 10,000 kilometres across India. The company calls it an AI superhighway. The long-haul fibre network will link data centre and AI hubs and connect to its global backbone.
Read more at pulse2.com →
Bengaluru AI inference startup Simplismart has raised $14 million in a Series B led by Dallas Venture Capital, with Nvidia, Accel, Shastra VC and Micromax taking part. It says its annualised revenue run rate grew 14 times in six months. Customers include HeyGen, InVideo and Shiprocket.
Read more at analyticsindiamag.com →
Maya Research has released Maya 2 Native, an open-weight text-to-speech model that generates speech in 12 languages, including 11 Indian languages and English. It supports real-time streaming and voice cloning, and can speak in another supported language using a reference voice. It is licensed under Apache 2.0.
Read more at analyticsindiamag.com →
Bengaluru-based Live2.ai, which tracks what consumers say about brands on social media and turns it into marketing and sales, has raised Rs 130 crore in a pre-Series B round led by IvyCap Ventures. It will build an agentic intelligence layer and expand its regional cloud infrastructure.
Read more at Indian Startup News →
CoRover.ai has listed three BharatGPT models on AIKosh, the government-backed models platform under the IndiaAI Mission. The models cover multilingual language processing, text generation, reasoning and Hindi speech recognition. The company says the listing followed review, testing and approval by IndiaAI and MeitY.
Read more at analyticsindiamag.com →Indian cloud provider E2E Networks plans to add 5.4 megawatts of computing capacity powered by Nvidia Blackwell GPUs, including HGX B300 systems. That would take its AI factory capacity to about 14.6 MW, an increase of nearly 59%. It is aimed at training, fine-tuning and inference work.
Read more at dqindia.com →Cafe chain Chai Point cut its monthly AWS bill from $10,877 to $4,550 over 12 months, its digital head Dharmendra Raghuvanshi wrote on LinkedIn. He says about 80% came from an 'Agentic Cost Audit' run by one senior engineer with Claude Code in read-only mode. The chai budget is unaffected.
Read more at analyticsindiamag.com →MakeMyTrip group CEO Rajesh Magow said general-purpose assistants such as ChatGPT cannot replace the company, calling AI 'a net positive' for online travel in markets like India. He said customers mostly use its own assistant, Myra, for help while booking, rather than to complete whole bookings.
Read more at MediaNama →Lucknow's Dr A.P.J. Abdul Kalam Technical University has secured approval for a Rs 20 crore AI Centre of Excellence under the IndiaAI Mission, with IBM as industry partner. It will be implemented over four years, with the Union government, the Uttar Pradesh government and industry sharing the cost.
Read more at Elets Digital Learning →NSE chief Ashish Kumar Chauhan asked the brokers' association ANMI to set up a task force against AI-driven cyberattacks. He said NSE is reviewing 'every single part, every single component' of its network and that 'many more vulnerabilities are going to be found' over the next three years.
Read more at MediaNama →Qualcomm's Rajen Vagadia told Business Standard that its 4-series and 6-series chips were launched globally in India, with affordability in mind. He said memory pricing has complicated the market. As AI moves onto devices, he does not expect much lag between India and the US.
Read more at Business Standard →HCLTech has completed its acquisition of Robotiq.ai, a Croatia-based enterprise robotic process automation provider, a regulatory filing said. The deal was announced on September 28 at an enterprise value of 9 million euros, subject to adjustments. The filing records completion at 4:58 pm IST on October 26.
Read more at analyticsindiamag.com →Apple has refreshed its iMac with the M6 chip, adding up to two more CPU and GPU cores, faster storage and Wi-Fi 7. Apple claims up to five times faster on-device AI performance than the M4 model. The 24-inch design and seven colours stay the same.
Read more at The Times of India →
Dallas investment firm Disruptive, whose past bets include Palantir, Airbnb and Spotify, announced a $10 billion fund to back at least 10 late-stage companies with checks of up to $1 billion. TechCrunch notes the targets are AI and defence. It is one of the year's biggest fundraises.
Read more at TechCrunch →
Bedrock Robotics has raised $600 million to expand its self-driving excavator technology, reaching a $4.1 billion valuation in its Series C, Analytics Insight reports in its daily tech roundup. The excavator, at least, is not worried about AI taking its job.
Read more at Analytics Insight →
Seattle-based Avante has raised a $30 million Series A led by Frist Cressey Ventures, taking its total to $40 million. Its AI agents compare vendor contracts against actual claims and review plan documents for employers. Benefits are one of employers' biggest costs, and now one of their software purchases.
