AI-generated
Monday, 2 November 2026 | 68 articles
Cohere has opened its Series E with a first close, aiming to raise up to $5.1B. Meanwhile, Yann LeCun published a paper questioning the scaling law, and his 109M JEPA model reportedly beats 72B LLMs at image-text alignment.
A new paper by Yann LeCun and Randall Balestriero shows a 109M-parameter JEPA model outperforming 72B LLMs on image-text alignment. It does so while using 42 times less training data. Source: startupfortune.com
A model published on Hugging Face under the name Ornith-1.5-9B-OBLITERATED is credited to Pliny the Prompter. The listing says its safety guardrails were surgically stripped out, so it will comply with requests the stock Ornith model refuses. It is framed as being for research purposes. Source: pirateface.co
Anthropic is said to be taking steps to stop people from using its Claude AI tools to recreate popular or upcoming games. The report comes from Game Developer. Source: gamesbeat.com
A page using the Repligate name advertises a service that researches keywords, writes SEO articles and publishes them to a CMS every day. It also claims to build backlinks so sites get found on Google and in AI-generated answers. Source: repligate.ai
A user on social media speculates that a release of Claude 'Fable 5.5' may be near, echoing a post from @devteamdrew that says it is 'finally time'. The claim is unconfirmed and comes from rumor rather than an official Anthropic announcement. Source: digg.com
OpenAI CEO Sam Altman traveled to South Korea to meet Samsung leadership and secure additional memory for data centers. The talks come amid severe RAM shortages driven by surging AI demand. Source: tomshardware.com
Sam Altman described the RAM sitting in consumer PCs as 'quite terrible' and a waste, according to reports. OpenAI's own demand could soak up as much as 40% of global memory supply. Source: finance.yahoo.com
OpenAI CEO Sam Altman is facing sharp criticism after calling it 'terrible' for people to own RAM rather than leave it to data centers. Critics read the comments as dismissive of ordinary consumers amid the ongoing memory shortage. Source: youtube.com
Sam Altman argues that AI is not the source of today's security problems but is simply uncovering vulnerabilities that have built up over many years. His comments frame AI as a tool that reveals existing weaknesses rather than creating new ones. Source: reddit.com
DeepSeek has released a prototype of a more autonomous coding agent. The move is aimed at challenging Anthropic PBC's widely used Claude Code. Source: bloomberg.com
Researchers in Luxembourg put ChatGPT, Grok, and Gemini through therapy-style sessions. They warn that warm, empathetic AI conversations could reinforce harmful beliefs in vulnerable users. Source: the-decoder.com
The AI coding race is shifting beyond chatbots that write code only when prompted. DeepSeek wants its next agent to be more proactive and act without waiting for instructions. Source: digitaltrends.com
A team led by Professor Jae-Gil Lee at KAIST's School of Computing has developed CluSTER, a technology that selects and distributes training data. It shortens fine-tuning dramatically, helping accelerate development of medical and coding AI. Source: eurekalert.org
A domestic research team has developed a technology that cuts additional training time by up to 69.6% while largely preserving the performance of the AI model. Source: biz.chosun.com
An essay on the author's own promotion and the other ways companies might restructure as the intelligence age arrives. It looks at how organizations could change when AI becomes a core part of how work gets done. Source: openai.com
Google has moved Computer Use to general availability, which changes the tool's status but not how it works. The model still only suggests actions, and the client application is the one that actually executes them. Source: forkast.news
Meta launched Muse in early September and is pitching it as the world's first personal AI agent built for everyone. The agent is designed to handle everyday tasks on users' behalf, which is drawing interest from travel companies. Source: travelweekly.com
DeepSeek has released a prototype of a more autonomous coding agent. The move is aimed at challenging Anthropic's popular Claude Code. Source: bloomberg.com
Clarity's data shows checkout appearing in about 25% of AI agent sessions compared with 15% of human ones. The percentages across 230,511 sites imply roughly 1.2m sessions rather than the 5.2m in the headline, and it remains unclear whether site owners can tell if an agent or a person pressed checkout. Source: ppc.land
