This Week in AI: Who Controls the Loop?

This week host and Turing Post founder Ksenia Se threaded the latest news into a single argument: AI is moving out of conversation and into the operational loops where real work happens. From SpaceX’s $60 billion acquisition in the developer tools market to the G7’s debate about frontier model access to image generation company Midjourney’s pivot to medical hardware, the stories all pointed in the same direction.

When agents own the loop, the IDE becomes infrastructure

SpaceX’s acquisition of Anysphere, the company behind Cursor, for a reported $60 billion in stock is the kind of deal that looks straightforward until you think about what Cursor actually is. On the surface, it’s a popular AI-assisted code editor. (It’s also one of many in a highly competitive market.) However, Ksenia argued that that’s thinking too small, especially for Elon Musk. SpaceX may be angling to position Cursor as the new center of software work, in the same way GitHub became the center of the previous era.

In the old model, GitHub owned the pull request. But in the new model, the question of who owns the full loop where agents read a repo, write code, open pull requests, run tests, handle failures, and enforce engineering standards is still open. GitHub still owns the system of record and is moving to defend it: Chief product officer Mario Rodriguez recently told Turing Post that GitHub’s mission has shifted from human-developer collaboration to developer-and-agent collaboration, with the platform becoming agent-native across its APIs, UX, and underlying infrastructure. But as Ksenia explained, “Cursor’s advantage is that it owns the developer’s active coding surface” where the work starts.

If agents write more code than humans, software infrastructure should be redesigned around agents from the start. Cursor was built for agents. GitHub was built for humans and is now playing catch-up. That architectural choice may matter more than any individual product feature.

Frontier AI access is becoming a geopolitical question

The G7 summit this week included discussions about a “trusted partners” framework that would give select allied nations access to advanced US AI models, following a US order that restricted foreign nationals from accessing Anthropic’s frontier systems on national security grounds. AI models that can write software, find vulnerabilities, and operate across tools are capability systems, not just productivity software. The access rules are catching up to that reality, although as Ksenia noted, things haven’t yet come into complete focus.

For a long time, AI regulation sounded like: How do we label synthetic media? How do we reduce hallucinations, prevent bias, make chatbots safer? Now the question is so much bigger. Who can use these capable systems? Can allies use them? Can cybersecurity firms outside the US use them? Can non-US employees at US labs use them? Can European companies use American models if those models are also strategically sensitive? This isn’t traditional software licensing anymore. This is capability access control.

The underlying tension behind the G7 conversation is the dual-use problem: A model capable enough to find software vulnerabilities for defense can also find them for offense. The “trusted partners” framework reflects the new geopolitics of AI as countries jockey with rivals to secure strategic benefits for themselves and their allies. It represents an alliance layer for AI access that applies access structures previously reserved for physical military hardware to capabilities too strategically important to make fully open and too useful to keep entirely locked down. As Ksenia noted, the alliance is “not literally NATO, but [it is founded on] the same kind of logic.”

But access restrictions might also impact the talent that built these systems, who are increasingly not citizens of the country trying to control it. For instance, AI researcher Andrej Karpathy, recently hired by Anthropic, is publicly described as Slovak-Canadian. If access controls apply to non-US citizens, he and others like him may be denied access to the very systems they’ve been hired to work on. It’s an area we’ll continue to watch closely.

AI is entering the measurement loop

Midjourney, the company you probably associate with AI-generated images, has announced a new medical division and a full-body ultrasound scanner built around water immersion, developed in partnership with medical imaging hardware maker Butterfly Network. The device is designed to scan the entire body in 60 seconds: A person descends into a shallow pool on a motorized platform, passing through a ring of roughly half a million ultrasound sensors, each functioning as both a transmitter and receiver. The system uses over two petaflops of processing power to reconstruct a 3D body map from the returning wave data. Midjourney says the resulting images look comparable to today’s MRI output at a fraction of the cost and time, though that claim still needs serious clinical validation before it can stand.

The current prototype uses 40 Butterfly ultrasound-on-chip devices per system, according to a disclosure from Butterfly Network, which confirmed its codevelopment and licensing agreement with Midjourney. Midjourney plans to open a facility in San Francisco in 2027, embedding its device in a spa environment alongside hot tubs, saunas, and cold plunges. Diagnostic medical uses will require FDA approval; the initial focus is body composition mapping.

If Midjourney can build a library of full-body scans taken over months and years, that longitudinal record would give doctors and AI health tools a level of baseline data that doesn’t currently exist at scale outside of clinical trials. That’s the same structural logic Ksenia traced through Cursor and GitHub: The value compounds inside the loop through repeated, precise measurement over time. Midjourney is positioning itself to own that loop in the health domain.

What’s next

The competition for AI advantage is moving from model capability to infrastructure position. Who owns the coding loop? Who controls access to frontier systems? Who builds the measurement environment where health data accumulates over time? Those questions are about where intelligence meets operational reality, not which model scores highest on a benchmark.

Hiring news from the week reinforces how seriously the labs are treating this phase. John Jumper, the Nobel laureate who shared the prize with Demis Hassabis for AlphaFold, left Google DeepMind for Anthropic. Noam Shazeer, one of the coauthors of “Attention Is All You Need,” reportedly left Google for OpenAI after Google paid approximately $2.7 billion to bring him back in 2024. The labs are betting on scientific talent at the same time they’re betting on infrastructure.

Next week, host Andreas Welsch will be back to discuss multi-vendor strategy with Conductor’s Matt Palmer. They’ll cover Sakana’s launch of Fugu, Qualcomm’s ~$4B move for Modular, Anthropic’s Claude Tag stepping into Slack as a virtual coworker, Samsung putting ChatGPT and Codex in front of its entire workforce, and more. Register here to attend live.

Starting in July, registration for the live event will be open only to O’Reilly members. (If you’re interested, try O’Reilly out for free.) We’ll continue to publish our takeaways here on Radar each Friday and share full episodes on YouTube, Spotify, and Apple.