Industry Roundup #4: O3 & O4-mini, LLama 4’s Rocky Release & Google’s Agent Ecosystem

Hi, I'm Maggie, VP of Curriculum at datacamp. And I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead, covering everything from data and AI literacy for busy executives to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com. All right, all right, all right. We are live. Richie, how are you?

Life is good. It is a sunny day here in New York. I'm glad winter is finally over and yeah, feeling summery. How about yourself, Adele?

Yeah, feeling summery as well. It's pretty sunny here as well here in Belgium. And it's Easter weekend, so I'm excited for the three day weekend weekend here. Do you have any Easter plans, Mr. Ritchie?

Yes, definitely only a two day weekend here, but yeah, the plan is mostly eating chocolate and probably going for a nap afterwards.

Yeah, that sounds exciting. A lot of. A lot of chocolate. I am personally going to a rave. Not a traditional Easter plan, but we're still going to do something. So excited for that. It's going to be a day rave out in the sun, so I'm pretty excited. Hopefully I'll be well recovered by Tuesday, but we shall see. So Richie, what are we going to be talking about today?

All right, so to begin with, we're going to talk about some of the New releases from OpenAI, from Meta and from Google. So lots of exciting tech just coming out or being announced. And then after that we're going to go global. We're going to talk about the AI Index report that just came out of one of the Stanford institutes. And then we're going to talk about whether or not AI is going to kill us all. Just to top it off, a cheery story to wrap up.

Yeah, very, very positive story. We're going to talk about the AI 2027 research report that was released a couple weeks ago, which is a really interesting read, but we'll stick around till the end to see how things could unfold. But yeah, what's the first story of the day?

So OpenAI have released the O3 model and the O4 mini model. Just talk us through what's new in these models.

So if people remember, O3 was actually released in December during the 12 Days of Christmas event and it was still under preview, it was not released. And I think then either at the time or a bit after they released O3 mini and it seemed to be doing relatively well. But when they showed it during the 12 days of Christmas, it seemed to have really good capabilities on reasoning, math and even did pretty well on the ARC AGI benchmark. A lot of people started saying, oh, this could be. So now people have access to it, you can Access it in ChatGPT and I think it's gonna be accessible soon via the API. And in a nutshell, O3 has much better reasoning capabilities and the two defining features that seem to give it a lot of this improved reasoning capability is a couple of things. One, native tool use O3 is able to do Google searches, do image generation, to be able to interpret images, so it's actually able to reason with images within the chain of thought and to use code interpreters. So they're both to execute code while reasoning. And an example that was shared here was the ability to do reasoning about an astrophysics research poster. So it actually took the picture of a research poster within the chain of thought. You could see the model zoom in on certain parts of the poster to be able to come up with a certain answer when prompted. Or for example, they will take a blurry image or an image that's upside down and manipulate it using Python to be able to better understand and reason with the image. This is quite exciting because it does open up a slew of different use cases. Think for example, being able to do a rough wireframe to code generation to ability to iterate on it. That's quite interesting. And I think there's going to be better use cases for vibe coding with these models down the line. And this does definitely translate to better benchmarks, right? Like OpenAI 03 and 04 mini. We're going to get to O4 mini. They do crush a lot of benchmarks on coding, math and sciences, visual reasoning especially, and that's pretty interesting to see. But we're definitely going to have to wait and see how these models behave in the wild. I've been using them a bit over the past couple of days. They're definitely better than GPT 4.0, definitely better than GPT 4.5 and then 01, but still encountering errors here and there, still have to fix stuff. It's not the AGI panacea that we're all hoping, or maybe if you're in Belgium, according to the Stanford AI Index Report, not hoping for, so. So we'll see. That said O3 is now that Frontier model from OpenAI. Greg Brockman was on the live stream saying, every now and then we release a model that is a foundational leap. GPT4 was that foundational leap. O3 is that foundational leap as well. So we'll see what's really interesting as well. And you can see here some of the reasons why OpenAI took such a long time in releasing that model is if you remember during the showcase of O3 during the 12 days of Christmas event, it was quite expensive to run. I remember kind of O3 high, which is a high end version of that model, cost around $3,000 to run inference on. So unless you were a multi billionaire, you're not going to be able to use that model every day. But now seems like on the API, at least what they shared, it's much more relatively affordable. So O3 apparently has $10 per 1 million input tokens, 2.5 dollars per 1 million cashed output tokens, and $40 per 1 million output token. It's definitely not cheap, but it is definitely not a $3,000 inference price tag. So pricing here is definitely complicated with reasoning. The cost depends on how long it thinks as well. So tokens here also include reasoning tokens. So that is something to think about. So you're not just getting the output. And this comes here why OpenAI also released 04 Mini, which is apparently looks like their next generation model but a much more smaller as well as cost effective version of it. And that is much more cost effective and also relatively cheap with good performance, which makes me wonder how good fully scaled 04 will be. And I think we'll find out in the next couple of months or so. So that's essentially it in a nutshell. Richie, I'd love your take on a couple of things. So according to Noam Brown, he's a researcher at OpenAI, he shared the graph. If you're listening in, we're going to link it in the description below. Scaling, inference and scaling Reinforcement learning seems to be holding up. So if the scaling laws continue. Simple question, easy one here for you, Richie. Are we going to achieve AGI this way?

