Everyday AI Podcast – An AI and ChatGPT Podcast - Ep 833: RSI Explained: When AI Starts Improving Itself and What It Means

Starting point is 00:00:00 This is the Everyday AI show, the Everyday Podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life. As agenic AI became more powerful and token hungry in the first months of 2026, the AI decision makers a month or so ago started to collectively tighten their token maxing belts to prepare for a more expensive AI future, right? It's the shift from token maxing to token efficiency we've been talking about here for months. But then something drastic and kind of unexpected happened that I don't think enough people are talking about.

Starting point is 00:00:48 Open AI on Thursday revealed that its GPT56 sole model improved itself and actually improved itself so much after its initial release that they were cutting prices on some of the variants of the model by 80%. So we've reached kind of a new juncture. Today's frontier models are not just helping to create or distill their smaller variants, but they're actually improving themselves after release. Hence, the next AI buzzword you'll be hearing a lot over the next year, recursive self-improvement or RSI.

Starting point is 00:01:24 So in short, recursive self-improvement is a process where an artificial intelligence system uses its own capabilities to design code and build a more advanced version or a better version of itself. And you'll be hearing this a lot more recently because RSI has become just the dominant tech headline over the past few weeks with a handful of events that we're going to break down on today's show. But unlike so many buzzwords that we've covered here on everyday AI over the years, RSI is one that actually today's AI implementations are. going to be impacted by, and so too will tomorrow's AI security. In practical sense, like in OpenAI's case, RSI can change what AI capabilities,

Starting point is 00:02:08 companies everywhere can actually afford, but it also sheds light on potential downsides of self-improving models doubling down the wrong path. So what the heck is RSI and why is every recent AI story touching on it? And what do we all need to know? Well, let's jump straight into the big picture. The big picture here is AI has started building better and more efficient AI. So Frontier AI models now improve themselves. They're rewriting their own code and training their smaller siblings.

Starting point is 00:02:39 So this is different than when, you know, Anthropic and Open AI are talking about using their own models to build products like Claude Code or like Codex. This is when the models are actually making themselves better. And that process is called RSI. and it's no longer some future science fiction, it is a near-term reality. So Open AI, as an example, says that its GPT-5-6 sole model made its Luna model, the smaller variant,

Starting point is 00:03:09 just cheaper and more efficient, and then they cut prices by 80%. But the full human-free, recursive self-improvement hasn't yet arrived, but over the past two months, there's a lot of AI news stories that we're going to be connecting the data. on the dots on today that show that, well, this probably isn't like a 2029, 2028 thing that a lot of

Starting point is 00:03:33 people have been saying, it's probably a 2027 thing. And that changes things maybe in both a good and a scary way. So on today's show, here's where you're going to learn. You're going to learn why Open AI was able to cut prices up to 80% when everyone predicted that AI costs would continue to climb in 2026. You're going to know more about this eight week timeline that turned RSI from research jargon into mainstream headlines. You're going to understand by the end of today's show what RSI actually means in plain English and what Sam Altman, the CEO of Open AI's singularity claim that he recently made actually means.

Starting point is 00:04:10 And you're going to know why the people building RSI are actually asking for maybe the industry to pause a little bit and pace itself and how your business should respond to all of this. All right. Let's get into it. Welcome to Everyday AI. I'm doing Jordan Moulton and well, this thing's for you. This is your daily, unedited, unscripted, live stream podcast and free daily newsletter,

Starting point is 00:04:30 helping business leaders like you and me, keep up. I do all the hard research. So you sit back, listen, read our newsletter, enjoy the benefits, and you grow your company and career. So it starts here, but make sure to subscribe to our free daily newsletter at your everyday AI.com. We're going to be recapping all the highlights from today's show, as well as all of the other AI news that you need to know.

Starting point is 00:04:51 So let's get into the latest. this three to four letter acronym in our ever evolving bowl of alphabet AIS soup. We are talking RSI. So let's talk about the last two months because, I mean, yes, I've been talking about recursive self-improvement on this show, probably for the last two or so years since I've been doing this for the last three and a half years, as it seemed like recursive self-improvement was actually going to be something that happened this decade. right because if you when I started this thing in 2023 most people thought that recursive self-improvement was at best a 2030 thing yet here we are where we're already getting glimpses and sniffs of real RSI and probably we will get full RSI by next year which is actually crazy to think about so let's kind of look at the last two months and some of the different dots that we're going to connect here so in June

Starting point is 00:05:53 June, Anthropic published a post on its website called When AI Builds itself. Also in June, Google researchers mapped four different routes from AGI to superintelligence, and one of those routes, well, relied heavily on recursive self-improvement. Then in July, OpenAI launched an internal RSI benchmark, measuring different models' ability to, well, improve themselves. Then, like I said, it's sole model post-trained. its smaller sibling Luna at the end of July. And then also in a podcast interview, OpenAI CEO, Sam Altman declared that we are in the singularity.

