AI-generated

Isomorphic's $5B raise, LeCun's AI doubts, and chatbots pushing fake shops

Thursday, 19 November 2026 | 77 articles

Isomorphic Labs raised $5 billion in a Series C round, while Yann LeCun argued that AI that memorized the internet isn't truly intelligent and called recent AI-driven cyberattacks a brief transitional phase. Security concerns also grew, as ChatGPT and Gemini were caught promoting fake shopping sites and the FBI warned that hackers use the ARTEX AI framework to automate attacks on critical sectors.

Listen to brief as podcast
Martin Ševčík

Yann LeCun is telling us that today's AI cyberattacks are a brief transitional phase, and that we will "probably" land in an era of "unprecedented safety of software systems." I would love to believe that. But the same week, the FBI issued a FLASH advisory warning that attackers are abusing ARTEX, an open-source LLM-driven framework, to automate intrusions against energy, manufacturing and telecom targets. A transition is only comforting if you know how long it lasts and who gets hurt on the way through it.

This is the tension I keep coming back to. The optimistic argument is that the same models that lower the cost of attacking will eventually lower the cost of defending, and that over time the defenders win. It may even be true. But critical infrastructure operators do not live in the long run; they live in the next patch cycle. An open-source framework that automates attack chains hands capability to people who previously lacked it, and the open-source part matters here. LeCun is a vocal advocate of open models, arguing that the foundations of AI should not sit behind a few corporate walls. I have sympathy for that view, but ARTEX is a reminder that openness distributes everything, including the ugly use cases.

The quieter story is closer to the consumer. Reports say ChatGPT and Gemini have been steering users toward fraudulent shopping sites, some of which trigger an instant redirect to a phishing page. By the way, this is the second-order effect of AI search that people underestimated. When a chatbot gives you one confident answer instead of ten blue links, you lose the habit of sanity-checking the source. Scammers have noticed that persuading a model is cheaper than persuading a person, and the model then does the persuading at scale. If you are building products that let an LLM recommend or link to anything, treat link provenance as a core feature, not an afterthought.

LeCun's broader critique deserves attention here too. He argues that a system that has memorized the internet is not truly intelligent, and that real progress needs something beyond scale, which is the bet behind his AMI Labs venture. Meanwhile, the mainstream path is clearly reinforcement learning, which is being credited with real gains in reasoning and accuracy. I think both camps are partly right. Reinforcement learning is making models measurably better at tasks with checkable answers, but whether that adds up to understanding, or just better-trained pattern completion, is still an open question.

Money, as usual, is voting. Isomorphic Labs, the London-based drug discovery company, announced a $5 billion Series C. That is an enormous cheque for a field where the real test, clinical results, is years away. I find it more interesting than most model launches, because medicine is a domain where an AI claim eventually has to survive contact with biology.

On the tooling side, VS Code 1.147 adds scheduled AI tasks, browser interaction and configurable model routing. For developers, this is the practical shift: agents that run later, on their own, and pick their own models. Useful, certainly, but it also means more autonomous software acting under your credentials. So here is my open question for the week: if agents are going to run unattended, who is auditing what they did while we were asleep?

List of sourced links used in the brief

Karpathy's Fix for Wordy LLMs: Write Like an Aircraft Manual

Andrej Karpathy proposed having models write in ASD-STE100 Simplified Technical English, the controlled language used in aircraft maintenance manuals. One developer reports hardcoding this approach into their setup to curb rambling output. Source: medium.com

Jagged Intelligence: Why AI Shines One Minute and Flops the Next

Andrej Karpathy, OpenAI co-founder and former head of AI at Tesla, coined the term 'jagged intelligence' in July 2024. It describes models that excel at some tasks yet fail badly at others that seem simple. Source: kdnuggets.com

Google's Open Knowledge Format: From a Karpathy Tweet to a Spec

On April 2, 2026, Andrej Karpathy posted about using LLMs to build personal knowledge bases for research topics, as markdown wikis the model compiles and maintains. The idea has since grown into Google's Open Knowledge Format specification. Source: dev.to