Let's go with the simple questions to open up. Are we going to achieve AGI? All right, so there are many different measures of like what constitutes AGI. Some of them we hit already, some of them we're not going to hit. I think the more important question is is performance going to continue to improve and are those improvements actually going to benefit users like either individuals or companies that want to make use of these things. So at this point I'm pretty much convinced that given the level of investment and the number of very smart people working on AI, we have the capabilities to create a better than human artificial intelligence. So the idea of an artificial superintelligence. So for me the two limiting things are first of all the economics. Is the money going to run out? Like all this sort of venture capital money, is that going to run out before we can get to that point? So we've seen a similar thing with self driving cars. So the self driving cars, we kind of know how to make this sort of level five completely autonomous car. It's just that it's probably going to be so expensive at the moment that the economics aren't quite there. And a lot of the self driving car startups sort of disappeared. So that's one sort of blocker. The other big blocker is going to be around energy usage. So generative AI is expensive to run already. It's very energy intensive. These reasoning models are just orders of magnitude more expensive. Once you bring in like agentic use cases, then the power grid infrastructure is not there in most countries in order to run these sort of things. So it's basically money and energy are going to be the limiting factors, not our ability to create artificial superintelligence.

Generally. I agree with the intuition. I think what's kind of the canary in the coal mine here is whether we're able to arrive at systems that are able to integrate or maybe take a step back. If you reach a point where you have systems that are both great at math, great at coding, and they are operating at near level human genius level, and then you're able to build agentic use cases where you can have recursive self improvement. I do see a future where you can reach a point where you have really strong intelligence that may even go beyond human intelligence. I do agree with the limiting factors of being energy and money, but I don't think if the opportunity is there, I cannot see governments and corporations not putting in all their capital and being able to reach that point. And we'll see here at the AI 2027 report. This actually touches upon a lot of these dynamics that we're discussing here, but it is quite interesting. I think the main question for me is will we keep seeing these scaling laws? A lot of the assumptions that we're discussing about of money, energy, investment, so on and so forth are predicated on the assumption that scaling laws will continue. And I guess we're going to have to see here whether we're going to need any algorithmic innovation on top of these scaling ones.

Actually, I have a question for you about the OpenAI releases, and it's kind of related to this in that there are so many different models available from OpenAI now. Do you think O3 is the one that's going to take over for everything? Has it made some of the older models like GPT4 obsolete?

No, I actually still use GPT4O quite a lot for, you know, maybe the number one use case for generative AI is writing emails. No, if you think about daily routine tasks that don't require reasoning, I still use GPT4O. It's still my default model. But I do think that OpenAI has reached a Google Gemini level of confusion in their user interface. If not more like I'm looking here at the ChatGPT interface. We have 034.0.4.0 scheduled tasks. Who uses scheduled tasks? GPT 4.50304 mini 04 mini high GPT4O mini and GPT4 is leaving on April 30th. Rest in peace GPT4. So we have what, like 6, 7, 8 models that you can choose from and I think this is going to be probably. Sam Altman has hinted at this. I think what will happen is that you'll just have one model, but you will have an orchestration engine. Depending on the task, the best model is assigned and you'll probably be able to access a developer mode where you can choose your favorite model based on the task. I think that's probably going to happen. But yes, indeed, the naming convention is a mess and I don't think that O3 will be the go to model for everything. I think it will be the go to model for developers, I think it will be the go to model for researchers. But if you are a casual user of AI, I don't think that you need O3 that much unless you're taking a stab at vibe coding, which is pretty interesting.

