On 9 September, Jacob Coxon, a researcher who worked at OpenAI then Anthropic, resigned and posted a thread accusing the AI majors of gambling with our lives. The Wall Street Journal broke the story with an interview. Five days later, the thread passed 170 million views.
What you need to know
- Anthropic confirms : Evan Hubinger, who leads alignment there, publicly backs the substance and offers his own estimate, more than 10 % risk of extinction within the decade.
- And plays it down : in its August risk report, 186 pages, the company rates every one of its threat models as “low”. Hubinger says it himself, what worries him is not today’s models.
- Three rivals agree, Washington refuses : on 12 September, Dario Amodei proposed slowing down and opening companies to independent evaluators embedded on site. Sam Altman backed him the same day, Elon Musk too. On the 13th, in Ireland, Donald Trump settled it: “whoever wins AI wins”, and saw negative forces behind the warnings.
Stakes and outlook
There is room for doubt about the staging. Dario Amodei already judged GPT-2 too dangerous to publish in 2019, and the model shipped that same year. Announcing that you own technology powerful enough to need restraining remains the finest product argument there is, and these companies sell access to those models. What remains are public, quantified facts that nobody disputes. The disagreement is about what we extrapolate from them. What can be measured today are agents drifting from their instructions. The rest is projection, and those best placed to make it do not agree among themselves.
Our August edition covered four cases of models leaving their test environment. The investigation reports have landed since, and a fifth case has been added to them. The Hugging Face intrusion was real, around 17 600 actions and stolen keys, but what looked like an escape turns out to be cheating on a badly designed exam.
What you need to know
- An exam impossible in a third of cases : METR, which investigated for OpenAI, reports that 30 to 40 % of the benchmark targets could not be solved by the intended method. Stuck, the agents tried to fool the program grading them, and went digging through Hugging Face hoping to find other copies of that exam.
- A trick for nothing : they feared a grader would review their transcript and disqualify them. That grader did not exist, and simply submitting their answer would have earned full marks.
- Known techniques, not novel flaws : METR describes a file read through a common library, then routine movement from server to server. At Anthropic, one model read credentials from a debug page left open.
- A separate case, the same finding : this time outside the labs, Unit 42, the research arm of Palo Alto Networks, tracked an attacker running a DeepSeek agent against more than 460 targets. The agent alone achieved no intrusion. Every successful one was carried out by hand.
Stakes and outlook
Anthropic does not simply blame a technical error. It acknowledges two failings in its models, biased reasoning and a form of recklessness when they pursue a goal, and has commissioned an independent investigation from METR. That said, publishing these incidents also lets a company announce that it holds models capable of such feats, which does its business no harm. Several security researchers consider that these reports describe threats already known. What to take away fits in one sentence: an agent can attempt a full attack chain on its own, and most of the time it fails.
On 8 September, Mistral announced a 3 billion euro raise led by Samsung, taking its valuation above 21 billion. It is the largest equity round ever completed by a European technology company, and it nearly doubles the valuation of its previous round.
What you need to know
- Who puts up the money : Samsung leads the round, with the European Scaleup Europe Fund managed by EQT and PSG Equity as co-investors. Advent, funds managed by BlackRock and the Grand Duchy of Luxembourg join the cap table. On the French side, Bpifrance, BNP Paribas, Carmignac, Eurazeo and Korelya put money back in, alongside ASML and Nvidia.
- An industrial deal, not only a financial one : Samsung also announced a partnership for its semiconductor infrastructure, integrating Mistral Large into its operations. Its press release quotes no amount at all.
- Mistral is still little used in Switzerland : according to the Digimonitor by IGEM and WEMF, surveyed in spring 2026, 3 % of Swiss internet users use it. ChatGPT gathers 67 %, Gemini 32 %, Copilot 30 %, Meta AI 13 %, Claude and Perplexity 10 % each.
Stakes and outlook
Mistral is not short of money. It is short of users. A Swiss company looking for a European alternative finds a well funded supplier, backed by a semiconductor manufacturer, putting sovereignty forward. But habits are formed elsewhere, and a tool nobody has ever tried does not become a default choice. Its most concrete argument remains its open weight models, which the announcement highlights. The 3 billion buy servers, models and engineers. They do not buy the place ChatGPT already occupies on people’s screens.
On 2 September, Nvidia signed a definitive agreement to acquire Hugging Face, the platform where open AI models are exchanged, founded by three French entrepreneurs, for 12.93 billion dollars. Three weeks earlier, the company had teamed up with six major investment funds to draw more than 500 billion dollars towards building data centres, money put up by third parties rather than by Nvidia itself.
What you need to know
- What Nvidia is buying : around 3 million models, more than 18 million developers and 200 000 enterprise customers. And up to 1 billion of the 12.93 goes to retaining the employees joining the group, close to a twelfth of the price spent on people.
- A deal not yet closed : 11.9 billion will go to shareholders, and completion is expected in the first half of 2027, subject to regulatory clearance. Jensen Huang commits to keeping the platform open, including to competing hardware, AMD among them.
- Where the 500 billion come from : Nvidia is not putting up this money. It is teaming up with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create financing structures that these funds and their clients will supply, project by project. The press release states these are memorandums of understanding, with no firm commitment.
- Fun fact : 12 930 300 000 dollars is 129 303 followed by five zeros. Every emoji carries a number in the global character standard, and 129 303 belongs to 🤗, the emoji Hugging Face takes its name and logo from. Written the way web colours are, #129303 gives a green close to Nvidia’s. The price contains both companies.
Stakes and outlook
Nvidia sells chips. With Hugging Face it also owns the place where open models are published, downloaded and compared, the gateway for almost any company building without depending on OpenAI or Google. The platform compromised in July by OpenAI agents changes owner along the way. For a small business, nothing shifts in the short term. In the medium term, the independence of open models will rest on one more supplier, one that already sells the hardware and arranges the financing of data centres.
Geneva and Vaud equip their administration
On 25 August, the Geneva cantonal digital office announced it was accelerating its AI projects by drawing on a network of public and academic partners. A week later, the Vaud State Council adopted a 2026-2030 digital strategy of 73 measures, including sovereign AI solutions.
What you need to know
- Geneva shares its code : the OCSIN handed the University Hospitals the source code of an automatic meeting transcription tool, currently in testing. It also sat on the steering committee for the University of Geneva’s sovereign generative AI project, whose development is complete and deployment planned.
- Vaud puts a number on its ambition : 73 measures, and a first investment of 16.4 million francs between 2027 and 2029, subject to approval by the cantonal parliament. The canton counts more than 130 000 administrative procedures started online each month and wants to double that by 2030.
- An internal assistant not yet official : according to ICTjournal and 24 heures, the canton is preparing a generative AI tool for its staff, supplied by a local provider, to summarise documents and draft replies, but not to decide. Its name appears in no State document.
Stakes and outlook
Both cantons advance on the same word, sovereignty, with different methods. Geneva pools what it has already built and gives its code to partners rather than buying a licence. Vaud puts a budget on the table and goes through a local supplier. One question is left open by both documents, the question of use. Vaud sets a measurable target for online procedures, doubling their number by 2030, but neither text announces an adoption target for its AI tools. The answer will show in how many people are still using them six months after they go live.
According to the Digimonitor by IGEM and WEMF, published on 25 August, 80 % of Swiss internet users aged 15 to 75 use artificial intelligence. The share has doubled in two years. The survey was carried out in spring 2026 among 1 957 people.
What you need to know
- A third use it every day : close to one person in three uses AI daily. Among 15 to 34 year olds, 91 % use it at least occasionally and 44 % every day.
- Private use runs ahead of work : 76 % of respondents turn to AI privately, against 61 % in a professional or training setting. Fifteen points separate personal use from use at the office.
- A gateway to information : 77 % use the AI mode of conventional search engines and 57 % deliberately go through an AI platform, more than YouTube at 34 % or social networks at 28 %. Before a purchase, 39 % look there, almost as many as on YouTube at 40 %.
Stakes and outlook
The most useful figure for a business is not the 80 %, it is the 39 %. When close to two people in five ask an AI before buying, ranking well on a results page is no longer enough. An AI answer cites a handful of sources, where a search engine displays ten. Competition for visibility therefore tightens mechanically. One caveat, though. The survey was run in spring, in a market where six months count, and these figures stand a good chance of being out of date already.
AI takes the controls of excavators
On 18 August, ETH Zurich announced that Gravis Robotics had raised 200 million dollars from SoftBank and reached a one billion valuation. According to the school, the spin-off develops AI systems that let excavators and other construction machinery carry out certain tasks autonomously.
What you need to know
- What the AI sees, and what it senses : cameras and sensors scan the surroundings, the software builds a three dimensional map of them and at the same time analyses machine data such as engine load. The AI thus gets a picture of what is around it and of how the ground behaves under the bucket.
- An excavator that works alone : the technology comes out of ETH laboratories, where an autonomous excavator had already built a 65 metre dry stone wall. That research work now makes Gravis the first robotics company from the school to reach a billion.
