From 15 to 17 June 2026, France hosted the G7 summit in Évian. For the first time, the heads of the world's three largest AI labs, Sam Altman (OpenAI), Dario Amodei (Anthropic) and Demis Hassabis (Google DeepMind), sat at the table of heads of state, alongside France's Arthur Mensch (Mistral AI). A historic photo that, above all, exposed how divided the democracies are on AI governance.
What you need to know
- An unprecedented lunch: on 17 June, leaders and a dozen AI bosses debated the “safe, fast and effective” deployment of the technology.
- A call for a US-led coalition: Amodei and Hassabis argued for structured access to frontier models and a chip trade that excludes China. Altman, for his part, handed responsibility back to governments: “Do not cede your responsibilities to labs like mine.”
- No binding text on AI: despite the event, no dedicated declaration was signed. The only digital text adopted concerns the protection of minors online.
- A sovereignty crisis in the background: five days earlier, Washington had forced Anthropic to cut off access to its frontier models for every foreign user. Leaders floated a “trusted partners” mechanism to restore access for allies, without finalising it.
Stakes and outlook
The Anthropic episode turned a theoretical debate on sovereignty into a brutal reality: the United States holds the bulk of the world's computing power, Europe barely 5%. This is exactly what is at stake for a player like Mistral, the only European voice at that table. Hence France's warning, calling for better regulation of these models without giving in to “non-cooperation between democracies”. The summit also laid bare the fault lines within the Western camp, between a Europe that wants to govern through risk and a United States that is loosening its rules to keep its lead over China. The real question is no longer whether democracies want to cooperate on AI, but whether they still can, when one of them alone holds the plug.
Switzerland ranks first in the world for the share of venture capital aimed at deep tech, with 63% of VC funding according to the Swiss Deep Tech Report 2026, published on 17 June. The challenge: turning its university research into AI, robotics and advanced computing companies that stay and grow at home.
What you need to know
- Highly concentrated funding: over 2020 to 2026, that 63% puts Switzerland ahead of China (56%) and the United States (54%), nearly double the share of Germany or the United Kingdom.
- A high effort per capita: with 1,470 dollars invested per person, Switzerland is first in Europe and in the global top 3, alongside Israel and the United States.
- A deepening talent pool: AI and machine learning now account for one in four new deep tech companies, and ETH Zurich and EPFL are Europe's leading universities for spinouts.
Stakes and outlook
Deep tech here refers to companies built on heavy advances such as precision sensors, robotics and chips, not simple consumer apps. Switzerland's strength is converting its research quickly into industrial companies, which draws in international funds. That is also its weakness: 88% of large rounds, above 100 million dollars, come from abroad, with only 12% from local funding. Without domestic capital for the later stages, the risk is clear: watching the very companies the country gave birth to leave just as they reach scale.
An anonymous conceptual artist, known by the pseudonym SHL0MS, posted a close-up of a real Monet on X, presenting it as an AI-generated image. The experiment triggered a wave of misguided criticism and reopened the debate on our ability to judge a work once it carries the “AI” label.
What you need to know
- A real Monet passed off as fake: the image came from a painting in the Water Lilies series, painted around 1915 and held at the Neue Pinakothek in Munich.
- Critics caught in the trap: several users described the image as an “incoherent”, “artificial” or “emotionless” AI production, without checking its origin.
- Experts identified the work: specialists pointed to material clues such as the brushwork, the impasto and the colours typical of late Monet.
Stakes and outlook
The whole affair comes down to one missing reflex: checking before judging. The experiment shows that a simple “AI” label is enough to flip perception, even faced with a museum-grade work. The problem is not the tool, it is the rush. For anyone handling images online, the defences exist and are concrete: reverse image search, consulting reliable catalogues, expert opinion. The real skill of the coming years will not be producing with AI, but keeping your judgement when everything pushes you to suspend it.
A “compute tax”, which would tax every unit of computing power used by AI, is back in the public debate. The real question: should we tax computation to cushion job losses and offset a shrinking tax base, or is it mainly the best way to throttle innovation?
What you need to know
- Where the idea comes from: nearly three quarters of US federal tax revenue rests on labour, through wages and payroll. If AI cuts demand for workers, that base shrinks, hence the idea of taxing computation to replace it.
- Who is for it, and why: in the lineage of the “robot tax” championed by Bill Gates, some economists and policymakers see it as a way to slow automation, fund social safety nets (even a universal basic income) and capture part of the wealth AI creates.
- Who is against it, and why: think tanks such as ITIF, Cato and Reason, along with many economists, call the tax counterproductive. It raises the cost of an input used everywhere, with a cascading effect on every downstream product, slows productivity and adoption (especially for SMEs) and pushes investment abroad, with data centres heading to Brazil, Mexico and the UAE, without slowing global AI. Many compare it to “taxing steel during the industrial revolution”.
