AI Agents Coming to OpenAI Dev Event?

0:00Today on the AI Breakdown, we're looking at reports that OpenAI might be bringing AI agents to their November developer event. Before that on the brief, a set of seemingly AI-related layoffs and a million-dollar contest to use AI to read ancient scrolls thought lost to time in history. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to Breakdown.net for more information about our YouTube channel, our Discord, and our newsletter. Welcome back to the AI Breakdown Brief, all the AI headline news you need in around five minutes. We start today with something that is not useful, not going to change the way you work or learn, but is just a mind-blowing example of things that we never would have thought possible revealing themselves to be, with ingenuity, creativity, and artificial intelligence, possible.

0:51So basically, the Vesuvius Challenge is a million-dollar competition to try to use AI to read papyrus scrolls that were burned in the eruption of Mount Vesuvius and buried in Pompeii. When Vesuvius erupted, in addition to the human toll, there was a villa with a vast library of papyrus scrolls that was also buried and lost to time. Now, these scrolls were carbonized by the heat of the volcanic debris, which was kind of a mixed bag. On the one hand, they are technically preserved, but on the other hand, almost as soon as they are exposed to air, they immediately decay and disappear. The villa was discovered in 1750, and scholars immediately realized just how significant the find was.

1:30Indeed, some people estimate that if we were able to open and read all these scrolls, it would more than double the amount of literature we had from that time. Of course, upon those first attempts, most of the scrolls that were attempted to be opened were destroyed. A few were unrolled by an Italian monk over a period of literally decades, but soon the project was abandoned with more than 600 unopened and, to all accounts, unreadable. However, zoom forward 265 years, and technology has advanced such that we now have the tools to virtually unwrap scrolls that would otherwise have been lost to antiquity.

2:02A version of this was used on a scroll found in the Dead Sea region of Israel that was found to contain text from the Bible without ever having to unroll it. And since then, people have been working on machine learning models that can detect ink that would otherwise be invisible from x-rays. This involves using a particle accelerator to scan these scrolls in order to pick up these extremely subtle surface patterns that theoretically AI models could interpret and decipher. Now, once this was shown as possible, a group of funders from Silicon Valley, including former GitHub CEO Nat Friedman, put together this challenge to get people working on actually reading these scrolls.

2:34Just last week, the competition announced that the first word in the scroll had been discovered. This was actually a multi-part process. In August, a contestant named Casey Hanmer found and wrote about what he called a crackle pattern that looked like ink. And from there, another contestant named Luke Ferreter, who is a 21-year-old college student and SpaceX summer intern, started training an AI model on that crackle pattern. This led ultimately to Luke's model being able to identify a first word meaning purple. Another contestant, Yusuf Nader, an Egyptian biorobotics grad student in Berlin used a totally different model to identify another set of letters in a similar area.

3:10I'll include a link for you guys to learn more about this, but since we're already going so long in what is supposed to be the brief, I'll move on. But I think it's just a really cool example of how AI, when combined with other technologies, in this case particle accelerators, can do things that literally seemed impossible until we actually figured out how to do them. Speaking of the ramifications of artificial intelligence, unfortunately a couple sets of layoff announcements today. Stack Overflow has laid off 100 people, which means that over the course of the last year, it's laid off about 28 % of its staff.

3:39Now, Stack Overflow has been one of the best examples of a company whose model was challenged by AI. Stack Overflow is a personal coding help forum where developers can post their questions and get responses or go search through other people's old problems to try to see if anyone else has encountered similar issues. But now the developers are relying more on AI-assisted coding, the use case has dropped pretty precipitously. LinkedIn has also announced layoffs of nearly 700 employees. That represents about 3 % of the company's staff and is part of a larger organizational restructuring effort. From what information has been shared, this will impact roles across all sorts of different departments from engineering to talent to product.

4:16Now, while LinkedIn has not confirmed that these layoffs have to do with AI specifically, that has been the working assumption of press who have reported on it, with a LinkedIn spokesperson saying to Gizmodo, this restructuring is to support LinkedIn's future broadly. Over in the world of new features, one of the big themes of this fall has been seeing AI actually integrated into the products that people are already using, and we got a couple of announcements like that over the last day or so. The first comes from CapCut, which is a video editor that is used most primarily by power users of TikTok.

4:45Now, just to get a sense of scale, this video editing platform was ByteDance's second product to hit 100 million in consumer revenue, after of course TikTok, and now they are introducing a set of new AI tools. One is an AI-powered script generator that helps advertisers come up with new ideas based on however they describe their product or their business. And their new CapCut for Business Suite also has AI-generated presenters who can be used for demos or explainer videos, and even a virtual try-on featuring AI models. Now, some of these features may remind you of what YouTube has announced over the past few months, but YouTube has also announced another set of AI tools.