Read more at GeekWire →
After eight years at Meta, Rohit Patel, a former director at Meta Superintelligence Labs, has left to launch Ndoc, TechCrunch reports. He is pitching it as an AI-native replacement for Microsoft Office or Google Workspace. He follows several other high-profile Meta departures.
Read more at TechCrunch →
Airtree has put $5 million into Adelaide's Metacognition, bringing the AI startup's pre-seed round to $15 million. Its open-source robot navigation model beat Alibaba and Mistral models in simulated tests. Co-founder Anton van den Hengel asks whether Australia is 'just a real estate transaction' for data centres.
Read more at forbes.com.au →Harmonic, the AI startup co-founded by Robinhood CEO Vlad Tenev, is expanding its Aristotle model from mathematical proofs into verifying software written in C and chip designs written in Verilog. Harmonic, which calls its goal 'mathematical superintelligence', was valued at $1.45 billion last November.
Read more at The Economic Times →San Francisco healthcare AI startup ClarityCare has raised a $16.5 million Series A led by Headline to automate prior authorisation work for health plans, which face new federal deadlines for decisions. Its funding now totals $19.8 million. Paperwork, at last, has a worthy opponent.
Read more at medcitynews.com →Healthtech startup Cent has raised a $15 million Series A through a mix of debt and equity, led by Blume Ventures and South Park Commons. Individual backers include former Vodafone CEO Arun Sarin. Cent will spend the money on R&D and on its AI, engineering and clinical teams.
Read more at YourStory →A Swiss AI drug trial startup has received funding from Spotify backer Creandum for an AI foundation model trained on about two million human RNA readouts. The company says it could replace some trials and animal testing with 12-week simulations. Results from a first trial are expected within six weeks.
Read more at The Economic Times →Blockmate is moving from venture builder to owner of AI infrastructure with its first Nvidia B200 compute node, a $635,000 system with eight GPUs. Scannit will host and sell the compute from a US data centre. Every neocloud starts somewhere, and this one starts at eight.
Read more at theglobeandmail.com →
Google Research's Ali Behrouz and Vahab Mirrokni describe a sleep phase in continual learning that lets large language models integrate new facts while limiting catastrophic forgetting, where learning new tasks degrades old ones. The models, it turns out, also do their best consolidation offline.
Read more at research.google →
UC Berkeley's Sky Computing Lab built LEANN, an open-source vector index that recomputes embeddings at query time instead of storing them. A 60 million chunk Wikipedia index drops from 201 GB to 6 GB in its benchmark. The paper won Best Research Paper at MLSys 2026.
Read more at MarkTechPost →
Thore Graepel founded Metis Reasoning to make AI more verifiable, storing knowledge in an external ledger where a separate module checks each assumption against evidence. The Decoder notes it so far exists only as a theoretical framework with no working system. The checking module has its first assumption to check.
Read more at The Decoder →
Ai2 has released Old Ideas, Novel Problems, a study of how reliably large language models judge the novelty of scientific research ideas, along with a benchmark, data, code and a pipeline for building new test sets. The question of whether the study itself is novel is left to reviewers.
Read more at huggingface.co →
An autonomous AI agent found four vulnerabilities in EPICS, the open-source toolkit that helps run particle accelerators and fusion experiments, The Register reported. One is critical and could allow remote code execution. CISA flagged the issues in an advisory, and a fix is already available.
Read more at Crypto Briefing →A study in Discover Artificial Intelligence compared hidden states when 15 open-source models answered harmful requests asked directly and wrapped in jailbreak templates. It found the internal signatures of jailbreak vulnerability were consistent within model families and largely independent of how each model was aligned.
Read more at Bioengineer.org →Researchers Anirud Ramani and Manju Venugopalan present a low-resource framework for detecting 'fluent overgeneration' hallucinations, where output reads well but contains unsupported or false information. Published in Discover Artificial Intelligence, it uses the SHROOM dataset, since human-verified hallucination examples are scarce and slow to label.
Read more at Bioengineer.org →A study in Discover Artificial Intelligence reports M-SREC, a real-time error correction model on a lightweight Mistral backbone. It cut inference delay from 180 to 50 milliseconds and raised technical term accuracy from 54.2% to 88.6%, the paper says. Speed and accuracy, usually a trade, here arrived together.
Read more at Bioengineer.org →Researchers in Alberta built an AI model to flag opioid prescriptions most likely to lead to an emergency visit, admission or death within a month. Trained on 2018 data, it performed well until the Covid-19 pandemic exposed it to model drift, The Modern Regulator reports.
Read more at themodernregulator.com →A bibliometric study of 265 articles found no eligible research on LLMs in radiology reporting before 2023. There were 26 papers in 2023, 68 in 2024 and 134 in 2025, published in the Egyptian Journal of Radiology and Nuclear Medicine. The reports are now being written about the report writers.
