← BACK TO DAILY

2026-04-03

AI Builders Digest — 2026-04-03

X / TWITTER


Andrej Karpathy (karpathy)

Former Tesla AI Director, OpenAI Founding Team

Andrej shared a detailed breakdown of his personal LLM workflow for building LLM Knowledge Bases — a system where large language models compile and maintain research wikis entirely on their own. He ingests source documents (papers, repos, articles) into a raw folder, then uses an LLM to iteratively build a structured .md wiki with summaries, backlinks, and categorized concepts. The wiki is maintained by the LLM — he rarely touches it manually. For Q&A, he queries the wiki with complex research questions and the LLM agent goes off to find answers. He runs "health checks" to find inconsistencies and gaps, and even vibe-coded a custom search engine over the wiki. His conclusion: raw data gets collected, compiled into a wiki by an LLM, then operated on via CLI tools for Q&A and enhancement, all viewable in Obsidian. The workflow compounds — every query and output feeds back into the wiki, building a growing knowledge base that outlasts any single session. No filler, just a dense, practical system for research-at-scale.

"I think there is room here for an incredible new product instead of a hacky collection of scripts."

He also floated a vision of frontier LLMs spawning entire ephemeral teams of sub-LLMs to fully automate complex research tasks — each question spawning a team that iteratively constructs a wiki, lints it, loops, then writes a full report.

https://x.com/karpathy/status/2039805659525644595
https://x.com/karpathy/status/2039808711452246261


Aaron Levie (levie)

CEO, Box

Aaron laid out the brutal loop of building AI agents — and it's not what most people want to hear. His core thesis: every time models improve, you have to ruthlessly rip out the scaffolding you built around the old model. The loop is: build systems around the LLM to handle specific tasks → model improves and those systems become redundant or harmful → remove them → new capabilities emerge → repeat. He cited his own Box Agent, where the architecture evolved multiple times between initial design and release because mitigations (like chunking strategies for context window limits) were actively hurting quality as models got better. The lesson: be unsentimental about your tech. Don't become nostalgic around the scaffolding you've built.

"The main lesson is always make sure you're taking advantage of the frontier capabilities and don't become nostalgic around the tech you've already built."

X (formerly Twitter)
Aaron Levie (@levie) on X
One of the biggest lessons thus far in building AI agents is you have to be brutally unsentimental in your architecture. The models get better and better at handling things you previously built scaffolding for, you need to ruthlessly jettison your prior tech to get those new

Ryo Lu (ryolu_)

Design Lead, Cursor

Ryo articulated Cursor's "glass" philosophy — the idea that AI tools should be transparent and controllable rather than opaque and addictive. The terminal was a black box (one line in, few out); AI kept the black box and made it worse (type a wish, pull a lever, accept or reject). Glass breaks that pattern: the PM writes a plan and watches it become real, the designer sketches and iterates in live code, the engineer breaks down architecture before a line is written, the beginner reads every diff and learns. The agents are visible, the diff is there, the plan is editable, the state is clear. You can Tab and make edits when you need to. Nothing hides, nothing is forced. As models get more powerful, glass gets more important — not because you need to watch every move, but because the best work happens when you know you can.

"Your shortcuts. Your patterns. Your taste. The tool learns your way, not the other way around."

X (formerly Twitter)
Ryo Lu (@ryolu_) on X
glass vs. black box we believe you should be able to see everything. and be in control of everything. the terminal was a black box. one line in, few lines out. you learned to think like the machine. AI kept the black box and made it more addictive. now you type a wish and pull

Sam Altman (sama)

CEO, OpenAI

Sam endorsed TBPN (The Business Podcast Network) as his favorite tech show, saying he wants them to keep going and expects them to hold him accountable — and offering to "help enable that with occasional stupid decisions." The post got 5,080 likes, reflecting genuine warmth for the independent tech media format.

X (formerly Twitter)
Sam Altman (@sama) on X
TBPN is my favorite tech show. We want them to keep that going and for them to do what they do so well. I don't expect them to go any easier on us, am sure I'll do my part to help enable that with occasional stupid decisions.

Peter Steinberger (steipete)

Polymath, OpenClaw contributor

Peter flagged a growing crisis: AI-generated slop is overwhelming open-source project maintainers. The Linux kernel security list went from 2-3 reports per week two years ago, to ~10/week last year, to 5-10 per day in 2026. Most of these AI-generated reports are correct, but the sheer volume required bringing in additional maintainers just to process them. The tool that was supposed to help is creating a new category of work.

"Now most of these reports are correct, to the point that we had to bring in more maintainers to help us."

