Google's sweeping AI search overhaul continued to dominate industry conversation this week, while an OpenAI reasoning model quietly made history by autonomously disproving a longstanding mathematical conjecture. Meanwhile, corporate AI spending reached new extremes — and the workforce consequences of AI adoption are no longer theoretical.
Google's AI Search Reshapes the Web
Google's new AI-forward search layout is triggering a measurable shift in how web traffic flows. The redesign reduces the visibility of traditional organic results, dealing a significant blow to publishers who depend on search referrals. A study of 846,000 search sessions found that users interact with AI Overviews and AI Mode differently than with classic results, with implications for brand traffic that marketers are still working to quantify.
The backlash has been swift in some corners: DuckDuckGo is reporting a surge in new users following Google's AI push, suggesting a meaningful segment of searchers would rather opt out of AI-mediated results entirely. Google, for its part, is pressing ahead — at Google I/O, the company announced what it's calling the "agentic Gemini era," positioning its AI assistant as an always-on, 24/7 presence across users' digital lives. A new "Daily Brief" feature in Gemini is part of that push, offering personalized morning summaries.
Sam Altman Recalibrates on Jobs and AI's Role
OpenAI CEO Sam Altman offered a more measured take on AI's economic disruption this week, saying the technology's impact on white-collar jobs has been slower than he initially predicted. That's a notable softening from earlier warnings that AI would rapidly displace knowledge workers. Separately, Altman articulated a vision of AI as a metered utility — more like electricity or bandwidth than traditional software — a framing that reflects OpenAI's evolving thinking about how AI value will be delivered and priced.
AI Makes a Mathematical Breakthrough
In a development that may prove historically significant, an OpenAI reasoning model autonomously disproved a longstanding mathematical conjecture — without human direction during the proof process. The milestone marks what researchers are calling the first autonomous mathematical breakthrough by an AI system, pushing the frontier of what these models can accomplish in formal reasoning beyond the well-worn benchmarks of code generation and language understanding.
The Corporate AI Spending Race
The scale of AI infrastructure investment is becoming difficult to fathom. Reports this week indicated that Anthropic is spending approximately $1.25 billion per month to access SpaceX's AI training clusters — a figure that underscores just how capital-intensive frontier AI development has become. OpenRouter, the AI gateway startup that manages multi-model access for developers, raised $113 million to scale its infrastructure as token usage continues to climb.
ClickUp made the week's starkest organizational statement, laying off 22% of its workforce as it pivots to what it's calling a "100x org" model built around AI. The move is part of a broader trend: companies are increasingly concluding that AI adoption requires system-wide organizational redesign, not just incremental automation bolted onto existing workflows.
ByteDance added a different kind of disruption at Cannes, where its AI-generated films are reportedly rattling budgets and forcing a reckoning with what low-cost AI production means for the industry's economics. Meanwhile, AI's physical footprint is generating its own backlash — data centers are drawing mounting conflicts over land use, power consumption, and community impact in regions where they're being built.
Geopolitics: China Restricts AI Researcher Travel
China has introduced new restrictions limiting overseas travel for top AI researchers at private firms, a move interpreted as an effort to prevent talent and intellectual property from leaving the country. The policy reflects growing government concern about the strategic value of AI expertise as competition with the US intensifies.
Robotics and Hardware
Hugging Face launched a $2,500 open-source humanoid robot leg platform this week, continuing its push to democratize robotics research. The affordable bipedal system is aimed at labs and developers who have been priced out of the hardware side of humanoid robotics.
On the chip front, Huawei claims a breakthrough with its "LogicFolding" stacking design, which the company says achieves performance equivalent to 1.4-nanometer processes — a significant claim if it holds up, as it would partially offset the impact of US export controls on advanced semiconductors. xAI, meanwhile, completed training on Grok V9-Medium, which scales to 1.5 trillion parameters — roughly triple its predecessor's count.
Imec fabricated what it describes as the world's first quantum dot qubit using High-NA EUV lithography, a technical milestone in the long road toward practical quantum computing. And Google added an open-source agent executor to its platform, a runtime designed to help enterprises deploy and scale AI agents in production.
Cybersecurity
The FBI issued a warning about Kali365, a phishing-as-a-service kit that bypasses multi-factor authentication by abusing Microsoft 365's legitimate device-code login flow to hijack accounts. The technique is technically sophisticated and difficult for users to detect. Separately, researchers determined that a widely circulated "OnlyFans data leak" affecting 340 million records is likely a compilation of recycled data from older breaches rather than a new intrusion.
Pope Leo's first encyclical included a warning about autonomous weapons, stating that AI-enabled weapon systems are evolving beyond meaningful human governance — a rare instance of the Vatican weighing in directly on AI ethics at a geopolitical level.
Quick Takes
ElevenLabs launched Music v2, an AI music model that supports mid-track genre switching and more granular editing controls. Anthropic introduced an AI Fluency scorecard to evaluate how effectively users interact with its models — a tool aimed at helping organizations measure their teams' AI literacy. CMU researchers published work suggesting large language models might benefit from "sleep" cycles that consolidate context, drawing an analogy to biological memory consolidation. A new benchmark called DeepSWE targets long-horizon software engineering tasks and was designed to prevent model pre-training data from contaminating evaluation results. Google also announced that Chrome now includes cost-efficient, offline-capable AI features built directly into the browser, and Cohere acquired Reliant AI to strengthen its position in regulated enterprise sectors.