LLMs can autonomously run experiments, tune hyperparameters, and optimize code. What they can't do: decide which question to investigate next, or recognize when an entire research track is a dead end and pivot to something fundamentally different.
Where this was said
At 2:14:00 · chapter starts 2:08:56
Dwarkesh presents an information-theoretic argument: naive RL learns near-zero bits per sample at low pass rates, while supervised learning learns negative log(pass rate) bits. The gap is enormous. [1] — Dwarkesh Patel "With a 1-in-100K pass rate, supervised learning gives you negative log(1/100K) = ~17 bits per sample. Naive RL gives you essentially zero. …" 2:09:35
With a 1-in-100K pass rate, supervised learning gives you negative log(1/100K) = ~17 bits per sample. Naive RL gives you essentially zero. You spend almost all of early training in a regime where nothing is learned.
Naive RL learns at the rate of the entropy of a binary random variable per sample, while supervised learning learns negative log(pass rate) bits — orders of magnitude more at low pass rates.
AlphaGo's elegance is that MCTS always produces a training signal — the policy is never stuck at zero pass rate waiting to stumble on a win.
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.
Inside SEO Bot, a single button labelled 'Boost My Domain Rating' routes users directly to Listing Bot. That one interaction converts a user of one tool into a user of two — without any marketing cost.
Directory listings are a powerful but underrated growth channel — but only if your product is genuinely interesting enough to earn the click. AI products have a natural advantage here because they're easy to package in a compelling, clickable way.
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.
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.
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.