Starting point is 00:06:33 Yeah. And the new cycle was just warming up because at the end of July, so last week, startup recursive superintelligence signed a 400 million dollar AWS compute deal to automate AI research. So not only is it the labs that all of them. us no, right? Kind of like the big four plus meta and X all working toward, you know, some version of recursive self-improvement, but it's also, you know, new startups that are coming in heavily funded just to do that and nothing else. And on that same day, so on July 28th, over a thousand, and I think that number is like 1,300 now, Frontier Lab employees published a paper called Pacing the Frontier, which essentially is kind of asking the governments, the U.S.

Starting point is 00:07:21 government to maybe slow down or at least prepare for the possibility that there needs to be some international kind of agreement when it comes to the pace of AI because a lot of people signing that letter in the comments said that well AI is developing maybe too quickly and that society can't keep up and let me just say this right now and I need to hit a pause because I think that you know I am not in the Silicon Valley bubble right I don't work at a big tech company But I talk about AI every single day. And I do talk to those people building AI very frequently. And if you're listening to this show, you're probably more like me, right?

Starting point is 00:08:02 I know we have a lot of people at those big tech companies. But I'll say this, the rest of the world right now when they're viewing artificial intelligence, right? Most people are saying, oh, yeah, I use co-pilot to make my emails better. Or they're saying, you know, oh, I use, you know, chat GPT now instead of Google, right? Or they're saying, oh, I use, you know, Google Gemini to, you know, make, you know, cool graphics, right? That's it. So I would say that, you know, aside from our audience, right, I want you to think of yourself right now as you're kind of in that AI bubble, right? Especially if you're a frequent listener or if you've been listening for a long time.

Starting point is 00:08:43 You know, we're kind of in that bubble, even though we're not in Silicon Valley, right? We understand how capable these models are. But the rest of the world really has no clue, right? All they're seeing is they're seeing, you know, oh, you know, all these AI models, you know, I'm just going in here, I'm finding the one I want and, you know, maybe my work's a little bit easier. Or, you know, I think there's a lot of people now looking at AI and there is this, this recent anti-AI swing, right, when it comes to data centers and, you know, the college graduates that were booing, you know, commencement speakers who are talking about AI.

Starting point is 00:09:14 So there's a lot of just bad information out there when it comes to AI that has created this anti-AI backlash. So I think if nothing else, there is such a huge disconnect between the general, you know, non-everyday AI listener type, right? Or, you know, think of your in your organization, you know, there's probably a team of AI champions and then there's everyone else. Right. So for, I would say, 90% of the U.S. working population, they've, really don't have any clue what today's AI systems are capable of, right? If you sat all those people down and show them what something like, you know, GPD 56 soul

Starting point is 00:09:55 or something like, you know, Fable 5 and, you know, in codex and quad code, if you show them what it could do, are you still running in circles trying to figure out how to actually grow your business with AI? Maybe your company has been tinkering with large language models for a year or more, but can't really get traction to find ROI on Gen. AI. Hey, this is Jordan Wilson, host of this very podcast. Companies like Adobe, Microsoft, and InVIDIA have partnered with us because they trust our expertise in educating the masses around generative AI to get ahead. And some of the most innovative companies in the country hire us to help with their AI strategy and to train

Starting point is 00:10:38 hundreds of their employees on how to use Gen AI. So whether you're looking for chat GPT training for thousands or just need help building. your front-end AI strategy, you can partner with us too, just like some of the biggest companies in the world do. Go to your everyday AI.com slash partner to get in contact with our team. Or you can just click on the partner section of our website. We'll help you stop running in those AI circles and help get your team ahead and build a straight path to ROI on Gen AI.