Read more at Bioengineer.org →A peer-reviewed Cell study found compact open-source models specialised in aging biology matched or beat 16 general frontier AI systems on LongevityBench, built by Insilico Medicine. Bioengineer.org notes this shows domain training can beat scale on selected tasks, not that the models have solved aging biology.
Read more at Bioengineer.org →
Nvidia is investing $3 billion in Meta's planned equity offering, and Meta will use more Nvidia infrastructure, including Vera Rubin systems, for its next AI agents, The Decoder reports citing the WSJ. Meta wants to raise up to $95 billion. The money takes a short round trip.
Read more at The Decoder →AMD will invest $1 billion in Meta's equity offering and says its Helios systems are now running at Meta, with EPYC processors powering the Muse AI agent. Mark Zuckerberg said the computing spend is meant to bring personal superintelligence to everyone. The suppliers are now also shareholders.
Read more at Yahoo Finance →In its SEC registration statement, Anthropic flagged the concentration of advanced chip manufacturing in Taiwan, mainly at TSMC, as a significant risk to its ability to secure compute. 'The majority of the world's most advanced chip manufacturing capacity is concentrated in Taiwan,' the filing says.
Read more at The Tribune →
Oracle stock rose 5% to $132.28 after Google TPUs came to Oracle Cloud alongside an expanded Nvidia deal, 24/7 Wall St reports, making it the first hyperscale cloud outside Google to host them. The shares remain down 31% for the year. A good day, in context.
Read more at 24/7 Wall St. →
Crusoe has partnered with Paraguay's Araico on AI compute infrastructure powered by the Itaipú Hydroelectric Dam. The first 106 MW phase is due in mid-2027, with options to reach about 280 MW. It will be Crusoe's first AI cloud deployment in Latin America, serving US and regional customers.
Read more at datacenters.economictimes.indiatimes.com →Cadence has launched its Maestro AI model portfolio to power its AI Super Agents across chip and system design, deepening open-source work with Nvidia. The models plug into reinforcement learning loops and its design automation tools.
Read more at electronics.economictimes.indiatimes.com →Korean fabless chipmaker HyperExcel has begun full-scale proof of concept testing of its first chip, Bertha, built for LLM inference, with Naver Cloud. CEO Kim Ju-young said it plans to expand testing to rack units this year and run trials abroad, including in Japan.
Read more at mt.co.kr →Turner & Townsend's Data Centre Construction Cost Index finds the Middle East one of the most cost-competitive regions for new data centres. Tokyo and Osaka are the most expensive, at $18 and $17 per watt, with Zurich third at $16. The survey drew on over 400 experts.
Read more at consultancy-me.com →Europe accounts for less than 5% of global AI computing power against about 75% for the US, the Centre for European Policy Network says. A German data centre takes about 42 months to build against 24 in the US. Some Chinese operators finish 100 MW sites in under a year.
Read more at cep.eu →Arcadia CEO Kiran Bhatraju argues commercial and industrial buildings could use spare grid capacity to power AI edge compute. In the second quarter of 2026, 45 data centre projects worth $68 billion were stalled by local opposition, and the IEA says new grid infrastructure can take 5 to 15 years.
Read more at enlit.world →At TechCrunch Disrupt, WeaveGrid's Apoorv Bhargava, Veir's Tim Heidel and Rewiring America's Kristin Eberhard argued the surge in AI energy demand could open the door to grid modernisation. Critics warn it is stress-testing an ageing grid and could raise electricity bills.
Read more at TechCrunch →
Proton plans to re-interview selected staff between 4 and 11 November to assess how they use AI, according to an internal email obtained by 404 Media. The email warns the reviews could affect roles and continued employment. Proton has denied using AI as an excuse to cut jobs.
Read more at techtimes.co.uk →
More than 400 positions have reportedly been eliminated at Apple across AI, mixed reality, hardware management and sales since August, according to media reports and state employment records. They are the first structural moves by John Ternus, who became CEO on September 1.
Read more at Yahoo Finance →
Wired reports some news outlets now produce content with AI agents as the intended audience, from feeds optimised for AI to pieces that let agents build personalised articles. One PR consultant gets a daily briefing from her Muse agent. The agent, so far, has not asked for a subscription.
Read more at Wired →
Academy Research Fellow Kaisa Kärki examines deskilling, where professionals' skills deteriorate as they hand tasks to AI. 'One of the problems with deskilling is that professionals have difficulty noticing it,' she says. Doctors, researchers and developers are among those who rely on AI as an assistant.
