Satya Nadella: models must be treated as an “insider risk” — with an emergency brake, full logging and independent oversight. In a long X article on Saturday (“Models as Insider Risks in the Super Intelligence Era”), Microsoft’s CEO wrote that we cannot trace the behavior of today’s most powerful models back to specific training data or weight settings, yet companies are giving them access to sensitive data and critical operations. In his view, both closed and open-weight models should be treated like the insider risk posed by highly privileged employees: not because they are malicious, but because they can make mistakes or be misused by someone. His proposals: controls over what a model can access and do should sit outside the model, and the model should be decoupled from the framework that orchestrates it; no single model should be the sole source of an important result or check its own work; every significant step should leave a tamper-proof, human-readable trail (“what cannot be observed cannot be trusted”); the whole system should be continuously tested against both failures and attacks; oversight should be independent of the model; and “we must assume a model is compromised” — an authorized person must always be able to pause or shut it down mid-task, like an emergency brake. Failures should be reported promptly and shared across the whole industry. His closing thought is that the most trustworthy system will not be the one whose model we trust most, but the one that demands the least trust in the model. According to Mustafa Suleyman, head of Microsoft AI, containing superintelligence is today an engineering and governance task, as it once was for aviation, nuclear materials or medicines; investor David Sacks also backed Nadella. The piece comes as more and more AI companies acknowledge cases where they lost control of their models. Source: Satya Nadella (X), TechCrunch, The Verge, Mustafa Suleyman (X).
Nvidia: talks to acquire open-weight model developer Reflection AI or take a larger stake. According to a Saturday report in the Financial Times, Nvidia is in early-stage talks to acquire US-based Reflection AI or further increase its investment in the company. One possible structure is a so-called “acqui-hire”: rather than buying the company, Nvidia would take on its staff and license its technology, which could avoid a lengthy antitrust review. According to the sources, a deal could come within weeks, but the talks could also fall apart. Nvidia is already one of Reflection’s main investors; according to the FT, it has put $800 million into the company. Founded in 2024 by former DeepMind researchers, Reflection unveiled its first open-weight model, Beam, earlier this week, aiming to compete with cheaper Chinese models from DeepSeek and Kimi on coding and agentic tasks; according to the FT, the Trump administration also sees it as a US rival to Chinese alternatives. Reuters could not independently confirm the report, and neither company commented. Source: Financial Times, Reuters (StreetInsider).
WSJ: a desperate scramble for compute — rivals are teaming up and executives are bargaining in person. According to a long article in The Wall Street Journal, the shortage of compute (the chips and servers needed to train and run AI models) is reshaping Silicon Valley. According to the paper’s reporting, Noam Shazeer, whom Google brought back in 2024 for $2.7 billion, left for OpenAI shortly after Google cut the compute allocated to his team in favor of another project. Anthropic co-founder Tom Brown personally visited Elon Musk at xAI’s offices in late March to rent capacity from him, because the unexpected success of Claude Code was causing frequent outages; this led to the SpaceX deal announced in May, later estimated at up to $45 billion. Earlier this year, Dario Amodei also called Meta’s AI chief, Alexandr Wang, about capacity, but Meta said no for now. At Google, Sergey Brin sometimes overrides the official allocation of compute, while Demis Hassabis is unhappy about how little capacity goes to long-term research. According to SemiAnalysis data, the one-year rental price of an Nvidia H100 chip rose 60% in a year, and cloud providers are increasingly asking for multi-year contracts and up to 30% upfront. According to a separate WSJ profile on Friday, Tom Brown’s Republican connections also played a role in resolving the June standoff over model safety. Source: The Wall Street Journal, Hindustan Times (WSJ), The Wall Street Journal (Tom Brown).
Alan Turing Institute: Britain must not depend on foreign AI that can be restricted or switched off at any time. George Williamson, the new director of the UK’s leading AI research institution, told The Guardian that in defense, national security, critical infrastructure and healthcare, the country cannot rely solely on systems it cannot inspect or control. As an example, he pointed out that in June Washington temporarily banned the release of Anthropic’s latest models outside the US and reportedly asked the major labs to let the UK AI Security Institute test their models only after its US counterpart. He said the institute cannot compete with the big labs’ money, but it can create “nationally owned IP” and smaller models that startups can use too; the main focus is “national resilience.” He considers percentage estimates of the likelihood of AI-driven extinction to be unfounded. In recent weeks, Christine Lagarde and Ursula von der Leyen have also urged similar self-reliance for Europe. Source: The Guardian.
