|
AI Builders Digest
Sunday, October 18, 2026
|
|
Yesterday we covered Aaron Levie on token consumption exploding. Today he's back, and the thing he's describing is bigger: AI that improves itself. Meanwhile, Anthropic quietly shipped something that 3,000 people apparently needed and nobody announced properly.
|
|
---
|
|
01
|
Box CEO Aaron Levie: we may be closer to self-improving AI than anyone thought
|
|
|
Levie posted this weekend that we've crossed a practical threshold where AI systems are becoming better at training future AI models than human engineers are. He points to model capabilities exceeding most human engineers, better infrastructure for long-running agents, more training data, and the ability to run thousands of parallel experiments simultaneously. His framing: one day recursive self-improvement feels like science fiction, and then suddenly you're looking at early signs of it in the wild.
|
Why it matters: If AI systems are now helping design the next generation of AI systems, the pace of improvement stops being limited by how many ML researchers a company can hire. Every major lab hiring freeze you've read about this year looks different through that lens.
|
|
Source →
|
|
---
|
|
02
|
Anthropic's Claude gets slides support, and people who work there actually use it
|
|
|
Two sources, same story. Thibault Sottiaux posted that Claude added slides support this weekend, currently rolling out to Pro, Business, and Enterprise tiers. Boris Cherny, who works on Claude at Anthropic, backed it up with a concrete example: he now uses Claude Slides to build decks instead of the usual ritual of hunting for iPhone frame SVGs and fiddling with templates. Cherny's post pulled 3,000 likes, which for a feature announcement with no press release is a meaningful signal.
|
Why it matters: PowerPoint and Google Slides have had roughly the same workflow for 20 years. If you can describe a deck and get something presentation-ready, the bottleneck moves from "building slides" to "knowing what you want to say." That's a different kind of work.
|
|
Source →
|
|
Source →
|
|
---
|
|
03
|
Y Combinator president Garry Tan is coding new projects with two AI agents at once
|
|
|
Garry Tan shared that his current workflow for starting new projects pairs Capy (an AI coding agent) with Aside AI to handle both writing the first version of the code and setting up the DevOps infrastructure in a single pass. No manual deployment configuration, no separate ops step.
|
Why it matters: When the president of the world's most influential startup accelerator publicly describes his personal dev setup, thousands of founders take notes. The "one agent for code, one agent for ops" pattern is about to show up in a lot of YC company codebases.
|
|
Source →
|
|
---
|
|
04
|
Swyx wants the prompts behind a Kurzgesagt-quality AI explainer video
|
|
|
Swyx shared an AI-generated explainer video he called "near-Kurzgesagt level clarity" and asked publicly for the prompts used to produce it. The video itself isn't linked cleanly in the post, but the comparison is notable: Kurzgesagt is the gold standard for making complex topics visually comprehensible. Swyx is saying something generated by AI is approaching that bar on explanation quality, even if the illustration isn't there yet.
|
Why it matters: If AI video explanation is approaching Kurzgesagt on clarity, the cost of producing high-quality educational content just dropped by several orders of magnitude. Every company that pays an agency to make explainer videos should be running experiments right now.
|
|
Source →
|
|
Follow builders, not influencers. A daily digest of what matters in AI.
Read online ·
Archive
|