Is AI the next great computing platform? ChatGPT vs. Google, containing AGI & RESTRICT Act

The episode opens with greetings and a long, light anecdote. One host says that at an upscale shop every flavor of shelled nuts was stocked except sea salt and vinegar, and he describes the empty shelf as proof that he had "moved the market".

The same host then recommends the product warmly, claiming it was sold out across the country. The others tease him for plugging his own salty nuts. This is banter with no analytical content, and the sell-out claim is his own, not verified.

The conversation turns to an op-ed in The Wall Street Journal by Joe Manchin, which one host reads as a sign that Manchin may run for president. A host also says a defection leaked that week, with a donor moving from Trump to "Santos", but this is passed along as hearsay and not checked. Asked about the op-ed, Sacks summarizes it as criticism of the Biden administration for failing to control costs.

The hosts think the headline did more damage than the text. It said Biden's inflation reduction act was betrayed because his administration subverts the law "for ideological ends". One host recalls pulling up CBO data earlier, which he says showed several hundred billion dollars of near-term spending and a hoped-for revenue boom later. On that account, he calls the deficit-reduction claim accounting shenanigans and the act a large spending package. These are his readings of the numbers, not figures the panel checked.

One host says he spent about two hours with Manchin on a Washington trip, along with meetings with Schumer and Mark Warner, and found him thoughtful, moderate and formidable. Asked where he would write a check, he says probably to both candidates. The panel then jokes about a cross-party ticket and credits Friedberg with the idea. The tone is playful throughout.

Sacks is skeptical. He recalls Manchin blocking a $3.5 trillion Build Back Better and then settling for a roughly $750 billion version. In his view many commentators said at the time that the promised deficit reduction was not credible, so Manchin should have known better. He argues that a candidate who appeals to centrists cannot win a party nomination without appealing to its base, and says Manchin is out of step with Democratic voters. Another voice disputes his example of Trump and Obama as outsiders. Sacks replies that Trump was an outsider but appealed to his party's base. Asked about Chris Christie, the panel dismisses him as pointless clutter.

One host formally opens the show, calling it "episode 100 something", and introduces the panel: Sacks, Friedberg in his garden in Paris, and Chamath. The others tease Chamath about his silver hair. The talk is about haircuts and has no argument in it.

The panel then jokes about Friedberg's virtual background, which viewers try to guess each week. One host says he used a reverse image search and a ChatGPT plug-in to identify it. After more banter, they agree to get started.

The first real topic is OpenAI's launch of ChatGPT plugins. The host introducing it highlights a blog post by Sacks, written as a back-and-forth with ChatGPT, and calls it one of the best conversations with the model he has seen.

Sacks explains his idea: AI startups could use a "gift to get" tactic to collect proprietary training data. His examples are an architect AI that needs many plans and a doctor AI that needs lab results or medical reports. He says such data is hard to get and that OpenAI does not necessarily have it yet. Users would earn points by uploading data and spend them on using the AI. This is his hypothesis, not tested here.

He says the tactic came from a company called Jigsaw almost 20 years ago, which he remembers though few others do. He then asked ChatGPT whether it knew Jigsaw and its approach, whether it would work for AI startups seeking proprietary datasets, and for more examples after he offered the architect case.

Sacks describes how the blog post came together. ChatGPT supplied about 20 more examples, helped flesh out details, and then summarized the thread into a first draft. He edited most of it heavily, used some passages verbatim, and had people at his firm review it. He pasted his edit back in, asked for suggestions, and told the model to incorporate its own edits before posting on Substack.

By his account the post would otherwise have taken about a week and was done in a day, and he expects to keep using ChatGPT as researcher, writing partner and sometimes editor. In another example, he asked it to turn text about Jigsaw's reward points into a table and says it did so instantly, saving what he calls a day of analyst work. The time savings are his own estimates.