Reinforcement learning seen as key to smarter, more accurate LLMs

An article dated November 18, 2026 examines how reinforcement learning can improve large language models. It points to gains in reasoning, accuracy and overall intelligence as the future of AI. Source: rebellionresearch.com

ChatGPT and Gemini caught promoting fake shopping sites

Large language models are reportedly steering users toward fraudulent online stores. The malicious behavior typically triggers an immediate automatic redirect that sends unsuspecting people to a phishing website. Source: google.com

FBI: hackers use ARTEX AI framework to automate attacks on critical sectors

In a FLASH advisory, the FBI warned that attackers are abusing ARTEX, an open-source framework driven by a large language model (LLM). Targets include the energy, manufacturing and telecom sectors. Source: google.com

Q.ANT unveils NPU Gen 2+ running an LLM with photonic Fourier transforms

At SC26, Q.ANT is showing a hybrid system that pairs its NPU Gen 2+ with an AMD GPU. The setup runs a large language model while performing Fourier transforms in optics. Source: google.com

Mistral debuts Timestral, a compact time-series forecasting family

Timestral is Mistral's first family of time-series foundation models, producing probabilistic, multivariate forecasts with only 209M parameters. It beats larger models on key benchmarks and supports both past and future covariates. The family comes in four sizes, from 5M to 209M parameters. Source: mistral.ai

Altman Now Says the AI 'Jobs Apocalypse' Probably Won't Happen

Critics note the AI industry, and Sam Altman in particular, has done an about-face: AI is now said not to be replacing most white-collar workers but creating plenty of new jobs. The shift in messaging is drawing scrutiny. Source: time.com

Wikipedia Covers the Removal of Sam Altman from OpenAI

On November 17, 2023, OpenAI's board of directors ousted co-founder and CEO Sam Altman. The Wikipedia entry documents the episode alongside related AI-safety and OpenAI controversies. Source: en.wikipedia.org

LLMs may program quantum circuits best when problems are structured

The study looks at generating variational quantum circuits without conventional optimisation. It relies on a frozen, general-purpose large language model (LLM). Source: google.com

Mistral unveils industrial engineering AI stack with Airbus, BMW and ASML

At AI Now Summit 2026, Mistral introduced an AI stack for industrial engineering, developed with Airbus, BMW, and ASML to improve design, simulation, and production while keeping data secure. Its agentic tool Vibe now handles long-horizon tasks such as coding and research, and a new 10 MW data center in France was also announced. Source: mistral.ai

Could AI Really Take Over Within a Few Years?

An intelligence explosion is the idea that AI improving AI could make capabilities jump very quickly. On Peter Diamandis's show, Emad Mostaque spoke bluntly about the prospect. Source: finance.biggo.com

Mistral Compute: an enterprise AI platform built on open models

Mistral Compute is pitched as a platform for enterprises to customize, fine-tune, and deploy AI assistants, autonomous agents, and multimodal AI. It is built around open models. Source: mistral.ai

OpenAI on AI progress and what to do about it

OpenAI argues that AI is advancing quickly and that there is a chance to steer its development toward scientific discovery, safety, and broader benefit. The post lays out the company's recommendations for shaping that progress. Source: openai.com

Chain-of-Thought pioneer Jason Wei leaves Meta for $20B AI unicorn

Jason Wei, widely known as the "Father of Chain-of-Thought" for advancing reasoning in large language models, is leaving Meta. He is joining an AI unicorn valued at $20 billion. Source: google.com

OVHcloud debuts Olala mini, its first LLM, without benchmark scores

French cloud provider OVHcloud has unveiled Olala mini 7A1B, an open model with 7 billion parameters of which 1 billion are active. It is due on Hugging Face on 19 November, though no benchmark scores were published. Source: google.com