X (formerly Twitter)
Peter Steinberger 🦞 (@steipete) on X
Prediction: This is gonna kill some oss projects. "On the kernel security list we've seen a huge bump of reports. We were between 2 and 3 per week maybe two years ago, then reached probably 10 a week over the last year with the only difference being only AI slop, and now since

Peter Yang (petergyang)

Product, Roblox

No notable posts this cycle.


Swyx (swyx)

Engineer, Latent Space

No notable posts this cycle.


Nan Yu (nan_yu)

Head of Product, Linear

No notable posts this cycle.


Amjad Masad (amasad)

CEO, Replit

No notable posts this cycle.


Dan Shipper (danshipper)

CEO, Every

No notable posts this cycle.


Garry Tan (garrytan)

President & CEO, Y Combinator

No notable posts this cycle.


Ryo Lu (ryolu_) on Cursor 3

Ryo also announced Cursor 3, a redesigned interface that strips away toggles and buttons cluttering the previous version. The new UI is intentionally minimal — it starts simple and reveals more tools only when needed, keeping the user in flow and in control. Peter Yang noted the old interface had "far too many buttons and toggles that got in the way of just talking to the agent" and questioned why the old UX required Cmd+Shift+P to access core features.

https://x.com/ryolu_/status/2039780768847958359
https://x.com/petergyang/status/2039850011044016291


Claude (claudeai)

Anthropic

Claude computer use is now available on Windows in both Claude Cowork and Claude Code Desktop.

X (formerly Twitter)
Claude (@claudeai) on X
Computer use in Claude Cowork and Claude Code Desktop is now available on Windows.

PODCASTS


Training Data — "How Autonomous Labs Will Transform Scientific Research: Ginkgo Bioworks' Jason Kelly"

Jason Kelly, Founder & CEO, Ginkgo Bioworks

Jason Kelly has been trying to make biology programmable since 2008. Bootstrapped for six years before YC, then Sam Altman convinced him to apply the Silicon Valley model to deep biotech. The mission: make engineering biology easier — DNA is code (ATCs and Gs instead of zeros and ones), and cells are programmable like computers, except cells move atoms instead of just information. The challenge: our ability to program cells today is still terrible.

His central thesis: AI + autonomous labs will change the fundamentals of how science is done — something he believes previous tech revolutions (social media, internet) never actually did for biotech. "Back office IT crap," he calls those. This is different.

On the OpenAI partnership: Ginkgo gave a reasoning model a robotic lab and asked it to optimize cell-free protein synthesis. After 6 rounds of 30,000 experiments, it beat the Stanford state-of-the-art by 40%. The key insight: the model didn't need to understand biology. It just needed to be logical — design an experiment, run it, interpret data, repeat. That's just programming. What let it win wasn't raw intelligence; it was 24/7 cycle time.

On why autonomous labs matter more than AI design tools: "More than half of your drugs today are produced by biotechnology." We already create enormous value with limited tools. The bottleneck isn't the science — it's the manual lab work. PhD scientists spending 5 years moving liquids by hand at the bench. That's the problem. Solve that first, then the science takes care of itself.

On humanoids in the lab: No. Humans walking between benchtop equipment is a solved problem — just put samples on a track system. Biology is a microscopic discipline. You don't send humanoids to etch TSMC chips, and you don't send them to handle microliters.

On the future vision: 100 AI scientists running experiments in parallel on a shared robotic lab, sharing raw data daily instead of publishing papers every two years. Failed results from one hypothesis become learning for 99 others. If you solve the reagent-cost problem (today, reagents are <5% of research spending — the rest is overhead, lab space, people), you get 10x more data per dollar. Even if the AIs are dumber than scientists, they win on speed, scale, and information sharing.

On what this means for America: China is already publishing more scientific papers. In biotech deal flow, 40%+ of drugs being developed now come from China — up from under 5% three years ago. "They have just as many scientists, they're just as smart, they get paid less, and now they're getting more research per dollar." Project Genesis — the DOE's AI for science initiative, which bought 97 of Ginkgo's robotic racks — is the US response.

The memorable quote:

"All of the previous revolutions in tech, Internet, right, like social media, like whatever, have been totally meaningless to biotechnology and biopharma. ... This is actually gonna change the fundamentals of how we do science, and our big science industries like biopharma are gonna get disrupted. I really believe that. And that's not been true for the last thirty years of tech."

And on GLP-1 drugs and the consumer biotech opportunity:

"What's the value of a biotech product that adds five years to lifespan? There's no limit on that."

YouTube
Training Data
Join us as we train our neural nets on the theme of the century: AI. Sequoia Capital partners host conversations with leading AI builders and researchers to ...

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders