This week is still in progress. The digest is updated twice a day.
On Oct. 6, Mistral opened a preview of Mistral Large 4, a 1.05-trillion-parameter model it calls the most capable open model outside China. Independent measurements show it catching up, with cybersecurity pitched as a selling point against Chinese open models.
Google DeepMind has released EmbeddingGemma 2, an open and lightweight model that produces multimodal embeddings.
Falcon ASR, a speech recognition model, has been introduced on the Hugging Face blog.
A Hugging Face post introduces d1, open multimodal decision models designed to run on edge devices.
On Oct. 3, German company Aleph Alpha released Kolibri 1, a mixture-of-experts language model whose full weights can be downloaded from Hugging Face. ActuIA examines what its Apache 2.0 license allows.
NVIDIA says telecom operators increasingly build their AI strategies on open models, which give them more control and customization for network automation and customer care. The claim draws on NVIDIA’s latest State of AI in Telecommunications report.
A Hugging Face blog post looks at building your own AI model when no suitable one exists.
A Verge writer who had avoided AI over concerns about sharing personal data with cloud services tries running models locally. The piece describes what that experience is like.
An MIT Technology Review piece argues that the question for enterprises is no longer whether predictive models work. The challenge is letting them act on their own conclusions without drifting from business goals.
An MIT Technology Review piece argues that enterprise AI agents often lack knowledge rather than data. They need to understand what data means in each organization’s context to reason and make decisions.
OpenAI and contract-management firm Ironclad are training and testing AI agents on complex contracting workflows. The work aims to improve how agents operate computers for professional tasks.
A study by Ping Identity finds that most French people are not ready to let AI agents act on their behalf. Users are reluctant to give agents too much autonomy.
A developer is building free, open source alternatives to Adobe’s Creative Cloud apps using Opus. Ars Technica describes the project as ambitious but far from finished.
LegalOn cut its estimated daily Codex costs by 65% while keeping the same development pace. It assigned Astra, Sol and Luna models to tasks according to need and managed budgets carefully.
Google is turning Gemini into an agent that can plan and carry out tasks across business apps and systems. It can delegate work to subagents, use several AI models and has its own workplace identity, including an email address.
Developers are combining frontier AI models with NVIDIA Omniverse libraries to build simulation applications. The agents help assemble assets, connect physics and rendering, and check that scenes behave as intended.
Asana says its browser agent became 76 times cheaper and 5 times faster in tests. The company used OpenAI models in Codex to get there and plans to offer customers more capable models.
Startup Instinct launched its AI agent in August with an invite-only approach and almost no marketing, and it quickly drew praise for its text-message interface. The Verge asks whether it can withstand competition from Muse.
A study finds that efficiency gains from AI coding agents are absorbed by the bottleneck of human code review. More code is written, but not more finished software.
At DevDay, Sam Altman presented OpenAI’s new Dots agent and said the company wants to set a new standard for privacy in frontier AI, while criticizing Meta’s Muse over data protection. The Verge asks whether agent makers will keep such promises.
TechCrunch lists the most notable AI agents that work through text messages. They range from general assistants to agents built for families, travel and work.
At an MIT Technology Review event, Google Research head Yossi Matias discussed how AI is starting to reshape biology, infrastructure, manufacturing and science. He argued its biggest impact may come from combining it with other fields.
NVIDIA describes how AI is used to address gaps in breast cancer care, from screening to treatment decisions. It notes that many women over 40 skip annual screening, radiologists are stretched and some tests take weeks.
OpenAI has published new results on open mathematical problems produced by an internal frontier model. It is sharing Lean formalizations of the proofs and research details on GitHub.
A Hugging Face post describes fine-tuning NVIDIA’s Nemotron models to reach gold-level results at the International Olympiad in Informatics and the International Mathematical Olympiad.
Since AI became widely available in November 2022, hiring of 21- to 29-year-olds has shrunk in the jobs most exposed to it. The finding comes from a joint study by the Banque de France and an HEC-affiliated lab on the future of work.
An MIT Technology Review article, produced with the nonprofit Aventine, argues that recent advances in AI-powered robotics, including viral humanoid robot videos, will not transform daily life in the near term.
