Jeff Dean leaves Google as DeepMind reshuffles its command
AI Watch, Thursday 6 August 2026

ZDNet reports how Aston Martin Aramco keeps a human in the loop to get value from its data and emerging technologies: according to its executives and partners, domain expertise remains the key to using AI, not the other way round. The case matters to any organisation deploying AI without wanting to take the decision away from its experts.

Since 2 August 2026, the commission has welcomed five members as part of a partial renewal of its board, the body that rules on sanctions and on the regulator's positions, including on AI and the GDPR. The board's makeup weighs on the French doctrine to come, as AI cases pile up.

Next.ink examines the spread of benchmarks, the standardised tests meant to compare models, and points out that their results, waved around by vendors at every release, are only worth something with the detail of the protocols, which is rarely published. For a team assessing a supplier, the lesson comes down to one rule: test on your own cases rather than trust the month's honour roll.

In its first quarterly results since its 12 June market debut, Elon Musk's group posts growth above Wall Street's expectations, carried by Starlink and AI. Investors nonetheless punished the amount invested in AI, in line with the market doubts hitting the sector's big budgets.

The maker of Claude is hiring silicon specialists to co-design hardware and models, aiming to run its technology faster and at lower cost. The move follows Google (the TPU chips), Amazon (Trainium) and OpenAI: control the chip to cut the inference bill and the dependence on Nvidia. Designing an in-house accelerator is long and expensive, but aligning the chip with the model's architecture promises a yield that generic hardware does not offer. For a player that sells access to its models, every gain in efficiency per watt feeds through to the margin.

According to Axios, the administration brought together OpenAI, Anthropic, Google, Meta, Microsoft and Nvidia to present a voluntary framework for assessing the cyber risks of the most powerful models, none of whose terms should restrict open models, which you can download and inspect. The companies that keep their technology locked up will therefore carry the compliance burden, not those that publish their weights. The Verge notes that these guidelines remain voluntary and vague, and that they explicitly leave open models outside the scope of the tests. The choice departs from the European logic, where the AI Act governs use whatever the degree of openness. For a French vendor, the doctrinal gap between Washington and Brussels becomes a product-strategy variable.

The online-commerce company says it is not suffering the cannibalisation hitting news publishers: traffic and orders coming from AI tools to Shopify stores tripled year on year in the second quarter. The finding tempers the idea that AI answer engines drain the merchant web. For a retail site, the practical lesson is to optimise its presence for these new entry points, rather than defend only classic search rankings.

In May 2026, users had spotted that Chrome was quietly installing a 4-gigabyte AI model on their machines. Google has since updated its documentation to explain these downloads and offer an option to disable them. The episode illustrates the tension between on-device AI, which processes data on the machine, and users' right to know what installs on their systems. For a corporate fleet, a model arriving unannounced raises a question of endpoint control and bandwidth consumption.

The FCC is reportedly considering blocking imports of new Chinese optical modules, the components that convert data into light pulses and feed AI data centres, on grounds of a risk of data theft or malware implantation. Zhongji Innolight shares fell as much as 16% on Wednesday, before analysts played down a measure judged "mild": current supply chains would stay intact and America depends heavily on Chinese manufacturing. The stock ended the day down 7.3% in Shenzhen. The sequence shows a market testing, in real time, the actual effect of US restrictions on the components chain.

The US bank lifts its estimate of annualised revenue for the Chinese AI model market from $10 billion to $13 billion by year-end, citing aggressive price cuts, technical advances and faster adoption at players such as DeepSeek and MiniMax. The revision measures the speed at which the Chinese ecosystem converts its progress into revenue. For a European company, the rise of a cheap and capable Chinese offer widens the choice of suppliers, subject to questions of data and hosting.