ChatGPT Launched 1 Year Ago Today: A Retrospective

0:00Today on the AI Breakdown, we're doing a one-year retrospective of ChatGPT and OpenAI on ChatGPT's first birthday. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, our Discord, and our newsletter.

0:24Welcome back to the AI Breakdown. Today we are doing something just a little bit different. In many ways, this is one of the more significant moments in the young history of the generative AI space, because on this day one year ago, OpenAI released fairly innocuously and without much fanfare a little chatbot they called ChatGPT. The blog post they used to release it was not particularly bombastic. They write, We've trained a model called ChatGPT which interacts in a conversational way. The dialogue format makes it possible for ChatGPT to answer follow-up questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests.

0:59They said we're excited to introduce ChatGPT to get users feedback and learn about its strengths and weaknesses. The limitations they give as ChatGPT sometimes writing plausible sounding but incorrect or nonsensical answers, ChatGPT being sensitive to tweaks in the input phrasing or prompting, the language being excessively verbose and overusing certain phrases, the model not asking clarifying questions but instead guessing at what users intend, and of course the ability for sometimes inappropriate questions or prompts to slip through. Now like I said, when the OpenAI team released this tool, they did not expect it to blow up the way that it did.

1:31However, almost immediately people were incredibly excited. Back in June of 2022, Sam Altman had tweeted about how OpenAI products had been increasing in the speed with which they got to 1 million signups. As he wrote, took GPT-3 24 months to get there, CoPilot around 6 months, and DALI only 2.5 months. Well, as Greg Brockman pointed out on December 5th of last year, ChatGPT crossed 1 million users in just 5 days. And that would certainly not be the end of it. Almost immediately, we started getting articles like this one from the New York Times. The brilliance and weirdness of ChatGPT, a new chatbot from OpenAI, is inspiring awe, fear, stunts, and attempts to circumvent its guardrails.

2:10Wrote Kevin Roos at the time, ChatGPT is quite simply the best artificial intelligence chatbot ever released to the general public. For most of the past decade, he writes, AI chatbots have been terrible. Impressive only if you cherry-pick the bot's best responses and throw out the rest. In recent years, a few AI tools have gotten good at doing narrow and well-defined tasks like writing marketing copy, but they still tend to flail when taken outside their comfort zones. But ChatGPT feels different. Smarter, weirder, more flexible. It can write jokes, some of which are actually funny, work in computer code and college-level essays.

2:41It can also guess at medical diagnoses, create text-based Harry Potter games, and explain scientific concepts at multiple levels of difficulty. The potential societal implications of ChatGPT, he writes, are too big to fit into one column. Maybe this is, as some commentators have posited, the beginning of the end of all white-collar knowledge work and a precursor to mass unemployment. Maybe it's just a nifty tool that will be mostly used by students, Twitter jokesters, and customer service departments until it's usurped by something bigger and better. Personally, I'm still trying to wrap my head around the fact that ChatGPT isn't even OpenAI's best AI model.

3:10That would be GPT-4, the next incarnation of the company's large language model, which is rumored to be coming out sometime next year. We are not ready. Well, apparently the world was ready, or at least it wanted to be. Because by the end of January, ChatGPT had become the fastest growing startup in history. It reached 100 million active users, something like 6 to 8 weeks after being launched. The previous fastest product to reach 100 million users was TikTok, and that took 9 months. Instagram, it took 2.5 years. Now, OpenAI was not alone in all of this. They had a big power player partner in Microsoft, and at the end of January, they announced that their deal was deepening even farther.

3:48After investing in both 2019 and 2021, Microsoft would be investing another$10 billion, giving OpenAI the capital it needed to compete in an extremely resource-hungry environment. Throughout the spring, OpenAI added capacities to ChatGPT and made it more useful for developers. At the beginning of March, they introduced APIs for both ChatGPT and Whisper, Whisper being their extremely advanced speech-to-text model which makes Siri look juvenile at best. And then, of course, on March 13th, came GPT-4. GPT-4 was obviously significantly better than GPT-3.5 that we had been using with ChatGPT, but what we didn't realize at the time was that even in spite of an incredible amount of effort, an AI technological arms race coming from all sides of the technology industry, GPT-4 would remain the undisputed state-of-the-art throughout 2023.

4:38It remains so today, leading some to recently wonder what it is about GPT-4 that's made it such an impossible barrier for other models to breach. At the end of that same month in March, ChatGPT started supporting plugins. Plugins were a way for ChatGPT to interact with other outside information and applications. Remember, at this time, ChatGPT wasn't natively connected to the internet. That wouldn't come until a little bit later. And so plugins were ways for all the various applications out there, as well as many new ones, to interact with ChatGPT in a bunch of different ways. For example, the Xpapers plugin allowed ChatGPT to go browse through Archive's database of research papers, making it extremely useful for content creators who have to summarize dense and abstract artificial intelligence research with some frequency.

