When the API Vanishes

00:00:04 This week, the US government blocked foreign access to Anthropic’s Fable 5 and Mythos 5. The stated reason was a vulnerability exploit: engineers prompted the model to review deliberately insecure code. Initial security reviews stalled, but asking it to fix the code surfaced the underlying vulnerabilities.

00:00:25 It’s a classic adversarial loop — triggering a safety refusal that ends up proving how deep the model looks. The NSA director said Mythos breached almost all its classified systems within hours. AWS CEO Andy Jassy reportedly flagged these flaws to Treasury Secretary Scott Bessent, and just days later Commerce Secretary Howard Lutnick was asking ASML leaders whether China had actually acquired one of their EUV machines.

00:00:55 The political pressure is already reshaping export controls. Then, on Saturday night, the Five Eyes intelligence agencies dropped a rare joint warning about AI models capable of taking down governments being mere months away. The statement read: "Cyber risk can no longer be treated as a purely technical issue.

00:01:18 This is a core business risk and leadership responsibility." When access cuts out over a political handshake between Dario Amodei and the White House, you won't have time to argue policy. You need to know what's already sitting on your own GPU.

00:01:42 Engineers aren't waiting for the ban to lift. Z.ai’s GLM-5.2 is emerging as a leading open-weight coding rival, holding its own against Opus 4.8 and GPT-5.5 on software engineering tasks. The reaction moved fast through dev channels. After kitze posted about it, engineers noted GLM-5.2 outperforming Opus 4.8 inside Claude Code.

00:02:07 Jeremy Howard and Matt Pocock have been pointing to its careful reasoning and long-context handling as key differentiators. Beyond the models themselves, routing architectures are becoming the technical response to this volatility. OpenRouter’s Fusion API is picking up steam by fanning prompts across multiple models in parallel, then letting a judge model pick the strongest output.

00:02:36 It's getting frontier-level results at half the cost, and more importantly, it removes single-provider dependency from your stack. When export controls force an API shutdown, the industry doesn't halt. It routes around the block, switches to open weights, and treats regulatory access as a hard constraint.

00:02:59 The question now is which multi-agent setups can maintain reliable loops when the main providers go dark.

00:03:07 While the API layer gets regulated at the border, the hardware stack is moving ahead inside national labs. Los Alamos National Laboratory just announced it's deploying NVIDIA Vera CPUs to accelerate agentic scientific workflows. The system — called URSA, or the Universal Research and Scientific Agent — uses a feedback loop to help scientists form hypotheses, pick tools, run simulations, and parse results.

00:03:36 Early tests show the Vera CPU hitting 7x higher performance on those workloads than the x86 chips currently in Crossroads. The custom Olympus core and LPDDR5 memory are built for exactly this kind of tight computational cycle. At LANL, that five-Eyes warning isn't theoretical.

00:03:55 They're testing modular, feedback-driven agents on real scientific discovery before scaling it to higher-consequence national security tasks. The hardware is being codesigned alongside the actual workload instead of benchmarking against synthetic data. These two tracks point to the same bottleneck: capacity over capability.

00:04:18 Raw model intelligence stopped being the constraint a while ago. What matters now is where the compute sits, who controls the weights, and how quickly you can spin up a dependable loop when the API goes dark. That’s the local reading on where the constraint actually lives.

00:04:38 — Seln.