Okay, that makes a lot of sense. You use the simplest model that you can get away with because it's going to be cheaper and easier. Also, yeah, I have not used scheduled tasks. If you want to schedule stuff, airflow is the way forward and there is a new airflow release. Airflow 3 is coming out I believe very soon. So I don't think we're going to talk about that. But yeah, exciting news on the date of runt at least.

Yeah, I mean here you're talking about data scheduling, but really GPT4O is just like with scheduled tests, it's just like glorified alarm. It's a glorified reminder tool. That is it. Just use your reminder, use your phone. We talked about 030 and 04 mini. Let's talk about additional releases. Richie. So seems like llama 4 isn't as good as people hoped and talks about Meta. Watching the release here. Richie, do you want to talk to us about what happened?

I feel slightly bad for Meta because it's like you spend however many billions of dollars creating this fancy new model, the benchmarks aren't quite there and suddenly half of social media is just complaining about it. So it's a little bit unfortunate on that part. There have been some accusations that because the performance wasn't quite there, Meta were trying to gain the benchmark. So because you train these large language models against a lot of Internet data, you can train them against like what's publicly known about the benchmarks and that gives you better performance on the benchmarks, but it's not actually benefiting real users. So little scandal there. But the biggest story is that Meta has a fundamentally different attitude or business model for developing these large language models compared to say OpenAI Anthropic, where it is existential for those companies to have the best model. For Meta, it's just kind of a side gig on top to help out the social network and then they give it away. So they're monetizing it less, but they're still providing stuff that's available for free to the world. So kudos to them is very cool stuff. Now there is one other complaint which is slightly more legitimate in that the Llama 4 models come in different sizes. The smallest one is called the scout model. Now this is actually still a pretty big model. This is a 17 billion parameter model and then some grumblers from researchers. So Maximilian Bonne, who is a future data frame guest, he was saying actually as an individual, really hard to work with this stuff because you need more multiple GPUs just to fit it in memory. So the sort of sweet spot at the moment for individual researchers or sort of smaller scale laboratories stuff you can put a single GPU like 4 billion parameters is kind of ideal, maybe 8 billion parameters at a push. So these smaller models are much more useful for researchers because you can play around with stuff, it's a bit faster and cheaper to do.

And speaking of that problem here as well, to give some context, like I tried playing around with Fine Tuning Llama 4, for example, you do need quite a bit of budget to run GPUs here. You can't just do like a Google Colab which with free TPU access for example to be able to fine tune it contrary to llama 3 or llama 2 for example. So it does definitely limit your ability to create useful new variants based off of this model as an individual, which is truly what the spirit of open source is about. And that's where it comes here. There are also grumblings on context length. Do you want to talk to us a bit about that ruching?

Oh yeah. So the other thing around this is Lamp 4 claims to have a 10 million token contact length, so it can remember basically everything that you've thrown into it if you want to get it to work on dozens of documents at once. Actually I have no idea how many documents. 10 million. It's a lot of documents you want to throw.

Like it's Most likely like 14 books if not more because I think 2 million is like the five Harry Potter books to a certain extent.

I love that you have that statistic on.

Like I did a code along on the Gemini model which has the 2 million context length window. I tried to run the math.

Yeah, okay, so it's big but so Andrei Berkov, also databarive guest was complaining that actually models aren't trained on any kind of import data that's long. So actually if you try and use the 10 million token context window, you're probably going to get garbage results just because the models just aren't optimized for that. There aren't any good pieces of import data that are that long. So it's probably going to give you jump results and it's more of a vanity metric than anything else.

So maybe a couple of discussion points questions that I have for you here Richie. If it's not a reasoning model, are we nearing the end of continual gains in LLMs?