5:21Now, plugins may never have fully taken off. It was always kind of a rough experience, there was never really a great browse feature, and of course custom GPTs, which would come much later, would kind of open up the question of whether they were really valuable in the first place. What's more, as Sam Altman said even at the time, it seemed like while a lot of developers thought they wanted their applications inside ChatGPT, what they probably wanted was ChatGPT inside their applications. Still around March and April, we started to see the peak of excitement and hype around ChatGPT. GPT-4 was out, plugins had been released, and one of the plugins that OpenAI themselves were working on was completely taking over Twitter slash X, and that was called Code Interpreter.

6:00On April 29th, John Backus, one of the lucky few who had access to it at that time, wrote, The Code Interpreter feature on ChatGPT is the most mind-blowing thing I've seen yet. All I did was upload a CSV of SF crime data and ask it to visualize trends, which Code Interpreter, by the way, did with aplomb. Now, when Code Interpreter rolled out a little more fully later that summer, many people said that it functionally turned GPT-4 into GPT-4.5. The reason for that is that by giving ChatGPT the ability to write code, it created a whole new pathway to solve problems that it previously had been unable to solve.

6:32That sort of step function change in what it could do felt to many like a difference in kind, not just a difference in scale. However, for our journey through ChatGPT and OpenAI's first big year, we're back in May, when OpenAI CEO Sam Altman was the star witness at the Senate's first AI hearing in the post-ChatGPT world. Now, at that hearing, Altman articulated the position that he's largely had in every interview, both before and since, which is that although these technologies were likely to do net good, there were some serious risks, including and up to an existential risk, and that because of that, the U.S.

7:04did need to put in society and state-level guardrails. This, of course, led to accusations of OpenAI and Sam Altman trying to promote regulatory capture, in which the introduction of new regulations and the difficulty of compliance cements the lead of companies who are already in a pole position and blocks out the ability for others to compete going forward. Meanwhile, right around the same time that Sam was testifying, ChatGPT was getting even more useful by moving to mobile. Part of the change there was that OpenAI's whisper speech-to-text made it actually a super useful experience when on the go.

7:33I used it, for example, as a live tour guide and art historian when I was traveling in London and Paris over the summer. Speaking of traveling, much of Sam Altman's May and June were spent crisscrossing the globe talking to everyone from open AI developers to heads of states. Indeed, Altman went on a 22-country tour, with the goal to, as they put it, spend significant time directly engaging people who are interacting with and who are affected by our technology. Now in their summary blog post, they wrote things like, Their users and developers are already building valuable applications. There are common hopes and concerns for AI's impact among communities.

8:07Policymakers everywhere are deeply engaged on AI, and that people want to know more about OpenAI's core values. One really interesting moment for there came from a blog post that was quickly deleted, that was basically the summary of what was supposed to be an off-the-record conversation with developers in London, in which among other things, Sam Altman talked about how the AI chip shortage was meaningfully and materially slowing the company down. They had wanted to release a multimodal version of ChatGPT that year, but they weren't sure they were going to be able to because of the chip shortage just as one for example.

8:36Now, like I said, that post came down pretty quickly, but of course, the leaks had already been digested by the broader public. Now, at the same time this was happening, the narrative had started to shift a little bit around not just ChatGPT, but around generative AI more broadly. One of the biggest catalysts for that was when SimpleWeb came out with statistics that suggested that ChatGPT's usage had actually gone down between May and June, the first month where that had happened. My point at the time, which I think history has validated, was that one, given that we didn't have any precedent for an application that grows to 100 million users in six weeks, we don't exactly know if it's actually normal to have some amount of a cool-off period, that two, school being over would have a material impact given how acutely there is a product market fit between ChatGPT and LLMs in general, and a wide variety of educational work.

9:23And three, more than anything else, after six months of totally fawning coverage, the media was just ready for a new narrative and a different story. Summer was comparatively quiet. One announcement that hinted that the future was custom instructions, in which users were able to specify a little bit more information about who they were and what format they liked their answers to be in by default to help ChatGPT have some amount of persistent knowledge of the user who was interacting with it. We also saw OpenAI join philanthropic Google and Microsoft to launch the Frontier Model Forum, which is a self-regulatory body responding to growing questions of regulation and policy and AI safety guardrails that were coming up in the public sphere more and more often.

10:00Over the summer, there was also the introduction of a new team at OpenAI called the Super Alignment Team. It was being led by Jan Leakey and OpenAI co-founder and chief scientist Ilya Sutskever and set out the very ambitious goal of solving alignment with super intelligent AI, or at least solving the core technical challenges of that alignment in the next four years. In addition to that lofty goal, they also announced that they would be dedicating 20 % of the compute that they had secured to date to that effort. Now, as we would learn later, the super alignment team was in some ways reflective of broader divisions starting to happen within the company, but at the time was sort of just taken as another example of OpenAI's commitment to this sort of goal of alignment.

