TL;DR
Claude landed computer use on Windows, Karpathy argued LLMs should build your wiki, and Amjad Masad pushed Replit deeper into enterprise sales. Peter Yang said Cursor 3 got out of the agent’s way, while Peter Steinberger warned AI slop was flooding kernel security with real bugs.
Computer use in Claude Cowork and Claude Code Desktop is now available on Windows. That’s a practical expansion for teams already using Claude to automate workflows and code tasks across more of the desktop stack.
He says the next step beyond chat is an LLM that continuously compiles your sources into a living markdown wiki, lints it, answers questions against it, and feeds new outputs back in. In his setup, Obsidian is basically the frontend while the model does the writing, organizing, and cleanup — and he thinks this is a real product, not a hacky script pile.
He says Replit just opened a sales office in Salt Lake City and is hiring there, a sign the company is leaning further into enterprise. He also teased an SEO audit tool and a no-setup enterprise auth solution, hinting at a broader push to make Replit easier to adopt and sell into bigger orgs.
He argues the best AI tools keep you in the driver’s seat: visible agents, editable plans, clear state, and diffs you can inspect instead of blindly accept. That’s the philosophy behind Cursor 3, which he says starts simple and opens up more power across local and cloud projects as you need it.
He says building agents is a loop of adding scaffolding, then ripping it back out when the models improve. In Box Agent, that meant pruning chunking, search, and other mitigations that started hurting once frontier models got better. The takeaway: don’t get sentimental about old guardrails — keep rebuilding around what the models can actually do now.
He says AI-generated reports are now hitting the kernel security list at 5–10 a day, up from a few a week, and most are actually valid. His prediction: this won’t just annoy maintainers — it’ll kill some OSS projects unless they adapt fast.
They’ve been testing Cursor 3.0 for a week and point readers to a full vibe check, so this is more hands-on review than hype. The rest is just lightweight cheerleading and UI appreciation — not much signal there, but the Cursor take is the one worth reading.
They’re pitching a sports version of TBPN — basically turning the live, internet-native sports conversation into a product. The follow-up replies frame it as a flexible collaborator model: sometimes a PM, sometimes a product marketer, depending on what you need.
AI labs could cut science costs and speed discovery 10x
The Takeaway: Jason Kelly thinks the real biotech breakthrough isn’t better models—it’s autonomous labs that let AI run experiments nonstop.
Kelly, founder and CEO of Ginkgo Bioworks, has been chasing the same mission since 2008: make biology easier to engineer. He bootstrapped for years, then found traction after YC in 2014, but the core idea never changed. DNA is code, cells are programmable, and the bottleneck has always been the painfully slow “compile/debug” cycle of biology.
What changed is where he’s placing the bet. He’s moved away from trying to solve the hardest design problem directly and toward the lab itself. His argument is blunt: “It was that it could run experiments.” The model didn’t need genius-level intuition; it needed access to a robotic lab, fast feedback, and the ability to share raw experimental data across many hypotheses every day.
That’s why he sees autonomous labs as a structural shift, not a tooling upgrade. Today’s science is fragmented, under-shared, and dominated by overhead. His vision is a fleet of AI scientists running 24/7, learning from each other in real time, and turning research into a much more efficient, usage-based system. For Kelly, that means biotech isn’t just getting digitized—it’s about to be reorganized.
Peter Steinberger called AI code review the killer dev workflow, while Thariq said Claude Code turned files into a workbench. Cursor bet on specialized models plus brutal RL infra, and Anthropic detailed how it contained Claude across products.
Steinberger said AI agents need lean skills, not bloated prose; Levie argued AI won’t kill jobs, it’ll raise the bar. Yang said Codex tested itself while Claude still owned frontend, and Garry Tan pushed hard evals as the way agents got better fast.
Levie warned AI demos fool CEOs; the hard work starts after. Tan said high agency plus taste beat everything, while Yang urged builders to build the system that builds the MVP first. LeCun kept the pressure on: LLMs help, but they’re not the road to real intelligence.
Levie said AI kills tasks, not jobs, so roles just expand. Masad said Replit’s AI stack can ship MVPs in a weekend, while Tan showed Qwen fine-tunes in hours and personal AI got real. No Priors said Cerebras won by betting on speed before the market cared.
Levie said AI will surface more bugs, making humans the bottleneck. Steinberger argued GitHub’s native PR limits kill bot policing, while Tan pushed the “ship the 60%” rule and Swyx said AI coding should harden codebases, not just move faster.