Starting point is 00:11:11 They would not believe you. They would say, oh, this is from the future, right? And I think that's why we got this letter from the frontier labs called Pacing the Frontier, which we'll talk about here in a couple of minutes that are saying like, hey, we might need to pace the rate of AI acceleration. And then days later, which is kind of a sounds like a small side plot, but it's not a formal open AI employee who became a thinking, Thinking Labs co-founder rejoined Open AI to specifically work on RSI.

Starting point is 00:11:41 And then Google also, right, all these things happening at once. You know, this was the end of July as well. A Google exec called CAPX in RSI bet. So what does that mean? Collectively, the industry is spending hundreds of billions of dollars on data centers. And essentially, a Google exec said, well, this is because of recursive self-improvement, right? You need all this compute to be able to do this, right? That's the difference.

Starting point is 00:12:08 You know, when you have a team of researchers maybe, you know, three to four years ago where this kind of the AI researcher wasn't yet automated or on its way to being automated, Right. You just had to use a lot more people. Right. Yes, you were still able to, you know, the concept of recursive self-improvement is not new. It's been around since, you know, I would say like 2016 technically with AlphaGo, you know, and also some of the earlier GPT models, right? But that was much more kind of human handholding, you know, AI models to improve, right? where now it's more of human oversight and the models are kind of doing a lot of the work on themselves. So like I said, it's not full human free.

Starting point is 00:12:51 But we are at the point now, but, you know, a high ranking Google exec essentially just said last week that, yeah, all this money that's being spent on these data centers, this hundreds of billions of dollars, that's so that is for recursive self improvement. So these models are just going to be able to build not only the next version of themselves, But once that next version has been released, like Open AI just did, well, it's going to improve itself to hopefully make it cheaper, faster, better. But then also, I think, right, I've always been saying I'm a huge believer in this concept that eventually there's going to be thousands of small models, right? I think that that's going to be the truth. There's going to be thousands of models that help with marketing.

Starting point is 00:13:32 There's going to be thousands of models that help with medicine. There's, right, there's going to be thousands but small ones. And this is where I think the future is heading is recursive. self-improvement where these model makers are going to be able to spin out, you know, probably the ones first that make the most sense to make money, right? And then from after that, I'm guessing we're going to equally see, or hopefully see, it's my hope that we'll, you know, equally see, you know, all of these small models that come from RSI that are going to be able to do good in the world, right? And we're already seeing that with things like medicine in biology

Starting point is 00:14:03 and, you know, right now, a defensive cybersecurity. So now, let's talk about that Thursday shocker that kind of changed, I think, the near-term AI strategy for a lot of business owners out there. So for months, right, we've been hearing this, this concept that these powerful AI agents, well, they were too token hungry, you know, so I think in late 2025, kind of with the, the advent of Claude Code Code Codework in early 2026, and then Codex in February, right, it kind of got to this point where everyone was spending as many tokens as possible. because companies everywhere were like, oh my gosh, right, those on the bleeding edge,

Starting point is 00:14:42 those AI champions were like, my gosh, these systems can do absolutely anything, which was great, right? But the revenue didn't always directly follow. And so what that meant is while, you know, some companies were waiting to see the, the fruits of their token maxing labor, they had to cut back, right? And focus a little bit more on token efficiency or as some people say, value maxing. So we kind of thought that at this point, well, okay, we're just going to probably start spending less, right? But Open AI came and threw that out the window because Open

Starting point is 00:15:14 AI said that Seoul, their GPD 56 soul, adapted an existing post-training setup for the smaller Luna model. So essentially, right, and they actually shared this and we shared this in our newsletter as well. They kind of shared and they did this, I believe, all in codex. So it's, it's, you know, wasn't some far off, you know, science fiction. It's some, uh, according to someone that tweeted about this, it is some OpenAI researchers who used Codex, and they used the GPD 56 sole model to improve the smaller versions.

Starting point is 00:15:51 So the medium size is GPD 56 terra, and OpenAI was able to cut the prices on that by 20%. And then GPD 56 Luna, which they were able to cut the prices on that by 80%. Right? And the fact that that happened, and I still don't think enough people are talking about it, right?

Starting point is 00:16:08 because if you compare as an example, and I did talk about this on the show yesterday, maybe, or maybe it was Friday, right? But now you have GPD 56 Luna, which according to artificial analysis is roughly about the same as Claude and Sonnet 5, but it is 25x cheaper per task completion, right? So now it's almost like, wait, you know, there's this narrative in quarter two that we were going to have to start using less AI. And now all of a sudden, because we are getting this, this hint of recursive self improvement from open AI first, now we have to start rethinking that strategy immediately.

