Series · Slide 11 · So what do you do?
Dwarkesh Patel — interviews on capability
Dwarkesh Patel
Dwarkesh keeps asking the question filmmakers should also be asking: what do these models actually learn, and what don't they? The Gorn-tagging conversation in particular reframes 'understanding' as a much weirder thing than the marketing implies.
Extended notes
Dwarkesh keeps asking the question filmmakers should also be asking: what do these models actually learn, and what don't they? The Gorn-tagging conversation in particular reframes 'understanding' as a much weirder thing than the marketing implies.
Use this as the technical counterweight to the cultural arguments. The third path in the talk only works if you can hold the engineering and the ethics in the same sentence.
Discussion prompts
- 01Pick one model you depend on. Write a paragraph describing what it actually learned, in your own words, with no marketing copy.
- 02Where in your workflow are you assuming the model 'gets it,' and where are you assuming it can only mimic the surface? What would change if you swapped those assumptions?
- 03If taste lives downstream of the model — in the human selector — what does your taste-training look like this quarter?
Quick questions (from the library card)
- What's the difference between 'the model gets it' and 'the model can produce it'?
- Where does taste live, if not inside the model?
- What does the technical detail change about the ethics conversation — and what does it leave untouched?