GuShiio.com Achieving more intelligence at lower costs does not mean that you will stop pursuing intelligence.
Author: MD
Produced by: Bright Company
Recently, Jacob Effron, partner of Redpoint Venture Capital, had an interview with David Luan on Redpoint Venture ‘s Unsupervised Learning podcast. From a technical perspective, they discussed the inspiration DeepSeek brought to research and practice in the field of large models, and shared their thoughts on current bottlenecks and potential breakthrough directions for AI models.
David Luan was an early employee of OpenAI. He graduated from Yale University in 2009 and first joined iRobot to work on robots. He then worked in many companies (including Microsoft) until 2017 when he joined OpenAI, which was still in its early stages. At that time, the R & D team had only 35 people. In this interview, he also mentioned that the reason for joining an artificial intelligence company came from his interest in robots. He believed that“The biggest limitation of robots lies in the intelligence of the underlying algorithms.
In 2020, David Luan left OpenAI to join Google, but not long after, he and two colleagues he met at Google co-founded Adept and served as CEO. In August last year, he joined Amazon as head of AGI’s San Francisco laboratory.
The following is the interview text compiled by “Bright Company”(slightly abridged):
Limitations of the Big Model and the Value of Reinforcement Learning
Jacob：David Luan is the head of Amazon’s AGI laboratory. He was previously co-founder and CEO of Adept, which raised more than $400 million to develop AI Agents. He was involved in many key breakthroughs during his tenure as vice president of engineering at OpenAI. I’m Jacob Effron.
On the show today, David and I discussed many interesting topics, including his thoughts on DeepSeek, predictions for future model progress, and we discussed the current state of agents and how to make them reliable, and when they will be ubiquitous. He also shared some interesting stories about the early days of OpenAI and the unique culture there. It was a very interesting conversation because David and I have known each other for over ten years. I think the audience will love it. David, thank you for joining our podcast.
David：Thanks for inviting me. This will be very interesting because we have known each other for over ten years.
Jacob：I remember when you first joined OpenAI, and I thought it seemed interesting, but I wasn’t sure if it was a smart career choice. Then it becomes obvious that you always see opportunities earlier than others.
David：I’m really lucky because I’ve always been interested in robots,The biggest limitation of robots (at the time) was the intelligence of the underlying algorithms. So I started working on artificial intelligence,It’s really cool to see these technologies progress in our lifetimes.
Jacob：I want to discuss many topics with you today. I want to start with the recent hot topics. Clearly, the reaction to DeepSeek has been strong in the past few weeks. People talked about it and stocks plunged. Some people say this is bad for OpenAI and Anthropic. I think people’s emotions have eased down from their initial panic. But I’m curious that people have a lot to do in broader discussionsWhat is right about the impact of this incident, and what is wrong?
David：I still remember that morning, everyone was paying attention to the news of DeepSeek. When I woke up, I looked at my mobile phone and saw five missed calls. I thought, what happened? The last time this happened was when SVB (Silicon Valley Bank) collapsed because all the investors were calling me to withdraw funds from SVB and First Republic Bank. So I thought, something bad must have happened. I checked the news and found that the stock fell due to the release of DeepSeek R1.I immediately realized that people were getting this thing completely wrong.DeepSeek did a great job,But it’s part of this broader narrative where we first learn how to make new big models smarter.，Then we learn how to make them more efficient.
So this is actually a turning point. What people misunderstand is thatJust because you can achieve more intelligence at lower costs doesn’t mean you will stop pursuing intelligence.On the contrary, you will use more intelligence. So when the market realized this, now we are rational again.
Jacob：Given that at least the basic model seems to have been trained on OpenAI, there are a variety of ways you can make the basic DeepSeek model behave like ChatGPT. So, looking to the future, will OpenAI and Anthropic stop due to knowledge distillation?Publish these models more openly?
David：I think what will happen is,People always want to build the smartest models possible, but sometimes these models are not always reasoning efficiently. So I think we’re going to see more and more, although people may not discuss this explicitly, people are training these huge teacher models in in-house laboratories, using all the computing resources they have access to.Then they will try to compress it into an efficient model suitable for customers.
