Khosla-Backed Startup Claims Breakthrough With Largest-Ever AI Model on an iPhone

There's something almost uncomfortable about Anthropic's new research this week: the company says it can now read Claude's "thoughts." Not metaphorically. The paper describes a "J-Space," a kind of global workspace inside the model where something resembling unified cognition happens, and Anthropic's researchers claim they can observe it directly. If that holds up to scrutiny, it's a bigger deal than most model releases this year, because it reframes what interpretability even means. We've spent two years poking at LLMs from the outside, inferring behavior from outputs. This is Anthropic saying they've found a window into the process itself.

And yet, on the same week Anthropic published that, the Future of Life Institute's AI Safety Index gave the company a C+. Anthropic still ranked first — which tells you everything about where the rest of the industry sits. The report's core complaint isn't really about any single lab; it's that safety commitments across the board have softened as competitive pressure has intensified. I find this pairing genuinely revealing: the company doing the most rigorous internal safety research is also, by an independent panel's judgment, still falling short of what "safe enough" should look like. That's not a contradiction — it's a fairly honest signal that the bar for AI safety right now is being set well below where anyone serious thinks it should be, and that even the best-resourced lab in the world is racing to keep up with its own capabilities.

Meanwhile, the infrastructure side of the industry keeps quietly solving problems that used to require brute-force scale. NVIDIA's new Nemotron-Labs-Diffusion model collapses speculative decoding into a single checkpoint — the same model now drafts and verifies its own outputs, rather than pairing a small draft model with a larger one. It's a genuinely clever piece of engineering, and it matters because inference cost, not training cost, is what determines whether AI products are actually profitable at scale. Similarly, the Khosla-backed startup claiming the largest AI model ever run natively on an iPhone is chasing the same underlying goal as NVIDIA and Apple: push intelligence to the edge, cut cloud dependency, and make privacy a selling point rather than a compromise. These aren't headline-grabbing breakthroughs on their own, but stacked together, they represent the industry's quiet second act — making models smaller, faster, and cheaper to run, after two years obsessed with making them bigger.

By the way, the enterprise agent stories this week point in the same direction from a different angle. KTern.AI's SAP agents on Amazon Bedrock AgentCore and the Southeast Asian logistics firm that cut vendor onboarding from five days to four hours aren't flashy, but they're the actual proof points skeptics keep asking for — measurable time saved, not vague productivity claims. Microsoft pushing Copilot further into autonomous, multi-agent territory fits the same pattern.

So here's my question heading into the second half of the year: as interpretability research like Anthropic's J-Space matures, will it actually inform how labs build and ship models, or will it remain a fascinating sidebar to a race that's still fundamentally driven by speed and market share?