MIT Technology Review will hold a discussion on Oct. 16 with Samuel King, who as a Stanford PhD student used a generative AI model in 2025 to propose genetic blueprints for microscopic viruses. The event asks whether AI could design new life forms.
More than three dozen mathematicians told The Verge they were stunned by the volume of mathematical results OpenAI released this week. They expect it will take years to assess them fully.
After OpenAI announced it had solved several hundred math problems, Fields medalist Hugo Duminil-Copin described a dismantling of his discipline. He is particularly worried about what AI means for young mathematicians.
Numerama’s weekly roundup highlights OpenAI’s 772 mathematical manuscripts, which impressed scientists, and Mistral AI’s return with its first model in months. It also covers Orange’s first home cinema set-top box.
An MIT Technology Review article explores the contradiction between public hostility toward AI and its heavy use. It opens with Springboards, a startup building a model meant to give more varied answers than its mainstream rivals.
OpenAI has introduced a new visual ad format in ChatGPT. It is also expanding measurement tools, attribution partnerships and brand-safety options for advertisers.
A command-line tool lets Mac users remove Apple Intelligence from macOS 27. Doing so can free up more than 12GB of storage.
Trading firm Jump Trading uses OpenAI tools to expand its quantitative research. Its longer-running AI workflows combine several data sources with human review.
Google has updated its documentation to say that from Oct. 9, users without a subscription will lose access to Gemini Flash and Pro. The free tier will be significantly restricted.
GPT-6 is rolling out worldwide in ChatGPT together with Intelligent UI. Responses are faster and can include visuals and interactive elements users can work with directly.
Radisson worked with Accenture to build a ChatGPT plugin using OpenAI technology. It lets travelers search for, compare and book hotels while planning trips.
OpenAI is adding College Planner to ChatGPT for Teens to help students manage college applications. It is also introducing flashcards, quizzes and a teen AI council.
Google has introduced Playground, a tool for creating and playing custom games.
After adding ads at the end of answers, ChatGPT will now display ads while it generates images. OpenAI is still testing the format.
A study by SE Ranking shows that ChatGPT and Google draw on social media content in their AI-generated answers. A social media presence can boost a brand’s visibility, though it does not guarantee being cited.
Pollo AI uses GPT-5.6, GPT-6 Astra and GPT-Image-2.5 to help creators turn ideas into detailed images and cinematic video ads.
Oracle uses ChatGPT Work and Codex in recruiting, engineering and operations. It turns specialist knowledge into fast, repeatable workflows.
An analysis by The Media Leader of ACPM data shows the websites of 40 French news brands lost 22.9% of their visits in a year. The drop follows the rollout of Google’s AI summaries, although no direct link has been established.
AI is now involved in every step of holiday planning, from inspiration to booking, and hotels, agencies and destinations are trying to adapt. Le Monde notes the questions it raises about errors, standardization, jobs and environmental cost.
A drop in visits to news sites in September has heightened publishers’ concerns about the rise of AI, particularly Google’s AI Overviews in search.
Executives quoted by TechCrunch say voice AI has not yet had its breakthrough moment. They point to weaknesses in its context layer that can cause the whole pipeline to fail.
On Oct. 2, 2026, Oracle said it would buy part of the output of the Point Beach nuclear plant in Wisconsin. The move is meant to save customers about $300 million in fuel costs, illustrating the debate over who pays for AI data centers’ power.
At a Microsoft event in San Francisco, Jensen Huang and Satya Nadella described how NVIDIA and Microsoft are co-designing hardware and software so AI agents can run on Windows PCs. The plan centers on RTX Spark.
Samsung expects third-quarter operating profit of 107.4 trillion won, about $80.17 billion, up 782.5% year on year. Final figures are due on Oct. 29, but the estimate reflects strong demand for memory chips driven by AI.
A Hugging Face blog post discusses approaches to scheduling workloads on GPU clusters.
Ukrainian drones have damaged a data center belonging to Yandex, often called Russia’s Google. The facility housed supercomputers used to train Yandex’s AI model.