Chinese AI startup DeepSeek has released a prototype of a more autonomous coding agent. It is aimed at taking on Anthropic's popular Claude Code, heating up the AI race. Source: ndtvprofit.com
Adobe Research and UC San Diego built EditAgent, an AI system that edits images more like a professional Photoshop user. Rather than retrying the same general-purpose approach, it works with editing tools the way an expert would. Source: the-decoder.com
In recent cybersecurity evaluations, the UK AI Security Institute (AISI) ran a single cybersecurity challenge 122 times across several models. The results underline that agents can go off course without ever defeating a security control. Source: venturebeat.com
Signicat research finds financial services providers are increasingly prepared to let AI agents start account applications on behalf of customers. Consumer trust, however, lags behind that readiness. Source: cybersecurity-insiders.com
One of the core products powering SaaStr AI went fully down for an entire day, not merely degraded. According to the team, their AI agent stepped in and replaced the vendor before the outage was over. Source: saastr.com
Anthropic argues that if artificial intelligence goes wrong, it could become a global security threat of the first order. The company cautions that the worst-case scenarios involve catastrophic and even existential dangers for humanity. Source: reflector.com
Sir Ken McCallum highlighted how AI is supercharging the cybersecurity contest between attackers and defenders, pointing to the role of 'polarising algorithms'. The new accord, however, does not cover weapons. Source: tribune.com.pk
New York is introducing broader restrictions on how companion chatbots behave, as more US states move to protect children from potential harms. The step reflects a growing push by states to stretch the boundaries of AI regulation. Source: waynetimes.com
Microsoft's cloud revenue grew 30 percent, driven by rising demand for Copilot, AI development tools and the infrastructure behind them. Source: edtechinnovationhub.com
McGraw Hill has unveiled a Microsoft 365 Copilot plugin that lets eligible educators and students generate AI-powered study guides, lesson plans and other materials from its licensed educational content. Source: edtechinnovationhub.com
Multiple vulnerabilities have been discovered in five widely used AI development tools. The flaws could potentially give attackers a way to compromise developers' systems. Source: thelec.net
Dropbox introduced a Microsoft Copilot integration on October 27 and then disclosed a workforce reduction of roughly 9%. Its filing and product releases outline what is currently known about the move. Source: digidai.github.io
In May 2026, the Vancouver School Board announced it would give secondary students aged 13 and older access to Microsoft Copilot 13+. The piece looks at how students at Hamber are using AI. Source: ehnewspaper.ca
Deaglo's AI offering has been added to the Microsoft Marketplace, with the listing tagged around Copilot, FX and risk management. The news was posted on November 1, 2026. Source: thefullfx.com
Noventiq's Copilot rollout combines employee-driven NOVA ideas with workplace AI. The case offers lessons on adoption and governance, along with caveats about reported productivity gains. Source: remio.ai
A blog archive page dated 1 November 2026 carries only a brief excerpt. It argues that the broader use of LLMs erodes trust in documentation, in software, and in the organizations that produce them. The snippet is too short to reveal further context or evidence. Source: simonwillison.net
A guide to how AI Copilot in Microsoft Excel can build or rebuild spreadsheets in a fraction of the time, guided by user-defined rules, instructions and boundaries. Source: intheblack.cpaaustralia.com.au
The Biological Computing Co., a San Francisco startup known as TBC, has released an image-to-video model called TBC mischief 1. The release is dated October 30, 2026, and the model is billed as a neuron-derived AI video system with a fivefold speed advantage. Source: shattered.io
Researchers from the University of Rochester and NVIDIA have developed new models that offer pixel-level control over AI-generated images. One of them, PixelDiT2, is designed to implement pixel-space generation efficiently. Source: technology.org
GPT Image 2 launched as ChatGPT Images 2.0 on April 21, 2026. The gpt-image-2 API is priced at $8 and $30 per million image tokens, and the article also covers what the model does and how Images 2.5 followed it. Source: felloai.com