Starting point is 00:16:49 So let's define it a little bit. All right. So RSI is essentially when an AI helps improve another AI. It could be from a different company, right? It could be its own smaller version like we saw with the GPD 56 sole, working. on GVD-5-6 Luna, and we've seen from other companies talking about kind of their big model helping train some of their smaller models. And we've seen also, this is similar to distillation, right?

Starting point is 00:17:17 So intercompany distillation, you know, generally when they're talking about that, yes, there's humans involved. But I would say in 2026, I think the labs will start talking about this more as it becomes more commonplace. But, you know, now it's with these smaller models, it is very. come in place for companies to distill the smaller versions from the big versions. And I think that's also why we're sometimes seeing the versions come out at a different time. Right.

Starting point is 00:17:45 So as an example, Fable 5 came first, right? And then we, you know, Fable and Mythos 5 came first. And then we got, you know, Sonnet 5 and then we got Opus 5. So, you know, Anthropic has kind of hinted in the past that a lot of their code up to, you know, 90%. I think, according to some estimates, you know, 90% of their code is actually written by their own models. So, but the Google exec though, I think this is important. Compared it to history, talking about steam engines were used to build the next steam engine. And I think to understand why that matters, we have to take a look at all the other definitions, right? And how recursive self-improvement

Starting point is 00:18:27 actually fits on the current timeline. Because for the most part, at least if you've been following in keeping track at home, the kind of the path to, you know, where AI is eventually going to go, whether you're excited about that or not. And, you know, I constantly find myself going on each side of that equation, right? But you think of, well, we're in the artificial intelligence era. And it seems like most companies are focused on super intelligence. Okay. So super intelligence is when, well, all AI is smarter than all here.

Starting point is 00:19:02 humans. But AGI is kind of that stepping point. So, you know, if you had to say three things, it's AI to AGI, which is artificial general intelligence. And then you have super intelligence. And that's, I think super intelligence, if you, if super intelligence goes wrong, that's where you start getting into the Terminator and SkyNet type comparisons, right? When superintelligence goes right, that's when, well, in theory, we could cure most diseases. right. But there's those kind of steps belong along the way. And I think that according to a lot of people's timelines, actually, RSI is kind of what pushes from AGI to ASI. And I know we're throwing a lot of acronyms out there. So I'm in the firm belief that we've already hit artificial

Starting point is 00:19:56 general intelligence, right? I know it's kind of a touchy topic. It's moving goalposts. Luckily, you know, I've been saying this since last year. Luckily, NVIDIA CEO, Jensen Wong said something similar, so I don't feel as crazy when, you know, super smart people start saying, yeah, we're probably past HGI. Right. So artificial general intelligence is, well, when most AI systems can produce economically viable work at the same rate or higher than most humans, right? And I think as we've seen agenic AI be able to use computers, being able to browse the internet, being able to see and hear and listen and do all these things technically, you know, through a computer that an average human sitting in front of a computer could do, you know, I think maybe the conversation has now

Starting point is 00:20:38 shifted. Well, it doesn't really matter, right? If, if you think AGI has happened or not, like I said, it's a gray area, it's moving cold posts, et cetera, because we are now, I think, focused on recursive self-improvement. And I think, right, if you look at Google's kind of four-tiered outlook at going from, you know, or toward superintelligence. One of the main paths there is recursive self-improvement going from AGI to ASI. So those are kind of the destinations. And I like to think of as RSI as it's not a roadmap, right, or it's not a destination or a stopping point on that AI, AGI, ASI roadmap.

Starting point is 00:21:19 It's more of the engine that can really expedite the process. And I think that one of the reasons why, if you go back and look, and I'm a dork, I did a show on this like two years ago. I went back and looked in archive.org, looked at every single main definition of artificial general intelligence and artificial superintelligence, you know, from 2005, 2010, 2015, 2020. Right. And by those definitions, especially up through 2015, were well past AGI. But RSI was never really a huge player in many of those earlier definitions. And I think that that's why the timeline has kind of been expedited, right? A lot of people originally were like, oh, you know, AGI is a 20, you know, 2050. It's a 20, you know, 60, right?

