Welcome to Eye on AI. Beatrice Nolan here. In today’s issue:
I spent last week in Paris for VivaTech, Europe’s largest startup and tech conference, which managed to draw 180,000 attendees, despite a record heatwave in France.
VivaTech’s biggest headliner, Jeff Bezos, was in techno-optimist mode, arguing that AI will create a labour shortage rather than mass redundancy. He paired that with his longer-term vision of moving heavy industry off Earth and building a permanent industrial presence on the Moon.
Executive Chairman of AMI Labs and former head of Meta’s AI efforts, Yann LeCun, was happy to provide a counterpoint, taking the time to tell CNBC that the AI industry could be heading for a “big bubble explosion” if companies fail to bring costs down fast enough.
The winner of the most memorable moment, however, goes to two humanoid robots. While attempting to perform a choreographed demo at a booth, the pair slowly drifted backward into a row of televisions and sent two of the screens crashing to the floor—a counter to some of the hype on display in the exhibition hall.
The more grounded conversations elsewhere at the conference revolved around immediate realities: who controls AI, how to counter some of the current risks, and what Europe, still reeling from the shock of being cut off from Anthropic’s Mythos model, needs to do to control it’s access to AI.
Perhaps unsurprisingly, the event was preoccupied by one of the more imminent risks posed by AI: cybersecurity. The debut of Anthropic’s Mythos and OpenAI’s GPT 5.5 Cyber has fueled anxiety among both business and government leaders due to their advanced hacking abilities.
These models can now, at speed and scale, do what once required time and specialized human knowledge: find unknown vulnerabilities in critical software and generate working exploits. The worry is that the same capability lands in the hands of criminal groups and nation-state actors before most organizations have had time to shore up their defenses.
Anthropic and OpenAI have both staggered the release of recent cyber products and models with the aim of giving defenders a head start—granting initial access only to vetted security firms and critical infrastructure operators so they can patch their systems before the capabilities spread more widely.
But there is some concern about what happens if that window—the period when defenders have access to better AI models than attackers—closes too quickly
“I think long term it favors defenders,” Peter DeSantis, SVP at Amazon, said in an interview. “But the concern is short term.”
At the moment, he warned, security teams are still scrambling to understand the new kinds of attacks these models enable and to update their practices and tooling, potentially creating a gap between how quickly AI can be used offensively and how quickly defenses are adapting.
Earlier this week, OpenAI expanded its Daybreak initiative, pairing GPT-5.5-Cyber with a new open-source patching effort called Patch the Planet to work with firms including Cloudflare, Cisco, and CrowdStrike with the aim of helping organizations find and fix vulnerabilities.
A few days prior, Thibault Sottiaux, OpenAI’s head of core platforms, told me the idea behind the staggered roll-outs was to “be responsible and help accelerate cyber defense.”
“We have invested a lot in our safety stack, to ensure that generally accessible models are responsible for broad use,” he said. “But then we have more specific models that advance different tiers in terms of cyber capabilities, those we restrict behind different levels of our trusted access program.”
The U.S. decision to abruptly cut off access to Anthropic’s frontier models last week gave the conversation around “sovereign AI” a new urgency. While it was clear the risk of dependence felt much more real for European companies and policymakers, it was also clear that sovereignty means different things to different people.
Cohere CEO Aidan Gomez told Fortune that sovereignty needed to start with domestically controlled infrastructure—chips, power, data centers, and private deployment all under national control. Failing that, he said, countries should form strategic alliances to counter U.S. and China’s grip on the infrastructure needed to power AI.
Big Tech executives took a less hardline approach. Amazon’s DeSantis said that no nation—including the U.S.—really has a truly sovereign infrastructure, and suggested that trying to achieve such a high bar was unrealistic.
Instead, he argued for a more pragmatic approach: keep sensitive data in-country and give governments or companies clear control over how AI is governed, rather than attempting to recreate the entire hardware and supply chain stack within each nation.
He added that, in his view, using large shared data centers in the cloud is the most practical way to let countries keep data local and control how AI is run, while still keeping capacity, costs, and power use manageable. (Of course, given that DeSantis is from Amazon, the world’s largest cloud provider, he would say that.)
At VivaTech, executives were less enthused by some of the loftier promises of AI, and keen to chat about more practical topics like training, workflow automation, and AI spend.
Philippe Rambach, Schneider Electric’s chief AI officer, told me he had recently made AI training mandatory for the company’s entire 42,000 staff. Rather than chasing shiny demos, he described being picky about AI pilots and only backing projects with a defined business case and a path to scale. According to Rambach, it’s all part of an effort to treat AI as an operational tool rather than an innovation toy.
From the product side, OpenAI was seeing a similar shift. Enterprise customers still want the best models, but they also want to know whether they are actually worth the money.
“We receive a lot of inquiries about: Is the ROI there? I’m running all these agents, are they actually providing value for us?” Sottiaux said. One strategy for the company, he said, is to focus on tighter cost control—making the models more efficient so companies can do more with less.
With that, here’s more AI news.