Report: a voice interface and two new menu items in Claude, and an unannounced “discovery” model on OpenAI’s price list. According to TestingCatalog, a hidden Voice menu item labeled “Preview” has appeared in the menu of Claude’s web interface; it points to Anthropic’s internal test environment rather than a public feature. A setting that is off by default also surfaced recently, letting users consent to their voice recordings being used to improve the models — this may suggest that Anthropic is working on its own speech models, but the company has not confirmed it. Two more menu items with internal codenames also appeared: Arborio (with a hand-holding-a-heart icon) may be a web version of Claude’s health features, while the purpose of Iggy (with a sun icon) is unknown. Meanwhile, the gpt-rosalind-discovery model appeared in the “Life sciences models” section of OpenAI’s API price list, at the same price as gpt-rosalind-research ($5 per million input tokens and $25 per million output tokens), and likewise only within OpenAI’s vetted access program. There is no announcement or description of the model; its name may suggest drug discovery use. In all three cases, these are traces found in code and on websites, not official announcements. Source: TestingCatalog (Voice), TestingCatalog (Arborio, Iggy), TestingCatalog (gpt-rosalind-discovery).
Apple: taking on the team and technology of personalized podcast maker Huxe. In a filing with the European Commission, Apple said it is making job offers to “certain employees” of Huxe AI and receiving a non-exclusive license to the startup’s intellectual property. This is a so-called reverse acquihire, which lets big companies obtain a startup’s team and technology without buying it, drawing less attention from competition authorities. Huxe was founded by developers who previously worked on the AI podcast feature of Google’s NotebookLM (now Gemini Notebook); the company announced on May 21 that it was shutting down, and Apple notified Brussels of the deal on June 9. The filing does not say who received offers or what Apple is planning; according to TechCrunch, Apple Podcasts could get similar features. Source: TechCrunch.
Elon Musk: Grok Bot can also shop, and “Digital Optimus” is already playing Diablo. On Saturday evening, Musk wrote on X that you just need to take a photo of your bank card and drop it into the chat, and Grok Bot will find the best deal for the product you want online and order it. One user replied that Amazon already blocks the bot; Musk said Grok finds another way, and that he himself ordered LEGO directly from the LEGO website this way. In another post, he wrote that “Digital Optimus,” playing like a human by looking only at the screen, has already got roughly halfway through the Diablo campaign, does well at Counter-Strike, is now learning League of Legends, and that the goal is for it to generalize to every game. Musk is hiring for TeslaSI. These are Musk’s X posts; there is no official product description for them yet. Source: Elon Musk (X), Elon Musk (X, LEGO), Elon Musk (X, Digital Optimus).
vLLM: now runs on Nvidia’s Vera Rubin chips — more than seven times the performance of GB200, according to the developers’ early benchmark. The developers of vLLM, the popular open-source model serving engine, announced that it now supports Nvidia’s new Vera Rubin platform. On SemiAnalysis’s AgentX benchmark, serving China’s MiniMax M3 model on a Vera Rubin NVL72 system, they were able to handle more than 7.8 times as many requests as on GB200 at the same response time; on the MLPerf Inference v6.1 benchmark, they achieved up to 3.7 times the throughput of GB300 NVL72. According to the developers, a Rubin GPU offers 3.5 times the FP4 compute, roughly 2.4 times the memory bandwidth and 1.7 times the NVLink bandwidth of GB200. The test version can already be tried with DeepSeek, Kimi, GLM and MiniMax models. The figures are the developers’ own early results. Source: vLLM (X), MiniMax (X).
What this means for you: Nadella’s proposals are a good practical starting point too: if you let AI agents into your company systems, build them sothat permissions, logging and the ability to shut them down sit outside the model, and a human can step in at any time. The capacity shortage and the UK debate are also a reminder that relying on a single provider is risky; with vendor-neutral consulting, we help you plan how to switch models or providers if access suddenly becomes tighter or more expensive.