AI-generated
Friday, 6 November 2026 | 72 articles
LeCun's JEPA paper reports a 109M-parameter model outperforming a 72B one at matching a vision model, fueling debate over whether the industry has taken a wrong turn. Meanwhile, Anthropic warns its AI models are improving nearly twice as fast with AI's help, and Tencent launches its open-source Hy4 LLM.
OpenAI has introduced GPT-6 Cyber, a cybersecurity-focused model built to find exploitable vulnerabilities and develop patches. It is trained to operate strictly within its authorized scope. Source: openai.com
A review examines how weight-space programming, where one LLM modifies another's weights directly, raised Qwen's MATH-500 score sixfold. The piece weighs the evidence, explains how the approach differs from conventional LLM fine-tuning, and discusses the conditions needed for recursive improvement. Source: medium.com
Tencent has announced Hy4, its latest-generation large language model, describing it as a significant jump in performance and capabilities. The company positions the open-source release among the world's leading open AI models. Source: google.com
In posts on X, OpenAI CEO Sam Altman said he wants the world to have more RAM. He was responding to remarks on the RAM shortage made during a Cheeky Pint interview. Source: google.com
Sam Altman reportedly claims he was misquoted after comments suggesting that the world's DRAM should belong to AI firms like OpenAI. The source frames his response sarcastically. Source: google.com
Sam Altman had said it made little sense for individuals to run their own AI models at home. According to this account, he changed his mind after the election; the framing is highly sensationalized. Source: futurism.com
A post promotes a 14-minute session with OpenAI co-founder Andrej Karpathy, claiming it explains how ChatGPT and Claude work better than most AI courses. The claim is promotional and unverified. Source: google.com
A video circulating online shows Andrej Karpathy using Claude to put together a website in roughly two hours. The snippet offers little detail beyond the video title and its comments section. Source: youtube.com
Gemini Omni 1.2 Flash now generates and edits video natively at up to 15 seconds, with improved consistency. Google is also lowering pricing by 30%. Source: blog.google
OpenAI CEO Sam Altman changed his display name on X to "Ram Altman." The move follows controversy over remarks he made on the October 29 talk show Cheeky Pint. Source: reddit.com
A Manifold prediction market asks whether a Netflix movie will mention janus/repligate, Joscha Bach or nearcyan before the end of 2026. The page offers an AI tool that analyzes probability shifts and searches for related news. Source: google.com
A Manifold market page asks whether repligate/janus will enter an arranged marriage. The page invites users to run an AI-powered probability-change analysis that searches for evidence. Source: google.com
A listing describes Andrej Karpathy opening YouTube late on a Friday to unwind. It promotes videos recommended by real people on Hacker News and ranked by AI. Source: google.com
Anthropic has struck its first licensing agreement with a news publisher, the Associated Press (AP). The deal lands just ahead of the US midterm elections. Source: the-decoder.com
Users can upload drawings to the Canvas tool in Gemini, which generates code from them and lets them refine it right away. The feature is available to try now. Source: blog.google
Researchers present a knowledge-grounded security assessment framework for large language models deployed in smart-grid environments. It improves risk quantification and enables cross-domain access verification. Source: google.com
The snippet is a list of video titles featuring Emad Mostaque, including a claim that AI is collapsing security and that one's economic life expectancy ends in two years. It contains no supporting detail. Source: google.com
Mistral Studio is pitched as a faster way to build and deploy enterprise agents, workflows and apps. Customers keep full control over their models, data and infrastructure. Source: mistral.ai
Franklin Templeton is building AI workflows, beginning with equity research. The tools are meant to help portfolio managers compare research views and probe investment risks. Source: openai.com
Emad Mostaque raises a troubling question about superintelligent AI: what happens when a system can continuously learn, persuade and predict human behavior. The framing compares it to a dictator that humanity cannot outlive. Source: threads.com
