Neural Computers, GameStop’s $55B eBay Offer | Diet TBPN

From MW2 on the big screen to Karpathy’s 'neural computer'

The episode opens in classic TBPN mode: joking about interns playing Modern Warfare 2 and being late because of it. Then it pivots fast into the real theme of the day — a Sequoia AI Ascent interview where Andrej Karpathy describes the arrival of a “neural computer,” a machine that generates the exact interface you need on the fly instead of relying on fixed software.

One prompt instead of Google Finance, Excel, and a dozen tabs

The host makes the idea concrete with his own workflow: trying to understand Ryan Cohen’s GameStop proposal for eBay without already knowing each company’s revenue, profit, or valuation multiples. In the old world, that meant Yahoo Finance, spreadsheets, maybe Bloomberg or CapIQ; now it’s one prompt that outputs a shareable image showing figures like eBay at about $50 billion revenue versus GameStop at $15 billion, and eBay’s $2.28 billion operating income versus GameStop’s roughly $232 million.

Karpathy’s menu app got replaced by a raw model in one shot

They play Karpathy describing how he vibe-coded “menu gen,” an app that OCRs a restaurant menu photo and generates pictures of the dishes so diners know what they’re ordering. Then Gemini plus “nanobanana” basically nuked the whole need for the app by taking the menu image directly and rendering the dish visuals right into the pixels — his point being that the old app “shouldn’t exist.”

The vibe-coding boom may be real — and still a temporary aberration

That leads into a sharper take: lots of people are building apps for problems frontier models can already solve in chat. The hosts say they keep hearing pitches from nontechnical builders where the product idea is something Claude Code, Codex, Gemini, or ChatGPT can already do in one thread, though they also acknowledge there’s still money in good packaging, marketing, and distribution.

'Fat protocols' comes back, this time for AI models

The conversation then zooms out to Union Square Ventures’ 2016 “fat protocols” thesis from crypto — the idea that value might accrue more to the protocol layer than the app layer. Their AI version is that models are getting “fatter” every month, swallowing more of the stack and making it harder for thin wrappers to defend themselves unless they have some real moat, liquidity, or distribution edge.

Walled gardens aren’t technical limits — they’re business limits

From there they talk about the “walled garden internet theory”: the useful information exists across Substack, X, Facebook, YouTube, SAP, and iMessage, but companies don’t want to interoperate. Their view is that models will keep going under, over, or through those walls anyway — whether via code, browser control, screenshots, or humans acting as the AI’s hands to grab the thing “off the high shelf.”

GameStop’s eBay bid runs into the arithmetic buzzsaw

The second half of the show shifts to the Wall Street Journal report that GameStop, led by Ryan Cohen, wants to buy eBay for $125 a share, valuing the deal near $55 billion. But the memorable moment is Cohen on CNBC struggling to answer Andrew Ross Sorkin’s very direct question about where the remaining money comes from, with the hosts comparing it to showing up to buy a house with only 80% of the funds and hoping the rest sorts itself out.

Gasoline-powered laptops, AI oversight, and an empty-theater side quest

The back end becomes a rapid-fire grab bag: a genuinely absurd gasoline-powered Windows XP laptop that gets about 90 minutes per tank, a report that the Trump White House may create a working group to vet AI models before release, and Trump’s unforgettable line that AI is “a beautiful baby.” They close by shouting out Riley Walz’s empty movie theater finder, built from the delightfully weird stat that 10% of AMC showings sell zero tickets at all.