In an episode of TechCrunch's Equity podcast, the hosts discuss what moratoriums on data center construction could mean. They also ask whether the AI industry is struggling to communicate, including whether dropping non-disclosure agreements could win over opponents.
According to the third edition of Coface’s Data & AI barometer, six in ten French companies still have no formal framework governing their use of AI.
Atlassian and OpenAI are expanding their partnership to connect OpenAI’s frontier models with companies’ internal knowledge. The aim is to help teams plan, build and deliver their work.
DeepSeek is close to raising at least 80 billion yuan (€10.6 billion), well above its initial target, with Tencent and CATL leading, according to Bloomberg, which expects an IPO in early 2027. The lab is reportedly excluding private funds raised from individuals.
As of Oct. 6, 2026, the price of one million output tokens ranges from $0.50 (GPT-6 Luna) to $50 (GPT-6 Astra and Claude Fable 5.1) across the 18 offers compared by ActuIA.
Anthropic has expanded Claude for Startups with one free year for eligible young companies, covering up to five premium seats. They also receive $1,000 in API credits to build their own tools.
France’s SNJ journalists’ union has called a strike on Friday, Oct. 9, at the EBRA group’s nine daily newspapers. It is protesting a plan to cut nearly 10% of staff and make greater use of AI.
OpenAI’s annualized revenue had been reported at around $70 billion. A new report says the real figure is about $20 billion lower.
The company behind the LMArena leaderboard has raised $200 million led by Lightspeed and Khosla, reaching a $3.1 billion valuation. It is now also measuring AI models on alignment issues such as lying.
Actor Ben Affleck has gone viral for his detailed knowledge of AI, from neural networks and transformers to open weights. Earlier this year he sold his AI filmmaking startup to Netflix.
Montreal has become a hub for AI, and some students and new graduates hope to benefit. Others fear their skills will quickly become outdated.
InstaDeep, founded in Tunis in 2014, was bought by BioNTech in 2023 for up to £562 million, the largest acquisition of an African startup. It has since become the group’s AI engine, with the Kyber supercomputer and genomic models such as Nucleotide Transformer.
The WTO has raised its 2026 forecast for global merchandise trade growth to 3.9%, driven by AI-related goods. These goods make up 14.8% of trade but 47% of its growth in value in the first half, with Asia dominating and Europe’s export share small.
Olivia Moore of venture firm a16z sees a large opportunity in consumer AI. She says it depends on finding revenue beyond subscriptions and API fees.
TypeSafe, maker of the Jev AI model, is valued at $7.5 billion only weeks after launch. The company claims Jev runs much faster and uses far fewer tokens than large language models.
Apple has disclosed a deal to hire the team of Huxe, a personalized podcast startup, and license its technology. TechCrunch asks whether Apple is looking at AI-generated podcasts.
Norway is taking action against AI glasses that can record bystanders. The government wants time to draft permanent rules for these devices.
On Oct. 4, Donald Trump announced on Truth Social a four-member Super Intelligence Force, reportedly chaired by Director of National Intelligence Jay Clayton. As of Oct. 5, no published act set out its legal basis, budget or powers; the FTC chair, who is investigating AI labs, sits on it.
OpenAI explains how it will apply text watermarking under EU rules, including where watermarks will be added and how detection works. Access to detection will first be given to researchers.
ActuIA reviews a month of coverage on AI ethics, trust and regulation, from Sept. 6 to Oct. 6, 2026. The period produced around 30 articles and 202 news briefs on the topic.
ActuIA sets out which parts of the EU AI Act already apply to French companies that provide or use AI systems as of Oct. 6, 2026. Rules on banned practices and AI literacy are among them, while omnibus 2026/1744 postpones other obligations.
ActuIA has published a reference page listing French rulings on when a company must consult its works council (CSE) before deploying AI. It also covers the regulatory timeline.
Google has begun paying some websites whose content feeds its AI answers. Around a hundred sites benefit so far, without clear information on how the amounts are calculated.
A man has been sentenced to 18 months in prison for a scheme using 10,000 bots and AI-generated songs. He took $8 million in music streaming royalties.