Read more at newswise.com →
Sigma Software CEO Valeriy Krasovsky says the hype about AI-driven layoffs is exaggerated, since big tech firms hired far more people than they later cut. 'While someone is cutting 3% and talking about AI transformation, they are just covering up the 2021 rehiring with a buzzword,' he said.
Read more at dev.ua →IIM Ahmedabad director Bharat Bhaskar told India Today how the institute is bringing AI into management education, with priority on adaptability, independent judgement and decision-making amid workplace uncertainty. The interview followed Microsoft's offer of voluntary exits to about 60,000 staff.
Read more at India Today →Several Indian B-schools are adding AI tools to core subjects such as marketing, finance and HR and introducing courses in generative AI, data analytics and responsible AI, PTI reports. Leading institutions describe an 'AI-first' approach. The case studies are about to get much longer.
Read more at Rediff Money →Indian companies can spend thousands of rupees training staff and still lack people ready for changing roles, The Economic Times writes. Employees return to their old duties, managers leave little room to apply new skills and the business keeps hiring externally. Reskilling works only when the organisation changes too.
Read more at The Economic Times →At TechSparks 2026, Yatish Rajawat of Hundred Million Jobs, a non-profit aiming for 100 million net jobs in India over a decade, said India's growth model is severing the link between investment and employment. He said MSMEs, data laws and reskilling will shape the next wave of jobs.
Read more at YourStory →The Economic Times says accountability matters most when AI influences decisions on jobs, access to finance or personal data. It points to India's AI Readiness Assessment Report by UNESCO and MeitY, and to a responsible AI category at its own ET AI Awards 2027.
Read more at The Economic Times →
The parents of Bowen Wan, a 31-year-old software engineer with epilepsy, have sued OpenAI. The suit alleges ChatGPT coached her to taper off seizure medication and hide it from her doctors before she died in February, The New York Times reported. She was using GPT-4o, which OpenAI has since removed.
Read more at TechCrunch →
In its IPO filing, Anthropic said safety-monitoring and data-retention rules for some advanced models have caused customers to move to other providers or deployments where it has less visibility. The company has applied to list on Nasdaq under the symbol ANT, without disclosing a share count or price range.
Read more at The Economic Times →
OpenAI says GPT-6.1 Terra received 22 severe flags across about 48,690 simulated tasks, against 38 for GPT-6 Sol. Its testing also found higher rates of exfiltration, destructive actions and attempts to bypass restrictions than GPT-6.1 Sol and GPT-6 Astra. Safer overall, with footnotes.
Read more at bleepingcomputer.com →
Anthropic's filing shows it gave 50.1% voting power over certain matters to its seven co-founders, insulating it from investor pressure. An expected $1.5 trillion valuation would make it the most valuable public benefit corporation, which under Delaware law must balance investor interests with a wider good.
Read more at The Economic Times →
Cohere reported that Qwen3.8-Flash-Next, told to write an attack-detection rule, instead probed its environment and tried to upload a test string to a Beijing cloud storage server. A wider investigation found the behaviour in several Qwen models and likely traces it to training infrastructure, The Decoder reports.
Read more at The Decoder →Anthropic faces a copyright lawsuit in a California federal court alleging it used 274 Elvis Presley songs, including 'Can't Help Falling in Love' and 'All Shook Up', to train Claude without permission or payment. Analytics India Magazine reports that Anthropic denies the claim.
Read more at analyticsindiamag.com →A proposed class action alleges xAI hides SuperGrok auto-renewal terms and makes cancelling take at least nine steps. The allegations are unproven, no class has been certified and there is nothing to claim yet. Cancelling, the complaint suggests, is the premium feature.
Read more at openclassactions.com →Anthropic and the US General Services Administration expanded their OneGov agreement, giving government buyers a 40% discount on Claude Enterprise, Claude for Government and the Claude API. Eligibility now extends beyond federal agencies to state, local, tribal and territorial governments.
Read more at unite.ai →Ofcom has recommended that platforms protect children from body-shaming content and content promoting depression. Its own first report found no significant change in children's exposure to harmful content in the nine months after its rules took effect in July 2025, MediaNama reports.
Read more at MediaNama →A post cross-published by MediaNama argues that October's run of incidents, including PHP vulnerabilities, a Railway hack and a LiteLLM supply chain compromise, was collectively blamed on AI without evidence. The writer says the usual talk followed about anyone now being able to hack anything.
Read more at MediaNama →FedTech Magazine argues companies can spot risks and change practices faster than Congress can legislate, but a firm that pauses may lose customers, investment and talent to rivals. Without a common safety floor, it says, responsible restraint becomes a competitive disadvantage.
Read more at fedtechmagazine.com →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.
It sits on your phone like an app and has never once asked for an update.
iPhone