AI has recently crossed a utility threshold, where cutting-edge models such as GPT-3, Codex, and DALL-E 2 are actually useful and can perform tasks computers cannot do any other way. The act of producing these models is an exploration of a new frontier, with the discovery of unknown capabilities, scientific progress, and incredible product applications as the rewards. And perhaps most exciting for me personally, because the field is fundamentally about creating and studying software systems, great engineers are able to contribute at the same level as great researchers to future progress. “A self-learning AI system.” by DALL-E 2. I first got into software engineering because I wanted to build large-scale systems that could have a direct impact on people’s lives. I attended a math research summer program shortly after I started programming, and my favorite result of the summer was a scheduling app I built for people to book time with the professor. Specifying every detail of how a program should work is hard, and I’d always dreamed of one day putting my effort into hypothetical AI systems that could figure out the details for me. But after taking one look at the state of the art in AI in 2008, I knew it wasn’t going to work any time soon and instead started building infrastructure and product for web startups. DALL-E 2’s rendition of “The two great pillars of the house of artificial intelligence” (which according to my co-founder Ilya Sutskever are great engineering, and great science using this engineering) It’s now almost 15 years later, and the vision of systems which can learn their own solutions to problems is becoming incrementally more real. And perhaps most exciting is the underlying mechanism by which it’s advancing — at OpenAI, and the field generally, precision execution on large-scale models is a force multiplier on AI progress, and we need more people with strong software skills who can deliver these systems. This is because we are building AI models out of unprecedented amounts of compute; these models in turn have unprecedented capabilities, we can discover new phenomena and explore the limits of what these models can and cannot do, and then we use all these learnings to build the next model. “Harnessing the most compute in the known universe” by DALL-E 2 Harnessing this compute requires deep software skills and the right kind of machine learning knowledge. We need to coordinate lots of computers, build software frameworks that allow for hyperoptimization in some cases and flexibility in others, serve these models to customers really fast (which is what I worked on in 2020), and make it possible for a small team to manage a massive system (which is what I work on now). Engineers with no ML background can contribute from the day they join, and the more ML they pick up the more impact they have. The OpenAI environment makes it relatively easy to absorb the ML skills, and indeed, many of OpenAI’s best engineers transferred from other fields. All that being said, AI is not for every software engineer. I’ve seen about a 50-50 success rate of engineers entering this field. The most important determiner is a specific flavor of technical humility. Many dearly-held intuitions from other domains will not apply to ML. The engineers who make the leap successfully are happy to be wrong (since it means they learned something), aren’t afraid not to know something, and don’t push solutions that others resist until they’ve gathered enough intuition to know for sure that it matches the domain. “A beaver who has humbly recently become a machine learning engineer” by DALL-E 2 I believe that AI research is today by far the most impactful place for engineers who want to build useful systems to be working, and I expect this statement to become only more true as progress continues. If you’d like to work on creating the next generation of AI models, email me (gdb@openai.com) with any evidence of exceptional accomplishment in software engineering.

For the first three years of OpenAI, I dreamed of becoming a machine learning expert but made little progress towards that goal. Over the past nine months, I’ve finally made the transition to being a machine learning practitioner. It was hard but not impossible, and I think most people who are good programmers and know (or are willing to learn) the math can do it too. There are many online courses to self-study the technical side, and what turned out to be my biggest blocker was a mental barrier — getting ok with being a beginner again. Studying machine learning during the 2018 holiday season. Early days # A founding principle of OpenAI is that we value research and engineering equally — our goal is to build working systems that solve previously impossible tasks, so we need both. (In fact, our team is comprised of 25% people primarily using software skills, 25% primarily using machine learning skills, and 50% doing a hybrid of the two.) So from day one of OpenAI, my software skills were always in demand, and I kept procrastinating on picking up the machine learning skills I wanted. After helping build OpenAI Gym, I was called to work on Universe. And as Universe was winding down, we decided to start working on Dota — and we needed someone to turn the game into a reinforcement learning environment before any machine learning could begin. Dota # Turning such a complex game into a research environment without source code access was awesome work, and the team’s excitement every time I overcame a new obstacle was deeply validating. I figured out how to break out of the game’s Lua sandbox, LD_PRELOAD in a Go GRPC server to programmatically control the game, incrementally dump the whole game state into a Protobuf, and build a Python library and abstractions with future compatibility for the many different multiagent configurations we might want to use. But I felt half blind. At Stripe, though I gravitated towards infrastructure solutions, I could make changes anywhere in the stack since I knew the product code intimately. In Dota, I was constrained to looking at all problems through a software lens, which sometimes meant I tried to solve hard problems that could be avoided by just doing the machine learning slightly differently. I wanted to be like my teammates Jakub Pachocki and Szymon Sidor, who had made the core breakthrough that powered our Dota bot. They had questioned the common