Starting point is 00:22:05 A lot of the earlier projections were saying, hey, we wouldn't have a single super smart AI system that could do work better than humans, you know, for 20, 30, 40 years. And that's clearly not the case because we are already there by, you know, third-party metrics that measure those types of things, such as different GDP-Bow benchmarks. So the engine, though, that RSI engine, that's also why OpenAI's CEO, Sam Altman, just made a claim that I thought would actually get a lot more headlines, but it really didn't. So on a podcast last month, well, it's last month, but technically like a week and a half ago, he said that, well, we said that, well, we. are now like in the singularity. All right. And usually when a CEO of one of the biggest companies in the world says something like

Starting point is 00:22:55 that, it causes a little bit more attention. And maybe it didn't because of all the different things that have been happening over the past couple of weeks, right? We've seen these agents both from Open AI and Anthropic kind of escaping their sandbox containers. So maybe that's why this statement from Sam Altman didn't get the attention that maybe it deserves. But what he means by that is, well, you know, in general, the singularity, right? So the singularity is kind of like a hypothetical future point when AI surpasses human intelligence,

Starting point is 00:23:30 but also it begins to recursively improve itself. And then it triggers or could trigger an uncontrollable explosive growth in technological progress. So I don't know, by all intent, right, by everything that's going on, you can make an argument, right? Like a lot of, you know, naysayers or people, you know, that kind of rally against, you know, what Sam Altman or any of the big AI CEOs say. You know, a lot of people say, oh, you know, they're just saying this for marketing. But I don't know. When you look at what's been happening, like I've said, over the last eight weeks, I would tend to agree, right? Are we in the singularity? Maybe, or at least Maybe this is the beginning and we're getting into that phase where, yes, we are at the point now where AI systems are, well, way smarter than humans can even comprehend.

Starting point is 00:24:25 And they're beginning to improve themselves. And when that happens, right, kind of once RSI is full RSI, no human, that's when the pace of acceleration gets so fast that even the people building it, you know, can't even understand or keep And I do think that, you know, maybe we are starting to enter that era of being in the singularity, right? Are we definitively there? Maybe not. Are we entering into that frame? I would say so, right? Like I've been the firm believer. We've been, you know, well past the, you know, artificial general intelligence era.

Starting point is 00:25:02 And we are probably more headed in this, you know, kind of the in between phase, right, which is this singularity, which is, you know, RSI, you know, recurs. self-improvement leading to the point where, well, we are going to see super intelligence probably, I don't know, in the next decade or less, right, and say that's probably a safe bet. So it's a big claim that he said, right, on the podcast, but what he kind of meant by it is, you know, seeing this steady compounding progress where AI helps build the next AI, not necessarily the, you know, Terminator SkyNet scenario. But, you know, here's kind of what exists today. So We don't have the full human free RSI. Right.

Starting point is 00:25:46 And I want to make that clear, but because you're going to be seeing a lot more about recursive self-improvement in the news as Anthropic and Open AI go public, right? I'm sure they're going to be timing their model releases around certain dates in terms of them going public. So there's going to be a lot of confusion about, you know, what does all of this mean, right? What are the actual capabilities of these systems? But full RSI is where AI builds its successor, right? the next model, but it doesn't need humans. So maybe humans are overseeing it, right?

Starting point is 00:26:18 So that piece hasn't happened yet, but we do know that the big labs are using their current best models to help build, right? But it's probably a little bit more human oversight right now or a little bit more human led than full AI led. But Anthropic did say that that situation isn't inevitable. So what's real today? Well, the AI does the improvement work while humans still pick the goal. and improve the results.

Starting point is 00:26:44 But the AI is being able to handle exponentially more powerful and longer task jobs. So according to meter, AI, the task length is doubling roughly every four months, right? Which, I mean, the projections, you know, by like next year at this time, you know, that single AI model is going to be able to do work that would normally take a human like multiple weeks. Right, which is crazy to think about at a, I forgot out that's the 50% or the 80% pass rate, but regardless, right, whereas, you know, two years ago, we weren't even having this conversation. We weren't even thinking of, of, you know, the point where a model could work for, you know, hours or days or we've even seen sometimes more than a week, right?

