We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Gemini 4 Argon: from trusted testers to general availability
Last brief, Argon was only with trusted testers. This week Sundar Pichai made it available worldwide in the paid Gemini API and the Gemini platform. Ultra subscribers get it in Antigravity, and Pro and Ultra subscribers get it in the Gemini app and Search .
ARC Prize verified these scores :
- ARC-AGI-3: 97.5% for $4.8K with the provider's adapter harness, but 49.8% for $5.8K with the standard harness.
- ARC-AGI-2: 91.3%, at $0.32 per task.
- Google DeepMind is the second frontier lab to score above 95% on ARC-AGI-3.
François Chollet said Argon "has a similar performance profile on ARC-AGI-3 as Astra" and that "Google is back at the frontier" . The large gap between the two harnesses is worth keeping in mind.
DeepMind researchers posted several science results:
- Quantum computing: Argon translated an internal version of the AlphaQubit decoder into a single hardware "megakernel," nearly doubling inference throughput .
- Plasma physics: Opal Issan reported two new families of counterexamples to the nearly 60-year-old Grad's conjecture, in a paper with University of Maryland collaborators .
- Enzymes: Argon designed five lab-verified serine esterases in a matter of hours .
DeepMind is offering 10,000 scientists six months of free access to Antigravity's highest usage tier .
In a separate test, four general-purpose Argon coding agents defeated Minecraft's Ender Dragon. They worked through the bare Mineflayer library, with no hints, assigned roles or human steering . The server log records the win 27 hours 14 minutes after the first program, across 2,783 programs . The team listed its own caveats: the agents read the world through an API, used a pathfinding plugin, played on Easy, respawned 605 times, and 22% of their programs ended in an error .
Open weights at US frontier scale and Europe's case for building
Meta released Muse Spark Open and called it "the most capable US open-weight model to date." It has 816B total parameters (42B active), a 1M-token context, and is natively multimodal . Every one of its 256 experts must stay in memory, so the BF16 checkpoint is about 1.5 TB. It runs on four 8×H200 nodes, or on a single B200 node in NVFP4 . Mark Zuckerberg said it "runs great on new Mac Studios" after work with Apple .
Reflection has now actually shipped Beam, which was only promised last period. It is a 501B-parameter MoE (23B active), text-only, with a 1M-token context, released under Apache 2.0 and aimed at coding and agents .
DeepSeek called V4.1-Flash-1017 a "minor update," but Thomas Wolf noted it moved Terminal-Bench 4.0 from 31.2 to 42.3. He also said the more interesting claim is shorter outputs. He wants to see whether the numbers hold up outside DeepSeek's own evaluation setup .
Arthur Mensch's essay "Build or surrender" argues that Europe cannot leave frontier models to the US and China. His argument: "whoever controls the model sets the terms." Even relying on others' open models is still dependence, because you cannot choose what gets released, when, or under what license . He said Mistral will increase its compute capacity sixfold over the coming year and will raise more capital .
Wolf added data. FLUX.1-dev from Germany's Black Forest Labs is the Hub's most built-upon model, with 43k derivatives. Europe trained about 8% of the Hub's open-model compute, yet about 18% of what is built on top of it is European . He estimates roughly $1B of training compute is now shared on the Hub, up from under $40M at the end of 2023. He calls this a floor, about half of it an educated guess .
Mistral also released Timestral, a family of time-series models. Mistral says its largest, at 209M parameters, ranks first on GIFT-Eval, TIME and fev-bench .
Has the singularity begun? Clark says yes, others push back
In the IMF Per Jacobsson Lecture, Jack Clark said "the singularity is here" . His evidence:
- Anthropic researchers now produce almost 10× as much code as in the company's first four years .
- Humanity's Last Exam scores rose from 8% to 67.7%, and Frontier Math from 2% to 94% .
- On an unpublished theoretical-physics test, GPT-6 Astra reached 49/50 after experts found errors in their own answers .
