← All-In

GPT-4o launches, Glue demo, Ohalo breakthrough, Druck's Argentina bet, did Google kill Perplexity?

Episode 2024-05-172024-05-17EnglishMembers only
Listen to this episode00:00 / 1:41:14
Reading settings
Reading preferences
Text size
Download in-depth read
Read offline

English · In-depth read · Current version

Downloads are not available yet.

Text language

In-depth read · Key ideas, reasoning, and examples, organized by topic.

§01Why the besties decline to play journalist

After a long run of poker and birthday banter, the hosts turn to the awkward timing of their Sam Altman interview, which ran three days before OpenAI launched GPT-4o rather than after it. Friedberg recounts that Altman had flagged a 'big announcement,' then texted the night before to say it would slip, confirming only that it was not GPT-5, so the show went ahead with a guest who was contractually mute about his biggest news. Friedberg draws a concrete lesson: push guests who have a pending announcement, and get them on the day after a launch so they can speak freely.

The group defends its interview philosophy as the real takeaway. Chamath frames the four of them as entrepreneurs, technologists and poker players rather than journalists, so they ask what interests them and deliberately avoid gotcha questions. Sacks contrasts this with a hostile conference interview he gave to promote Glue, where the second half became 'fastball questions.' Jason adds his fixed conditions for interviews: he does not clear questions in advance and does not let subjects edit the result, while conceding that listeners who want adversarial journalism should watch Fox or CNN instead.

§02GPT-4o: the leap is interface, not just benchmark scores

The hosts argue GPT-4o's significance lies less in raw capability than in interaction. Jason notes the 'o' stands for Omni because the model ingests audio, text, images, the desktop and camera video together, and that OpenAI largely fixed the interruption problem that breaks tools like Siri when two speakers overlap or talk over each other. The demos they cite are concrete rather than abstract: a translator instructed to convert English to Italian and back that worked smoothly, and a counting demo where the model sped up and slowed down on command.

The tutoring demo is the one they dwell on. A Khan Academy clip has Sal Khan ask the model to coach his son through a right-triangle problem without giving the answer, talking to the boy directly. The transcript shows the model correcting a specific mistake, telling the student that side AC is the adjacent side to angle alpha and defining the hypotenuse as the longest side opposite the right angle, then confirming his revised answer that AB is the hypotenuse. Jason points to the same app joining Zoom calls and explaining charts as a 'guide on the side,' and notes Duolingo's stock took a hit during the translation demo.

Keep reading as a member.

You've read the free preview. Become a member to read the rest, with side-by-side reading and downloads where available.