Podcast·2024
Why Google failed to make GPT-3 + why Multimodal Agents are the path to AGI — with David Luan of Adept
About this source
~42-min single-subject deep dive — densest framework source. Luan on OpenAI/GPT-2, why Google Brain couldn't reach GPT-3 critical mass, and Adept's reliability-over-generality agent thesis. Full verbatim transcript on the page. Transcript-intake target for the ASH-269 child.
Frameworks extracted
6 totalSTRongoing
Reliability Over Generality
Demos work 60% of the time; enterprises need the nines. Prioritize reliability over everything, then climb the abstraction ladder.
INNongoing
Multimodal for Knowledge Work, Not Cats and Dogs
Economic value is in knowledge work, so train multimodal models on charts, tables, invoices, PDFs and UIs — not COCO.
INNongoing
Drive by Vision, Not APIs
Itemize the workflows where every step has an API and the count is near zero — so the agent must see and act on the screen like a human.
LEADongoing
Research Leader as Portfolio Allocator
A research org wins by concentrating compute on a few big bets and disbanding the rest — not by letting a bottom-up credit market allocate it.
INNongoing
Behavioral-Cloning Shortcut to AGI
You don't have to re-run evolution's compute to reach AGI — you can behaviorally clone everything humans already know.
STRongoing
The Augmentation Company
Always solve tasks just a little too hard for the model — keeping a human in the loop is what builds the data flywheel.