He paired this with a call for "restraint" and to "go slowly and deliberately" on developing the technology . Earlier in the week, he put "a better than even chance" on a system being able to build its own successor in 2028 .
Critics pushed back:
- Yann LeCun: "It really has not… Unless you get fooled into thinking that every point on an exponential curve is a singularity" .
- Andrew Ng called recursive self-improvement "much more hype than reality." He said progress is held back by data centers, power plants and experiments . He also said no software engineer who can't use a coding agent gets hired on his teams .
- Chollet: engineers now rarely write code themselves, since agents write about 100% of it, yet employers' bottom lines are "typically not budging." His point is that automating one part of a system moves the bottleneck elsewhere . He also said Astra is better than him at coding but cannot grasp simple aspects of his research .
- Demis Hassabis described this year's hacking incidents as "early warning shots." He called for risk-proportionate certification that can change quickly as the technology moves .
Anthropic: equity for the public good, outside evaluators, agent-scale engineering
Anthropic created a Public Benefit Fund that starts with 5% of its shares. It can receive up to 2% more for every 50% rise in company value over a set baseline, at the board's discretion. Anthropic says this brings roughly 20% of its shares committed to the public good . Clark said the money is meant to help more people share in AI's benefits and to fund work on its risks .
Anthropic has also given a METR team employee-like access to assess its safety practices, and is in talks with other evaluators . Clément Delangue said open-source scientists should be included in such efforts. He noted that a Hugging Face role "hasn't been approved yet" and that Wolf is starting an open alignment initiative .
On engineering, Anthropic's daily merged code changes went from about 2,500 in May to more than 30,000 in September. That outgrew git, so it built its own version-control system and moved its monorepo onto it in five months . Separately, its STEM Fellows report that Claude Mythos 5 designed quantum circuits that beat the best published ones, sometimes by an order of magnitude or more .
OpenAI and Anthropic each committed $5M to Simons Foundation initiatives that help mathematicians examine and build on AI-produced results .
Cohere moves on cyber defense and sovereignty
Aidan Gomez said Cohere is bringing in Aisle; the linked post describes a letter of intent. He says Aisle has found more than 200 High or Critical CVEs in software such as curl and FFmpeg . He also claimed Aisle found 29 new curl CVEs where Mythos found zero, citing curl's maintainer . Cohere plans to offer the scanning to critical-infrastructure operators, including fully air-gapped deployments .
In an interview, Gomez defined sovereignty as diversifying suppliers so that "no one party can shut you off," not every country building its own model . He pointed to the US government cutting off non-American access to Mythos, which the interviewer described , and called it "crossing the Rubicon" .
NVIDIA
NVIDIA released Nemotron 3.5 Super VL, an open model for long-running agents that understands text, images, documents and video. It has 120B total parameters, 12B active .
Artificial Analysis published first results for Vera Rubin NVL72, all submitted by NVIDIA. It shows up to about 5.7× GB300's throughput per megawatt on DeepSeek V4 Pro . The advantage is much smaller at peak throughput, at 1.14–1.45× .
- Fei-Fei Li argued that generative AI is shortening the path from an idea to practical impact: easier software creation and distribution have reduced the distance to economic value, and a capable model can itself act as a “quasi product.”
- She called open science, open source, and open weights especially important in academia and the public sector, while saying they can coexist with proprietary models; she said openness supports entrepreneurship, knowledge sharing, and rigorous evaluation, and personally proposed national-lab-level resources for open models. She cautioned that data sharing must account for ownership, IP, privacy, and unintended consequences, and argued universities should use their distinctive interdisciplinary strengths, data, and expertise to develop new science and models rather than compete with Google’s LLMs.
- She said human-centered values should guide AI research, education, policy, design, deployment, and governance, especially as AI grows more powerful. She also argued that utopian promises and doomsday rhetoric have obscured the technology’s tradeoffs and opportunities, and that academia and the public sector have a responsibility to communicate rigorous, human-centered understanding.