The biggest problem I see right now is that I imagine artificial intelligence use cases as concentric circles of complexity. The innermost complexity may be like having a simple chat conversation with the basic language model, which we have been able to do well in GPT-2. And every additional level of intelligence, such as the ability to perform mental arithmetic, programming, or later Agents, or even drug discovery, requires smarter models.But almost every previous level of intelligence has become so cheap that it can be quantified(Quantize refers to reducing the numerical accuracy of the model to reduce resource consumption).
Jacob：This reminds me of the trend of computing while testing (test-www.gushiio.com compute). This seems like a very exciting path forward, especially in easy-to-verify areas such as programming and mathematics. How far can this paradigm take us?
David：A series of papers and podcasts have documented my discussions over the years about how to build AGI (General Artificial Intelligence).
Jacob：Let’s add something new to these discussions.
David：We started thinking about GPT-4, and we lived in a world where people were not sure if they just needed to predict the next token prediction.that can solve all AGI problems.
My opinion, and the opinions of some people around me, is actually no.The reason is that if a model is trained to make the next token prediction, it will essentially be punished for discovering new knowledge because the new knowledge is not in the training set.So what we need to do is we need to look at other known machine learning paradigms that can truly discover new knowledge.We know reinforcement learning (RL) can do thisRL can do this in search, right? Yes, or like AlphaGo, this may be the first time that the public has been aware that we can use RL to discover new knowledge. The problem has always been thatWhen will we combine large language models (LLMs) with RLs?, in order to build a system that both has knowledge of all mankind and can be built on this basis.
Jacob：So for areas that are not easy to verify, such as health care or law, can this test-time computing paradigm allow us to build models that can handle these issues?Or will we become very good at programming and mathematics, but still can’t tell a joke?
David：This is a topic worth debating, and I have a very clear point.
Jacob：What is your answer?
David: These models are more generalizing than you think.Everyone is saying, I used GPT-1, which seems to be better in terms of mathematics, but while waiting for it to think, it may be a little worse than ChatGPT or other models.I think these are just small twists and turns to greater strength.Today, we have seen signs thatSolve the problem by explicitly verifying that the model correctly (as we saw in DeepSeek), does lead to migration on some slightly ambiguous issues in similar areas. I think everyone is working hard, my team and other teams are working hard to address the issue of human preferences in these more complex tasks to satisfy those preferences.
Jacob：Yes. And you always need to be able to build a model to verify, like, hey, this output is good legal opinion, or this output is a good medical diagnosis, which is obviously much more difficult than verifying whether a mathematical proof or code works.
David: I think what we are taking advantage of is the gap between the good and bad models——The ability of the same set of neural network weights to determine whether they have done a good job is compared to their ability to generate the right answer. We always see that these models are better at judging whether they have done a job well than at generating good answers. To some extent, we are taking advantage of this, through some RL tools (stuff), to give it a feel for itself as to whether it has done something well.
Jacob：In order to truly launch a model like this, what research issues need to be solved?
David：There are so many questions that I think I may only list three questions we need. First of all, I think the first issue is that you need to really know how to build an organization and process to model it reliably.
I keep telling my team and the people I work withToday, if you run a modern artificial intelligence laboratory, your job is not to build models, but to build a factory that can reliably make models.When you think like that,This completely changes the direction of your investment.Before reproducibility is achieved, I don’t think there is much progress to some extent. We have just gone from alchemy to industrialization, and the way these models are built has changed. Without this foundation, these models cannot work.
I think the next part is,You have to go slow as fast.But I think this is the first part. I always believe that people are always attracted to algorithms because they look cool and sexy. But if we look at what really drives all this,It’s actually an engineering issue.For example, how do you perform large-scale cluster computing to ensure that they can run reliably long enough?If a node crashes, you won’t waste too much time on your tasks.To push the front edge of scale, this is a real problem.
Now, throughout the field of reinforcement learning (RL), we will soon enter a world where there will be many data centers, each data center doing a lot of reasoning on the underlying model, and perhaps testing in new environments brought by customers to learn how to improve the model and feed this new knowledge back into a central place where the model learns to become smarter.
Jacob：Some people like Yann LeCun have been criticizing the limitations of large language models (LLMs) recently. I want you to summarize this criticism for our audience, and then talk about your thoughts on those who say these models can never allow for truly original thinking.