Emad Mostaque argues that AI development is moving toward a permanent two-tier system. As frontier models grow far more capable, he says AI labs will gain an advantage that others cannot match. Source: google.com
Earlier this year, footage of retirees and young people queuing for hours to install the OpenClaw AI agent went viral in China. The enthusiasm, however, appears to be running into obstacles despite government support. Source: bloomberg.com
Developers were promised that AI agents would lighten their workload, yet running several agents in parallel has turned into a job of its own. AWS is aiming to cut that supervision burden. Source: theaieconomy.substack.com
A guide to calculating the cost of serving an AI agent, with a full formula that accounts for inference, retries, tool calls and context tokens, items many startups overlook. Source: startupfortune.com
Airbnb says its upcoming AI agent will help travelers plan trips, compare options, manage bookings and get support during their stay. Source: techcrunch.com
Cyberhaven's Agentic Data Security is designed to let enterprises adopt AI agents without losing control over their data. Source: helpnetsecurity.com
A survey by FIS Inc. found that 50% of U.S. adults want to sign off on every transaction an AI agent might make on their behalf. The poll also asked about the conditions under which people would trust such agents. Source: digitaltransactions.net
Anthropic disclosed that its AI systems may be nearing recursive self-improvement (RSI), a threshold that could speed up AI development while creating new safety risks. One expert quoted in the coverage calls 2026 the 'last normal year'. Source: google.com
A study reported by Reuters finds that Anthropic's AI models are advancing almost twice as quickly as before. The speed-up comes as AI takes on a larger role in building more powerful systems. Source: google.com
Clément Delangue, CEO of Hugging Face, one of the most influential open-source AI platforms, downplayed fears of existential AI risk. He argues that open source brings the danger close to zero. Source: google.com
Anthropic's S-1 filing and other transparency measures may be commendable. Ultimately, though, the public will have to be protected by legislation rather than company disclosures alone. Source: news.bloomberglaw.com
On Tuesday, voters slammed the brakes on data centers and elected governors and legislators who promised to protect individuals from AI-related threats. The results signal a period of tougher scrutiny for the technology. Source: google.com
Since early September, when Jacob Coxon resigned from Anthropic over AI safety concerns, a wave of dire warnings has followed. The piece weighs the good and bad news on AGI and climate change. Source: google.com
An explainer on the idea that artificial intelligence could evolve beyond human control and wipe out humanity. The piece argues that signs of such a scenario are increasingly pointing in that direction. Source: google.com
European organizations will be able to use Claude models inside Microsoft Copilot while keeping data within the EU Data Boundary. The rollout promises consistent security and compliance standards. Source: microsoft.com
The GitHub app for Microsoft Teams is now generally available, bringing GitHub Copilot, repository management and development workflows into a shared collaboration space. Source: google.com
Adobe's CX Enterprise Coworker is arriving in ChatGPT, Claude, Gemini and Microsoft Copilot. Adobe argues that marketers too often use AI tools in isolation. Source: thenextweb.com
Microsoft is giving Windows 11 users a way to reclaim the keyboard function displaced by Copilot. The dedicated AI key can now act as Right Ctrl or the Menu key. Source: google.com
The new alpha release of datasette-apps adds support for hosting and editing custom HTML applications directly inside Datasette. They can be written by hand, though the plugin is designed to make creating them easier. Source: simonwillison.net
Users without the paid add-on face undisclosed limits in Copilot Chat that are now expanding. When features run out, upgrade prompts appear. Source: theregister.com
According to Caddi data, the largest RIAs mention Copilot, ChatGPT and Claude in AI-related job postings much more often than advisor-specific fintech tools. Source: investmentnews.com
MEMASA ($MEMASA) has outlined the AI copilot at the heart of its chat-first crypto trading app. The announcement came from Dubai on November 5, 2026. Source: streetinsider.com
The platform integrated the new Sunburst and Flare models and now runs them next to rival image generators. It also ran a side-by-side text-rendering test comparing them against other image models. Source: markets.businessinsider.com
