Lightcap: AI model scaling laws have held predictably consistent for six years
Brad Lightcap · Sam Altman & Brad Lightcap: Which Companies Will Be Steamrolled by OpenAI? | E1140 · 20VC with Harry Stebbings · Apr 15, 2024 · at 48:29
Brad Lightcap, COO of OpenAI, discusses the scaling of OpenAI models with Harry Stebbings, emphasizing the reliable performance gains observed as model sizes increase.
0:00 / 0:17exact quote · 17.9s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“I think it's how consistently the scaling of models has worked. It still breaks my brain. Like I don't, maybe I've been, if we've, I've watched the same trend line for six years now, but I still find it incredible that you can make these models bigger and they get predictably better.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Brad Lightcap
Lightcap: Current AI models are honestly not that good
“I think right now expectations are extremely high. Reality is still pretty bad, honestly. These models are not that good.”
Lightcap: AI expectations will dip before rapid model improvements outpace them
“I think very quickly expectations will start to come down as people come into contact with today's models, but then very quickly also these models will get really, really good, and you'll see this inversion of expectations reality where all of a sudden then ex…”
Lightcap: Enterprises criminally underrate the ROI of granting broad AI access
“I think people though criminally underrate how important it is actually and how much like return you really get on just giving people access to the technology. And that there's this kind of, because you can'queries quite quantify exactly how it works, but like…”
Lightcap: Many enterprises mistakenly assume GPT-4 is the peak of AI
“A lot of companies think it's static, so a lot of companies think GPT-IV is the best the models will ever get. That's understandable. Every technology they've ever had to adopt has been relatively static.”
Brad Lightcap: 100x model improvement is the key test for startup defensibility
“Ask the company whether a hundred X improvement in the model is something they're excited about. It's actually, we can tell pretty well because we know the companies that come to us saying, we want the next model. When is it coming out? When is it coming out? …”