David：I think we already have counterexamples, and AlphaGo is an original thinking. If you look back at the early work of OpenAI, where we used RL to play Flash games, if you’re that age group, you probably remember MiniClip and things like that. These used to be a pastime in middle school, but it’s really interesting to see them becoming the cornerstone of artificial intelligence.。We were looking at how to use our algorithm to play these games at the same time,You’ll soon find that they learn how to passUse loopholes to pass through walls and other methods to quickly clear customsThese are things humans have never done before.
Jacob：In terms of verification, it is mainly about finding clever methods to find verification methods for these different fields.
David：You just use the model.
How to build reliable Agents
Jacob：I want to turn the topic to the world of Agents. How would you describe the current status of these models?
David：I am still extremely excited about Agents. This reminds me of 2020 and 2021, when the first wave of truly powerful models such as GPT4 came out. When you try these models, you will feel the huge potential. They can create excellent rap songs, make wonderful complaints, and basically pass three-digit addition. But when you ask it to order a pizza for me, it will only imitate the conversation model of Domino’s Pizza customer service,It’s impossible to complete the actual task。This obviously exposes major flaws in these systems, right?
Since then, I have firmly believed that the problem of Agents must be solved. When I was working at Google, we started working on what became known as tool use, which was how to expose the operator interface to a large language model (LLM) and let it decide when to take action. Although the academic community has always called it an intelligent agent, the public had not yet formed a unified understanding at that time.To this end, we tried to create a new term Large Action Model to replace the Large Language ModelThis concept has sparked some discussion.But in the end, the industry chose the term “Agent”, it is regrettable that this term has been abused to the point where it loses its true meaning.But it’s cool to explore this area as the first modern Asian company.
When we founded Adept, the best open source LLMs at the time performed poorly. Since there were no multimodal LLMs at the time (like LLMs for image input, like later GPT-4v), we had to train our models from scratch\We had to do everything from scratch, which was a bit like starting an Internet company in 2000 and having to call TSMC to make your own chips.
So along the way, what we learned was that large language models, without today’s RL technology, are essentially behavioral clones., they do what they see in training data, which means,Once they enter a situation they have never seen before, their generalization becomes poor and their behavior becomes unpredictable.So Adept has always focused on useful intelligence. So what does practicality mean? It’s not about launching a cool presentation that triggers viral spread on Twitter。Instead, put these technologies in people’s hands so they don’t have to do the tedious tasks that most knowledge workers have to do, such as dragging files on their computers. So these knowledge workers are concerned about reliability.So one of our early use cases was: Can we process invoices for people?
Jacob：Everyone likes to process invoices (laughs). This seems like a natural start for these universal models.
David: This is a great Hello World. So no one really did these things at the time, and we chose an obvious Hello World use case. We did Excel and other projects. If this system deletes one-third of your QuickBooks entries once in seven times, you will never use it again. Reliability is still an issue, and even today, systems like Operator are very impressive and seem to be superior to other cloud computer Agents. But if you look at these two systems, they are both focused on end-to-end task execution, such as you type I want you to help me find 55 places for weekend vacations, and it will try to complete the task.But end-to-end reliability is very low，A lot of manual intervention is needed.We still t reach a point where
Jacob：We must solve this problem. Maybe we can explain to our listeners thatIf you start with an existing basic multimodal model and transform it into a large-scale action model, what really needs to be done behind it?
David：I can discuss this issue from a higher perspective, but basically there are two things that need to be done. The first is engineering issues, how to demonstrate what can be done in a way that the model can understand. For example, here are the APIs you can call, and here are the UI elements you can call. Let’s teach it a little bit about how Expedia.com or SAP works. This is the content of some research projects. This is the first step, which is to give it a knowledge of its own capabilities and basic ability to act.
The second part is the interesting part,which is how to teach it to plan, reason, replan., and follow user instructions, and can even infer what the user really wants and complete these tasks for it. This is a difficult R & D challenge that is very different from regular language modeling work,Because normal language modeling work is for us to generate a piece of text”，Even today’s reasoning work, such as mathematical problems, has a final answer.
So it’s more like a one-step process, and even if it involves multi-step thinking, it just provides you with the answer. This is a complete multi-step decision-making process that involves backtracking, involving trying to predict the consequences of your actions, and realizing that delete buttons can be dangerous, and you have to do all of this in basic settings.
