ExoBrain weekly AI news

4th July 2025: Missionaries versus mercenaries, the open web’s last stand, and Microsoft’s medical superintelligence

Author

Joel Miller
July 04, 2025

Welcome to our weekly email newsletter, a combination of thematic insights from the founders at ExoBrain, and a broader news roundup from our AI platform Exo…

Themes this week:

  • Meta's $300m packages to recruit OpenAI's top AI researchers

  • Cloudflare's new tools protecting publishers from AI crawlers

  • A new medical AI achieving 85% accuracy versus doctors' 20%

Missionaries versus mercenaries

Just weeks after we reported on the eye-watering introductory sums being offered by Meta, the talent war has exploded into open conflict. With Chinese labs dominating the open-weight model landscape, Meta has launched an acquisition and recruiting offensive aimed at forcefully regaining lost ground.

Meta is reportedly offering select OpenAI researchers packages worth up to $300 million over four years, with immediate stock vesting. To put this in perspective, that's nearly four times what Microsoft's CEO earned last year. These aren't signing bonuses; they're full compensation packages designed to strip OpenAI of its top talent.

Meta's recent haul includes seven OpenAI researchers with expertise spanning computer vision and multimodal AI. Lucas Beyer, Alexander Kolesnikov and Xiaohua Zhai, all from OpenAI's Zurich office, bring deep experience in vision transformers and large-scale image-text models. Beyer and Kolesnikov were previously at Google Research where they contributed to breakthrough work on Vision Transformers (ViTs), while Zhai has published extensively on scaling laws for visual models. The four additional hires - Shengjia Zhao, Jiahui Yu, Shuchao Bi and Hongyu Ren - represent a further cross-section of OpenAI's research talent.

Meta's aggression extends beyond individual hires. After failing to acquire Safe Superintelligence Inc. (SSI), the company successfully recruited its CEO, Daniel Gross. Ilya Sutskever, SSI's co-founder and former OpenAI chief scientist, confirmed Gross's departure with a terse statement: "Daniel Gross's time with us has been winding down, and as of June 29 he is officially no longer a part of SSI."

The spat between Meta and OpenAI has turned personal. OpenAI's Mark Chen described Meta's recruiting as feeling like "someone has broken into our home and stolen something." Sam Altman responded with barely concealed contempt, declaring "missionaries will beat mercenaries" and dismissing Meta's efforts as creating "deep cultural problems." Mark Zuckerberg, meanwhile, has promised recruits unlimited GPU access and resources.

MSL, co-led by Scale's Alexandr Wang and Nat Friedman, represents Meta's attempt to leapfrog the competition. The focus in this high-level race has shifted in recent months from artificial general intelligence (AGI) to superintelligence. Meta's new division is called Meta Superintelligence Labs (MSL), not AGI Labs. The industry now sees superintelligence as the more immediate, achievable goal. Perhaps building AI that exceeds human intelligence in specific domains appears more tractable than creating truly general intelligence. Sutskever signed off his note to investors: “We have the compute, we have the team, and we know what to do. Together we will keep building safe superintelligence." Not if Zuckerberg can get there first.

Takeaways: The AI talent war has entered a new phase where money is essentially no object. Meta's willingness to pay CEO-level compensation to individual researchers shows how critical human talent has become to AI progress. The shift from AGI to superintelligence terminology suggests a more focused, perhaps more achievable target. This concentration of talent and resources in a handful of companies will likely accelerate progress but raises questions about the healthy development of AI.

The open web’s last stand

The death of the open web has been predicted many times, but AI might actually finish the job by destroying the economics that made the web work. This week Cloudflare has taken a stand.

For decades, publishers allowed search engines to crawl content in exchange for traffic. This quid pro quo created the internet economy. AI has shattered this arrangement. Google's AI Overviews alone cut click-through rates by 30-35%. Travel sites have lost 20% of search referrals year-on-year. News sites are down 17%. Where Google crawls websites 14 times for every visitor it sends back, OpenAI's ratio is 1,700:1. Anthropic's? A staggering 60,000:1.

Cloudflare, currently powering 20% of the web, has launched two defensive strategies. First, a managed service that automatically blocks AI crawlers. Second, intelligent blocking that targets AI bots only on pages showing advertisements, protecting monetised content whilst allowing access to documentation. More radically, they've proposed "Pay per Crawl", a marketplace where AI companies pay micropayments for each page scraped. Publishers including Conde Nast and TIME have signed on. The vision is an economy where AI pays to acquire information.

But these solutions might accelerate the web's fragmentation. Premium publishers will lock content behind paywalls or payment systems. AI companies, facing costs, become selective. The once open web may split into tiers, premium content for paying AI and human subscribers; ad-supported content farms; free corporate documentation; and grey markets of repackaged content. We are seeing a move away from what the founders of the web envisioned to stratified access based on ability to pay.

Takeaways: AI crawlers aren't sustainable entities but value extractors operating at an increasingly large scale. Cloudflare's defensive tools and payment marketplace represent the first serious rebalancing attempt, but they might simply formalise the web's fragmentation. The likely future isn't death but transformation into economic zones where information flows according to payment. The open web as we knew it, a vast, interconnected repository of human knowledge, is already gone. What emerges next depends entirely on who's paying whom.

Microsoft’s medical superintelligence

This image shows a new medical research project from Microsoft called the AI Diagnostic Orchestrator. MAI-DxO achieves 85% accuracy on complex medical cases whilst spending less on diagnostic tests than human physicians, who averaged just 20% accuracy. The purple line traces MAI-DxO's performance across different budget constraints, consistently outperforming individual AI models from OpenAI, Google, and others. The orchestrated AI system doesn't just diagnose better, it does so more efficiently, ordering fewer unnecessary tests. This Pareto frontier visualises how AI has crossed a threshold, delivering both superior medical reasoning and potentially reducing the pressure on healthcare systems.

Weekly news roundup

This week's developments reveal AI's accelerating impact across industries, from fashion to sports, while regulatory challenges intensify and infrastructure investments reach unprecedented scales.

AI business news

AI governance news

AI research news

AI hardware news