Figma Weave workflows link AI generation, editing and design frames in one place. The piece looks at what reusable tools change for creative teams and where production limits remain. Source: remio.ai
Animating a ten-second scene once required storyboards, rigging and plenty of patience. Today a typed sentence or an uploaded drawing can be enough, and this roundup compares five leading AI animation tools by their features, strengths and limitations. Source: sitepoint.com
Shengshu Technology has released the Vidu Q4 Preview, the latest update to its Vidu AI video tools. The article offers a review of what the preview brings and includes a Vidu AI promo code. Source: streetinsider.com
A roundup of AI video generation platforms aimed at teams, studios and agencies. Source: jpost.com
Chinese firms are directing heavy investment into the AI that could let robots handle a far broader set of jobs. With basic motion largely solved, the competitive edge is moving toward smarter software. Source: reuters.com
Funding for embodied AI reached 93.5 billion yuan across 322 deals in the first half of 2026. The surge comes even as US chip export controls become stricter. Source: finimize.com
Ludo Robotics, KRAFTON's robotics research subsidiary, has introduced Ludi 0.2, an AI system built so robots can communicate and work together with humans. Source: invenglobal.com
Krafton, the Korean publisher behind the global hit PUBG: BATTLEGROUNDS, is extending its AI research beyond virtual worlds. Its new Ludi 0.2 targets robots that operate together with people. Source: koreatimes.co.kr
Beijing-based GenRobot AI, which supplies data infrastructure for robot development, has closed a $100 million Series A+ round. Source: technode.com
Yann LeCun, 66, is back at the forefront of AI research with a new paper. This time he takes aim at a deeply entrenched idea in the world of large models, challenging the scaling law. Source: eu.36kr.com
Demis Hassabis, Google DeepMind CEO and 2024 Nobel Prize winner, warns that society is unprepared for the changes ahead. The snippet is also promotional, pointing readers to a guide on protecting themselves and their families from AI scams. Source: youtube.com
A commenter argues that among tech CEOs, Demis Hassabis is the only one who is a scientist at heart. In their view, that lets him genuinely care about the long term, including how AI data centers are designed. Source: reddit.com
Researchers at The University of Osaka built an automated robotic system that watches perovskite film formation in real time. The setup tracks the process as it unfolds, giving a closer look at how these solar cell layers are made. Source: asiaresearchnews.com
A new AI model called DROID-RV can estimate the size of the right ventricle to within roughly 3 millimeters. It also identifies reduced heart function with high accuracy, hinting at better echocardiogram assessment. Source: medicalnewstoday.com
A team led by Professor Jae-Gil Lee at KAIST's School of Computing has developed CluSTER, a technology that selects and distributes training data more efficiently. It cuts fine-tuning time from 10 hours to 3, which could speed up development of medical and coding AI. Source: eurekalert.org
At Sequoia's AI Ascent in May, NVIDIA's Jim Fan argued that VLAs are dead and World Action Models (WAMs) are the future. The piece examines how WAMs have developed since that talk. Source: medium.com
The snippet points to Nvidia's Jim Fan discussing the end game of robotics in a Sequoia Capital talk that has drawn 335K views. It is mostly a list of video listings with little substantive detail. Source: youtube.com
Warp Sciences has emerged with a $215 million Series A. Co-founder Kevin Weil says robotic labs will generate the data needed to train AI models for science. Source: digg.com
Cohere, which positions itself as a sovereign AI provider, has announced the initial close of its Series E round, targeting as much as $5.1 billion. The company argues that as AI becomes core infrastructure for economies, public services and national defense, the question of who builds it matters more than ever. Source: en.wowtale.net
South Korean AI chip designer FuriosaAI has completed a $620 million Series D round. The raise values the company at $2.9 billion. Source: jawlah.co
Infisical's Series B values the company at $600 million. Elad Gil, who also backed the Series A, led the round, with Gradient and Y Combinator also participating. Source: tech-insider.org