Google has released Gemini Omni 1.2 Flash, a video generation model that produces native 15-second clips. Pricing is set at $0.07 per second of video. Source: marktechpost.com
Google announced on November 5, 2026 that its Gemini Omni 1.2 Flash video generation and editing model is generally available. The update raises the length of natively generated clips. Source: unite.ai
Dreamina AI announced an informational guide on how an AI video generator can support creative projects. The announcement was issued on November 6, 2026 from New York. Source: google.com
A developer describes wanting automatic cover images for a side project, a small tool that turns blog drafts into shareable posts. The piece walks through adding image generation without reworking existing code. Source: techguide.com.au
Pangram now supports video uploads and wider image screening, moving beyond its text-detection roots. The change brings automated AI-content verdicts closer to everyday browsing. Source: remio.ai
Pangram has introduced Pangram Image 1.0, a detector for AI-generated images and video. Source: pangram.com
Beijing has blocked SoftBank's $5.4 billion bid to acquire ABB's robotics division, a setback for the global tech investor. The decision underscores China's tightening grip on strategic AI-related technology. Source: wsj.com
Agility Robotics, billed as the first pure-play humanoid robotics company, starts trading Thursday morning on the Nasdaq under the ticker AGLT following a SPAC merger. It positions itself as a direct competitor to Tesla's Optimus. Source: investors.com
The US gets its first publicly traded pure-play humanoid robotics company as Agility Robotics begins trading on the Nasdaq under the ticker AGLT. Source: forbes.com
JEPA, or joint embedding predictive architecture, is an AI design proposed by Yann LeCun that predicts outcomes in an abstract representation space rather than in raw data. A new paper puts the approach to the test against far larger models. Source: mindstudio.ai
NVIDIA researchers have developed a visual planning method called VFlash System One, introduced by Jim Fan. It is presented as a new approach to vision-based planning. Source: tldr.tech
Yann LeCun has long warned that the AI industry is making a major mistake. A new paper may be the strongest evidence yet that he was right. Source: youtube.com
A model with 109 million parameters outperformed one with 72 billion at matching a vision model. The result is cited as support for Yann LeCun's view, while the industry pours billions into the opposite approach. Source: google.com
Google DeepMind co-founder and chair Demis Hassabis says the world is not yet prepared for the next stage of AI. In conversation with Reid Hoffman, he discusses how AI could become dangerous. Source: youtube.com
Demis Hassabis says Google was probably mistaken to hold back LaMDA. The piece draws 10 lessons for founders and investors on launch bars, objective functions and AI safety. Source: the-ai-corner.com
Demis Hassabis argues that AI systems are great tools for analyzing data and surfacing insights but cannot ask the right question themselves. The Google DeepMind co-founder sees this as a key limit of current AI. Source: google.com
Jim Fan, Director of AI and Distinguished Scientist at NVIDIA's GEAR lab, argues that robotics is missing a key kind of model. He made the case on stage, explaining why teleoperation has reached its limits. Source: google.com
Maximum inner product search (MIPS) is a core building block of recommendation engines, search, and retrieval-augmented generation (RAG). Making this step faster directly improves how efficiently these systems run. Source: machinelearning.apple.com
Gemini Omni 1.2 Flash introduces native generation and editing of 15-second videos, along with improved consistency across outputs. The update also lowers pricing by 30%, making the model cheaper to run. Source: deepmind.google
Nearly every AI system we know starts out the same way: by absorbing huge amounts of real-world data such as books. This new work explores a different path, where a model learns to predict the world without direct exposure to it. Source: google.com
Demis Hassabis says no one knows what will happen with AI, and that anyone claiming otherwise is lying or has ulterior motives. He spoke on the uncertainty around AI consciousness. Source: google.com
South Korea's URobotics, a physical AI startup, has raised a Series A of 40 billion KRW (about $29.9 million), it said on November 5. The round reflects a roughly tenfold jump in the company's valuation. Source: app.dealroom.co
