Sub-$2,500 humanoid, a $4B Taiwan chip hub in Texas, and Anthropic IPO doubts

The most telling story today isn't a launch. It's a quiet cluster of them, all circling the same question: who do you trust when AI touches something that matters?

Start with the unglamorous news. Researchers have flagged high-risk vulnerabilities in a widely deployed LLM gateway tool, the plumbing that organisations use to connect their systems to language models. I find this more concerning than most headline-grabbing model releases. Gateways sit in the middle of everything: credentials, prompts, internal data. When the middleman is compromised, it doesn't matter how good the model behind it is. Perplexity seems to have read the same room. After several high-profile agent incidents, it is open-sourcing SPACE, a secure sandbox for running AI agents. Whether SPACE holds up under real adversarial pressure is an open question, but open-sourcing it invites exactly the scrutiny it needs. If you're building agents in production, the practical lesson is that isolation is no longer optional.

The NASA Ames work points the same way, from a different angle. Instead of asking one model to answer aviation questions and hoping for the best, the team routes natural-language queries through nine validated expert models, so answers are grounded in verified tools. This is the pattern I expect to win in high-stakes domains: the LLM as an interface and orchestrator, not an oracle. By the way, IBM Research's result fits here too. Its "agent factories" automated the architectural design of quantum error correction decoder chips and cut inference cost by 11.5x. That's a striking number, though I'd want to see how it generalises beyond this specific design problem before calling it a broad shift.

Then there's the geopolitical layer. Taiwan is reportedly preparing to mobilise around $4 billion for industrial parks in the US, with a Texas complex at the centre, to anchor a full supply chain around Nvidia and its partners. Think about what that means: the world's most important AI hardware dependency is being physically relocated, at least in part, closer to its biggest customer. Meanwhile, a Chinese team says it has used a homegrown LLM to help crack a major mathematical problem, a first for domestic models. The details matter here, and I'd like to know how much of the heavy lifting the model really did. Still, the signal is clear: China wants proof that it can do frontier-grade work without foreign tools.

Not everything today deserves the same weight. Faraday Future's FF Nova, billed as an embodied AI humanoid under $2,500, comes with a developer programme and a curriculum, which is a smart way to build a community. But at that price, I suspect we're looking at an educational platform, not the general-purpose humanoid the framing implies. I'll gladly be proven wrong.

Finally, a comprehensive study on training data compliance, covering copyright, personal data and trade secrets, is a reminder that the legal bill for how models are trained hasn't fully arrived yet. As gateways, agents and supply chains all get hardened, will data provenance be the next weak point we're forced to fix?