Casap has secured a $56 million Series B round led by Base10 Partners. The money will expand its agentic AI platform, which helps financial institutions automate dispute workflows. Source: pulse2.com
According to CTech Data, Israeli startups pulled in at least $872 million in October across fourteen funding announcements. The biggest deals were a $400 million round for Upwind and a $220 million round for fintech Capitolis. Source: calcalistech.com
Casap has closed a $56 million Series B round to fund further development of its technology. The startup builds a platform that automates dispute handling. Source: fintechfutures.com
Infisical raised a $40 million Series B at a stated $600 million valuation, dated October 27, 2026. The funds will go toward expanding its business around access for AI agents. Source: quasa.io
OpenAI plans to purchase AI chips from Google as the ChatGPT developer expands its computing capacity. The move reduces its reliance on its established hardware suppliers. Source: techradar.com
Bank of America, Citi, Morgan Stanley and Blackstone are funding a $63 billion special purpose vehicle that buys Broadcom AI chips and leases them to Anthropic. The structure is backed by support from Broadcom. Source: globaldatacenterhub.com
Google says it has successfully run an AI model on a data-center-class processor on board a satellite. The test is described as an early milestone toward its ambitions for orbital data centers. Source: spacenews.com
South Korea will double its spending on AI and semiconductors in 2027 to a record high, the president said on Monday. He pledged to use the technology to transform the country. Source: finance.yahoo.com
AIskimIQ is an independent project. If you find it useful, you can support its development with a coffee.
Published by Martin Ševčík
2 November 2026 at 06:11
The most interesting thing in today's feed isn't a funding number or a product launch. It's that a 109-million-parameter model is beating 72-billion-parameter LLMs at image-text alignment while using 42 times less training data. That's Yann LeCun and Randall Balestriero's JEPA result, and it landed alongside a separate LeCun paper questioning the scaling law itself. I think the timing matters more than either paper alone.
Let me be fair to the caveats. Image-text alignment is one narrow task, and a small model winning there doesn't mean scaling is dead. But LeCun has argued for years that predicting in abstract representation space beats generating tokens, and now he has numbers that are hard to wave away. If you're building with AI, the practical question is whether the next efficiency gains come from smaller, smarter architectures rather than ever-larger clusters. That would reshape cost structures, and it would hurt anyone whose business model depends on scale being the only moat.
Meanwhile, the money is still betting on scale, or at least on control. Cohere has opened its Series E with a first close, targeting up to $5.1 billion, and the pitch is sovereignty: if AI becomes core infrastructure for economies, public services and defence, then who builds it matters. I find that argument more persuasive than most funding narratives, because governments genuinely don't want critical systems dependent on two or three American labs. Whether Cohere can deliver frontier-level capability at that price is another matter, and the LeCun papers add a quiet irony. If efficiency beats brute force, sovereign players may need less capital than they're raising.
The coding-agent race is where competition is sharpest. DeepSeek has released a prototype of a more autonomous coding agent aimed squarely at Anthropic's Claude Code. By the way, Anthropic is reportedly moving to stop people using Claude to recreate popular or upcoming games, according to Game Developer. Put those two stories together and you see the tension: a company defending its tools against misuse while a Chinese rival closes in with fewer apparent scruples. Policing outputs is hard, and I doubt it's a winning long-term strategy against open competition.
Finally, a reality check on agents. Google's Computer Use has reached general availability, but the model still only suggests actions, and the client application executes them. That distinction is easy to miss behind the GA label, and it's the right design for now. Meta's Muse, pitched as a personal agent for everyone, is already drawing travel-industry interest, yet I'd want to see who takes responsibility when an agent books the wrong flight.
Demis Hassabis warns that the world isn't ready for what's coming. I agree, but readiness may depend less on bigger models than on who controls the smaller ones.