China May Lock Down Its Best AI Models

This week on the AI Inside podcast, Jeff Jarvis and I dug into a strange role reversal, with China weighing whether to wall off its best AI models from the rest of the world while its national champion races to build its own chip. We also got into OpenAI offering Washington a five percent stake to ease the political heat, the first ransomware attack run entirely by an AI agent, and Meta’s genuinely contradictory week on privacy.

But first, a huge thank you to our patron of the week, and newest patron Gary Hunt, for supporting us directly at patreon.com/aiinsideshow. You keep this show going, THANK YOU!

For the last year, the story has been that Chinese models are the cheap, open alternative to the American labs. Reuters reports that Beijing held meetings led by the Ministry of Commerce with Alibaba, ByteDance, and the startup Z.ai about restricting overseas access to their most advanced models, both closed and open-weight. They even discussed making a leak of proprietary AI tech a national security offense. If Beijing decides those models are a strategic asset to guard, the cheap floor everyone’s been standing on goes away.

And the same day, Reuters also reported that DeepSeek is building its own inference chip to cut its reliance on Nvidia and Huawei. Rounding it out, Alibaba is banning its own employees from Claude Code starting July 10 after a hidden geo-detection routine turned up in the tool, and told staff to switch to its own Qoder. A lot of tightening is happening in China, which is a direct response to the tightening happening in the US. So I suppose I shouldn’t have been that surprised after all.

Anthropic brought Claude Cowork to the web and mobile this week, and they also shared some very interesting stats about who is using it most. When they looked at 1.2 million sessions, more than 90 percent of the use was not software development. The biggest category was business process and operations at about a third (exactly how I rely on Cowork these days), then content and copywriting at 16 percent, with actual coding under 9 percent. Agents made their name writing code, and it turns out most people reach for them to handle the boring work around the work.

And tying right back to the China story, Anthropic also redeployed Fable 5 globally on July 1 after the US government lifted the export controls it had imposed in June. You have until this Sunday, July 12, before Anthropic switches it to paid usage, so get in there now!

OpenAI might be trying to buy its way out of a headache in Washington. CNBC reported that OpenAI offered the US government a five percent equity stake, worth somewhere around 42 billion dollars, to address political blowback and secure good relations. OpenAI is operating under a setup where it has to send engineers to DC and wait for the Commerce Department to sign off before it can ship a model, and this week Commerce cleared GPT-5.6 for a wide launch, about six days after the stake offer surfaced. Nobody has shown a quid pro quo, but the timeline is right there in the open.

Meanwhile, the Wall Street Journal’s opinion page has been hammering the whole idea as state socialism, even as Bernie Sanders argues the government should take far more. Everyone seems to agree that Washington should own a piece of AI, and the fight is quickly becoming just how much.

The cloud security firm Sysdig documented what it believes is the first ransomware attack run entirely by an AI agent, which they’re calling JadePuffer. The agent handled the whole chain itself, from recon to credential theft to encrypting over 1,300 configuration items, and it adapted in real time, in one case going from a failed login to a working fix in 31 seconds. The real tell that it was AI was that it left natural-language code comments explaining its own reasoning. It was also pretty sloppy, though, since the encryption key was never saved anywhere, which means victims probably couldn’t recover their files even if they sent payment to the included Bitcoin address. Those tells might not last forever as these systems continue to improve, so don’t get too cozy.

Meta launched Muse Image and a Muse Video preview out of its Superintelligence Labs, and the interesting part is that Muse Image works like an agent, writing and running code to get a chart or QR code right and refining its own output mid-generation. One feature lets anyone at-mention a public Instagram account and pull that person’s public photos to generate new images of them, with no consent, no notification, and on by default. You can opt out, but it only stops future images.

As for Meta’s smartglasses, The Financial Times reports Meta is testing always-on super sensing glasses where the recording light might not come on, in the same week Meta announced it is hardening that exact light against tampering.

The Wall Street Journal reports that high-earning families are pulling their kids out of traditional school for AI-first alternative schools. The flagship is Alpha School in Austin, with two hours of AI tutoring a day, then project workshops, up to 75,000 dollars a year, and teachers rebranded as guides. Jeff’s take was that years from now these poor children will tell therapists their parents tried to turn them into capitalists instead of letting them be kids. A study of more than 26,000 students found AI-assisted homework scored higher, but exam performance dropped by as much as 24 percent. Also worth considering the equity piece, because AI tutoring is arguably the closest thing we’ve had to free one-on-one help for any kid with internet, and here it is getting sold as a luxury upgrade.

Nvidia’s Kyber Rack Slips to 2028. Nvidia’s next-gen Kyber rack, the system built to house 2027’s Rubin Ultra chips, is reportedly slipping more than a year to 2028, according to SemiAnalysis. The holdup is a 78-layer circuit board, one of the most complex ever designed for a commercial product. Nvidia’s spokesperson says the roadmap is intact and denies the delay, so file this as a report versus a company denial for now.

Claude Has a Silent Inner Workspace. Anthropic put out interpretability research finding that Claude has a small internal workspace called “J-Space” where it seems to think without speaking, holding a few dozen concepts it isn’t saying out loud. Delete that workspace, and Claude still talks fine, but its multi-step reasoning drops to almost nothing. Apparently they can read it, and they caught the model privately noting when a test scenario was fake and even flagging its own dishonesty. Anthropic is explicit that this is not a claim that Claude is conscious.

Cloudflare Makes AI Pay for Content. Cloudflare is changing its defaults so that starting September 15 it will block crawlers that mix search and AI scraping from ad-supported pages unless the site owner opts back in. CEO Matthew Prince said that now that most traffic on the internet is non-human, they have to move faster so a sustainable ecosystem can survive. They are turning last year’s Pay Per Crawl into Pay Per Use, where publishers get paid when their content actually shows up in an AI answer. Jeff was not happy about this on the show and didn’t mince words either.

xAI Is Now SpaceXAI. xAI finished its rebrand this week, and it is now SpaceXAI, complete with a merged logo. Not much more to say about that word salad.

Tilly Norwood Lands a Movie. Tilly Norwood, the AI-generated actor that triggered Hollywood last year, is getting her first feature film, a comedy-drama called Misaligned, where she plays an AI being living in a digital world. SAG-AFTRA’s line has been blunt, that she is not an actor but a character trained on the work of countless real performers without permission or pay. But more of this is coming, just wait.

Big thanks to our Executive Producers: DrDew, Jeffrey Marraccini, Radio Asheville 103.7, Dante St James, Bono De Rick, Jason Neiffer, Jason Brady, Anthony Downs, Mark Starcher, and Karsten Samaschke.

Thanks for reading and watching. You can find audio, video, show notes, and subscribe links at aiinside.show.