TL;DR
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.
CEOs are especially prone to AI psychosis because they’re far from the last mile, so they see the happy-path demo and miss the 10–20 follow-on steps needed for real enterprise value. The point: use AI a lot, but come out with a clearer sense of both the upside and the ugly integration work agents still need.
He says the current unlock is pairing high agency with high taste — basically, being able to spot what matters and move fast on it. The other posts are more vibe than substance, but the throughline is clear: YC’s Garry Tan is still pushing the idea that builders on the frontier win by moving early and with judgment.
He says the Bun rewrite points to a bigger shift: old codebases will become valuable raw material for distilling software into new forms. The bigger bet is that models will get good enough to turn legacy apps into cross-platform, web-native software — no COBOL-era baggage required.
He says the old startup playbook is flipping: instead of hiring fast, founders should set up AI agents to handle email, meetings, sales outreach, and even coding. In his chat with Ryan Carson, the big idea was simple — spend upfront on docs, skills, and access, then one founder can run like a 10-person team.
He argues the real trick with Claude Code isn’t DOM scraping — it’s sniffing network requests, reverse-engineering the underlying APIs, and automating against those instead. That leads to a bigger prediction: websites will need to be headless, and the web will eventually need something like a `tools.txt` for agents to discover what they can use and pay for.
CEOs are feeling AI FOMO, but too many are still running the company from arm’s length — which leads to vague mandates and fake-progress demos instead of real adoption. The warning: if leadership won’t get hands-on with AI, a startup with tighter execution will eat their lunch.
He argues that if model weights stopped improving today, better tooling and orchestration around them could still make people feel AGI across every domain. It’s a neat reframing from OpenAI’s Yann Dubois: the missing piece may be the harness, not just bigger models.
LeCun says LLMs are useful, but not the road to real intelligence
The Takeaway: Yann LeCun thinks the future belongs to world models that plan, not chatbots that predict tokens.
LeCun, the Meta AI veteran and Turing Award winner, has now spun out Ami Labs to push what he calls “AI for the real world.” His argument is blunt: language is special, but reality is messier — continuous, noisy, high-dimensional — and that’s where current architectures break. He points to JEPA-style systems as the better path because they learn representations by predicting one view from another, rather than reconstructing every pixel. In his view, that shift matters because intelligence isn’t about regurgitating inputs; it’s about anticipating outcomes and choosing actions through search.
That’s also why he’s skeptical of vision-language-action models and imitation-heavy robotics. They can look impressive, but they’re brittle and data-hungry. “Why can’t a 17-year-old learn to drive in twenty hours?” he asks. If a machine needs endless demonstrations for every new task, it’s not really generalizing. His near-term targets are industrial control, robotics, and other complex systems like jet engines, power plants, and even patient modeling. The long game is bigger: “what we’re designing are systems that are capable of thinking.”
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 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.
Altman shipped Codex while asking what AI should fix. Levie saw AI pricing split into tiers, Masad pushed buy-first rewards, and Anthropic added Auto Mode to Claude Code as the reliability bar jumped.