wisdom within OpenAI that reinforcement algorithms didn’t scale. They wrote a distributed reinforcement learning framework called Rapid and scaled it exponentially every two weeks or so, and we never hit a wall with it. I wanted to be able to make critical contributions like that which combined software and machine learning skills. Szymon on the left; Jakub on the right. In July 2017, it looked like I might have my chance. The software infrastructure was stable, and I began work on a machine learning project. My goal was to use behavioral cloning to teach a neural network from human training data. But I wasn’t quite prepared for just how much I would feel like a beginner. I kept being frustrated by small workflow details which made me uncertain if I was making progress, such as not being certain which code a given experiment had used or realizing I needed to compare against a result from last week that I hadn’t properly archived. To make things worse, I kept discovering small bugs that had been corrupting my results the whole time. I didn’t feel confident in my work, but to make it worse, other people did. People would mention how how hard behavioral cloning from human data is. I always made sure to correct them by pointing out that I was a newbie, and this probably said more about my abilities than the problem. It all briefly felt worth it when my code made it into the bot, as Jie Tang used it as the starting point for creep blocking which he then fine-tuned with reinforcement learning. But soon Jie figured out how to get better results without using my code, and I had nothing to show for my efforts. I never tried machine learning on the Dota project again. Time out # After we lost two games in The International in 2018, most observers thought we’d topped out what our approach could do. But we knew from our metrics that we were right on the edge of success and mostly needed more training. This meant the demands on my time had relented, and in November 2018, I felt I had an opening to take a gamble with three months of my time. Team members in high spirits after losing our first game at The International. I learn best when I have something specific in mind to build. I decided to try building a chatbot. I started self-studying the curriculum we developed for our Fellows program, selecting only the NLP-relevant modules. For example, I wrote and trained an LSTM language model and then a Transformer-based one. I also read up on topics like information theory and read many papers, poring over each line until I fully absorbed it. It was slow going, but this time I expected it. I didn’t experience flow state. I was reminded of how I’d felt when I just started programming, and I kept thinking of how many years it had taken to achieve a feeling of mastery. I honestly wasn’t confident that I would ever become good at machine learning. But I kept pushing because… well, honestly because I didn’t want to be constrained to only understanding one part of my projects. I wanted to see the whole picture clearly. My personal life was also an important factor in keeping me going. I’d begun a relationship with someone who made me feel it was ok if I failed. I spent our first holiday season together beating my head against the machine learning wall, but she was there with me no matter how many planned activities it meant skipping. One important conceptual step was overcoming a barrier I’d been too timid to do with Dota: make substantive changes to someone else’s machine learning code. I fine-tuned GPT-1 on chat datasets I’d found, and made a small change to add my own naive sampling code. But it became so painfully slow as I tried to generate longer messages that my frustration overwhelmed my fear, and I implemented GPU caching — a change which touched the entire model. I had to try a few times, throwing out my changes as they exceeded the complexity I could hold in my head. By the time I got it working a few days later, I realized I’d learned something that I would have previously thought impossible: I now understood how the whole model was put together, down to small stylistic details like how the codebase elegantly handles TensorFlow variable scopes. Retooled # After three months of self-study, I felt ready to work on an actual project. This was also the first point where I felt I could benefit from the many experts we have at OpenAI, and I was delighted when Jakub and my co-founder Ilya Sutskever agreed to advise me. Ilya singing karaoke at our company offsite. We started to get very exciting results, and Jakub and Szymon joined the project full-time. I feel proud every time I see a commit from them in the machine learning codebase I’d started. I’m starting to feel competent, though I haven’t yet achieved mastery. I’m seeing this reflected in the number of hours I can motivate myself to spend focused on doing machine learning work — I’m now around 75% of the number of coding hours from where I’ve been historically. But for the first time, I feel that I’m on trajectory. At first, I was overwhelmed by the seemingly endless stream of new machine learning concepts. Within the first six months, I realized that I could make progress without constantly learning entirely new primitives. I still need to get more experience with many skills, such as initializing a network or setting a learning rate schedule, but now the work feels incremental rather than potentially impossible. From our Fellows and Scholars programs, I’d known that software engineers with solid fundamentals in linear algebra and probability can become machine learning engineers with just a few months of self study. But somehow I’d convinced myself that I was the exception and couldn’t learn. But I was wrong — even embedded in the middle of OpenAI, I couldn’t make the transition because I was unwilling to become a beginner again. You’re probably not an exception either. If you’d like to become a deep learning practitioner, you can. You need to give yourself the space and time to fail. If you learn from enough failures, you’ll succeed — and it’ll probably take much less time than you expect. At some point, it does become important to surround yourself by existing experts. And that is one place where I’m incredibly lucky. If you’re a great software engineer who reaches that point, keep in mind there’s a way you can be surrounded by the same people as I am — apply to OpenAI!