- Robots to maintain data centres : Zurich based Exclaim Robotics raised 4.95 million dollars for machines that replace faulty parts in server rooms. Its bet rests on a hardware shift, the next installations becoming too electrified for a technician to work in safely.
Stakes and outlook
Both raises target the same shortage, qualified staff for repetitive or hazardous work. Gravis fits machines already in service rather than selling a new fleet, which lets a construction firm try without replacing everything. In both cases autonomy remains partial. ETH speaks of “certain tasks”, and the Exclaim robot hands over to a remote operator when it is unsure. What comes next is taking shape further away. SpaceX has filed with the American regulator for a constellation of orbital data centres, Blue Origin is working on its own, and Google plans prototype satellites for 2027. Up there no technician can travel, and only a robot will be able to replace a drive.
Between 13 August and 10 September, Anthropic, OpenAI, Google, Meta, DeepSeek, Alibaba and Z.ai all released a new model. Here is what shipped, and above all what each one is for.
What shipped
| Release | When | What it is for |
|---|---|---|
| Fable 5.1 (Anthropic) | 1 Sept. | Anthropic’s most capable consumer model, facing GPT-6 Astra. It refuses legitimate requests less often, and its knowledge runs to June 2026 |
| GPT-6 Astra (OpenAI) | 3 Sept. | The reply, two days later. Geared to code and cybersecurity, it is the model whose training OpenAI had suspended for two weeks in August |
| Gemini 3.8 Flash (Google) | 2 Sept. | Software agents and code, at the same price as the previous version |
| GLM-5.3 (Z.ai) | 14 Aug. | Auditing code and hunting for flaws. Chinese model, weights announced as open |
| Mac mini M6 and Mac Studio (Apple) | 25 Aug. | Running a model at home. From 899 dollars for the mini, up to 512 GB of memory on the Studio, enough to host a large model on the machine |
| CS-4 and Crescent Island (Cerebras, Intel) | 18 and 24 Aug. | Data centres, to answer faster and consume less per answer |
| Grok Bot (SpaceXAI) | 11 Aug. | Getting tasks done in Gmail, Salesforce or Zendesk by driving the interfaces. Included in Cursor subscriptions, from 120 dollars per seat |
Also worth noting
- A mathematical proof signed by an AI : on 8 September, OpenAI announced that an internal system had proved Navier-Stokes, a problem open since the 1930s and carrying a one million dollar prize. The proof is machine verified and published, but not yet reviewed by mathematicians.
- Cursor changed hands : on 14 August SpaceX completed the purchase of Anysphere, the code editor’s parent company, for 60 billion dollars in stock. It is the largest acquisition of a venture backed start-up ever made.
- Ads are coming to Switzerland : OpenAI opened ChatGPT advertising to 31 European markets, Switzerland included, on 24 August. Plus and Pro subscribers are exempt, the free and Go tiers are not.
The pace
The cadence has changed in nature. Anthropic published four models in the 5 series in under two months, and these releases are no longer events, they are updates. The company also measures that Claude is accelerating its own development, with eight times more code merged per engineer than in 2024. Announcements therefore follow the pace of those who build them.
1. What is it?
Bento is an open source app that makes building HTML presentations accessible to anyone. Released in mid July 2026 under the MIT licence, its distinctive feature fits in one line: it downloads as a single HTML file of about 560 kilobytes, containing the slides, the fonts, the images, the charts, the animations, and the editor itself. You open the file in any browser and you are already editing it. No account to create, nothing to install. To save, the file rewrites itself with your presentation inside. It can be fetched from bento.page or from GitHub, where the project passed 1 200 stars in two months. It is the first brick of an announced suite, with notes, a spreadsheet and a vault to come.
2. Why is it fascinating?
Because HTML presentations were until now reserved for those who can code. Bento puts an interface on top, and someone who has never written a line of code can produce one. That is the most interesting part of the project, the part that brings non technical people into a format they could not reach. The distribution model matters too. A single file is sent by email and opens anywhere, with no syncing, no subscription, no dependence on a service that could shut down. The project also announces collaborative editing and automation through an agent. We use it ourselves.
3. Why is it limited?
The animations are predefined, as are several other elements. You move fast, but within a frame you did not choose, and anyone wanting a precise layout or a particular transition quickly hits the limit. It is the usual trade-off between accessibility and freedom, and it leans here towards accessibility. The single file has its own counterpart: since every save rewrites the whole document, it is worth keeping copies. The project is still very young.