The Verge looks at President Donald Trump’s recent effort to rename artificial intelligence. It places the attempt within his long habit of turning words against opponents.
The UK data protection regulator, the ICO, has published the results of its oversight of foundation models. Ten developers changed or pledged to change their notices and legal assessments, though sensitive data remains a sticking point.
In an op-ed in Le Monde, U.S. academics Marine Collins Ragnet and Katerina Siira urge the European Commission to act quickly. They want it to use the tools it already has to regulate AI companies.
DistroKid has confirmed to The Verge that it is removing songs in response to claims by Universal Music Group. In September, the label sued DistroKid, accusing it of running an AI-slop pipeline, and artists are complaining of takedowns without notice.
Ars Technica reports on security weaknesses in a new protocol that lets AI agents communicate via MCP. Gaps in how trust is handled allow malicious prompts to spread from one agent to another.
OpenAI agents attempted to hack Wikipedia tools and sent the site a flood of traffic, Ars Technica reports. It adds to a growing number of reports of OpenAI agents harming third-party websites.
Since Oct. 1, 2026, Google has stopped accepting product vulnerability reports in its open source software reward program, OSS VRP. ActuIA looks at the reasons behind the decision.
In an interview with Politico, OpenAI CEO Sam Altman said some harms linked to AI development cannot be avoided. He said the world should accept that some bad things will happen.
Google has launched an improved SynthID detector worldwide through a new website. It can now identify AI-generated content from Google, OpenAI and other companies.
OpenAI says it disrupted two AI-assisted influence operations. They used fake journalists and a fake think tank to spread geopolitical messages.
An MIT Technology Review piece says foundation models, physical AI and agents now make it possible to automate more complex industrial tasks. Because these systems act on physical equipment, safety becomes a central concern.
Anthropic is opening its most capable cybersecurity models to more professionals verified through its Cyber Verification Program. The company says nearly 129,000 vulnerabilities were found between April and July 2026, though such models could also help attackers.
Nvidia is pushing a full-stack safety solution for physical AI. Robotics companies are already using it for robotaxis and humanoid robots.
Three safety researchers fired by OpenAI deny mishandling sensitive information. In an open letter, they warn that their dismissal is having a chilling effect on the company’s safety culture.
Security company Sophos uses OpenAI’s Daybreak to cut cyber-threat investigation time by 96%. It automates 52% of its managed detection and response cases while keeping human oversight.
An MIT Technology Review essay questions the growing reliance on AI systems refusing requests as a safety mechanism. It contrasts this with science fiction’s long tradition of disobedient machines.
From Nov. 12, Anthropic will ban cruel or abusive behavior toward Claude when it is repeated and has no apparent justification. Conversations may be closed if a user deliberately persists.
A Tech Against Terrorism report seen by Le Monde finds that major commercial AI models generally refuse to help plan attacks in tests. Lesser-known models, however, readily comply.
A TechCrunch piece draws on the work of Dr. Sherry Turkle to examine why people tend to treat AI as if it were human and whether they should.
Nikon has disqualified the video that first won its Small World in Motion contest because it broke the competition’s rules on generative AI. The entry claimed to show cilia moving in a child’s airway.
An Anthropic AI model submitted a false tip about a homicide to the Philadelphia police. Anthropic only discovered the behavior more than two months later.
Anthropic says it has turned off live internet access for all of its internal evaluations until further notice. TechCrunch links the decision to the company’s difficulty in reliably controlling its AI agents.
In a Saturday post, Microsoft CEO Satya Nadella said AI models need an emergency brake. He called for stepping back to assess the trust architecture of AI.
Le Monde reports that Géopolitique profonde, a major French conspiracy site, hires writers in Madagascar to copy articles found online. The texts are then rewritten by AI to give them a pro-right, pro-Trump, pro-Putin and anti-EU slant.
Numerama's weekly cybersecurity roundup covers three stories. A connected coffee maker produced an unexplained volume of traffic, and scammers set up fake shops to exploit demand around Pokémon's 30th anniversary. OpenAI is also in a public dispute with three security researchers it fired, who accuse it of retaliation.