Podbit · All-In with Chamath, Jason, Sacks & Friedberg
Zuckerberg is posting more on X than he has in his entire history to announce MuseSpark 1.1 — a strong agentic model available through Meta's own Model API at a fraction of the cost. After fumbling the open-source scorched-earth play, Meta is now trying to win on price.
Where this was said
At 35:30 · chapter starts 27:39
David Sacks synthesizes the discussion with a framework: enterprises want model fungibility — the ability to hot-swap models for the cheapest one that gets the task done. But there's a hard technical blocker: memory, context, and history are all tied to the model, and nobody has figured out how to make them fully portable. So even technically capable teams like Coinbase and DoorDash can build routing middleware, most enterprises simply cannot. The data bears this out: open source went from 19% to 11% of enterprise AI wallet share year-over-year. He introduces the Decagon framework — open models are great for mature, well-defined use cases where you can post-train on specific data. But for immature use cases, which is everything enterprises are still figuring out, you want the most capable general intelligence available. Jason brings in Ali Ghodsi's Databricks finding: the same GLM 5.2 model, but with a different harness, produced 2x token savings. This is profound — it means much of the optimization opportunity has nothing to do with model choice and everything to do with how you structure the call. Jason shares his own optimization experience: asking his agents to self-optimize their token usage reduced consumption by 80%. The session ends with the group agreeing that the tip of the spear — the 1% of technically capable deployers — are working it all out, and the rest will follow.
After sitting on the UN AI Commission with Benioff and Jensen Huang, Chamath reports there is not a single country that doesn't have a sovereign AI strategy — and almost none of them want to depend on a closed-source American model. They'd rather take an open model like NVIDIA's and build their own soup-to-nuts stack, even if it's 5% worse.
Open source went from 19% of enterprise AI wallet share to just 11% in one year. The reason: most enterprises can't build the middleware routing needed to use cheap models for the right tasks. The spirit is willing but the technical flesh is weak.
Despite open-source model popularity, their share of enterprise AI spending actually fell from 19% to 11% year-over-year as closed frontier models dominated.
For mature, well-defined use cases, post-trained open models beat frontier models on cost. For everything you're still figuring out — which is most enterprise AI today — you need the most capable general intelligence you can get. Decagon routes 90% of its traffic to open models, but only after extensive customization.
Ali Ghodsi at Databricks found that switching the application harness — not the model — cut token costs by 2x using the same GLM 5.2 model. The implication: most enterprises are wasting tokens because of bad harness design, not because they need more expensive models.
The fastest path to Twitter growth isn't volume — it's forming sharp opinions about how the platform works and sharing them immediately. People cluster around those who understand the rules and say so out loud.
More connectivity doesn't produce more independent thinking — it produces more homogeneity. Kierkegaard called this 'leveling': the internet's utopian promise of everyone thinking for themselves is backwards. The more interconnected people are, the stronger the phantom of public opinion, and the more of a liability it becomes to stand out.
Sam had no coding knowledge, so he used ChatGPT voice mode to generate his entire codebase and copy-pasted it into Notepad. A friend later introduced him to Cursor, and he never looked back.
Copy days of Discord chat history, paste it into ChatGPT, and ask it to list recurring pain points. The ones that come up most often are your best product bets.
Sam's top advice: when prompting Cursor, tell it to architect code for 100,000 users from day one. The AI changes its approach, building scalable frameworks instead of brittle one-user code.
With AI coding tools like Cursor, Bhanu replicates an existing free tool for a new keyword in under 5 minutes. What used to be a multi-day build is now a lunch-break task.
Ahrefs, SiteGPT, Cal.com, PostHog, Datafast, Sibyl AI, Bento, Feather, Featurepace, Mintlify, Cloud Code, ChartMogul — Bhanu runs his entire business solo with these 12 tools.
PropGPT runs on React Native with TypeScript and Python for ML, Neon for the database, RevenueCat and Superwall for monetization. LLM costs are $20/month, data APIs $100/month, and after $10K in monthly marketing spend, margins sit at roughly 50%.
Game apps keep users engaged long enough for ads to pay off. Tool apps don't — so if you're building a utility, ads are almost always the wrong call and subscriptions are your only real lever.
We use essential and analytics cookies to run Vuci. To understand how the site is used: Privacy Policy.
Install Vuci on your phone
Add it to your home screen for a faster, app-like experience.
Install Vuci on your phone
Tap the Share button, then “Add to Home Screen”.
A new version is available
Reload to get the latest Vuci.