This post is co-written by Greg Brockman (left) and Ilya Sutskever (right). We’ve been working on OpenAI for the past three years. Our mission is to ensure that artificial general intelligence (AGI) — which we define as automated systems that outperform humans at most economically valuable work — benefits all of humanity. Today we announced a new legal structure for OpenAI, called OpenAI LP, to better pursue this mission — in particular to raise more capital as we attempt to build safe AGI and distribute its benefits. In this post, we’d like to help others understand how we think about this mission. Why now? # The founding vision of the field of AI was “… to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it”, and to eventually build a machine that thinks — that is, an AGI. But over the past 60 years, progress stalled multiple times and people started thinking of AI as a field that wouldn’t deliver. Since 2012, deep learning has generated sustained progress in many domains using a small simple set of tools, which have the following properties: Generality: deep learning tools are simple, yet they apply to many domains, such as vision, speech recognition, speech synthesis, text synthesis, image synthesis, translation, robotics, and game playing. Competence: today, the only way to get competitive results on most “AI-type problems” is through the use of deep learning techniques. Scalability: good old fashioned AI was able to produce exciting demos, but its techniques had difficulty scaling to harder problems. But in deep learning, more computational power and more data leads to better results. It has also proven easy (if costly) to rapidly increase the amount of compute productively used by deep learning experiments. The rapid progress of useful deep learning systems with these properties makes us feel that it’s reasonable to start taking AGI seriously — though it’s hard to know how far away it is. The impact of AGI # Just like a computer today, an AGI will be applicable to a wide variety of tasks — and just like computers in 1900 or the Internet in 1950, it’s hard to describe (or even predict) the kind of impact AGI will have. But to get a sense, imagine a computer system which can do the following activities with minimal human input: Make a scientific breakthrough at the level of the best scientists Productize that breakthrough and build a company, with a skill comparable to the best entrepreneurs Rapidly grow that company and manage it at large scale The upside of such a computer system is enormous — for an illustrative example, an AGI following the pattern above could produce amazing healthcare applications deployed at scale. Imagine a network of AGI-powered computerized doctors that accumulates a superhuman amount of clinical experience, allowing it to produce excellent diagnoses, deeply understand the nuanced effect of various treatments in lots of conditions, and greatly reduce the human error factor of healthcare — all for very low cost and accessible to everyone. Risks # We already live in a world with entities that surpass individual human abilities, which we call companies. If working on the right goals in the right way, companies can produce huge amounts of value and improve lives. But if not properly checked, they can also cause damage, like logging companies that cut down rain forests, cigarette companies that get children smoking, or scams like Ponzi schemes. We think of AGI as being like a hyper-effective company, with commensurate benefits and risks. We are concerned about AGI pursuing goals misspecified by its operator, malicious humans subverting a deployed AGI, or an out-of-control economy that grows without resulting in improvements to human lives. And because it’s hard to change powerful systems — just think about how hard it’s been to add security to the Internet — once they’ve been deployed, we think it’s important to address AGI’s safety and policy risks before it is created. OpenAI’s mission is to figure out how to get the benefits of AGI and mitigate the risks — and make sure those benefits accrue to all of humanity. The future is uncertain, and there are many ways in which our predictions could be incorrect. But if they turn out to be right, this mission will be critical. If you’d like to work on this mission, we’re hiring! About us # Ilya: I’ve been working on deep learning for 16 years. It was fun to witness deep learning transform from being a marginalized subfield of AI into one the most important family of scientific advances in recent history. As deep learning was getting more powerful, I realized that AGI might become a reality on a timescale relevant to my lifetime. And given AGI’s massive upside and significant risks, I want to maximize the positive parts of this impact and minimize the negative. Greg: Technology causes change, both positive and negative. AGI is the most extreme kind of technology that humans will ever create, with extreme upside and downside. I work on OpenAI because making AGI go well is the most important problem I can imagine contributing towards. Today I try to spend most of my time on technical work, and also work to spark better public discourse about AGI and related topics.

The text of my speech introducing OpenAI Five at yesterday’s Benchmark event: “We’re here to watch humans and AI play Dota, but today’s match will have implications for the world. OpenAI’s mission is to ensure that when we can build machines as smart as humans, they will benefit all of humanity. That means both pushing the limits of what’s possible and ensuring future systems are safe and aligned with human values. We work on Dota because it is a great training ground for AI: it is one of the most complicated games, involving teamwork, real time strategy, imperfect information, and an astronomical combinations of heroes and items. We can’t program a solution, so Five learns by playing 180 years of games against itself every day — sadly that means we can’t learn from the players up here unless they played for a few decades. It’s powered by 5 artificial neural networks which act like an artificial intuition. Five’s neural networks are about the size of the brain of an ant — still far from what we all have in our heads. One year ago, we beat the world’s top professionals at 1v1 Dota. People thought 5v5 would be totally out of reach. 1v1 requires mechanics and positioning; people did not expect the same system to learn strategy. But our AI system can learn problems it was not even designed to solve — we just used the same technology to learn to control a robotic hand — something no one could program. The computational power for OpenAI Five would have been impractical two years ago. But the availability of computation for AI has been increasing exponentially, doubling every 3.5 months since 2012, and one day technologies like this will become commonplace. Feel free to root for either team. Either way, humanity wins.” I’m very excited to see where the upcoming months of OpenAI Five development and testing take us.