- An alternative path: rather than taxing computation, which is an input, several economists prefer taxing the consumption of AI services, such as digital-service fees or taxes on AI-generated content, which raises revenue without discouraging investment.
Stakes and outlook
The debate pits two readings against each other. For its supporters, AI will concentrate value in capital and drain the labour tax base: without a new tax, the state would lose the means to fund social protection just when it is most needed. For its critics, the problem is badly framed. Capital's share of income has not really risen, the risk of mass unemployment is overstated, with at most 8% of jobs highly exposed, and taxing computation would be self-defeating, handing the AI field to other countries. One point of agreement remains, even among sceptics: if AI does end up displacing labour for good, governments will have to find something to tax, and the real issue will be less about throttling computation than about rethinking what we tax.
The Trump administration is in talks with AI companies about taking equity stakes that would let Americans profit financially from their growth, according to Bloomberg and CNBC. OpenAI is the leading candidate cited, with the idea of feeding a public fund from the company's shares.
What you need to know
- A direct public gain: CNBC reports that the Trump administration has discussed taking an equity stake with OpenAI. The company could hand shares to the state to seed a “Public Wealth Fund” whose returns would be redistributed directly to citizens.
- A recent precedent: the US government has already become a shareholder, taking a 10% stake in Intel last year, a chipmaker in difficulty.
- A debate that crosses party lines: Senator Bernie Sanders, for his part, proposes a binding route, an exceptional 50% tax paid in shares by OpenAI, Anthropic and xAI if they go public.
Stakes and outlook
The idea is simple: turn part of the value created by AI into a direct financial benefit for the public. But it changes the very nature of the relationship between Washington and the AI labs. The state is no longer just a regulator, it becomes a partner, even a shareholder, in an industry it is also supposed to oversee. That is precisely the risk flagged by David Sacks, the former White House AI lead: a tighter merger between big tech companies and government. For the companies, handing over shares means sharing their future gains and letting the state into their capital, at the very moment when several are preparing to go public.
UNIDIR, the United Nations research institute on disarmament, held the second Global Conference on AI, Security and Ethics on 18 and 19 June 2026 in Geneva and online. The aim: to bring diplomats, the military, researchers, industry and civil society together around concrete rules for AI applied to security.
What you need to know
- A UN framework moving into action: after two resolutions laying out the broad principles, the conference is now seeking concrete and verifiable measures. It also launched a Centre of Excellence on AI, Peace and Security, tasked with keeping this work alive between summits.
- Very concrete defence uses: the sessions covered model bias, dual-use civilian-military AI, agents able to chain actions together, cybersecurity and satellite imagery.
- An openly public-private dialogue: Microsoft co-sponsors the event, and industry players such as Palantir, Helsing and Kongsberg were among the speakers, alongside governments, universities and international organisations.
Stakes and outlook
The heart of the matter is turning ethical rules into verifiable requirements for systems used in sensitive contexts. A tool for analysing satellite images or supporting cyber defence must be auditable before it shapes a security decision. Yet several NGOs are sounding the alarm: on the ground, military adoption of AI is already moving faster than the tests meant to guarantee its lawful use. UNIDIR is not a global regulator, and that is the whole ambiguity of the exercise: creating a common language between states with diverging interests and defence industry players, without the power to impose it.
On 2 June 2026, at Build, Microsoft unveiled MAI-Thinking-1, its first advanced reasoning model. The signal is clear: Microsoft is accelerating the development of its own models, after years of heavy dependence on OpenAI's technology.
What you need to know
- An in-house reasoning model: MAI-Thinking-1 is described by Microsoft as a mid-sized model, trained from scratch on its own data, able to match reference models on certain software engineering and mathematics tests.
- A broader lineup: Microsoft announced seven models in all, covering reasoning, image, transcription, voice and code.
- Tools already wired into professional use: MAI-Code-1-Flash, the model dedicated to code, is built into GitHub Copilot and Visual Studio Code, used by developers to write, fix and complete code.
Stakes and outlook
Microsoft is reducing its dependence on OpenAI by bringing a key part of its technology in-house. The real issue is structural. For years, Microsoft was the best distributor of someone else's intelligence. By building its own models, trained on data it controls and deployable beyond Azure, it is seeking to own the engine rather than rent it. For corporate customers, this opens an alternative to OpenAI's models, better integrated into the Microsoft ecosystem.
On 19 May 2026, Google devoted its I/O keynote to a simple idea: pushing Gemini, its most advanced AI, into the heart of all its products. Until now, Google's AI was scattered, added in small touches here and there. From now on, it becomes the common foundation of Search, Workspace, YouTube, Android and future XR devices.