5:19Specifically, they announced a new advertising package that they're calling spotlight moments. Basically, the product uses AI to automatically flag the most popular YouTube videos related to specific cultural moments, such as Halloween, award shows like the Oscars, or sporting events, and then give the advertiser the ability to serve ads across videos referencing that topic or event. Now, as someone who has a fair bit of experience in the advertising field, this is an absolute no-brainer and feels like a guaranteed hit of a product. So much of what you do as an advertiser is trying to connect your brand to these cultural moments.

5:52And so having the ability to automatically target content that relates to those moments is a huge upgrade over just the traditional keyword style advertising. Between these announcements this week from CapCut and YouTube and the absolute slew of tools announced by Adobe at Adobe Max last week, it is very clear that the theme for this fall continues to be the integration of AI into real business workflows. But my friends, we will end there for the day. Next up, the main AI breakdown. Welcome back to the AI breakdown. Today, we are discussing a set of information about OpenAI, some things that are actually known, some things that are just rumors.

6:31And I think the context for this is that obviously this fall, we have seen a lot of announcements from many of the big players in the space about their next big moves in AI. Now, of course, what makes OpenAI different is that ever since the launch of ChatGPT, they have been in the pole position in the industry. GPT-4 really reinforced that, and as of today, there just aren't any other models that can compete with GPT-4. Now, many anticipate that Google Gemini might be the first to actually do so or even exceed GPT-4, but as of yet, all the information we have about Gemini is mostly in the realm of rumor.

7:05Last month, OpenAI got people excited when they announced that on November 6th they would be holding their first ever developer conference. OpenAI CEO Sam Altman in his announcement post made sure to make it clear that there was not going to be any sort of GPT 4.5 or GPT 5 announcement, but that they would still have quote great stuff to show developers. The announcement that they posted on their website didn't give any more information either. They just said it would be a one-day event that would bring together hundreds of developers from around the world, and that they would be previewing new tools.

7:32Now, towards the end of last week, Reuters posted some sources suggesting a few things about the update. The way that they framed it was that OpenAI plans major updates to lower developers with lower costs. So again, these were quote sources briefed on the plans, and the updates that they discussed with Reuters included one, the addition of memory storage to developer tools, which Reuters said could theoretically cut costs for application makers by up to 20x. Another development would be the addition of vision capabilities to the API. Now obviously GPT-4 Vision has been one of the biggest feature updates of the year, and really in many ways a fundamental change to the ChatGPT experience.

8:07If you're interested, I just recently did a video about the seven categories of use cases for GPT Vision, so go check that out, but it would only make sense for those capabilities to be brought to the developer tools as well. Another update was around Stateful and Vision API releases, as summed up by Linus Ekinstam, the Stateful API will make applications more cost-effective by recalling conversation history, but the big one was this little line, the new features are designed to encourage companies to use OpenAI's technology to build AI-powered chatbots and autonomous agents that can perform tasks without human intervention.

8:38This is definitely the part of the announcement that has gotten the most attention, specifically the idea of OpenAI native autonomous agents. Now, of course, agents have been one of the biggest focuses for this show, a theme that we've been watching ever since April when AutoGPT and Baby AGI started to jump into people's attention. As of right now, of course, there isn't any real breakout winner in the AI agent space. Indeed, for the average observer, the promise and potential still far outweigh the actual applications in practice right now. Now, of course, that doesn't change the fact that there are an incredible number of developers exploring this exact space, and so the introduction of developer tools specifically focused around building autonomous agents in the OpenAI ecosystem would be a just massive development in that space.

9:18Now, the way that a lot of folks are framing this is in terms of an evolution for OpenAI, From a company that is just ChatGPT, to a real platform for developers. Reuters sums it up,

9:38Now of course, just how new this is, I think, is a reasonable question. Altman, for example, has clearly been focused on developers for some time. And indeed, one of the big questions for developers has been trying to understand where OpenAI is itself likely to compete with them. Basically, people don't want to build tools on OpenAI's APIs that OpenAI is going to turn around and offer a native version of. There is a lot of competition for OpenAI, especially when you get beyond just the consumer chatbot and ChatGPT itself. On the enterprise side, you have companies like Microsoft, Amazon, who are all leveraging their existing relationships and the very high burden of trust that comes with providing an AI platform with access to enterprise data to secure customers that might otherwise have flowed to a new startup like OpenAI.

