Follow builders, not influencers.
TL;DR
Karpathy said free ChatGPT lagged while frontier coding models didn’t. Albert pushed cheap-to-smart escalation, Rauch said cloud infra went agent-native, and OpenAI’s next leap looked like autonomy—not chat.
Letting Sonnet “phone a friend” by calling Opus boosts performance and cuts total cost, because it stops wasting tokens on hard problems. It’s a clean routing pattern for Anthropic-style systems: use the cheaper model first, then escalate only when the task actually needs it.
He says prompting won’t fade — it’ll stay a high-leverage skill, like writing or public speaking, because it’s really how humans talk to agents through the harness. The bigger goal, from Anthropic’s Claude Code side, is to widen the bandwidth between people and agents so they understand each other better. He also points to the Monitor Tool as a strong example: you have to ask Claude Code to use it, but it can watch a dev server and catch errors.
He says people are arguing past each other because they’re using very different versions of AI: the free, older stuff that still fumbles obvious questions versus the paid frontier models that can spend an hour restructuring codebases or hunting vulnerabilities. His bigger point is that technical domains like programming and math are pulling ahead fastest because they have clear rewards, so RL can actually push them hard.
Google’s Gemini app is opening up Lyria 3 so anyone can generate up to five full songs a day, with 30-second clips still available after that. The pitch is simple: images, video, and now music — let users bring the ideas, and Gemini supplies the tools.
He says the cloud is shifting from serving developers to serving coding agents: think long-running compute, sandboxes, token delivery, and infrastructure that’s self-configuring, self-healing, and self-securing. The punchline is that Vercel’s next act is helping agents both build and run software, not just ship websites.
He says companies are massively underestimating how much software and automation AI will create in “non-software” work — from pharma simulations to healthcare workflows to bank analysis. The big shift is agents lowering the cost enough that long-running, background automation becomes practical, not just chat. At Box, he sees the bottleneck now moving to compliance, security, and messy legacy data, not demand.
They say Claude Code now sets up much faster with Bedrock and Vertex, which should make it easier for teams to wire Anthropic’s coding tool into enterprise cloud stacks. Small tweak, but the kind that removes friction where adoption usually stalls.
He says Claude Code can be a decent personal assistant replacement, but it still doesn’t feel as “mine” as OpenClaw — even though no model matches Opus for that workflow. He’s basically drawing the line between useful automation and tools people actually want to live in.
She says she rarely uses AI for writing because she actually enjoys it — and the back-and-forth usually makes it slower, not faster. Her real win is using AI outside her comfort zone, like coding, which is the cleaner take on where these tools actually save time.
He says GBrain gives OpenClaw and Hermes Agent “perfect total recall” across 10,000+ markdown files, basically turning them into a mini-AGI setup. It’s MIT-licensed, works with the same install script on Hermes Agent, and he’s already using it in his own stack.
OpenAI sees autonomy, not chat, as the next leap in AI
The Takeaway: The real frontier isn’t smarter prompts — it’s models that can work autonomously for days and learn from the world.
Key Insights
The Story
Jakub Pachocki, OpenAI’s chief scientist, is thinking less about flashy demos and more about what it takes for models to become real collaborators. He says the company’s internal shift is already visible in coding: “we use Codex for the majority of actual coding,” which he sees as evidence that autonomy is moving from theory into daily work.
For him, math benchmarks were never just trophies — they were a “North Star” because they’re brutally clear about success and failure. That same logic now extends to research. OpenAI is watching for models that can discover new things, not just answer questions, and he believes the jump from short tasks to long-horizon work is the key transition.
His view on alignment is equally pragmatic: the hard problem is generalization. Models need to learn what “good partial progress” looks like, especially when the task is messy, open-ended, or tied to the real world. That’s why he thinks the future of AI won’t be a single universal harness, but systems that “meet you where you are” — whether that’s Slack, code, or a scientific workflow.
The punchline: the next wave won’t just be more capable models. It’ll be models that can stay on task, adapt to context, and eventually run for “a couple days” with enough autonomy to produce genuinely useful work.
Woodward gave Gemini a second brain with Notebooks, while Anthropic shipped Managed Agents to move Claude from prompt to production. Rauch called the web AI’s native OS, and Levie, Masad, and Shipper all bet agents will do the work, not the people.
Albert teased Anthropic’s Mythos Preview, Cat Wu juiced Claude Code’s CLI tricks, and Peter Steinberger patched CodexBar with 2 providers plus billing fixes. Levie said agents are eating knowledge work, while Nikunj Kothari preached retention over launch hype.
Levie said agents won’t erase work, just push it up a layer; Yang argued they’ll shrink teams, not ambition. Garry Tan flagged an unpatched file leak in Claude’s coding env, while Kothari called Anthropic’s revenue ramp absurdly fast.
Rauch said v0 now builds physics, not just UI, while Karpathy noted GitHub Gists have weirdly good comments. Levie argued AI efficiency creates more work, not less, and Tan called open source’s golden age.
Karpathy pushed “your data, your files, your AI.” Levie argued context beat raw model IQ in enterprise AI. Garry Tan said GStack kept shipping security fixes fast, while No Priors spotlighted Periodic Labs’ bet on atoms, not just text.
Claude plugged into Microsoft 365 everywhere, Swyx said Devin one-shot blog-to-code, and Peter Steinberger called out GitHub’s API as still not built for agents. Aaron Levie hit the context wall, while Garry Tan shipped a DX review tool from his own stack.
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.
Steinberger called plan mode training wheels, while Thariq gave Claude Code a mouse-friendly renderer and Cat Wu showed sessions jumping phone-to-laptop. Masad framed Replit as an OS for agents, Rauch said Vercel signups compounded fast, and Anthropic’s infra tweaks swung coding scores by 6 points.
Levie said AI productivity hit the enterprise risk wall, while Weil argued proofs got cleaner, not just better. Agarwal floated public source code as the new prod debugging, and Data Driven NYC claimed one founder could run a company if agents handled the layers below.
Karpathy warned unpinned deps can turn one hack into mass pwnage, while Rauch and Levie said agents still need human guardrails and redesigned workflows. Meanwhile Claude Code got enterprise auto mode, Replit added built-in monetization, and Swyx spotted “Sign in with ChatGPT” already live.
Andrej Karpathy highlighted how LLMs can argue any side, suggesting we use it as a feature. Guillermo Rauch finally shipped his dream text layout, bringing his vision to life. Meanwhile, Amjad Masad claimed AI is democratizing app building and elevating top engineers.
Andrej Karpathy suggested leveraging LLMs' ability to argue any side as a feature. Guillermo Rauch turned text layout dreams into reality with Vercel's latest feature. Meanwhile, Amjad Masad claimed AI is democratizing app building, liberating top engineers for bigger challenges.