Then you put it in a sandbox environment and let it learn under its own conditions. The best analogy is that Andrej Karpathy (note: a member of the OpenAI founding team who founded the AI+ education institution Eureka Labs in 2024) said that modern AI training is a bit like the organization of textbooks. First, you have a complete explanation of a physical process, and then some example questions. The first part is pre-training, the example questions are supervised fine-tuning, and the last step is open-ended questions, perhaps with answers at the back of the textbook.We are just following the process.
Andrej Karpathy’s description of the large model (Source: www.gushiio.com, Bright Company)
Jacob：I think you must have thought a lot about how these smart agents can really enter the world. I want to ask you a few questions. First, you mentioned that part of the problem is letting the model know what it can access.So how will the model interact with browsers and programs over time?Will this be similar to the way humans interact? Or just through code? Is there any other way?
David：If I’m going to comment on this area, I think the biggest problem right now is that peopleWe lack creativity in how to interact with these increasingly intelligent large models and agents.You remember when the iPhone first came out, the App Store came out, and people started making apps, such as pressing buttons to make a hiccup, or tilting your phone to pour beer into your mouth.Our interface is like that now, and it feels terrible because chatting is a super-restricted, low-bandwidth way to interact, at least in some ways.For example, I don’t want to go through seven rounds of conversations to decide the ingredients for my pizza.
This lack of creativity frustrates me.I think part of the reason is that the good product designers who can help us solve these problems don’t really understand the limitations of these models yet.That is changing rapidly, but in turn, so far, those who have been able to drive technological progress have always seen it as I am delivering a black box here, rather than I am delivering an experience here.
When that changes, I look forward to seeing systems like this, which when you interact with an agent, will actually synthesize a multimodal user interface for you to list what it needs to get from you and establish a shared context between humans and AI, rather than just chatting with it like the current paradigm. It’s more like you’re doing something on the computer with it, looking at the screen,It’s more parallel than vertical.
Jacob：I think you mentioned that the Operator, while impressive now, is sometimes imperfect. So, when do you think we will have reliable smart agents?
David：I think the Operator is great, but the entire field is currently missing the last piece of the puzzle.
Jacob：I think, considering the history of autonomous driving, they probably had a demonstration of autonomous driving as early as 1995, where vehicles could travel across the country and complete 99% of the journey.
David：Yes.
Jacob：Do we need to wait another 30 years?
David：I don’t think so because I think we actually have the right tools.
Jacob：As you mentioned before, AGI (General Artificial Intelligence) is actually not far away.
David：The main milestones in the field I am looking for in Agents are, I could give this agent any task during training and come back a few days later and it was 100% completed.Yes, it’s like humans have given us a 5% improvement in reliability, but this agent has learned how to solve this problem.
Jacob：As you mentioned before, when you founded Adept, there was no truly open source model, let alone a multimodal open source model. Do you think if someone started a company like Adept today, could a startup succeed here? Or will it be basic model companies and hyperscale cloud service providers that ultimately push the ball forward?
David：I have great uncertainty about this issue.But my current view is that I personally think AGI is not really far away.
Jacob：When you mention AGI, how do you define it?
David：A model that can perform any useful task that humans do on a computer is part of the definition. Another definition I like is thatIt’s a model that can learn to do these things as quickly as humans.I don’t think these are too far away, but I don’t think they will spread quickly into society. As we know, according to Amdahl’s Law, once you actually accelerate one thing, other things become bottlenecks, and the overall acceleration you get is not as great as you think.
So, what I think will happen is,We will have this technology, but the ability of humans to use these technologies truly efficiently will last for a long time.Many of my colleagues call this Capability Overhang, a huge Overcapacity.
Jacob：Have you given any preliminary thoughts about possible accelerating factors once we have these capabilities?
David：I think it depends on the person. This is about how to jointly design interactions with models and how to use those models. This will be a matter of social acceptance. For example, imagine you have a model coming out tomorrow that says: I have invented a whole new way of doing things and everyone should use it. rdquo; Humanity needs to reconcile with it and decide whether this is really a better solution, which will not be as fast as we think.
Jacob：As you said, even if the laboratory is the first place to develop these models, there may be an opportunity for startups to truly bridge the gap between the capabilities of these models and the actual interaction that end users want.
