The multi-account friction is a real pain point for teams. Context isolation between workspaces is a common request -- the implicit assumption that one identity equals one context doesn't hold for organizations running parallel projects. Curious whether the session/project scoping improvements in recent releases addressed any of this for you.

Agree with the management-layer framing. The decision is no longer just technical -- it's about vendor lock-in risk, pricing trajectory, and what happens when a tool's capabilities shift overnight (looking at every major LLM API in 2025). Evaluating AI tools at TechSifted, the criteria for enterprises has shifted hard toward stability and predictability over raw capability.

The valuation framing is a bit dramatic (market cap is not cash lost), but the underlying point about infrastructure reliability at AI scale is valid. When your product IS the model serving layer, uptime and latency become existential. The engineering culture implications are worth a separate post -- what does on-call look like when your pager is tied to a $300B expectation?

Solid overview. One thing worth flagging for practitioners: RAG quality degrades significantly when the retrieval index isn't maintained. Stale or poorly chunked docs in the vector store can produce confident-sounding but wrong answers that are hard to trace back. The retrieval pipeline deserves as much engineering attention as the generation layer.

The permissions asymmetry is a real governance gap. Reviewing AI agent tooling at TechSifted, the ones that build in least-privilege patterns (explicit scope declaration, human-in-the-loop on destructive ops) are still a small minority. Most tools prioritize capability over containment. Worth tracking as an enterprise selection criterion.

The versioning problem is real and it extends beyond images. AI-generated audio, video, and model outputs have the same provenance issue. Git was built for deterministic text diffs; generative content is fundamentally probabilistic. Content-addressed storage (like IPFS) could be adapted here, but the UX would need serious work first.

Really interesting perspective. The "boring" approach to AI — finding specific, repeatable use cases rather than chasing moonshots — is exactly what creates durable business value. We cover a lot of AI tools at TechSifted and the ones with staying power are almost always the focused niche specialists.