Well, there are enough companies trying to make better models. The competition is incredibly, incredibly fierce. Like it's got to be brutal being a Frontier model creator. So yeah, we're going to see gains but there are very much limits in terms of can you generate better marketing copy. Like this stuff kind of it just works already. Like it's like my main use. It's like I hate writing marketing copy. I will have to source that to AI. So yeah, you can't get that much better at that. One thing I have noticed is is some fun like story generation apps you can do like choose your own adventure with AI and those kind of work better. And the benefit there is not like that the LLMs are getting better. It's more that they remember stuff that you said, so they don't just forget characters that were in the story like two minutes ago. So long context window and better like memory handling is a really great feature for that fun use case. And there are benefits. Like it's not reasoning, it's just other parts of pretend brain. Like memory is very different from thinking.

Yeah, I think you allude to something that's quite important here. Right? Because the ability for a model to perform well on benchmarks is something, and the ability for a model to be useful in the real world is something else. I don't think benchmarks, for example, measure memory. They measure the ability to solve problems. There may be specific benchmarks that are able to measure memory, but generally reasoning benchmarks or math benchmarks or coding benchmarks or something along those lines. Lines don't necessarily do that. But indeed, as you mentioned, one of the things that makes GPT4O image generation so useful is the ability to retain context and the ability to edit an image and iterate on one particular image. The more large language models have the ability to kind of keep that context, be cohesive, create and provide value for the user and however they query it, that's I think actually goes a much longer way than the ability to just do really well on benchmarks. But that said, the vibes on llama4 and what I discussing have not been so positive as well. So it seems like it's a losing battle on both sides here. Maybe a last one given the discussions here and the accusations. And we've also seen this slightly alluded to during the 12 days of Christmas. So this is not just a meta problem. It was alluded to during the 12 Days of Christmas release of 03 when O3 did really well on the ARC AGI benchmark. But our LLM providers trying to game benchmarks and is this the game that we're seeing today?

Almost certainly there is a strong incentive to rank well on benchmarks because that's one of the measures that people use in order to choose which model should they be using. I mean, obviously if you're a product manager or an engineer, you want to try lots of different models and see how well does this perform on your own use cases. Because just because the benchmark score was high doesn't mean it's going to be good for what you want to do. But benchmarks are like they're kind of like a shopping tool. How well does this model perform? It's like, oh, maybe I'll take a look at that. So yeah, there's obviously a strong incentive for all these model creation companies to do well on benchmarks. So we're likely to see scandals like this thing with llama4 over and over again.

Yeah, I'm not surprised here either. So third story of the day, I think this is the last one covering technical releases. So Pritchi, do you want to walk us through it? Google and its new AI agent tooling ecosystem.

Absolutely. So Google had their Cloud Next conference and they made some announcements around agent tooling at the same time. So agents obviously like the hottest thing this year in AI and there were two releases. So the first thing is the Agent Development Kit. So this is just an open source framework for creating agents. Now there are a few of these around already. So LangChain was the first AI development framework and just added some agent features. Suddenly Llama Index, which was also spun after Meta, that's been around for ages and that has agent features. There are a few others. So Anthropic's got tool use, Microsoft's got Autogen, you got other tools like Crewai, There's a ton of these things. So it's a very crowded space. But if you are working in a Google Slack like I guess if you are primarily working with Gemini, you probably want to use Agent Development Kit and yeah, let's see how it pans out. Obviously it's just been released, it's early days but it's worth taking a look at if you're building agents. Now the other thing which may be more exciting is a framework called Agent to Agent. So actually it's a protocol rather than framework. So it is about standardizing communication between two agents as the name suggests. I think the closest thing that exists already is model context protocol from Anthropic. So that is about standardizing how agents communicate with tools. And Agent to Agent is about how two different agents talk to each other. I'm not quite sure what the use case is yet, but I presume in general when you have agents you're probably gonna have like one agent that does stuff and then you have another agent that monitors the quality of the original agent's work because obviously they get things wrong. And so you wanna make sure that you have some kind of quality control in place. So it's probably gonna be standardizing communication in that kind of multi agent setup. And obviously you can have more, more complicated agents. So let's talk about having like if you want a scientist then there's lots of different things you want Someone to do like background research. You've got an agent for that and you've got an agent to like, I don't know, come up with new hypotheses and all these sort of things. So different stages of your workflow, they're going to require different agents. You can probably get very complicated setups. And so this is going to be an important tool for that.

