Back to the era of ideas
A one-pager of the Ilya Sutskever podcast, read aloud: AI is somehow only ~1% of GDP; models are the student who memorized 10,000 hours and can't generalize, versus the 100-hour student who understood; humans learn to drive in ten hours on evolutionary priors while Waymo panics at a box. His periodization lands hardest: 2012-2020 was research, 2020-2025 was scaling, and now, with the internet already absorbed, ideas are the bottleneck again. The pair add their own theorem to Amodei's '12 months' for programmers: Silicon Valley bubbles run years ahead, so add three or more before it reaches Latin America.
// trace: where this idea came from
- ↳ video diary @ 24:41 (los dos estudiantes, memorizar no es generalizar)
- ↳ video diary @ 26:52 (la data es finita, las ideas son el cuello de botella)
- ↳ Entry 214-2: Differentiate or de-limit (la especialización de Anthropic en código, ya observada aquí)
The one-pager habit gets its richest subject yet: the Ilya Sutskever podcast episode ▸ 23:32. Opening paradox: for all the hype, AI is only about 1% of GDP ▸ 24:11. The diagnosis is a parable he loves: the student who drilled 10,000 hours and memorized every problem passes the test but can’t generalize, while the 100-hour student who understood the concepts can ▸ 24:41. Today’s models are the first student: a human learns to drive in ten hours ▸ 25:37, Waymo trains on millions and still loses its mind at an unexpected box ▸ 25:57, the difference being sample efficiency and evolutionary priors, millions of years of tuning we ship with for free ▸ 27:33.
The periodization is the line worth keeping: 2012 to 2020 was the era of research, AlexNet, Transformers, new architectures; 2020 to 2025 was the era of scaling, more data and more compute ▸ 26:20; and now, with the internet effectively absorbed, the data is finite and the field returns to research ▸ 26:52, ideas as the bottleneck ▸ 27:15. The future he sketches isn’t a static artifact trained once but a blank learner you teach like an apprentice ▸ 28:21.
se acabó la internet; vuelven las ideas →
Two local annotations earn their place in the one-pager. First, specialization survives even inside AI ▸ 29:15, and his example is the one this diary already tracks: Anthropic hyperspecializing in code until “ya son los vencedores en código” ▸ 30:15. Second, when Dario Amodei’s clip promises programmers upheaval in 12 months ▸ 30:35, Julia supplies the correction they both sign: those clocks run inside the Silicon Valley bubble, and like everything else it will take years more to reach the wider world ▸ 31:01, his version being 12 months in the Valley, three or more to Latin America ▸ 31:37, and even then his father’s dream prompt, “hágame un programa para ganar la lotería,” will still need someone who understands what’s underneath ▸ 32:13. The closing quote he keeps from the episode: the models seem smarter than their economic impact implies ▸ 36:42…
// continued in
no entry has continued this idea yet: the arc is still open