|
AI Builders Digest
Wednesday, July 15, 2026
|
|
The enterprise AI conversation keeps circling the same drain: companies want AI to know everything about their business, but the information that would actually make it useful is exactly the information they can't hand over. Today's feed makes that tension concrete, from two different angles.
|
|
---
|
|
01
|
Why your company can't just "train a model on everything"
|
|
|
Box CEO Aaron Levie posted a sharp reality check on the "every enterprise needs its own model" thesis that's been making the rounds. His argument: the data that would make a custom model genuinely valuable is also the data with the strictest access controls. Your most sensitive customer records, your deal terms, your internal strategy docs. You can't bake that into a model and still enforce who gets to see what. The security layer has to live outside the model, which means the model will always be working with a partial picture. Levie's take is that context retrieval and agent architectures will dwarf fine-tuned models by at least 100x in terms of real enterprise use cases.
|
Why it matters: If your company is budgeting for a "proprietary model trained on our data" project, you should read this before the next planning cycle. The thing you want the model to know is probably the thing your legal and security teams won't let you put in it. That's not a technical problem with a technical solution.
|
|
Source →
|
|
---
|
|
02
|
OpenAI publishes its framework for measuring AI ROI in the agent era
|
|
|
OpenAI posted a guide aimed at enterprise buyers trying to figure out whether their AI spending is actually working. The core proposal is to stop measuring inputs (seats, tokens, licenses) and start measuring "useful work per dollar," which they define as completed workflows with verifiable outcomes. The piece is clearly aimed at CFOs and IT buyers who are getting pressure to justify AI budgets without obvious metrics.
|
Why it matters: The fact that OpenAI is publishing this suggests enough enterprise customers are struggling to answer the "is this working?" question that it's becoming a sales problem. If your company is evaluating AI spending right now, the benchmark your vendor proposes will shape what you optimize for. Optimizing for "useful work per dollar" sounds right until someone has to define "useful."
|
|
Source →
|
|
---
|
|
03
|
Swyx's model stack for serious projects
|
|
|
Swyx, the AI engineer and writer behind Latent Space, posted his current working setup for large projects: GPT-5.6 Sol Ultra for planning, Fable 5 for critique, and a rotating cast of Claude Sonnet 5, Terra Ultra, and SWE 1.7 for actual code generation. He also uses Devin for review, with a tool called Kakuna on top, and recommends running a "grill-me" or "interview-me" prompt variant at the start to surface decisions before the work begins.
|
Why it matters: The fact that a serious practitioner is running five different models on a single project is its own story. No single model wins every task. If you're building a workflow that routes everything through one provider, you're probably leaving quality on the table somewhere.
|
|
Source →
|
|
---
|
|
|
|
**Vercel's CEO promotes agent-controlled feature flags** — Guillermo Rauch flagged a new building block letting AI agents configure and tune A/B experiments autonomously, framing it as infrastructure for self-optimizing applications. Yesterday we covered Rauch's case for owning your AI stack. This is what that looks like in practice.
|
|
Source →
|
|
**The "AGI-level" hype check** — Aditya Agarwal, posting with some genuine bewilderment, noted he couldn't tell whether he was using Codex or ChatGPT, and that he was using a self-described "AGI-level coding agent" to look up what necklaces a pop star wears for his daughter. It's a small joke, but it lands. The gap between the marketing and the use case is real, and most of the time the use case wins.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|