Person
1 episodes · 19 quoted claims
David Luan says he led the research and engineering team at OpenAI from 2017 to mid-2020, during which the team developed GPT-2, GPT-3, CLIP and DALL-E.
“I led the research and engineering team at OpenAI from 2017 to mid 2020, where the teams did GPT two and three and CLIP and Dolly.”
David Luan says he later led the large language model effort at Google, which trained a model called PaLM.
“I led the LLM effort at Google, uh, where we trained a model called Palm, which was a, uh, quite strong model for its time.”
David Luan says Adept was the first AI agent startup and effectively invented the computer-use agent.
“It was the first AI agent startup. We ended up inventing the computer use agent effectively.”
David Luan says Amazon brought in him and his Adept team to run agents work at Amazon about a year before the interview.
“Amazon brought us in to go, uh, run agents for Amazon about a year ago at this point.”
The Platonic Representation Hypothesis holds that there is one underlying reality and that AI models trained on enough data converge toward representing it.
“So the Platonic representation hypothesis is this idea, similar to Plato's cave, which is really what it's named after, that there is one reality.”
David Luan says OpenAI's working definition of AGI was a system that outperforms humans at economically valuable tasks.
“The OpenAI definition for AGI we had was a system that outperforms humans at economically valuable tasks.”
David Luan says Google had an internal chatbot project, called Lambda or Mina, in 2020 that was available only to Google employees and resembled ChatGPT.
“There was a project called Lambda or Mina at Google in the, uh, in 2020 that basically was ChatGPT before ChatGPT, but only available to Google employees.”
David Luan says Alexa Plus uses a remote browser powered by his team's technology to use Thumbtack and book a plumber.
“Then what happens is Alexa Plus spins up a, uh, remote browser powered by our technology, basically, that then goes and uses Thumbtack like a human to go get you a plumber to your house”
Amazon released Nova Act, its agent toolkit, as a research preview in March.
“We released Nova Act, which was the research preview that came out in in in March.”
David Luan says enterprises using Nova Act report over 95% reliability, compared with an average of about 60% for other agent products.
“they're seeing ninety five plus percent reliability, which is, as I'm sure you've seen from the coverage of, you know, uh, other agent products out there, is like a material step up from the average 60% level of reliability”
David Luan says Navan, formerly TripActions, uses Amazon's agent technology to automate backend travel bookings for its customers.
“we have another customer called Navan, which is formerly TripActions, which uses us basically to automate a lot of backend bookings for travel for their customers.”
David Luan says some Nova Act customers have automated quality-assurance workflows of 93 steps using a single Act script.
“Uh, we've got companies that basically have like ninety three step QA workflows that they've automated with a single act script, et cetera.”
David Luan says self-play, a technique where a model plays against itself, was popularized by DeepMind in the mid-2010s when its system beat humans at Go.
“a technique that really DeepMind made popular in the mid twenty tens when they beat humans at playing Go.”
David Luan says Amazon calls its training environments 'gyms' after OpenAI Gym, an early OpenAI project of video game and robotics tasks.
“we call them gyms actually, because at OpenAI, we had a very useful early project called OpenAI Gym.”
David Luan says Peter Abbeel, a Berkeley professor who came from Covariant (which he calls Clearpath), works at Amazon, and that he or his students created most of the reinforcement learning algorithms that work well today.
“Here we also have Peter Abiel, who came from Clearpath and is a Berkeley professor that basically created or his students ended up creating the majority of the RL algorithms that work well today.”
David Luan says DeepSeek surprised the world because it jumped directly to the reasoning-model training recipe without every intermediate step.
“That's why DeepSeek was able to surprise the world, because they straight quantum tunneled to that instead of doing every stop along the way.”
David Luan says Google's Gemini caught up partly by betting on native multimodal training.
“I think that was actually part of how Gemini was able to catch up was that they bet on certain interesting ideas on native multimodal that turned out well for them”
David Luan says Adept considered terms such as 'large action models' and named its first product Action Transformer, before 'agents' became the standard term.
“We played with things like large action models and action transformer. So our first product was called Action Transformer.”
David Luan says Adept raised $450 million, which was a very large sum at the time.
“And Adept, you know, we raised 450 million, which at the time was a very large number.”