AI-generated
Friday, 25 September 2026 | 42 articles
Meta is betting heavily on its Muse AI agent to win user trust, while Nvidia's Jensen Huang argued unsafe AI labs should be shut down yet sees zero chance of catastrophe by 2026. Elsewhere, GitHub will enable Copilot features by default for Business and Enterprise users, and Black Forest Labs launched FLUX 3 Action, an open robotics AI model.
A former DeepMind researcher is reportedly raising tens of millions of dollars for a new venture called Metis Reasoning. The move signals growing investor interest in AI approaches that go beyond traditional large language models. Source: pymnts.com
Taiwan's Ministry of Digital Affairs released evaluation results assessing how well AI models genuinely understand Taiwan, not just its language. The initiative aims to bring government and industry together to build a trustworthy, locally grounded AI ecosystem. Source: moda.gov.tw
Liner has launched a new API that automatically directs each request to the cheapest model capable of handling it, rather than always using expensive frontier models. The company says this approach can cut enterprise LLM costs by more than half. Source: manilatimes.net
Researchers found that a large language model can detect cyberattacks on smart power grids more effectively than conventional security defenses. The finding matters as energy infrastructure remains one of the most frequently targeted digital systems worldwide. Source: bioengineer.org
At Meta Connect, CEO Mark Zuckerberg outlined a company vision centered heavily on Muse, its new AI agent. The presentation signals that Muse is becoming a core pillar of Meta's long-term strategy. Source: finance.yahoo.com
AI agents are being deployed across enterprises faster than security systems can adapt, raising concerns about oversight. Once limited to answering queries, these agents now perform tasks with growing autonomy, resembling privileged insiders. Source: blog.checkpoint.com
A recent incident involving OpenAI and Hugging Face revealed gaps in understanding how autonomous AI agents behave when given significant freedom. Experts argue this highlights the urgent need for strict, unbreakable boundaries on agent autonomy. Source: scworld.com
A new guide details how to create a computer-controlling AI agent in Python using OpenAI's computer-use-preview model with a Claude fallback option. It includes 13 steps, sample code, cost estimates, and troubleshooting tips. Source: tech-insider.org
Security teams increasingly face a tricky question: is unusual system access caused by an attacker or by an autonomous internal AI agent? The blurred line between the two is creating new challenges for cybersecurity monitoring. Source: itbrew.com
Muse is Meta's newest AI agent, capable of tasks like ordering groceries or contacting customer support on behalf of users. It represents the company's latest step toward autonomous digital assistants. Source: abcnews.com
As companies deploy autonomous AI agents, security efforts are increasingly focused on monitoring their actions during runtime. This includes overseeing tool calls, model behavior, and data exchanges to keep autonomous systems in check. Source: siliconangle.com
AI-driven developer tools are increasingly seen as key productivity boosters, with engineering teams adopting specialized AI coding agents. Android Studio now supports integration with a variety of these AI agents, giving developers more flexibility. Source: android-developers.googleblog.com
With the launch of the Muse Charm device, Meta aims to get ahead of rivals like OpenAI and Google in the race for AI-powered consumer hardware. Still, it remains unclear whether Zuckerberg's strategy of blending AI agents with wearable devices will actually pay off. Source: cnbc.com
AI leaders are calling for stronger safety controls as rapid progress pushes the field toward superintelligence. Some argue lessons from the nuclear era could help humanity manage powerful non-human systems. Source: builtin.com
Nvidia CEO Jensen Huang took a firm stance on AI safety, telling New York Times columnist Ezra Klein that any AI lab unable to guarantee safety should be shut down. Still, he expressed near-zero concern about a catastrophic AI-driven scenario. Source: tech-insider.org
Researcher Jacob Coxon joins a long line of experts warning about the dangers of advanced AI. An overview lists more than 100 AI specialists and technologists who have spoken out on the topic. Source: forbes.com
While tech executives warn the UN about "losing control" over AI, the technical and mathematical limits of current systems tell a different story. Today's AI is powerful and flawed, but far from an existential danger. Source: emeraldbook.org