CesiumAstro closed a $730 million Series D led by Janus Henderson, valuing the company at more than $4.5 billion. The round follows a fully successful Element-1 mission, and the company also revealed its Synchronicity satellite constellation. Source: spacenews.com
New York-based Kalepa, an insurance AI startup, raised $75 million in a Series B led by Updata Partners at a $470 million valuation. The round comes five years after its Series A. Source: techfundingnews.com
Shield AI plans to announce on Thursday that it is acquiring a radar company for $5.3 billion. It is also raising $2.5 billion in fresh funding tied to the deal. Source: bloomberg.com
The acquisition of Echodyne will expand Shield's Mission Systems business into the fast-growing radar and counter-drone markets. The deal comes alongside a $2.5 billion raise at a $28.8 billion valuation. Source: shield.ai
A newsletter published in November 2025 argued that Google would sell TPUs to outside customers. The authors say events since then have repeatedly proven that prediction right. Source: newsletter.semianalysis.com
Chinese AI lab Moonshot AI has reportedly pursued leasing deals with European neocloud providers to get access to advanced Nvidia chips. U.S. export controls block the direct import of such hardware into China, pushing the company to seek capacity abroad, sources told CNBC. Source: google.com
NVIDIA says it deliberately co-designs its GPUs, CPUs, networking, storage and software as one system. The goal is to squeeze more useful work out of memory at every layer of the AI factory. Source: blogs.nvidia.com
AIskimIQ is an independent project. If you find it useful, you can support its development with a coffee.
Published by Martin Ševčík
6 November 2026 at 06:09
A model with 109 million parameters just beat one with 72 billion at matching a vision model, and the industry's response should be more uncomfortable than it is. That's the headline result from Yann LeCun's latest JEPA paper, and I think it deserves more attention than the usual benchmark noise. JEPA, the joint embedding predictive architecture, predicts outcomes in an abstract representation space rather than reconstructing raw data pixel by pixel or token by token. LeCun has argued for years that the industry's bet on ever-larger generative models is a dead end. Now he has a result that is hard to wave away: a model roughly 660 times smaller coming out ahead on a specific task.
I want to be fair here. One paper, one task, and a matching benchmark is not proof that scaling is finished. But the second-order effect matters. Billions of dollars are flowing into data centres on the assumption that bigger is the only road. If representation-space prediction keeps delivering at a fraction of the compute, the economics of the whole sector start to look different, especially for anyone building products who can't afford a 72-billion-parameter inference bill.
By the way, a different kind of efficiency story is hiding in the weight-space programming research, where one LLM directly edits another's weights. The reported sixfold jump on Qwen's MATH-500 score is striking, and the piece rightly spends time on the conditions needed for genuine recursive improvement. I'd treat the number with caution until others reproduce it, but the idea is genuinely different from conventional fine-tuning. It treats a model's parameters as something to be programmed rather than merely trained. If that holds up, it changes who gets to improve models and how fast.
Meanwhile, the deployment side is showing its own friction. Tencent launched Hy4, an open-source LLM it positions among the global top tier, which keeps up the pressure on Western labs from the open-weights direction. Yet inside China, the OpenClaw agent craze, with retirees and young people queuing for hours to install it, is reportedly hitting obstacles despite government backing. Enthusiasm and infrastructure are different things, and agents seem to be learning that everywhere. AWS is now going after what may be the most honest complaint in the field: developers were promised relief, and instead found that running several agents in parallel is a job of its own. Supervising your automation is not automation.
Put these together and a theme emerges. The next phase may be less about raw size and more about whether systems are efficient, reliable, and cheap enough to run unattended. So here is the question I'll be watching: if a tiny model can match a giant on perception, how long before the giants have to justify what they cost?