10:21And so in many ways, on the one hand, OpenAI is dealing with that sort of competitive pressure from enterprise software companies, while they're also dealing with skepticism from developers around whether the OpenAI ecosystem is a good one to build in, given that it's not exactly clear what OpenAI will want to build themselves in the future, and given that there just isn't that much of a history that developers can content themselves with in terms of where those competitive dynamics might play out. Now, adding to that pressure is, of course, the fact that Meta has taken a totally open approach and is racing to catch up to the state of the art and give developers the ability to build things with AI tools that are close to as good as what they can get in the OpenAI ecosystem with potentially fewer of those concerns.

10:59The point I'm trying to make is that in many ways this is more than a developer event that's coming up next month. It is sort of a referendum and reflection of, if not a shifting strategy for OpenAI, a clear rededication at least to appealing to that developer audience. It kind of makes sense then that it's happening almost a year exactly after ChatGPT was released, as OpenAI figures out what its next big act is going to be. There are other indications that OpenAI is consolidating around its focuses heading into next year. The company recently updated its core values on its website. They used to be audacious.

11:32We make bold bets and aren't afraid to go against established norms. Thoughtful. We thoroughly consider the consequences of our work and welcome diversity of thought. Unpretentious. We're not deterred by the boring work and not motivated to prove we have the best ideas. Impact-driven. We're a company of builders who care deeply about real-world implications and applications. Collaborative. Our biggest advances grow from work done across multiple teams. And growth-oriented. We believe in the power of feedback and encourage a mindset of continuous learning and growth. Now, at the end of last week, the company very publicly updated those core values.

12:00Logan, who does developer relations over at OpenAI, posted the new values, which are led with AGI focus. We are committed to building safe, beneficial AGI that will have a massive positive impact on humanity's future. Anything that doesn't help with that is out of scope. Next value is intense and scrappy. Building something exceptional requires hard work, often on unglamorous stuff and urgency. Everything that we choose to do is important. Be unpretentious and do what works. Find the best ideas wherever they come from. Third is scale. We believe that scale in our models, our systems, ourselves, our processes, and our ambitions is magic.

12:31When in doubt, scale it up. Make something people love. This one borrowed, of course, from Y Combinator. Our technology and products should have a transformatively positive effect on people's lives. Finally, team spirit. Our biggest advances and differentiation come from effective collaboration in and across teams. Although our teams have increasingly different identities and priorities, our overall purpose and goals have to remain perfectly aligned. Nothing is someone else's problem. Now, interestingly, I've seen some people suggest that this is a commercialization of OpenAI's core goals. I don't exactly see it like that.

13:00Of course, you hear echoes of that first set of goals in these newer updated goals, but frankly, the first ones were much more wishy-washy and kind of just generic about how OpenAI imagines its teams working together. The new core values, particularly with the anchor of AGI Focus, is a much more tangible and clear articulation of what this company is trying to do. To the extent that these core values are about attracting a certain type of person, which I believe that is their whole purpose, anchoring around the actual objective of the company, that AGI focus, the fact that anything that doesn't help with that is out of scope, feels to me a lot clearer and a lot more transparent than what that first set of goals represented.

13:36Now, it's worth here as we round the corner actually checking in on what OpenAI's timelines for AGI really are. Sam Altman was recently on the Joe Rogan Experience, and when Rogan asked him what sort of timeline he imagined for AGI and how it had changed since the beginning of OpenAI, Altman said, AGI isn't the endpoint. To accomplish ASI, artificial superintelligence, that'll take us until 2030 or 2031. That has felt to me all the way through kind of a reasonable estimate with huge error bars. I think we're on the trajectory I would have assumed. Now, as the AI Safety Memes account points out, however, it seems likely that OpenAI actually thinks that superintelligence could arrive even sooner than that, given that their superalignment team that was announced earlier this year, or as this account calls it their AI-not-kill-everyone-ism team, set a four-year deadline for solving alignment, which would of course bring us to 2027.

14:25The AI Safety Memes account also points out that in the wake of GPT-4 Vision, prediction markets have increased how fast they think AGI is coming. Metaculous markets now put a 25 % chance at AGI in the next 1.5 years. For me, I am looking forward to this November 6th Developer Day as a way to better understand 1. How OpenAI views its relationship with developers and how it sees its most important competitive challenges in the year to come. And 2. For anything it suggests about changed timelines or updated senses of the speed with which things are developing. I'm sure as November 6th approaches, we will get even more leaks and hints and eventually schedules, and I will, of course, keep you posted on all of those as they come up.

15:04For now, I want to say a big thank you to you guys for hanging out and watching or listening, however you are consuming this content. If you are enjoying it, I would so appreciate it if you would take a time to leave a rating or review. Silly as it may seem, it actually makes a huge difference for people discovering the show. And until next time, peace.

15:29Thank you.