|
AI Builders Digest
Friday, October 23, 2026
|
|
Two themes collide today: cheap AI getting cheaper, and the jobs question refusing to die. Guillermo Rauch is practically daring you to run a model into the ground for pennies, while Aaron Levie is making the case that AI creates work as fast as it eliminates it. Both can be true, and the tension between them is where the interesting products get built.
|
|
---
|
|
01
|
Vercel drops DeepSeek v4.1 Flash into its Token Pass at near-zero cost
|
|
|
Vercel CEO Guillermo Rauch announced that DeepSeek v4.1 Flash is now included in Vercel's Token Pass subscription, with zero-data retention for privacy-conscious teams. His framing: use it as the cheap executor while a larger, smarter model does the planning. "You can port TypeScript to Rust 10 times and not run out of credits."
|
Why it matters: The "orchestrator plus cheap subagent" architecture just got a lot more accessible to teams building on Vercel. If you've been paying frontier model prices for every step in a multi-step workflow, this changes your cost calculus. The interesting design question is now which tasks actually need the expensive model and which ones don't.
|
|
Source →
|
|
---
|
|
02
|
Aaron Levie points to cybersecurity as AI's best jobs argument
|
|
|
Box CEO Aaron Levie is making his third consecutive appearance in this digest, and this time the argument is specific: cybersecurity is where you can see clearly how AI produces jobs rather than eliminating them. His case is that even as security work becomes more automated, you need more people to direct agents at the right problems, review what they find, and respond when things go wrong.
|
Why it matters: Every company right now is running some version of the "agents will replace headcount" pitch internally. Cybersecurity is the counterexample worth bookmarking. The more autonomous your security tooling gets, the more the human role shifts to triage and judgment, not less human involvement. If your team is buying security agents and assuming the analyst headcount follows, this is the argument you'll need to answer.
|
|
Source →
|
|
---
|
|
03
|
Microsoft Research ships Flint-interactive for AI-generated charts you can actually explore
|
|
|
Microsoft Research published Flint-interactive, a new language that lets AI agents build charts users can click, filter, zoom, and cross-reference. Instead of agents spitting out a static image, they can now generate a visualization where selecting a data point highlights related text elsewhere in the document.
|
Why it matters: Static AI-generated charts are already annoying. You ask a question, get a bar chart you can't interrogate, and then have to ask five follow-up questions to dig into the parts that look interesting. Flint-interactive is an attempt to compress that loop. The practical test will be whether it lands in Copilot and actually works under real conditions, not just in a research demo.
|
|
Source →
|
|
---
|
|
04
|
Thibault Sottiaux clarifies Pro plan users are exempt from weekly usage caps
|
|
|
Following yesterday's coverage of Sottiaux's caching work, he confirmed that whatever he's shipping is fully included in the Pro plan and doesn't count against weekly usage limits.
|
|
Context from yesterday: Sottiaux extended prompt caching to 12 hours via API, with permanent cache for logged-in users. Today's clarification means Pro subscribers get all of it without watching a usage meter. The mystery project is starting to look like a real pricing play.
|
|
Source →
|
|
---
|
|
05
|
Swyx's quant post goes viral
|
|
|
AI researcher and Latent Space co-founder Swyx posted something about "my quant" that picked up over 1,100 likes. The post itself links out but the content isn't self-explanatory from the tweet alone.
|
|
Worth noting because Swyx rarely posts things with no substance behind them, and the engagement suggests something landed. Worth clicking through if you follow the AI x quantitative finance overlap.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|