TL;DR
Altman landed Codex in ChatGPT mobile, Rauch made Terminal AI spit out images, and Amjad Masad turned any site into a free mobile app. Meanwhile, Peter Yang said hackathons now mean waiting on agents, and Peter Steinberger saw AI ship the feature then fail the review.
He says Codex is now in the ChatGPT mobile app, which is a pretty direct signal that OpenAI wants coding help to feel native everywhere, not just in a desktop workflow. The other post is just a link with no extra context, so the only real takeaway here is the mobile rollout.
You can render images straight in the terminal with `ai-cli`, then tap into image, video, and text models through Vercel AI Gateway. It’s a neat little demo, but the real point is bigger: Vercel is pushing AI access into the same fast path developers already use for everything else.
He says AI hackathons have turned into a weirdly passive experience: you spend half the time waiting for agents to do the work. He also warns that letting AI generate screens without a design system or reusable components quickly turns into slop — basically, designers were right all along.
He says the real signal that a company is getting "agent-pilled" is whether leadership uses Codex, Claude Code, or Cowork every day. He’s been quietly helping top tech companies get exec teams deep into those tools, and the barber-chair anecdote — crypto trading with Claude, jiu-jitsu with Gemini, 1,300 Claude seats — makes the adoption curve feel very real.
He says AI is temporarily jamming up job boundaries — PMs can mock up launches, engineers can think more about business, GTM can ship more of the workflow. But his bet is the dust settles back into specialization, just with much higher leverage for true experts and a better toolchain around them. The other jab: a year of agent scaffolding can get wiped out by one model update.
He says you can import a website built anywhere else into Replit and get a free mobile app — a pretty aggressive funnel for pulling vibe-coded projects into his platform. He also says Replit just got back into the App Store after four months of Apple wrangling, with more updates coming.
He’s hammering the anti-datacenter crowd, arguing the real issue isn’t energy, water, or jobs — it’s local politics choking off infrastructure. The YC chief is basically saying markets should decide where power and compute get built, not loud opposition.
He built a new media-storage feature into Discrawl, had Codex declare it done, then ran his own Codex review skill and caught the gaps. That’s the real workflow shift here: AI can get you to “done” fast, but human review still decides whether it’s actually shippable.
He says @GradiumAI is already trouncing the voice AI field on third-party TTS benchmarks — beating OpenAI, ElevenLabs, Cartesia, and Deepgram. That’s a loud signal that the voice stack is still very much up for grabs, even with the big names in place.
Every agent needs a machine, not just a model
The Takeaway: Agents stop being toys the moment you give them their own computer, permissions, and isolation.
Ivan Burazin, CEO of Daytona, has spent years building developer infrastructure, first in cloud IDEs and now in agent sandboxes. His core thesis is blunt: “every agent will need at least one sandbox, sometimes more.” Why? Because if an agent is supposed to do real work — fetch bank data, run scripts, browse the web, or operate across multiple steps — it needs a machine the way a human worker needs a laptop.
That’s why he rejects the idea of running serious agents on your own laptop. It breaks security, kills concurrency, and makes long-running work fragile. His bank-data example is telling: Claude asked him to “log in and give me access,” and that was the moment the whole setup failed. The fix was to give the agent its own Daytona account, its own phone number for 2FA, and tightly scoped permissions. The point isn’t trust; it’s containment.
Burazin also draws a sharp line between the pieces of the agent stack: models are the brain, tools are the hands, memory is still messy, and orchestration is basically management. The big frontier labs are bundling more of this together, but Daytona’s bet is that the sandbox remains the execution layer underneath it all. The deeper insight: agents don’t just need intelligence — they need infrastructure that lets them work like employees without becoming liabilities.
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.