
AI Scaling Limits and Future Paths

Combining AI and Human Work 🤖 — with Marco Trombetti
Topic 3 of 9
Published Feb 20, 2025 · Clip 16:37–23:28 (6:51)
Marco breaks down whether pre-training is plateauing. He notes diminishing returns are normal, but content is running out (about 10-15 trillion quality tokens from the internet), architecture is still simple compared to the brain's plasticity, and crucially the cost is becoming prohibitive — battles now cost tens to hundreds of billions and even Google doesn't have a trillion in the bank. This forces a new era where ingenuity matters more than brute force, which he sees as a great opportunity. He bets AGI is a trillion-dollar problem that will become feasible in years as compute costs drop, and describes two paths forward: inventing new training algorithms and verticalizing AI like Translated does to reach 'singularity' in one domain.