Yeah, I'm really happy that Google is taking the lead on building these types of ecosystems and frameworks and protocols, because we will reach a point where you have quite a few different AI agents, frameworks and kind of tools and working in the wild. Right. And to be able to run these communication protocols that are standardized makes development much easier. So maybe given this, and the tooling framework, especially on the agent side, is maturing. And we saw even on kind of protocols like the model context protocol is also maturing. What do you think this means for developers and organizations?

Richie, if you are a business where agents are not your primary thing, so you're not like a tech company, then you don't really want to care too much about the tech stack. You just want it to be easy because you want to care mostly about your business processes. It's like, how do I automate these things? How do I do stuff? How do I achieve things with these agents rather than like worrying too much about building them? So having agents that are really, really easy to create is going to be huge. And that means better tooling, better middleware. So we really need to reach a stage where creating an agent is as easy as creating a dashboard.

That will be an exciting time. That will be. I mean, there are already low code tools to be able to build these types of mini AI agents to a certain extent that perform tasks. But indeed, if it becomes really streamlined to build agents that can communicate with other agents relatively easily without having to think about it, that will be super exciting. Maybe also a really easy one here. Where do you see the AI agent space going in the next 12 months?

You keep giving me these big questions. It's like I just came across the wall. So there's only a few things you care about. Like basically at the moment you can do really simple narrow agents. It is actually as easy as creating a dashboard at the moment. You want to do something really hard like, okay, I'm creating an AI doctor or I'm creating an AI software programmer. That's like crazy research and it doesn't work very well at the moment, or it's not that reliable. So basically the way we're going is it has to be more complex in its ability to reason. It can do broader tasks and going faster as well because sometimes you don't want something to be thinking for a minute before it performs the task. And then also more reliable as well because sometimes you want it be correct pretty much 100% of the time, not just 90% of the time.

Okay, so better reliability, better ease of creation and just more raw intelligence.

Yeah, yeah. Smarter, faster, more reliable.

So I think this wraps up this story. Next up, we're going to talk about the 2025 AI index report. So maybe give us a bit of background, Richie, on the report and I'll go a bit more deeper into the results.

Yeah, sure. So this comes out of the Human Centered AI Institute at Stanford University. So this is an institute founded by Fei Fei Li of AI Data Collection Legendary fame. And yeah, it's just a great piece of thought leadership about what is going on with AI all around the world. So, yeah, you've gone into it in depth, but tell us more.