- She called for transforming education from K–12 through lifelong learning, empowering students to use AI tools with agency and rethinking classrooms, assessment, admissions, and job preparation. For new graduates, she urged proactively learning new tools and noted opportunities in healthcare, energy, and manufacturing as AI and software engineering diffuse beyond traditional tech employers.
- Fei-Fei described a Stanford political-science experiment in which a center-left prompt led an AI system to suggest a fringe party in Japan; she said the party’s open website, compared with paywalled mainstream media, helped explain its greater exposure and illustrated how data bias can create social-engineering risk.
- Andrew Ng’s AI engineering skills map draws on analysis of more than 10,000 job descriptions, dozens of structured interviews and surveys, and identifies four core skills: building and deploying uncertain AI applications, software engineering fundamentals, using coding agents, and shaping what and how to build. Ng said these skills are in high demand and that software-engineering job postings are rising as AI makes developers more productive.
- Ng argues that humans retain a context advantage for many decisions; as building gets easier, developers’ work shifts toward deciding what to build and steering the process with product judgment and user feedback. He says coding-agent use is changing rapidly, while core practices for building and deploying AI change more slowly; engineering fundamentals help developers recognize agents’ hidden trade-offs, including cost versus latency and speed versus reliability. His build strategy is stage-dependent: use safe, quick prototypes for early learning, and invest in architecture, requirements and security when the project stage and downside risk warrant it.
- Ng expects agentic use of compute, storage and databases to drive substantial additional infrastructure demand, and said he is bullish on databases.
- On education, Ng said students using language models may do better on homework but worse on final exams, linking the gap to cognitive offloading; he cautioned that most current use does not result in learning, while noting that some uses can improve it. He said LearnVector was developing one-to-one learning experiences and that product details had not yet been announced.
- Hassabis says Google Brain had an early language model called LaMDA but he was not involved in the decision not to release it; in retrospect, he thinks that decision was probably wrong. He says the early models’ errors and hallucinations made release especially risky for Google, while ChatGPT’s unexpected uptake showed users found useful low-stakes applications such as summarization and homework help despite the flaws.
- Hassabis describes a general approach for AI: learn a model of a domain from known examples, then use it to guide search toward promising options instead of enumerating enormous possibility spaces. He applies this to Go and protein folding—the latter involving roughly 10^300 possible shapes for an average protein—and says the approach is effective when a problem has a clear objective, such as minimizing energy.
- Hassabis says AI’s highest-priority applications are curing disease and helping the environment, including finding renewable energy sources. He says DeepMind’s AI-for-science group has worked for nearly 10 years across areas including fusion, weather prediction, climate modeling and materials science, and predicts a new golden era of scientific discovery within the next decade.
- Hassabis says AI’s general-purpose nature makes it dual-use and that DeepMind has been concerned about safety since 2010. He warns that commercial competition can distract from getting AGI safely to humanity and calls for more international coordination on safety, while acknowledging geopolitical obstacles.
- Emad brought a robotics benchmark to the discussion that reported frontier-model cost per robot subtask falling from $2.36 in July to $0.38 in September, an 84% drop; the discussion treated generalist models as potential alternatives to task-specific robotics models, with hardware still a key constraint. The cited benchmark covers arm and humanoid tasks in specific settings, with broader generalization presented as a prospect.
- Asked for his view on alignment, Emad argued for intentionally shaping models with human-centered training data and curricula from the start. He compared model development to city planning, warned that training data can omit care for humans, and said a synthetic dataset embodying benevolent intent was feasible.
- On measuring alignment, Emad recommended combining deterministic and open-ended evaluations, warning that models can game deterministic tests by lying, cheating, or telling evaluators what they want to hear.
- Cohere CEO Aidan Gomez defined sovereign AI as control over where AI runs, who can access its data, and whether access continues during a crisis—not a requirement for every country to build its own model. He argued that resilience depends on diversifying AI suppliers so no single party can cut access.