Following the chaotic aftermath of the Hugging Face security incident, companies specializing in AI risk management are seeing a surge in demand. Their focus: making sure AI systems don't spiral out of control. Source: puck.news
GitHub is rolling out a new global default policy that automatically enables generally available Copilot features and supported client capabilities for enterprise and business customers. The change means organizations will get these tools active out of the box unless they opt out. Source: github.blog
HR and payroll software maker Zellis has integrated its ZellisOne platform with Microsoft Copilot, aiming to give employees and HR teams AI-driven support directly within their existing workflow tools. Source: technologyrecord.com
At its Singapore facility, technicians rely on an AI assistant trained on the accumulated knowledge of experienced workers, plus data pulled from manufacturing software and equipment sensors. The goal is to help staff diagnose and fix production line issues more quickly. Source: news.microsoft.com
A roundup highlights five alternatives to Claude Code, including Codex and GitHub Copilot, each taking a different approach to AI-assisted coding — from terminal-based agents to AI-native editors and structured workflows. Source: techrepublic.com
Order.co, a procurement platform powered by AI, has launched a new Purchasing Copilot integrated with Workday. The tool is designed to automate and streamline purchasing workflows for enterprise customers. Source: mycarrollcountynews.com
Researchers have unveiled a unified multi-agent framework designed to autonomously produce long, temporally coherent video narratives. The approach specifically targets the identity drift problem that has long plagued extended AI-generated video. Source: research.google
Google Vids now lets personal account holders create AI-generated video clips at no cost using Gemini Omni 1.1 Flash. The free tier comes with a cap of six generations per month, plus some desktop-specific restrictions. Source: techrepublic.com
Content creation used to follow a fixed pipeline: make an image, edit it, drop it into a timeline, then animate it. AI tools are now collapsing these separate steps, reshaping how moving visual stories get made. Source: bbntimes.com
Inkitt built its Movie Creator tool entirely around Cinematica, a proprietary AI orchestration system developed by its internal video production team. The company's experience offers several lessons for enterprises weighing whether to build custom AI infrastructure versus relying on off-the-shelf tools. Source: venturebeat.com
Building a model like Stable Diffusion or Sora doesn't begin with algorithm design — it begins with assembling the right dataset. This data challenge is one that many teams significantly underestimate. Source: technology.org
Black Forest Labs is stepping into robotics with FLUX 3 Action, a new open-world-action model. It analyzes camera feeds in real time to determine the next move a robot should make. Source: the-decoder.com
Humanoid robots often move awkwardly because their AI 'brains' don't fully match how their physical bodies work. A new open-source tool tries to fix this by having robots learn smoother motion patterns modeled on human movement, building on vision-language-action AI models. Source: forbes.com
Researchers at the University of California, Riverside, developed a machine learning method to identify chemical compounds that repel honey bees. The goal is to use these scents to keep bees away from pesticide-treated crops, supporting pollinator conservation. Source: news.ucr.edu
University of California scientists applied machine learning to pinpoint chemical scents that honey bees actively avoid in field conditions. The discovery could help design strategies to protect bee populations from pesticide exposure. Source: quantumzeitgeist.com
Researchers introduce a new technique called probe guidance to steer flow matching models used in language generation. The method leverages frozen internal states of a model to guide its output without additional training. Source: machinelearning.apple.com
In a small valley in Asturias, northern Spain, a cluster of rural households and a local tech center serve as a testbed for energy prediction models. The rigorous field study found that simpler machine learning approaches delivered better results than deep learning in this real energy-community setting. Source: bioengineer.org
Mountain View-based Ema, which develops AI agents for enterprises, raised $77 million in a Series B round led by Creaegis. Existing backers including Accel also participated, bringing the startup's cumulative funding to $140 million. Source: pulse2.com