Beatrice Nolan
beatrice.nolan@fortune.com
@beafreyanolan
Exclusive: Seltz, a startup rebuilding web search for AI agents, raises $12.5 million in seed funding — Jeremy Kahn
Getting past the pilot: Why so many AI test projects have trouble scaling — Alexei Oreskovic
Robert Wright sees an ‘earthquake’ coming from AI that goes far beyond jobs: ‘cultural, political, personal, family, psychological’ — Nick Lichtenberg
Asia’s defense boom is rewiring the global arms supply chain — Chris Oberoi
AI companies spent $27 million on a Manhattan primary. A Democratic primary in Manhattan's 12th congressional district became an unlikely battleground over AI regulation over the last few months, with super PACs linked to Anthropic and OpenAI pouring a combined $27.4 million into the race (Anthropic-linked PACs supported State Assemblyman Alex Bores while OpenAI-linked ones supported his opponent, Assemblyman Micah Lasher). Bores—and his pro-stance on state AI regulation legislation—was where the fight was concentrated. In the end, Bores finished ahead of John F. Kennedy's grandson, Jack Schlossberg, and Trump critic George Conway, but lost to Lasher by roughly four percentage points. That's despite pro-Bores groups backed by Anthropic-connected money outspending Leading the Future—a deregulation-focused PAC funded partly by executives from OpenAI, Palantir, and Andreessen Horowitz—by more than two to one. The race was seen as a bellwether for how AI regulation might fare in the broader midterms, but some political analysts cautioned against drawing national implications from the result. Both sides have now spent more than $50 million across 19 states as the midterm campaign intensifies. Read more in The Verge.
Google delays its next frontier AI model to July. Google has pushed back the launch of Gemini 3.5 Pro by around a month, according to Business Insider. The company is using the extra time to gather feedback from early testers and tweak the model, which has already been made available to some users on Google's Antigravity platform and the AI benchmarking site LMArena. CEO Sundar Pichai had told developers at Google's I/O conference in May that the model would arrive "next month." The delay comes at a moment of intense competitive pressure, with Anthropic and OpenAI continuing to pull ahead of Google in coding—one of the major enterprise use cases for AI. The new model is expected to improve on long-horizon tasks and agentic capabilities, and Google has reportedly incorporated feedback from its Flash 3.5 model into the build, including criticisms that Flash consumed tokens too quickly. Read more in Business Insider here.
Google DeepMind researchers keep leaving. Several prominent researchers from Google DeepMind are heading to rivals such as Anthropic and OpenAI. Bloomberg reported that Jonas Adler and Alexander Pritzel, both considered significant contributors to the Gemini model, are set to depart for Anthropic. Separately, DeepMind researcher Arthur Conmy announced he was joining the lab to work on alignment during training, writing that Claude's capabilities are "extraordinary" but that current models remain insufficiently aligned to safely delegate AGI development. Last week, Noam Shazeer, regarded as one of the most influential researchers in AI, also announced he was leaving Google for OpenAI, and John Jumper, who led DeepMind's work on protein structure prediction and shared the 2024 Nobel Prize for Chemistry with DeepMind CEO Demis Hassabis, announced he was leaving for Anthropic. Lucas Beyer, a researcher who previously worked at Google Brain and DeepMind before moving to OpenAI and then Meta, said on X that the departures may point to a deeper structural shift within Google DeepMind, with most of the high-profile exits appearing to be long-tenured researchers based in London specifically. That, he said, is consistent with what he has heard about the center of gravity for pretraining work quietly moving from DeepMind's London hub to Google's Mountain View headquarters. Read more from Bloomberg.
Anthropic accuses Alibaba of AI model theft. Anthropic has accused Alibaba of illicitly extracting capabilities from its Claude AI model in what it described as the largest known attack of its kind. According to a letter, which was reviewed by multiple media organizations, the alleged campaign ran from April 22 to June 5 and generated more than 28.8 million exchanges with Claude through nearly 25,000 fraudulent accounts. Anthropic said the activity was tied to Alibaba and its AI lab, Qwen, and was aimed at helping China accelerate access to Anthropic’s advanced model capabilities. The accusation underscores how aggressively American AI companies are trying to protect their edge in the global AI race as China increasingly closes the gap. Read more in Reuters.
OpenAI and Broadcom unveil new AI chip. OpenAI and Broadcom have unveiled Jalapeño, a new inference chip designed specifically for large language models, as OpenAI pushes deeper into the full stack behind its products. The companies say early testing shows the first-generation accelerator delivers substantially better performance per watt than current state-of-the-art chips, and that it was co-developed from design to tape-out in just nine months. OpenAI said the goal is to make advanced AI faster, more reliable, and more affordable by tightening control over the infrastructure underneath it. The chip is part of a longer-term platform that OpenAI says will scale with data center partners over multiple generations. Read more from OpenAI.
That's the proportion of human content moderation reviews Meta has already replaced with AI this year, according to the Financial Times. The figure marks a significant acceleration of plans the company first announced in March, when it said it would deploy more advanced AI moderation systems "over the next few years." Meta is aiming to reduce human review further by the end of the year, potentially by more than 90% for some content types. The company argues the shift is about accuracy rather than cost savings, citing internal tests showing its LLMs make 13% fewer mistakes than human reviewers and catch 10% more actual violations. But some staff warn that the rollout is moving faster than the technology can handle, with errors including the wrongful removal or suppression of harmless content.
The push is already triggering lay-offs among contractors, and raises concerns particularly around advertising, where errors have higher stakes on Meta's business model. Meta is already contending with multiple lawsuits alleging it has failed to adequately police fraudulent ads on its platforms.
July 6-11: International Conference on Machine Learning (ICML), Seoul, South Korea.
July 7-10: AI for Good Summit, Geneva, Switzerland.
Aug. 4-6: Ai4 2026, Las Vegas.
Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.
Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.
Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。