David：I’m pretty sure that’s what’s going to happen. Because at the end of the day, I still believe that inIn a world with AGI, the relationship between people is really important.Ultimately, understanding and owning customers,And getting closer to them and understanding their needs will be more important than just controlling this tool that is owned by many other laboratories.
Jacob：How do you think humans will use computers in the next decade? All of these models meet your definition of AGI. Will I still sit in front of the computer? What is your vision for the future way humans will interact with these technologies?
David：I think we will get a new toolbox for interacting with computers. People still use the command line today, right? Just like people still use graphical user interfaces (GUIs). In the future,People still use voice interfaces. But I think people will also use more ambient computing。And, one indicator I think we should focus on is thatThe leverage per unit of energy that humans gain when interacting with computers。I think this indicator will continue to increase as these systems develop.
Jacob：Maybe it’s possible to talk a little bit about the world of this future model and whether we’ll end up having models in any particular domain.
David：Let’s look at the hypothetical legal expert model. You might want this hypothetical legal expert to know some basic facts about the world.
Jacob：Many people take a general degree before going to law school.
David：That’s right. So I think there will be some domain-specific models, but I don’t want to hide the point, just say there will be some domain-specific models. I think there will be domain specific models for technical reasons,But there will also be policy reasons.
Jacob：This is interesting. What does this mean?
David: It’s like some companies really don’t want their data mixed up.For example, imagine you are a large bank, you have sales and trading departments, you have investment banking departments, and AI employees or LLMs support these departments, just as these employees cannot share information today, the model should not be able to share information through its weights.
Jacob：What else do you think needs to be solved? In terms of models, it seems that you are confident that if we just expand current computing power, we can be very close to solving the problems we need to solve.。But are there other major technical challenges that need to be overcome to continue to expand the intelligence of the model?
David：In fact, I don’t agree with this view:Just migrate existing technology directly to a computing power cluster two years later, and everything will work miraculously.Although scale will still be a key factor, my confidence stems from our review of the current core open issues and we need to assess the difficulty of solving these issues. For example, are there super problems that must be overcome through disruptive innovation? For example, completely replace the gradient descent algorithm (Note: gradient descent, the core algorithm for parameter optimization of current deep learning models, iteratively updates parameters by calculating the negative gradient direction of the loss function.), Or you must rely on quantum computers to achieve General Artificial Intelligence (AGI). But I don’t think these are inevitable technical paths.
Jacob：When new models come out, how do you evaluate them? Do you have any fixed questions to test on, or how do you judge whether these new models are good or bad?
David：My assessment methodology is based on two core principles: Methodological Simplicity: This is the most fascinating trait in the field of deep learning——When a study comes with methodological documentation (which is increasingly rare today), you just need to examine the implementation path and you may find a solution that is simpler and more effective than traditional solutions.Such breakthroughs tend to be loaded into deep learning canons and bring about epiphany moments that truly demonstrate the beauty of algorithms.
Benchmark Misalignment:The hype in the current field has led to a large number of benchmark tests that are out of touch with the actual needs of the model, but are overvalued in the R & D process.These tests are essentially a game. The complexity of assessments and measurements is seriously underestimated, and they deserve more academic reputation and resource investment than many current research directions.
The accumulation of differentiated technologies is actually very small
Jacob：It seems that everyone has their own internal benchmarks that they don’t publish publicly, such as what they believe more. Just like you can see OpenAI’sModels perform better in many programming benchmarks, but everyone uses Anthropic’s models and they know they are better.It’s interesting to see how this field evolves. I want to hear about your current situation at Amazon. What do you think of Amazon’s role in the broader ecosystem?
David：Yes, Amazon is a very interesting place. In fact, I learned a lot there. Amazon is very serious about building general intelligent systems, especially general intelligent agents.I think what’s really cool is that I think everyone at Amazon understands that computing itself is transforming from the basics we know about it to calls to large models or large agents, which may be the most important computing basics in the future.So people care a lot about this, which is great.
I think it’s interesting that I’m in charge of Amazon’s Agent business, and it’s cool that you can see how broad the reach of agents in a big company like Amazon. Peter and I opened a new research laboratory for Amazon in San Francisco,This is in large part because many people at Amazon really believe that we need new research breakthroughs, to address the main issues we discussed earlier on leading to AGI.
Jacob：Are you interested in any of these alternative architectures, or more cutting-edge research areas?