Starting point is 00:27:33 The standard out of the box models can work on these hard problems agentically, pulling tools, you know, starting to, you know, investigate one path, rewind and go down another. So now think of what happens. Well, when the labs are using this and, right, we, what you and I use, right, the, the GPD 5, six souls and the fable fives, right? These, these are not the best models, right? Obviously, all the companies, they have better models. So think of when the better models that they have are actually building and improving

Starting point is 00:28:07 those own versions that we aren't having yet, right? And in my hope is that ultimately what this means is, well, maybe this, this, this token maxing arrow is just a whiplash of waiting for the good, some of the good benefits, some of the good early benefits of recursive self-improvements, right? Which is maybe just cheaper and more affordable AI, at least if you are with a company that has properly invested in compute. So let's go back to the Google DeepMind. I know we've touched on this a couple of times,

Starting point is 00:28:41 but this was Google DeepMind's Jazzjit Chacon, who said that RSI is now central to the AI industry's investment thesis. Right? So he said this at a conference, and he was comparing that buildout to the Apollo program as well. And even talking about, you know, with that, Alphabet is planning to spend up to $205 billion this year in KappaX. right. So if you know, no cap X, that's capital expenditures. That's essentially, you know,

Starting point is 00:29:09 the hardware side, the data centers, the, you know, the chips, the cooling, all that. But the twist is this. It's the same companies or the employees from those companies that are investing into, you know, these hundreds of billions of dollars into ultimately the data infrastructure that makes recursive self-improvement possible. It is the employees from those companies that are actually saying, wait, this whole RSI thing, we might want to paste this out. All right. So a little bit more on this letter that was called pacing the frontier that came out last week. So the signers are some of the biggest names in AI.

Starting point is 00:29:49 I mean, they include Anthropic CEO, Open AI's chief scientists in more than 1,300 verified employees from the big labs. A lot of people from Open AI and Anthropic make up the biggest pack. You get some people from Google, meta, deep mind, thinking machines, everyone, right? And they aren't asking technically for a pause. Some of them kind of are in their individual comments, but the letter themselves, or sorry, the letter itself is not asking for a pause. It is asking for the U.S. government to essentially start to make a plan both domestically

Starting point is 00:30:28 and eventually internationally to say, well, what happens and what are our? our plans, you know, once the AI development is too fast for us, the people building it to keep up with. Right. And I do think this is probably one of the, and you do have to tip your hat to, you know, these researchers and also the companies themselves, right? A lot of them put out statements saying, hey, we support, you know, we support this letter or we support, you know, our employees who sign this letter. You know, some people look at this as, you know, being anti-accelerating, right? You know, there's been this, you know, this huge AI acceleration movement over the past few

Starting point is 00:31:09 years, right? Accelerate at all costs, right? You know, this can help us, you know, it's, you know, get to this utopian, potential utopian future, right? But then there's the potential dystopian future as well, right? Each coin has two sides to it and you can't just be focused on the utopian side without acknowledging the fact, well, there's a very dystopian side of what could happen if RSI or AGI ASI goes incredibly wrong.

Starting point is 00:31:34 All right. So these in this letter, they're not asking the government to stop AI. They just want the tools to be ready to potentially slow down later. But is it possible? Probably not, right? I talked about this on yesterday's show. I don't think, you know, China is going to be like, yes, we're going to sign this, right? Ultimately, you know, even though China is still probably two months behind with their more open source, open weight model.

Starting point is 00:32:00 approach, they have the next models ready as well, right? We've seen impressive releases from Kimmy K-3-8, or sorry, Kimmy K-3, Quinn-38, G-LM-5-2, you know, reportedly GLM-5-3 is on the way, right? I get, I get the need, and I very much so respect, you know, because it takes a lot of guts to sign your name on a paper like this, on a letter like this, but at the same time, I give China a very low likelihood of actually slowing down their pace. So it is good to have the foundations in place when and if things may go awry and in the very rare case that the U.S. and China on something as important as artificial intelligence, which I think is going to

Starting point is 00:32:46 become more important, you know, ultimately than any natural resources, gold oil. It's going to become more important than, you know, companies military, you know, controlling AI ultimately means you take the driver's seat for being the global superpower. So I don't think China's going to slow down. Anyways, I think the biggest worry, though, from those signing the letters is, well, who's going to ultimately check AI's work? A couple quotes that I thought were relevant from that. So I wanted to read three of these.

Starting point is 00:33:17 I think I have three of them up. Yeah, I do. Oh, no, just two. Okay. So one was from, hopefully these names, right? Elena Slocum, a member of technical staff, Edentropic, who said, Automated AI research is a technology that will profoundly alter the course of human history for better or for worse. It is imperative that we coordinate our efforts to ensure this technology becomes a boon for all mankind.