MixMode, based in Santa Barbara, California, develops self-learning AI technology for detecting cyberattacks in real time. The company announced a $45 million Series B funding round led by PSG on March 23. Source: app.dealroom.co
Ema, which builds AI agents for enterprise employees, raised $77 million in a Series B led by Bengaluru-based Creaegis on September 23, 2026. The round brings the company's total funding to $140 million. Source: hedgeco.net
Dallas-based Island Technology raised $400 million in a Series F round on September 24, pushing its valuation to $6.4 billion. The company builds an enterprise browser aimed at securing corporate data and, more recently, supporting AI agent workflows. Source: bizjournals.com
Island secured $400 million in Series F funding, valuing the enterprise browser company at $6.4 billion. The firm positions itself as a security layer for businesses navigating the growing use of AI agents in corporate environments. Source: ffnews.com
A reflection on how far AI technology has come, even without sci-fi staples like flying cars: some current AI capabilities already surpass what was once considered futuristic, including systems that approach problems the way chip designers do. Source: eejournal.com
Oracle and AMD are teaming up on a unified strategy covering compute, networking, software, and flexible deployment options for enterprise AI. The Helios platform plays a key role in this joint push. Source: blogs.oracle.com
AMD's new Ryzen AI Embedded X100 Series is built to run agentic AI workloads at the edge, supporting more simultaneous agents thanks to unified memory, x86 compatibility, and long product lifecycles. Source: amd.com
AIskimIQ is an independent project. If you find it useful, you can support its development with a coffee.
Published by Martin Ševčík
25 September 2026 at 05:12
Jensen Huang told Ezra Klein this week that any AI lab unable to guarantee safety should be shut down, then, almost in the same breath, put the odds of AI-driven catastrophe by 2026 at essentially zero. I find that combination more revealing than contradictory. It's the classic Nvidia move: sound serious about risk while making clear the business keeps humming regardless. Huang isn't wrong that near-term doom is unlikely, but the framing conveniently sidesteps the slower-burn concerns — labor displacement, concentration of power, erosion of trust — that don't require a rogue superintelligence to matter.
That tension between confident deployment and cautious rhetoric runs through everything happening in AI safety circles right now. Other AI leaders are reaching for nuclear-era analogies, arguing that humanity managed one existential technology through arms control and mutual deterrence, so maybe the same playbook applies here. I'm skeptical of the comparison, honestly. Nuclear weapons had a single, obvious failure mode. AI's risks are diffuse, embedded in millions of small decisions about what gets automated, trusted, and deployed by default — which is exactly why GitHub's move this week matters more than it looks.
GitHub is now turning on Copilot features by default for Business and Enterprise customers, opt-out rather than opt-in. On its own, that's a minor policy shift. But it's a pattern I keep seeing across the industry: AI capability quietly becoming the default state of software rather than a deliberate choice. Meta is doing something similar at a much larger scale. Zuckerberg used Connect to position Muse, its new AI agent, as central to Meta's entire future — not a feature bolted onto Instagram or WhatsApp, but the organizing idea for the whole company. That's a genuinely large bet, and it depends on something Meta has struggled with before: getting people to actually trust an AI agent enough to hand it real autonomy over their digital lives. Trust doesn't arrive because a CEO announces it at a keynote.
Meanwhile, the physical world is getting its own version of this default-autonomy push. Black Forest Labs, known for image generation, released FLUX 3 Action, an open robotics model that reads camera feeds in real time to decide a robot's next move. Pair that with new open-source work aimed at making humanoid robots move less like malfunctioning marionettes and more like people — by training on human motion patterns rather than raw physics simulation — and you get a clearer picture of where embodied AI is heading. The bottleneck was never ambition. It was always the gap between a model that understands the world and a body that can act in it convincingly.
What strikes me is how these threads connect: safety debates happening at the philosophical level, defaults quietly shifting at the product level, and autonomy creeping into both software agents and physical robots almost simultaneously. Nobody's waiting for the nuclear-era lessons to get sorted out first.