David：Let me think. I always focus on things that may help us better map model learning to computation.Can we use more calculations more efficiently?This provides a huge multiplier effect on what we can do. But I actually spend more time focusing on data centers and chips because I find it very interesting. There are some interesting movements going on right now.
Jacob：It seems that one of the main factors driving the development of the modelThe data is annotated, and obviously, all laboratories spend a lot of money on this. Is this still relevant in the computation-at-test paradigm?How do you view this issue?
David：The first thing I can think of are two tasks that need to be solved for data annotation.The first is to teach models the basics of how to complete a task by cloning human behavior.If you have quality data,Then you can use it to better inspire what the model has seen during pre-training。Then I think the second task is to teach the model what is good and what is bad for those vague tasks. I think both are still very important.& hellip;…
Jacob：You have clearly been at the forefront of this field for the past decade. Is there one thing you have changed your mind about in the past year?
David：What I have been thinking about is the construction of team culture.I think we knew all along, but what I became even more convinced is that recruiting really smart, dynamic, and internally motivated people, especially early in their careers, is actually a major engine of our success. In this area, every few years, the best strategies change. So if people get too used to the best strategy before, they can actually slow you down.So I think that compared to what I thought before,It would be better to bet on the newcomers.
Another thing I changed my mind was thatI used to think that building AI would actually have real long-term technical differences, and you could build on that.I used to think that if you did a good job in text modeling, it should help you become a natural winner in the multimodal space. If you do a good job at multimodal, you should be a winner in the fields of reasoning and agency and these advantages should continue to accumulate.But in practice, I see very little accumulation。I think everyone is trying similar ideas.
Jacob：The implication is that just because you break through A first does not mean you will have an advantage in B. For example, OpenAI has made breakthroughs in language models, but that doesn’t necessarily mean they will make breakthroughs in reasoning.
David：They are related,But that doesn’t mean you will definitely win the next opportunity.
When will robots enter the home?
Jacob：What I want to ask is, you first entered artificial intelligence through the field of robotics. So,What do you think of the current situation in the field of artificial intelligence robots today?
David：Similar to my opinion of Digital Agents, I think we already have a lot of raw materials. And, I think it’s interesting that Digital Agents provide us with an opportunity to solve some tough issues before physical Agents.
Jacob：Let’s talk about how the reliability of digital agents continues to physical agents?
David：To give a simple example, suppose you have a warehouse that needs to be rearranged, you have a physical Agent, and you need toAsk it to calculate the best plan for rearranging the warehouse.This can be difficult if you study in the physical world, or even in a robotic simulation environment.But if you’ve done this in digital space, and you already have all the knowledge of the training recipes and adjustment algorithms to learn from simulated data, it’s like you’ve done this task on a training round.
Jacob：This is interesting. I think when people think of robots, there are two extremes. Some people think that the rules of scale we find in language models will also be found in the field of robotics.We are on the verge of huge change.You often hear Jensen talk about this issue. Then there are others who think it’s like the self-driving car in 1995, a great demonstration, but it’s still a long time to actually work. Which end of the spectrum are you at?
David：I go back to what I mentioned earlier,What gives me the most confidence is our ability to build training recipesThis allows us to complete the task 100%.We can do this in digital space.Although it is challenging, it will eventually migrate to physical space.
Jacob：When will we have robots at home?
David：I think this actually goes back to the issue I mentioned before. I think the bottleneck of many problems lies not in modeling, but in the diffusion of modeling.
Jacob：What about video models? Obviously, there are many people entering this field now, which seems to be a new cutting-edge area that involves understanding world models and physics for more open exploration. Maybe you can talk about what you see in this area and what you think about it.
David：I’m very excited about this. I think it solves one of the main problems we mentioned before, which is that we discussed before that today we can make reinforcement learning work on problems with a Verifier, such as theorem proving.
Then we talked about how to extend it to the Digital Agents realm, where you don’t have a validator, but you probably have a reliable emulator, because I can launch an application’s staging environment and teach the agent how to use it. But I think the main question that remains is, what happens when there is no explicit validator or explicit simulator? I think World modeling is our way to answer this question.
OpenAI’s organizational growth path
Jacob：That’s great. I want to change the subject a little and talk about OpenAI and your time there. Obviously, you were involved in a very special period for the company and played a similar role in many of its progress. I think we will see a lot of analysis of OpenAI culture in the future and what was so special about the era when GPT-1 to GPT-4 was developed. What do you think those analyses will say?What makes this organization so successful?