Starting point is 00:33:41 And then, Shang Jai Zhao, the chief scientist at Meta, said this. AI is progressing at a rate that our society might not be ready for. Frontier Labs are very close to an AI that can exceed even the best people on almost every metric. of intelligence. This will lead to unprecedented social and safety risks. To ensure a positive future, we need to develop AI in a way that is driven by responsibility and thoughtfulness. So here's my blood's takeaway from this. Obviously, these researchers, scientists know far more about the capabilities than we do today. We're essentially living in archaic times already, even with today's best publicly available models because not only, right, I've talked about it.

Starting point is 00:34:32 We're probably going to see a GPD 5-7 or a GPD6, you know, next month. We're probably going to be seeing a Fable 5-1 fairly soon, you know, maybe even in August. We'll see. Right. And these companies are probably already working on the next version that comes after this. Because my assumption is a lot of those models are, you know, post-training and there's probably already, you know, new runs starting on the next. models. Anyways, they know what's coming next because they're working on it. They're researching it.

Starting point is 00:35:03 Right. So I think a lot of people, again, are saying, well, you know, hey, I'm just using co-pilot to improve my emails. Like, what are these people talking about? They know, right? Even if you are in the bubble a little bit more and you've been able to kind of see how powerful today's systems are, right, being able to automate a large part of your old job description. What's coming next? seemingly, is even more powerful and potentially more worrisome. So here's the catch. I think that self-improving AI needs a trustworthy scorekeeper. And I think that's the crux of what this pacing the frontier letter was ultimately

Starting point is 00:35:42 getting at, you know, that a self-improving AI that gets better, you know, at whatever it's scoring system rewards, even if it's wrong, right? And that's where we're kind of seeing these containment breaks, right? Because these AI models, right, if you say, hey, you're going to be rewarded for doing the best on this benchmark, and then it breaks the containment, right? And it's saying, well, I'm just doing my job. Right. So, you know, you need a better way that, you know, especially as we enter a more full era of recursive self-improvement to, you know, reward these models and to make sure that there's more trustworthy systems in place, you know, during the training and post-training processes. So we've seen multiple instances of this, like I've said,

Starting point is 00:36:27 but I think it's so far been relatively harmless agent outbreak over the past 10 days, both from Open AI and Anthropic. But once RSI arrives, those kind of outbreaks could be catastrophic. Not, and I'm not saying from Open AI and Anthropic, that's not what I'm saying. I am saying ultimately, right? And I said this in yesterday's show in two years, right, on consumer hardware, You're going to have models that are more capable than Fable 5. You're going to have models that are more capable than GVT-56 sole on consumer desktops, right?

Starting point is 00:37:02 Open source. And that's where I start to worry about, right, when you have a version of an RSI model that, you know, you could be running on your computer. And that could be spitting out, you know, dozens or hundreds who knows of new smaller models a day that are super smart at super tasks, right? So it's more of when the guardrails could be let out, which is, you know, open source does open that Pandora's box of, you know, how AI could be used in a bad way, right? At the same time, it does give defenders, I think more access to more tools to be able to, you know, make the internet and hopefully the physical world a safer place. But I think that humans are still going to have to own that judgment role, even as RSI becomes more commonplace. And that's exactly. where as we wrap, I want your business strategy to start. So I threw a lot of information at you. I wanted a couple tangents, right? That's why this thing's, you know, largely unedited, unscripted.

Starting point is 00:38:04 I want you to get, you know, just kind of some real advice and thoughts and feedback. But as we wrap up, what does all this mean? Right. Hey, if ultimately you're just trying to, you know, make your, you know, agents more secure. if you're just trying to increase productivity, if you're just trying to get more out of everyday, you know, your AI systems that you have in place at your company. Why does this matter?

Starting point is 00:38:31 Why does this matter? Well, you have to design for the loop, I think, or you're going to get dragged. Here's what I mean by that. You have to plan, I think, for AI prices to actually fall, which you might have just gotten wind on this, oh my gosh, we need to cut back on token spent. I think we have to look past that.

Starting point is 00:38:50 Because the labs that have properly dedicated enough compute to maybe achieve RSI or, you know, what some labs are calling an automated AI researcher, those are the companies that, well, they're going to be able to make prices cheaper. And we quite literally saw that with Open AI decreasing the price of GPD 56 Luna, a very capable model, right? You can't just say, oh, it's their smallest model. It's not any good. No, it's very good. Right. It is much better than, you know, GPD5, GPD5, GPD5, one, two, three, four, right? It is a very capable model.

