Best of your X follows: June 5
Today: ChatGPT gets a major memory upgrade and a web-app publishing feature; Ethan Mollick maps the widening frontier gap and raises open-weights sustainability risk; Naval calls agent-first the next full platform rebuild; and Paul Graham points to a YC startup that built an MRI in 101 days and explains why large companies failing to profit from LLMs is actually a signal for startups.
Today's window (June 4 18:00 UTC – June 5 18:00 UTC) pulled in a tighter but sharper-than-usual set: ChatGPT gets smarter memory, Ethan Mollick maps where the frontier is actually going, Naval calls the next platform shift, and Paul Graham notices startups doing things incumbents can't.
AI tools & products
ChatGPT memory gets a big upgrade — Sam Altman announced a major memory improvement rolling out today, alongside a new feature that lets users build and publish web apps directly from ChatGPT. Altman tied the nostalgia for HyperCard to the new capability: the goal is to let anyone make and share software without a development environment in the way.1
Cargando tarjeta de contenido…
Cargando tarjeta de contenido…
What it means: the memory upgrade is incremental; the web app publishing feature is a bigger structural bet — putting ChatGPT in the same lane as Replit and Glitch, but for people who never thought of themselves as builders.
Business & competitive landscape
The Big Three are pulling further ahead. Ethan Mollick ran through where the frontier stands after Microsoft and Meta's recent model drops: both were "fine, but not frontier." SpaceX (Grok) has not recaptured its position. Chinese labs are improving but still lag. His read: until rapid improvement stops, it's hard to see anyone catching OpenAI, Anthropic, and Google.2
Cargando tarjeta de contenido…
In a follow-up, Mollick raised an underappreciated structural risk: open-weights may not be a viable business model as compute costs rise. If Chinese labs stop releasing open weights because the economics don't work, the gap between frontier closed models and anything you can run locally will keep widening.3
The two points together form a compound bet: frontier improvement is still real even when individual tasks feel similar, and the open alternative may quietly become unavailable for reasons that have nothing to do with capability.
Developer ecosystem
Agent-first is the next platform shift. Naval posted the cleanest version of an argument that's been circulating at conference levels for months: "Software platforms are going to be rebuilt for agent-first." He's made adjacent points before (desktop → mobile → agent-first in May), but this phrasing is sharper — rebuilt, not just updated.4
Cargando tarjeta de contenido…
For developers, the practical question this raises: which platforms are already building agent-native APIs versus which are bolting on MCP connectors to existing architectures. Those are different bets.
Startups & economics
A YC startup built an MRI machine in 101 days. Paul Graham flagged this as notable without naming the company. The implication is about what AI-enabled teams can now compress: hardware development cycles that used to take years. It's a data point, not a trend — but it's the kind of data point that makes previously-implausible timelines feel less implausible.5
Cargando tarjeta de contenido…
Incumbents can't profit from LLMs — that's a startup opening. Graham also made a pointed observation about the current ROI debate: if large companies can't make a net return on their LLM token costs, the correct interpretation isn't that LLMs are overvalued — it's that incumbents historically can't adopt new technology well. "Upstarts who can" is where the return shows up.6
That framing cuts directly against the "AI bubble" narrative. The question isn't whether current enterprise deployments are net-positive; it's whether the companies building new things with LLMs from scratch are. Graham thinks the answer is yes, and the evidence is already forming in YC.
Añade más opiniones o contexto en torno a este contenido.