Three separate stories today are all circling the same uncomfortable question: what does it actually mean to "own" an AI-powered product when the model is a black box, the product boundaries are dissolving, and your most sensitive data is the thing you can't put inside any of it?
01
Why enterprise AI is harder than the demo suggests
Box CEO Aaron Levie posted a thread that cuts through a lot of the "just fine-tune a model on your data" advice circulating in enterprise circles. His argument: the information that would make a company's AI genuinely powerful is also the information you absolutely cannot pack into a model. It changes constantly, it's often subject to access controls (not everyone in the company should see everything), and your security logic has to live outside the model anyway. His conclusion is that there will be far more use cases for custom-trained models than most people expect, but the architecture looks nothing like "give the model all your data." ---
Why it matters: If your company is planning an AI rollout that assumes a fine-tuned model solves the "it doesn't know our stuff" problem, Levie is describing why that plan hits a wall. The answer isn't a smarter model. It's a permission layer, a retrieval system, and a security architecture that treats the model as one component, not the solution.
OpenAI publishes a guide to measuring AI ROI in the agent era
OpenAI's blog put out a piece aimed at enterprise buyers trying to justify AI spend. The core argument: stop measuring AI by activity (queries run, tokens consumed) and start measuring by "useful work per dollar." The post walks through how to identify high-value workflows, run efficiency benchmarks, and scale from pilot to production. ---
Why it matters: This piece exists because enterprises are getting their AI budget renewal conversations now and the old metrics aren't working. If your CFO is asking why the ChatGPT Enterprise invoice went up 40% and you can't point to an output, this framing is what OpenAI wants you to bring into that meeting. It's a sales document dressed as a strategy guide, but the underlying measurement advice is sound.
Vercel's CEO shares a building block for AI agents that run A/B tests
Guillermo Rauch, CEO of Vercel, posted about a new capability: giving AI agents the ability to set up and tune feature flag experiments autonomously. The idea is that an agent doesn't just build and deploy a website, it can also configure the experiments that optimize it over time. ---
Why it matters: Yesterday Rauch was arguing you should own your AI stack rather than outsource it. Today he's showing what that stack looks like in practice. An agent that can write code, ship it, and then tune the experiments testing it is a meaningfully different thing than a code assistant. This is what "autonomous application" actually means, and it's further along than most people's mental model.
**Swyx's multi-model workflow for serious projects** — Swyx posted his current setup for large coding work: GPT Sol Ultra for planning, Fable 5 for critique, and Claude Sonnet 5 or SWE 1.7 for the actual code generation. The interesting part is using a separate model specifically to review and challenge the work before shipping. If you're still using one model for everything, this is the direction power users are moving. **Aditya Agarwal on what we're actually using AGI-level tools for** — Aditya Agarwal posted a screenshot of asking a "AGI LEVEL coding agent" what necklaces Benson Boone wears, for his daughter. He didn't even know if he was in Codex or ChatGPT at the time. Funny, but the point lands: the interface consolidation is real and so is the gap between the marketing and the actual queries.