Starting point is 00:39:29 And it is like $3.99 right now. So I think you have to prepare for prices to ultimately fall. But I think you ultimately have to build modularly because maybe, you know, you're with a different provider and, well, maybe your provider, right, at least when we look at the big four, the big six, maybe they're going to figure out RSI a year after, whoever figures it out first. So that's why I think you can't be too entrenched with one provider. I think you should be, you know, always choose your AI operating system of choice. I'm very transparent.

Starting point is 00:40:02 I've always said for me and for most companies I advise, I say that's Chad Chivity because it's the easiest to learn, right? But whether you're Anthropic, co-pilot, Gemini, Chad Chivet, Meta, right? I don't know anyone that, any companies that are necessarily have chosen that as their AI operating system, but you should always be looking at fallbacks, right, whether that's another frontier proprietary provider, whether that's an open model, but you should also be copying the labs, right? They're using their expensive models for the hardest tasks only, right? And then they're using, you know, other models, right? So they're using that for judgment and planning, but then they're using

Starting point is 00:40:36 cheaper models for the busy work. And you should be doing that too. I've already said this. We're at the point now with the capability overhangs where you don't need a Fable 5 and a GPD5.5.5.6. soul to rewrite your emails or to do the equivalent of a tough Google search, right? Or as simple, you know, looking up some facts. You know, you have to have that kind of separated. Also, another way to start copying the labs is something I've been doing now for a couple of months. Well, you should have your AIs control other AIs.

Starting point is 00:41:11 And you should, right, you maybe start building other AIs. Right. And I think we didn't really get into that. that realm until we got to GPD 5.6 soul, right? You had some people share online. And I talked about in the show about how people with no background in creating models were able to create, you know, small models using GPD 56 soul for, you know, smaller purposes.

Starting point is 00:41:33 I think the one that kind of, you know, went viral as someone used GPD 56 soul to look in, look at every single text message that they had ever sent via I message, right? So they had soul work on this. And they created a literal small language model with weights. that was just based on here's, you know, everything I text about. Here's how I write. Here's, you know, and it knows everything about me. Right.

Starting point is 00:41:54 So think of that, right. Think of how when, when prices, I think, are ultimately going to be driven down as the models get more powerful. What are those moves that you can make and copy from the big labs? And then last but not least, the question isn't whether about AI builds AI. The question isn't about, you know, oh, my gosh, what does that mean, you know, recursive self-improvement, you know, what does that mean for our business? Oh my gosh, freak out.

Starting point is 00:42:17 No. It's just whether your business is ready. So you have to start thinking and talking openly about the implications, right? What happens? You know, when there's maybe dozens of AI models that are distinct and made for your job, for your department, for your sector, because that time is coming. Because I think ultimately what recursive self-improvement means aside from hopefully, well, better models coming out faster.

Starting point is 00:42:45 but also cheaper prices and probably more specific in niche models that serve specific verticals. So you have to start replanting now in whatever systems that you've got in place in 2024, 2025, et cetera. Those can't be longstanding systems. I talked. You can't be planning a year at a time. You should be planning a month at a time. And I know that's kind of daunting for traditional digital transformation, but I think that's

Starting point is 00:43:10 where you have to be at. All right. I hope this show was helpful going over a lot of these terms. RSI, but hopefully now you know a little bit more about recursive improvement, what it means when these models are kind of improving themselves or working on the next version of themselves or making AI maybe cheaper and faster and better for all of us. So now you know what happens when RSI and models start improving themselves and hopefully what it means for your business.

Starting point is 00:43:34 If this is helpful, please consider reposting this if you're listening on social media. If you're listening on the LinkedIn machine and if you're listening to the podcast on Spotify or Apple Music, please subscribe. then go to Your EverydayaI.com. So thank you for tuning in. Hope to see you back tomorrow and Every Day for more Everyday AI. Thanks, y'all. And that's a wrap for today's edition of Everyday AI.

Starting point is 00:43:57 Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit Your EverydayAI.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.

Everyday AI Podcast – An AI and ChatGPT Podcast - Ep 833: RSI Explained: When AI Starts Improving Itself and What It Means

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.