What you need to know
- Assistants that carry out tasks: Gemini Spark is presented as a personal agent that works in the background to sort an inbox, organise documents, track confirmations or prepare a morning brief from Gmail and Calendar.
- Video made simpler to produce: Gemini Omni Flash lets you create and edit a video through simple instructions, from text, photo, video or audio.
- Search becomes a space for action: Google adds more conversational searches, monitoring agents that work continuously, and a universal basket that tracks the best deals across the web.
Stakes and outlook The giants that hold a frontier AI are beefing up their work tools with centralised, more powerful AI. The catch is that keynotes sell dream scenarios of full automation that, in practice, demand a great deal of day-to-day adjustment. Part of the hype therefore remains an empty promise. And behind the demo, one topic often slips into the background at some of these giants: the privacy of our own data.
Sina Weibo, the Chinese social media giant known for its microblogging platform, has published a report on VibeThinker-3B, an open source AI model with 3 billion parameters. Its high scores in maths and code reignite the debate on the reliability of benchmarks and the race for ever-larger models.
What you need to know
- A small model that performs strongly: VibeThinker-3B reaches 94.3 on AIME 2026, a demanding maths competition, level with DeepSeek V3.2, a model roughly 224 times larger.
- Limits beyond reasoning: on GPQA-Diamond, a scientific knowledge test, it tops out at 70.2 against 91.9 for Gemini 3 Pro. Its authors own up to it: it reasons, but does not replace a large model rich in knowledge.
- Doubts about real-world use: users point to weaknesses in everyday development and suspect “benchmaxxing”, a model tuned to shine in tests more than in real work.
Stakes and outlook
The real issue is economic: a 3-billion-parameter model is far cheaper to train and run, to the point of fitting on a laptop. If it lives up to its promises, it cracks the idea that frontier reasoning requires colossal infrastructure. But a good benchmark score says nothing about how it holds up on a real workstation. The most credible path is not one giant model good at everything, but a constellation of small specialised models, each cut out for a precise task, backed by large models for general knowledge.
According to Reuters, Kirkland & Ellis will devote 500 million dollars to developing its own AI platform. For a corporate law firm, that is an investment of rare scale.
What you need to know
- A massive, phased investment: Kirkland plans to spend 500 million dollars over three to four years, including 100 million as early as 2026, to build an internal AI platform.
- A legal player, not a Big Tech: Kirkland & Ellis is the most profitable corporate law firm in the world, with 10.6 billion dollars in revenue last year, specialising in major corporate deals.
- A strategy of control: the platform is built to measure from the know-how of 250 of its lawyers, even if the firm will keep licensing certain third-party tools. The idea: no longer depending entirely on outside vendors.
Stakes and outlook
On paper, the lever is productivity: reviewing contracts, comparing versions, finding precedents in large case files. AI does not replace legal judgement, but it can redraw how work is split between partners, associates and support functions. The real signal is strategic. When every firm can license the same AI from the same vendors, that tool stops being an advantage. By building its own, Kirkland is betting that value shifts towards those who own their layer of intelligence, not those who rent it. It is also a fault line opening up: only firms able to line up hundreds of millions can keep pace.
✻ App under the prism: Microsoft 365 Copilot Cowork
1. What is it?
Microsoft 365 Copilot Cowork is an autonomous agent built into Copilot, rolled out worldwide since 16 June 2026. The difference with the classic Copilot is clear: where the latter answers a question or generates text on demand, Cowork strings actions together on its own until it delivers a finished result. A bit like a coding agent, but in the cloud and for office work. It can draft and send emails, schedule meetings, create Word, Excel, PowerPoint or PDF documents, post in Teams and organise files. Because its tasks run in the cloud, they keep going even with the computer switched off, and every important action must be approved before it runs.
2. Why it's fascinating?
Cowork pushes Copilot beyond writing assistance: it no longer just suggests, it acts like a digital colleague able to chain several steps of work. You can ask it to pull figures from an Excel sheet, turn them into a Word summary, then post that in Teams and schedule a follow-up, all in a single conversation. It is especially valuable for repetitive administrative tasks, calendar management, preparing communications or tracking down information scattered across the organisation.
3. Why it's limited?
Cowork depends heavily on the Microsoft 365 ecosystem, access rights, available data and the policies set by administrators. The more the agent acts alone, the costlier a mistake becomes: a botched draft is annoying, but a corrupted spreadsheet or a sensitive piece of data shared by mistake is an incident. The pay-as-you-go billing model (Copilot Credits, charged per action) adds a measure of budget unpredictability. Finally, its effectiveness will depend on the quality of internal data, on permissions and on how clearly users can frame their requests.