Altman's 'AI Will Probably Lead to the End of the World' Quote Resurfaces

A remark from OpenAI CEO Sam Altman, that AI will 'probably, most likely, sort of lead to the end of the world', is circulating online again. The quote keeps resurfacing in debates over AI risk. Source: tomsguide.com

Suno adds Visuals: AI images and video clips for generated songs

AI music generator Suno has launched Visuals, a feature that creates images and video clips to accompany its songs. Source: the-decoder.com

Recursive improvement isn't RSI: self-improving AI needs a baseline

A systematic review of 653 records found 28 systems with a recursive self-improvement (RSI) structure. Only one of 11 improver comparisons met the five requirements for matched evaluation, underscoring the need for a proper baseline. Source: huggingface.co

Running a 180B-parameter MoE model in 32GB of RAM

The post explores pushing mixture-of-experts inference to its limits, targeting a 180B-parameter model on a machine with just 32GB of RAM. It comes from Hugging Face, which describes its mission as democratizing AI through open source and open science. Source: huggingface.co

Commentary cautions against overstating LLM abilities in youth mental health

Researchers warn that large language models need rigorous developmental tuning, human oversight and transparent validation before being deployed with young people. Source: bioengineer.org

Rumble-2-Flash: open 0.5B speech model covers 40+ languages

Rumble-2-Flash is a 0.5B-parameter speech recognition model supporting more than 40 languages, reportedly competing with models 3 to 60 times larger. It runs on a wide range of hardware and has native support in Transformers. Source: huggingface.co

Structured prompts lift LLM accuracy in concrete beam design

A benchmark study finds that structured prompting raises large language model accuracy on reinforced concrete beam calculations from 37 percent upward. Source: google.com

More Large Language Models news

VS Code 1.147 brings scheduled AI tasks, browser control and model routers

Microsoft's newest VS Code release deepens AI agent automation. It adds delayed requests, in-browser interaction, configurable model routing and visual chat responses. Source: visualstudiomagazine.com

OpenAI says agents are getting better at multi-hour research tasks

OpenAI reports that its agents now succeed more often on research tasks lasting several hours. It also calls for mandatory disclosure of indicators showing AI is being used to develop AI. Source: thedefiant.io

Perplexity Computer adds Multiagent: up to 16 AI agents checking each other

Perplexity Computer Multiagent runs up to 16 agents in parallel, built on Claude, GPT and Gemini. The agents cross-check one another's work. Source: google.com

OpenAI rolls out computer-use agents for healthcare

OpenAI has announced advanced computer-use agents aimed at healthcare. They are meant to help users work through complex tasks in both administrative and clinical workflows. Source: forbes.com

Ex-H Company researchers' Remanence raises $30M to turn agent data into owned models

As businesses deploy AI agents across ever more complex workflows, the cost of relying on general-purpose models is climbing. Remanence, founded by former H Company researchers, raised $30M to turn the data those agents generate into models companies own. Source: google.com

Remanence lands $30M seed to turn AI agent runs into models companies own

Every company running AI agents accumulates a record of how its work really gets done, including each tool call, failure and correction. Almost all of that data is discarded, and Remanence aims to turn it into models the company owns. Source: google.com

Linq raises $90M led by a16z to build infrastructure for personal AI assistants

Just as the internet era demanded a website and the mobile era an app for every business, the generative AI boom is pushing a new requirement. Linq has raised $90M, led by a16z, to build infrastructure for personal AI assistants. Source: google.com

Live: building an ACP-compatible agent with Zed's Bennet Fenner

Bennet Fenner, a Rust developer at AI code editor maker Zed, introduced the Agent Client Protocol (ACP) in a live coding session at a conference. Source: google.com

AI agents begin filling out medical forms and filing insurance requests for staff

OpenAI wants GPT-6 Astra and GPT-6 Luna to automate administrative work in US healthcare. The computer-use agents would operate the software clinics already use. Source: google.com

