Today's throughline is that the interesting AI work is happening a level below the megacaps, an Australian startup fixing model groupthink, Alibaba slashing agent token costs, Square quietly routing AI orders around delivery apps, and India betting on offline models as public infrastructure. Meanwhile the giants spent the day fighting over who owns the value chain, Anthropic clawing back its export ban and eyeing its own Samsung silicon, Karp needling the labs from the enterprise layer, and fresh billions pouring into the compute and defense-drone startups actually building what the AI boom runs on. The surprising consequence is that leverage keeps draining from the model layer itself toward whoever controls chips, distribution, or the applied use case.
Australian startup Springboards built Flint, a model tuned on Qwen 3 that injects variety only at key decision points so brainstorming stops converging on the same cliches.
Raised 800 million dollars at an 8.3 billion dollar valuation, led by Aramco Ventures with Nvidia and Vista Equity Partners joining in.
A live terminal proxy that shows every real tool call passing between your AI client and MCP servers, like Wireshark for MCP.
Anthropic's models are unblocked but the ad hoc approval fight shows Washington still has no real rulebook for releasing frontier AI.
