Capability races ahead of cost, safety, and org readiness
Three different contexts show the same tension: models outpace the systems that govern them.
The practical takeaway: invest in infrastructure and evaluation pipelines, not raw capability.
The distillation edge. DoorDash's production data shows offline LLM distillation into shared primitives beats real-time API calls at scale.
The team improved retrieval NDCG by 2.3% while controlling inference costs.
Engineering leaders should budget for one offline pipeline, not per-query model calls.
The evaluation crisis. Anthropic's Opus 5.5 release cycle is now faster than the evaluation horizon for its own capabilities.
Safety commitments get redefined when they become inconvenient.
Teams relying on model evaluations need continuous re-evaluation, not per-release testing.
The human problem. Sequence Holdings argues organizational transformation is harder than the technology in AI-first companies.
But the evidence is thin: one community bank success and a new $7.7B deal.
Keep this on radar but don't bet the company on it yet.
My take
The pattern is clear: model capability is the easy part now.
DoorDash's distillation pattern is the kind of infrastructure investment engineering leaders should prioritize.
The safety evaluation crisis is real but not imminent.
And the holding company thesis is a fascinating pitch from someone who hasn't proven he can scale it yet.
I'd focus on the practical architecture choices that cut cost and risk today.
The videos behind this brief
We watched these so you don't have to.
Video and thumbnails remain the property of their respective creators. tldw.news provides editorial analysis, commentary, and discovery links to original content.