David：When I joined OpenAI, the research community was still very small. That was 2017, just over a year after OpenAI was established. I knew the founding team and some early employees who were looking for someone who could blur the boundaries between research and engineering, and I fit that need.
So joining OpenAI is a very lucky thing. There were only 35 people in the team at that time, but they were all extremely outstanding talents. They had done a lot of work in supercomputing, and there were many other people that I could list one by one. They were all very outstanding people on the team at the time.
Interestingly, at the beginning, my job was to help OpenAI build an expanded infrastructure, from a small team to a larger scale. butSoon, my work began to shift into how to define a differentiated research strategy that would allow us to make the right judgments for machine learning during this period. I think we realized earlier than anyone else that the era of the previous research model where you and your three best friends wrote a world-changing paper was over.。What we really need to think about is this new era,We try to use larger teams, combining researchers and engineers, to solve major scientific goals, whether or not the solution is defined as novel by academia.”。We are willing to take responsibility for this. When GPT-2 was first released, people said it looked like a Transformer, and yes, it is a Transformer. And we are proud of it.
Jacob：So, what were your considerations when you joined OpenAI?
David：I was very excited because I wanted to be at the forefront of research.The choices at the time were OpenAI, DeepMind or Google Brain.……As I mentioned before,Betting on people who are truly internally motivated, especially those early in their careers, is a very successful strategyThere were many other people who defined a field at that time who didn’t actually have a PhD degree or 10 years of work experience.
Jacob：Have you found any qualities these outstanding researchers have in common? What makes them so good? What did you learn from it about how to put them together into a team to achieve goals?
David: A lot of it is intrinsic motivation and intellectual flexibility.There was one person who was very excited and devoted to the research he was doing in our team and I won’t mention his name for a moment. About a month and a half later, I had a on-one conversation with him, and he suddenly mentioned that he had moved to the Bay Area to join us, but before he had time to install Wi-Fi or power his apartment, he spent all his time in the office, doing experiments all the time,It doesn’t matter to him at all。
Jacob：This enthusiasm is really impressive. I’ve heard you mention before that Google is not making progress with a GPT breakthrough, even though Transformer was invented at Google. It was clear at the time how much potential the technology had, but it was difficult for Google as a whole to gather around it. What do you think about this?
David：Thanks to Ilya, our scientific leader in basic research and later responsible for the birth of GPT, CLIP and DALL E. I remember he used to go to the office,Like a missionary, tell people: Man, I think this paper is important.” rdquo; He encourages people to experiment with Transformer.
Jacob：Do you think these basic model companies are doing a lot of things right now? Will there be another formula that will emerge at some point in the future?
David：I thinkLosing focus is very dangerous.
Jacob：You may be one of Nvidia and Jensen’s biggest fans. In addition to the achievements that everyone knows about, what else do you think is Nvidia that hasn’t been widely discussed but is actually very important to the company?
David：I like Jensen very much, he is a true legend. I think he made a lot of right decisions over a long period of time, and the past few years have really been a huge turning point for Nvidia, and they willInternalization of interconnects，and choose to build a business around systemsThis is a very wise move.
Jacob：We usually have a quick question and answer session at the end of the interview. Think this year’s model will progress more, less or the same?
David：on the surfaceProgress may be similar, but it is actually more.
Jacob：What do you think are currently overhyped or underestimated in the AI field?
David：What is over-hyped is“Skills are dead, we are completely finished, stop buying chips”。What is underestimated is how we can really solve very large scale simulation problems so that these models can learn from them.
Jacob：David, this was a very wonderful conversation. I’m sure everyone will want to learn more about your work at Amazon and some of the exciting things you’re doing, where can you find more information?
David：For Amazon,You can pay attention to the Amazon SF AI Lab in San Francisco. I don’t actually use Twitter much, but I plan to start using it again.So you can follow my Twitter account @jluan.
Welcome to join the official social community of Shenchao TechFlow
Telegram subscription group: www.gushiio.com/TechFlowDaily
Official Twitter account: www.gushiio.com/TechFlowPost
Twitter英文账号：https://www.gushiio.com/DeFlow_Intern