10:41In fact, I think I remember even Eliezer Yudkowsky saying that his PDoom had gone down just a little bit after reading the announcement. Now, by the end of August, in the beginning of the fall, we were heading right back into feature mode. One of those came on August 28th when OpenAI introduced ChatGPT Enterprise. ChatGPT Enterprise was of course meant to address questions of security and privacy and data integrity and give big business users a version of ChatGPT that they didn't have to be scared would leak proprietary information or otherwise mess up their privacy and security in any particular way.

11:13Throughout the year, we've seen the enterprise AI battle increase. Amazon has been extremely active in that space. Of course, just this week, we got Amazon Q, which is a chatbot dedicated entirely to the enterprise use case. But so too has Microsoft. So too has Google. This is clearly an area where many people are focusing, and ChatGPT and OpenAI were no exception. On almost the same date of that ChatGPT enterprise announcement, the information and others also reported that OpenAI had passed a$1 billion revenue run rate. Now, on the one hand, that was ahead of revenue projections that they had previously shared with investors, but at the same time, it still did not make them profitable.

11:48The cost of training these models, the cost of compute, was still keeping them in the red, even with that significant milestone. At the beginning of September, the company really got people excited with the announcement of their first ever developer conference. OpenAI's Dev Day was slated to take place on November 6th, and pretty immediately, speculation turned to what would be coming. Now, in his announcement of the event, Altman had made sure to say that there would be no GPT 4.5 or GPT 5, so don't even get your head in that way. But people were still speculating on everything from better prices to longer context windows to potentially the beginning of AI agents.

12:22Interestingly, however, even with that big day coming, OpenAI didn't slow down, but in fact increased the speed of announcements. On September 25th, they announced what would come to be known as ChatGPT Vision. Basically, this was the first walk of ChatGPT into the realm of multimodal. With ChatGPT Vision, users could take a picture and use that as input for ChatGPT, which opened up a whole new set of use cases. For example, imagine trying to fix a bike, pointing your camera phone at the troublesome area, and asking ChatGPT for help. The voice capabilities meant that people could interact with ChatGPT primarily by talking and listening.

12:53A couple weeks later, we got the latest version of the DALI 3 image generation model, and importantly, it was integrated directly into ChatGPT. It had a bunch of differentiations over other competitors like Midjourney, a notable one of which is that it could actually handle and recreate text. But for sure, the most significant thing about it was that it again added multimodal dimensionality to the ChatGPT product itself. Instead of having to learn a whole new lexicon of prompting, users could interact with DALI 3 in natural language, and when the results came back and they weren't perfect, get specific in describing what needed to be changed and how.

13:27I know personally that shift in interface from having to learn these specific prompting norms of a tool like MidJourney to just being able to interact in natural language has seen me shift a significant amount of my behavior from MidJourney to Dali 3, and I am a diehard MidJourney stan. Still, in many ways, the most significant and exciting announcement of the fall was, of course, the introduction of custom GPTs at the OpenAI Dev Day. Now, there was a ton of other stuff that happened at that Dev Day as well, including the anticipated decreased prices and longer context windows and improved API performance and all that sort of stuff, but the big show was GPTs.

14:01GPTs represented OpenAI's very first steps into an agent future and were basically ways for people to create versions of ChatGPT that were refined and focused to a specific purpose. The core of it was effectively a setup prompt that was embodied and kept persistent in the GPT, but then users could also combine it with external actions to make it a more robust tool. When Altman and his team introduced GPTs at DevDay, they made very clear that while they weren't AI agents, they were a very first fledgling step into that arena. Now, even if that had been the only thing that happened in November, it still would have been a big month for ChatGPT.

14:36And yet, it pales in comparison to what happened next. On Friday, November 17th, out of the blue, at the very end of the day, although not before the end of the trading day, much to Microsoft's chagrin, the OpenAI board announced that they were firing Sam Altman because they had lost trust in him and that Greg Brockman was also being demoted from the board of directors. Now, I will not go into everything that happened in the subsequent five days. I've obviously done nearly endless content about that. But of course, ultimately, Sam Altman would return, the board would get a new starting composition, and the company begins now its second year in a very different place than it might have thought it was going to just a couple weeks earlier.

15:13Altman now has the new task of rebuilding trust among investors, customers, the board, of figuring out how governance of OpenAI is going to work going forward. In the meanwhile, they have numerous contenders who are nipping at their heels, even if none of them have actually outperformed GPT-4, and the policy and regulatory conversation is likely to do nothing but increase in significance in the year to come. Taking a step back, ChatGPT is singularly, whether it is the dominant chatbot in the future, whether OpenAI makes the most important AI models of the future, whether it's actually the company to achieve AGI or someone else's, it will forever be the case that ChatGPT was an inflection point moment for the world.

15:52A zero-to-one moment that divides the world crisply and cleanly into a before and an after. It's very rare for it to be so precise that something like that happens. But I think in this case it's fairly undeniable. And somehow, someway, it was just one year ago that it happened. I, for one, cannot wait to see what the next year, the second year of ChatGPT's life brings, you can bet I will be here to help you navigate through it all as it happens. Thanks as always for listening or watching, and until next time, peace.