Comments by "Gilad Barlev" (@GSBarlev) on "" video.

  1. The YouTube submission system already has dropdowns for settings like license, caption certs and embedding settings—having a per-upload toggle to set permissible uses would be incredibly seamless, and most people would gladly opt their content into being included in model datasets, especially if there was at least a token amount of revenue-sharing.
    16
  2. Hey Hank. Data Scientist and AI policy wonk here. Some notes. - A lot of companies' legal arguments around "fair use" hang heavily on the use of the word "train." That's an anthropomorphism, and NIST has called out that we should stop anthropomorphizing this tech for that very reason (so train →fit, hallucinate → confabulate) - at 16:30 you talk about "bootstrapping," which is a common technique in machine learning. All indications right now point to bootstrapped datasets requiring heavy human review in order to be useful - 21:00 if I were a betting man, I'd bet on the bubble bursting, and soon. The Transformer framework is actually a step down from the previous generation of deep learning. What made it explode on the scene was that it was comparatively cheap to deploy. But that means there's a hard ceiling on how good GenAI can be without being prohibitively expensive, and ChatGPT is already costing OpenAI far more to run than they're taking in from revenue. I'm sure you have an arsenal of experts at your disposal, but if you have any technical or policy questions about any of this, feel free to ping me, and I'll be glad to point you to some authoritative resources.
    2
  3. ​ @DanknDerpyGamer True, "anthropomorphism" might not be the best term here. But the point is, the way these models become more accurate is vastly different from how humans and animals learn.
    2