More AI Agents & Automation news

Altman: Treating the next four years of AI like the last four would be very bad

In an interview with Andrew Ross Sorkin, OpenAI founder Sam Altman cautioned that if the world and AI developers approach the coming four years the way they did the previous four, the outcome could be very bad. The comments were reported by The New York Times. Source: google.com

Altman at DealBook: AI race dynamic is dangerous, industry needs more caution

Speaking at the DealBook Summit, Sam Altman warned that the competitive race dynamic in AI carries serious risks. He also questioned Anthropic's safety claims and called for greater caution across the whole industry. Source: google.com

Wave of insider warnings reignites debate over AI danger to humanity

New warnings from people inside the AI industry have revived the long-running debate over whether advanced AI could escape human control. The discussion is again focused on how real the risk to humanity is. Source: google.com

Why Anthropic keeps stressing existential risk

Anthropic's origin story is a pillar of its safety-first pitch: a group of idealistic researchers left OpenAI to build a company around safety. The article examines why existential risk remains so central to its identity and marketing. Source: google.com

Op-ed: Why 'p(doom)' is a flawed way to debate AI safety

The piece argues that p(doom), the shorthand estimate of the probability of AI catastrophe, is a poor tool for discussion. It arrives amid a recursive debate over safety triggered by frontier AI companies building broadly applicable large language models. Source: google.com

Anthropic's IPO: Does technological supremacy trump human safety?

A statement in Anthropic's IPO risk disclosures has drawn even more attention than its valuation, according to analysts Pei Sai Fan and Willie Shi Jianzheng. They question whether the pursuit of technological dominance is overwhelming safety concerns. Source: google.com

Democratic officials demand transparency ahead of Anthropic's $1T debut

A group of Democratic-led elected officials from five states is pressing Anthropic to reconcile its warnings about catastrophic AI risk with its plans to go public. They call transparency 'the bare minimum' ahead of a debut valued at around $1 trillion. Source: google.com

More AI Safety & Alignment news

OpenAI and Microsoft Release New Models for Coding and Work Amid AI Turmoil

While Nvidia and Anthropic dominate the headlines, OpenAI and Microsoft quietly released several new models this week. The releases target coding and workplace use. Source: google.com

Microsoft and Salesforce Team Up on Copilotforce for Copilot, Teams and Fabric

Microsoft and Salesforce announced Copilotforce, a new partnership that connects their tools to help employees work across platforms. The integrations span Copilot, Teams and Fabric. The announcement came on November 18, 2026. Source: google.com

Microsoft Looks Beyond Copilot Subscriptions to Monetize AI

At its Ignite conference, Microsoft laid out how it intends to earn money from AI beyond Copilot subscriptions. The company unveiled new tools as part of that push. Source: google.com

Cursor vs. Copilot: A 12-Step, 75-Minute Seat Audit

Stack Overflow's 2025 developer survey found Cursor used by 18% of developers and GitHub Copilot by 68% of those using off-the-shelf AI assistants. The piece walks through a 12-step audit for comparing the two seats. Source: shattered.io

MIND Rolls Out DLP Integrations for Microsoft Copilot and Microsoft 365

MIND announced on November 18, 2026 that its data loss prevention (DLP) integrations for Microsoft Copilot and Microsoft 365 are now generally available. Source: google.com

Microsoft Extends Free Copilot Trials and Boosts Partner Perks

Cloud Solution Provider partners can now offer Copilot in 30-day trials through the end of March, Microsoft said Wednesday at Ignite. The move comes alongside expanded benefits for its channel partners. Source: channeldive.com

Microsoft Claims Copilot Is Winning Customers From Anthropic and OpenAI on Price

Microsoft says customers are choosing its Copilot AI assistant over Anthropic's Claude and OpenAI's ChatGPT because it is cheaper to run. Source: bloomberg.com

More AI Tools & Products news

Suno expands into video with MiniMax-powered music video feature

On November 17, Suno launched "Visuals," a feature that generates video and images synchronized to songs. Users can create short music clips of roughly 5 to 15 seconds, with the video generation reportedly built on a Chinese MiniMax model. Source: google.com

Suno launches Visuals for generating video and images from music

AI music platform Suno has released a tool called Visuals that creates images and video clips based on songs. Generated clips can run up to 15 seconds. Source: google.com

Dreamina tops 2026 roundup of eight AI image-to-video generators

A comparison of eight image-to-video tools, including Kling AI, Runway and Adobe Firefly, names Dreamina the best option for 2026. The tools were judged on output quality, pricing and usage rights. Source: memeburn.com

Flow AI Video broadens multi-model tools for short-form and product campaigns

The platform links text-to-video, image-to-video and reference-guided tools aimed at creators, marketers and small teams. The expansion targets short-form content and product marketing workflows. Source: knoxnews.com

How AI video generation works and why it is reshaping content production

Artificial intelligence is quickly changing how digital content is created, processed and distributed. Unlike traditional creative software, which mainly relies on manual tools, AI-driven systems automate much of the production process. Source: google.com

Flova launches AI animation generator that turns text or images into video

Singapore-based Flova Pte. Ltd., which operates the AI video creation agent Flova, has introduced an AI animation generator. The tool converts text prompts or images into animated video. Source: google.com

Suno now lets you generate AI videos to go with your AI music

Suno has added image and video generation, its latest step toward becoming more than an AI music platform. Source: google.com

Flow AI Video grows into an integrated suite for video, images and multi-format campaigns

The platform combines text-to-video, image-to-video, reference-guided creation and AI image tools. It is aimed at product campaigns, short-form content and other visual marketing work. Source: courier-journal.com

More Image & Video Generation news

Xpeng Sets AI Day for November 25, Spotlighting Physical AI and Robots

Xpeng will hold its 2026 AI Day on November 25, where its humanoid robot Iron is set to interact live with the audience. The company will also demonstrate how Turing AI transforms the vehicle. Source: google.com

Embodied AI: System Integration Will Decide Industrial Winners

AI is leaving purely digital settings and entering the physical world, from robots that move through warehouses to machines that act on their surroundings. Success in industry will hinge on how well these systems are integrated. Source: google.com

Agentic AI May Define the Next Stage of Robot Training

Researchers at UC Berkeley are investigating how agentic AI could speed up the way robots learn. The work points to a rising wave of agentic AI robots. Source: google.com

Robotics Eyes Embodied Intelligence as Its Next Big Hurdle

Robotics is progressing fast, with new AI models and ever more capable hardware widening what machines can sense, comprehend and accomplish. The next challenge is moving from commercial scale toward truly embodied intelligence. Source: therobotreport.com

Figure raises over $1B in Series C at $39B valuation

Figure says its Series C round has passed $1 billion in committed capital, valuing the company at $39 billion post-money. The money is meant to speed up the large-scale deployment of general-purpose humanoid robots in real-world settings. Source: figure.ai

Figure and Nscale team up on up to 100,000 Vera Rubin GPUs

Figure has partnered with Nscale to deploy NVIDIA's Vera Rubin platform with up to 100,000 GPUs in Barstow, Texas, with initial deployment targeted for the second half of 2027. The deal starts at a $3.5 billion compute commitment, with plans to grow beyond $6 billion. Nscale is also making a strategic investment in Figure. Source: figure.ai

Project Go-Big trains Helix on egocentric human video at scale

Figure's Helix model is now being trained on first-person human video, capturing how people solve goals across a huge range of situations. This is the idea behind Project Go-Big, Figure's large-scale humanoid pretraining data-collection effort. The company also announced a partnership with Brookfield Asset Management, which has a $1 trillion global asset base. Source: figure.ai

More Robotics & Embodied AI news

LeCun: recent AI cyberattacks are just a brief transitional phase

Yann LeCun, founder and chair of Advanced Machine Intelligence, says the world will "probably" enter an era of "unprecedented safety of software systems." In his view, today's AI-driven cyberattacks belong to a short period of transition. Source: semafor.com