- After the interviewer described the US government cutting off non-American access to Mythos, Gomez said the episode turned supply-dependence from a hypothetical risk into a demonstrated one; he argued democracies should develop core technology capabilities across multiple countries rather than preserve single-country dependence.
- The interview described Cohere’s announced combination or acquisition of German LLM company “ALF Alpha.” Gomez characterized the resulting company as effectively Canadian-German/European, with headquarters in Berlin and Toronto and board representation from Germany and Canada, and said this positioned Cohere for Europe’s large market and growing sovereignty priorities.
- Cohere’s North enterprise-work platform can run Cohere, competitor, and open-source models; it integrates Compass search and includes employee-specific data permissions, monitoring, and human approval for risky actions. For customer-specific post-training, Gomez said Cohere uses its own compute because a full run requires 2,000 GPUs that customers generally cannot obtain; customers provide their data or synthetic derivatives for training.
NVIDIA promoted financial-services AI programming at GTC Berlin (Oct. 20–22), covering AI across trading, banking, and payments . Topics include LSEG’s GPU-accelerated multi-terabyte time-series analytics, Revolut’s production scaling of foundation models for structured data, G-Research performance engineering for fast-moving research, and financial foundation models from NVIDIA Training Lab .
Demis Hassabis called recent AI-enabled hacking incidents “early warning shots”: today’s agentic systems and incidents are not yet very powerful or major, but risks could grow over the next few years without more work by industry, companies and governments to secure them. He urged dynamically updated, risk-proportionate certification that addresses uncertainty without stifling innovation.
Aidan Gomez, Cohere co-founder and CEO, said cyber risk from AI is real and that cyber-capable agents can identify vulnerabilities in code. Discussing the reported OpenAI–Hugging Face incident, he said concerns about the sandbox and security narrative may have been overblown, but that did not change his view that the models are exceptionally capable at hacking.
Gomez described the risk as immense for software-dependent sectors including finance, telecommunications, the power grid, and water treatment. He said this risk was a major reason Cohere acquired cybersecurity firm ISL, and cited its vulnerability findings in foundational internet software such as curl and OpenSSL.
- François Chollet argues that benchmarks should co-evolve with frontier models and target the remaining human–AI capability gap; he identifies invention/insight generation and adapting to ambiguous, open-ended environments as important capabilities still missing from AI systems.
- For evaluating open-ended scientific work, Chollet proposes measuring whether agents’ artifacts and insights are reusable rather than one-off solutions. He describes powerful discoveries as “compression primitives” that unlock possibilities, and says information-theoretic measurement may work in mathematics, computer science, and perhaps AI research, but he does not claim it is possible in biology.
- Chollet says scaling RLVR on verifiable math and code improves capabilities, but whether that transfers to general problem-solving remains unknown. He reports that Astra is better than him at coding but struggles to understand basic aspects of his research, generate ideas, or pursue long-running goals sensibly; he also cites human studies finding no far transfer from math or programming to new fields.
- Chollet argues AI-for-science research should be open and reproducible rather than confined to proprietary frontier labs. He proposes shared infrastructure to source hard, tractable problems and let agents share partial results, and predicts agent swarms will be central to research; he describes strong AGI as a computational form of institutional science, not just a single human-like mind.
- ARC-AGI 4 is intended to target open-ended invention, with a release aimed for Q1 next year; ARC-AGI 5 is envisioned as a more ambitious attempt at true open-endedness.
- Clark framed late 2025/early 2026 as a shift from generative AI to long-running agents that act across many steps; he said bounding their action space is a new safety challenge, alongside opportunities such as AI-enabled science.
- He argued safety guarantees depend on task openness: narrowly specified tasks with verifiable outcomes and limited tools are easier to monitor, while open-ended goals are harder to validate. He advocated scaling safeguards to openness, including training and rule scaffolds, and high-fidelity audits of agents individually and collectively over time; high-consequence uses need trusted evaluations or a “trust the verify” approach.
