TL;DR
Altman pushed Agents SDK 2.0, while Levie said enterprise agents will spawn a new services layer. Karpathy said software is shifting from code to context; Masad said Replit’s agent swarm is already doing real work.
He says OpenAI’s Agents SDK 2.0 is underrated, which is a pretty direct signal that the company thinks the agent tooling story still has room to run. The rest is a warm public shoutout to Greg Brockman’s technical grit and importance to OpenAI’s decade-long arc.
He says the real bottleneck isn’t the model — it’s everything around it: secure data access, entitlements, logging, process docs, evals, and new human+agent workflows. That complexity means a big market opens up for consultants, internal agent teams, and vertical AI vendors that can actually get agents into production. He also argues AI should be treated like a utility, not a being — less mysticism, more systems design.
He shipped RepoBar 0.4.0, and the tiny menubar app is starting to feel genuinely handy: persistent SQLite caching, fewer API calls, visible rate limits, better Issues/PR loading, and archive fallback support. It’s a classic power-user polish pass — less noise, more signal, and a tool that should get better the more you use it.
He says Replit now has 10 active, 198 draft, and 700+ done tasks, and claims there’s more agentic parallelism happening there than anywhere else online. The bigger signal: marathon vibe coding can still ship a lot in a day, which is exactly the kind of founder energy he’s trying to productize.
He says personal AI only matters if you control the stack: your prompts, your data, your context. That’s the pitch behind GBrain being open source — a bet that AI should make individuals harder to capture by institutions, not easier.
He shared a quick Mac workaround for keeping AI agents running after you close your laptop: install Amphetamine, then tweak Session Defaults so nothing gets suspended. It’s a tiny but useful ops tip for anyone running long-lived workflows on a Mac.
Karpathy says software is shifting from code to context
The Takeaway: The real shift isn’t faster coding — it’s that LLMs are becoming the new runtime, and humans now steer with context.
Andrej Karpathy, cofounder of OpenAI and former Tesla AI lead, has moved from explaining modern AI to naming its next phase. His core claim is blunt: software is no longer just explicit code or trained weights — it’s becoming a system where “your programming now turns to prompting,” and the context window is the control surface.
That’s why he says he felt behind as a programmer in December, when agentic tools stopped feeling like assistants and started feeling like competent collaborators. The change wasn’t gradual. “The chunks just came out fine,” he said, describing the moment he stopped correcting the model and started trusting it. From there, he argues, the right question isn’t how to speed up old workflows, but what new things become possible when the model itself does the work.
His examples are telling. A menu-scanning app he built suddenly looked unnecessary once he realized Gemini could take the photo and render the result directly. That’s the bigger thesis: many products are still built for humans to click through, when they should be designed for agents to act on. Even then, Karpathy isn’t handing over judgment. Agents can handle the “intern” work, but humans still own taste, specs, and oversight. The ceiling is rising — fast — but the winners will be the people who learn how to direct it.
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.