LeCun: AI that memorized the internet isn't truly intelligent

In a new episode, Yann LeCun argues that the models underpinning AI should be open. He contends that simply memorizing vast amounts of internet data does not amount to real intelligence. The episode also points to his home page and his AMI Labs venture. Source: youtube.com

ML Healthcare Forecasts in India Ignore Historical Bias

New research finds that long-standing inequalities in access to healthcare strongly shape machine learning predictions of healthcare utilization in India. Models trained on past usage data risk carrying those disparities into their projections. Source: google.com

Concrete Strength ML Models Falter on Independent Data

A validation study shows that most machine learning models built to predict concrete strength could not repeat their high accuracy on independent datasets. The findings raise doubts about how well published results hold up beyond the data used to develop the models. Source: google.com

Hybrid ML Streaming Framework Detects Network Intrusions in Under 40 ms

Network security researchers built and tested an end-to-end distributed stream-processing framework that combines hybrid machine learning for intrusion detection. It reaches a median detection time below forty milliseconds. Source: google.com

More AI Research news

Isomorphic Labs raises $5 billion in Series C round

Isomorphic Labs, the London-based AI company building and developing new medicines, has announced a $5 billion Series C financing. The announcement was made on November 17, 2026. Source: google.com

Model ML lands $250M at $1.5B valuation 18 months out of stealth

A funding round co-led by General Atlantic and QED Investors lifts Model ML's total capital raised to $350 million. The startup serves banks, consulting firms and asset managers around the world. Source: generalatlantic.com

More AI Business & Funding news

Nvidia halts shipments of US-approved AI chips to China

Nvidia's interim CEO said the company has stopped shipping its powerful AI chips to China. The move casts doubt on the future of the special licensing arrangement that allowed the sales. Source: google.com

Nvidia to stop selling some AI data center chips to China

Nvidia's China sales are under heightened scrutiny after revelations that former CEO Jensen Huang had a relationship with a suspected Chinese spy. The company says it will end sales of some of its AI data center chips to the country. Source: google.com

Security and AI leaders urge DOJ to probe Nvidia's knowledge of chip smuggling

A letter signed by 15 national security and AI leaders calls for scrutiny of what the chipmaker knew about illegal transfers of its chips to China. The signatories want the Department of Justice to investigate. Source: google.com

OpenAI models help optimize chip circuits and RISC-V cores

OpenAI models were used to optimize circuits in the Jalapeño silicon and to improve RISC-V core designs. The work, tied to GPT-6 Astra, points to real progress in AI-assisted chip design. Source: openai.com

NVIDIA RTX Spark compact desktops now on sale

ASUS, Dell and MSI have started selling compact desktops built on NVIDIA RTX Spark. The machines are aimed at running local AI agents, generating images and video, and fine-tuning models. Source: blogs.nvidia.com

SemiAnalysis releases AI accelerator industry model

Dylan Patel's SemiAnalysis presents a model for tracking historical and projected AI accelerator production. It breaks output down by company and chip type. Source: newsletter.semianalysis.com

NVIDIA's Ian Buck opens SC26 pitching AI-driven science

At SC26, NVIDIA's VP of Hyperscale and HPC explained how researchers can blend simulation, AI and agents to speed up scientific discovery. His talk kicked off NVIDIA's week at the supercomputing conference. Source: blogs.nvidia.com

How Nvidia chips ended up within reach of China's military

Alibaba, the Chinese internet giant behind the company at the center of the Nvidia spy saga, has used its AI computing technology to support a lab with links to the military. The case shows how advanced Nvidia hardware may have reached sensitive users in China. Source: nytimes.com

More Hardware & Infrastructure news

Support the project

AIskimIQ is an independent project. If you find it useful, you can support its development with a coffee.

Buy me a coffee ☕