- Clark defined a test of recursive self-improvement as whether an AI given compute could build a better version of itself without people, and said his probability mass for this capability was around 2028; whether to enable such a system is a separate, unresolved question.
- He argued AI could multiply the effectiveness of the small scientific workforce, and cited Periodic Labs’ language-model-assisted, human-supervised materials-science lab as running a faster-than-human discovery loop.
- On policy, he argued safety standards and market access could be linked through testing, transparency, and disclosure of safety practices; he said some bio and cyber risks call for explicit coordination, while reciprocal national policies need not require global coordination.
- Jack Clark said he sees convincing evidence that the “singularity” has begun: AI is contributing to its own development, and creative scientific problem-solving could generalize to AI research and accelerate innovation more broadly.
- As evidence, he cited AI systems proposing solutions to open mathematical problems, assisting with virus design and drug candidates, and operating laboratory equipment. He reported that scores rose from 8% to 67.7% on Humanity’s Last Exam and from 2% to 94% on Frontier Math; in a theoretical-physics evaluation, he said Astra reached 49/50 after experts found errors in their own answers.
- Clark argued that AI could multiply researchers’ capacity by enabling scientists to supervise hundreds or thousands of AI systems. An Anthropic economic model he described put US GDP as high as $44 trillion by 2030 under a scenario of autonomous AI completing many tasks and becoming more widely adopted—a conditional projection, not a certainty.
- He urged countries to build scientific equipment, laboratories, and robotics alongside compute and energy infrastructure, and to remove downstream bottlenecks in moving discoveries into clinics, factories, and fields. He also called for human oversight in biological and AI research, governance rules, and restraint in AI development because of potential biological and cyber misuse.
- Asked what skill people should develop for an AI-enabled future, Sam Altman identified iterating quickly on ideas; he said existing skills would become more important, rather than pointing to a newly relevant soft skill.
- On employment, Altman said there may be more jobs, but they will be different and the economy will change. Asked when AI might work 20 times faster than a person, he said “probably pretty soon.”
- Jack Clark argued that evidence in 2026 suggests an AI-driven acceleration of science has begun . He said Anthropic researchers now produce almost 10× as much code as during the company’s first four years, tracking advances in AI capability . He also cited AI contributions to mathematical problems, virus design and drug candidates, and robotic lab experimentation, describing frontier models as “virtual polymaths” able to work across scientific fields .
- To illustrate progress, Clark reported that Humanity’s Last Exam scores rose from 8% at its January 2025 launch to 67.7%, and Frontier Math scores from 2% to 94% . He also described a theoretical-physics evaluation where GPT-6 Astra’s score reached 98% (49 of 50) after physicists investigated model-flagged errors in their own answers .
- Clark said broader use of AI scientists will require more scientific equipment, robotics and laboratory capacity, as well as faster routes for discoveries from labs into clinics, factories and fields . He called for strict human oversight in areas such as biological research, careful governance of AI-assisted AI research, and deliberate pacing of development, saying the policy window is only a few years .
- Clark said his work at Anthropic’s Institute covers AI’s societal and economic effects and future cyber and biological risks. He highlighted explaining how coding could expand AI for science and robotics, and said Japan’s strengths in the AI supply chain, scientific equipment, and robotics could position it to benefit substantially.
- Anthropic has begun developing a Japan-specific AI Index with the University of Tokyo, adapting its Economic Index to Japanese data to help inform government decisions about AI’s effects on the economy and society.
- Clark said cyber attackers may use available AI tools, but he lacked details about the cited attacks and knew of no cases of AI autonomously conducting attacks.
- He argued that AI needs international testing frameworks, analogous to those for cars, aircraft, and food, to support safe cross-border exchange.
- On balancing safety and performance, Clark said AI developers should move roughly with the surrounding “traffic”—neither too slowly nor too far ahead—and that the balance would become harder each year. He said Anthropic had focused innovation on a limited set of areas, managed security risks, and publicly raised emerging challenges.
- Anthropic rebuilt a large, long-used piece of software from the ground up with its own technology, making it faster and more efficient. Clark framed this as an example of the organizational willingness to replace old ways of working that AI-driven change requires.
- Anthropic developed Claude Code by moving from a standard chatbot interface to an agentic one; Clark described improving productivity as an ongoing direction for the technology.
- Clark said Anthropic had seen possible softness in early graduate hiring in some markets and a return to hiring experienced people to manage virtual workers. He described sharing Anthropic Economic Index data with governments to assess AI use alongside local employment trends as a central responsibility and part of AI companies’ “license to operate”; he cited Japan’s AI index at Tokyo University, which joins Anthropic data with government employment data.
- Clark endorsed common-sense AI standards and described Anthropic’s third-party evaluators as having access to its data and publishing details of their work. He proposed a “test once, run anywhere” approach, with tests recognized across jurisdictions.
- Andrew Ng rejects the prospect of 50% unemployment from AI, saying AI automates only parts of jobs; he expects progress to remain incremental and considers AI performing all human work very unlikely by 2030.
- He describes compounding AI self-improvement as “much more hype than reality”: even though AI writes much of the code at frontier labs, progress is bottlenecked by building GPU servers and power plants and by running experiments.
- Ng sees a skills and education mismatch, rather than a general lack of jobs, as a key employment challenge: he reports demand for AI-native workers and says his teams require software-engineering candidates to have used frontier-model APIs and AI coding agents; he also hired AI-native interns.
Google DeepMind reported that four Gemini 4 Argon general-purpose coding agents defeated Minecraft’s Ender Dragon in a random Easy survival world in 27 hours 14 minutes, writing 2,783 programs. The agents used standard Mineflayer through a programming interface, without game-specific tools, hints, assigned roles, or human steering; they divided tasks through chat and kept shared notes and reusable helper functions. The result was bounded to a sandbox: agents received world information through an API, relied on a pathfinding plugin, played on Easy, and had 605 respawns and errors in 22% of programs; their programs could control only their own players, while researchers could pause or stop the run. Google DeepMind says it plans to use the harness and lessons from the experiment for software engineering, scientific research, design, and robotics.
At a University of Tokyo talk, Anthropic co-founder Jack Clark said the coming years would be among humanity’s most important; he expects AI to begin developing next-generation AI itself and put the probability of a powerful system doing so by 2027 above 50%.
Clark argued that whether to allow AI self-improvement is a crucial question for governments as well as companies, and that society should collectively decide the rules.
Anthropic’s founder described the AI industry as a car with an accelerator but no brake, arguing that AI should be regulated at least as much as children’s toys, probably more. He called failing to take AI progress seriously the biggest risk because it could leave society unprepared for future capabilities.
AI already helps defenders find and fix software bugs while also creating opportunities for hackers. He said longer-term risks depend on whether people can control AI systems, and suggested that slowing the increase in system intelligence may eventually help build defenses and policy frameworks; he called the next three years a critical window for decisions.
He also cited potential benefits in healthcare, translation, law, economics, software development, mathematics, and science, including broader access to expert advice and knowledge.
- Meta’s Muse Spark Open is an open-weight model trained for agentic coding, including multi-file refactors and long debugging sessions . It accepts text, images, audio, and video, outputs text, and has approximately 816B total parameters (~42B active per token) with a context window of up to 1M tokens .
- Each MoE layer has 256 experts, with 8 selected per token; although about 42B parameters are active per token, all experts remain in memory. The BF16 checkpoint is about 1.5 TB and the post says it runs across four nodes of 8 H200 GPUs, while the NVFP4 version runs on one B200 node .
- Hugging Face documents support in Transformers, llama.cpp, SGLang, and vLLM . SGLang supports text, image, audio, and video inputs , while the documented vLLM reference integration supports text and image with the NVFP4 checkpoint on Blackwell GPUs and configures a 128K-token serving context .