Yeah, so personally, it is one of my favorite reads of the year. That and the State of AI report by First Point Capital I think are my definitely two reads of the year. So a lot of the results are not surprising. Models are getting better, they're becoming more intelligent. And you see this in the benchmarks and you see that large language models, I think we experience this every day. AI models are becoming more and more useful as well. So they're able to take on even more tasks than the previous year. Adoption is increasing. There is a lot of lot of excitement on adoption, especially by businesses, and a lot of money is going in that are all investing in AI. One of the things that I'm particularly thinking about these days, especially when you consider the state of the world and the state of politics today, is the geopolitical angle of AI. So currently the USA is leading on notable model releases according to the report. So 40 models in the US according to the report, and 15 to China and three to Europe. So Europe is pretty behind here. As a resident of Europe, that makes me sad. But China is closing the gap on performance as judged by the benchmark. So we can see this definitely with deep seq R1, and I wouldn't be surprised if we see deep seq R2 in the next couple of months. And there's definitely more publications in patents. There were also notable launches in terms of models from regions such as the Middle East, Latin America, Southeast Asia, so on and so forth. And if you look at that geopolitical dimension here, maybe taking a bit of an aside, regardless of what you think of the latest craziness when it comes to the tariff policies and what's going on here, but if you look at a lot behind the tariff policy and you listen to a lot of folks that are close to the presidency in the US that are talking about the tariff policy, a lot of it seems to be related to shore up the semiconductor industry and to better compete and solidify the US's advantage in AI. So this is definitely something to look out for in the next couple of years and we're going to touch upon AI 2027. That heavily emphasizes here the geopolitical competition between the US and China on AI, but definitely something very interesting. Now one thing that also in the report that surprised me is the degree of excitement and skepticism across different nations about AI. So countries like China, Indonesia, Thailand, the population here, I'm not talking governments or officials, are pretty AI enthusiastic. They seem to generally have a positive attitude towards AI in the sense that they think it is beneficial and useful. On the other hand, Western countries actually have less of a favorable view towards AI. For example, Belgium, the Netherlands and the US are some of the most AI skeptical nations. Surprising on the us, not surprised in Europe, France and Germany had the biggest increases in AI optimism, which is pretty interesting, and India had the biggest drop compared to last year. That's also quite interesting here as well. One thing that the report also covers here, data from the Raspberry PI Computing Education center, which covers and tracks countries that provide AI education across the schooling system. So There are about 20 countries here, Britain, Australia, most of Eastern Europe, Saudi Arabia, South Korea, Argentina, a few countries I hadn't expected here, including Kazakhstan, Thailand, Ghana. So good on these countries that provide AI education. And it will definitely, I think it may not be as important today, right. But we'll definitely see generational gaps being created over the years to come between these different countries if they do not invest in AI education. Maybe here a shameless plug to check out datacamp for classrooms. It is free for teachers, but the responsive AI situation is a bit sketchy. So that's another one that also caught my eye. There are many new really good responsible AI benchmarks, but not all large language model providers are actually testing their models on these benchmarks before release. The increasing arms race around AI, whether it's between the US and China or between OpenAI and anthropic, and all of these different frontier model providers are creating dynamics where safety is not as important as it used to be. And now it's more about deployment and making sure that these frontier models drive value for users. Personally speaking, I do not think like doomerism necessarily is a good path forward when it comes to kind of the responsible AI discussion. A lot of the discussion here seems to be couched by discussions over that even if you release a large language model today, it's going to destroy everyone. But definitely responsive AI is extremely important. And the report does show that there are notable increases in AI related incidents. Right. So as AI becomes even more embedded in the workflows and businesses and daily life, and as AI related incidents increase, there may or may not be one big incident that will change people's minds, I think in the next two years. So that's essentially in a nutshell, the overview of the report. I don't know. Richie, did you read it? What was also cap came out to. Came out to you is interesting.

Yeah. So those cultural differences are very fascinating, I have to say. The fact that Asians are more enthusiastic about AI, that doesn't surprise me. But the particular countries, like I hadn't realized, like Thai people and Indonesians would be like the most enthusiastic about AI. I would probably have had Japan, actually, because Japan has like a big culture or historically at least of like being enthusiastic about robots. So that would have been my guess. But yeah, there's certainly a lot of skepticism around AI in many Western European countries, I think.

Yeah. And I think it was quite surprising for me to see such skepticism about AI in the west especially. That I think shows kind of the myopia that I have about AI, because I'm relatively in the know when it comes to the industry. I use AI every day. I talk about AI for a living. So it's relatively exciting for me. But the data here does show that if AI veers to becoming a. You know, I'm not saying intentionally. Right. But if we have some form of externality or unintended consequences, which we will certainly have when it comes to AI. And AI is increasingly seen as a force for harm than a force for good. I do think that we're going to. We may have a social reckoning when it comes to AI, especially in the West. A really good example comes from social media. So if you remember in the early 2000s, 2000s, social media was often looked at as a generally net positive thing. And then 2015, 2016, Cambridge Analytica scandal. Social media is rotting people's brains. Jonathan Haidt's book recently on the anxious generation. Right. And now today you see schools banning funds in the school in the US you see Australia pushing for age verification on social media. I'm not saying this is a bad thing thing. Right. I'm just saying that you see a reversion to the mean. I think the backlash on AI, given that we are experiencing the growth of AI and potential backlash of AI in a much more compressed timescale, could even be sharper if not well managed. And depending on how the backlash goes and it goes forward to the discussion we're going to have on AI 2027, the prisoner's dilemma of do we accelerate or do we slow down becomes very real when your competitors also have to make the decision here. So if the United States or if Europe or the west slows down, but other countries do not, then you will be in quite a disadvantageous position compared to your competitors. So a lot to unpack, a lot to think about, not necessarily a lot of clear answers.

Absolutely. So yeah, AI is going to have some big, big geopolitical implications and I think that leads nicely into the next story about the AI 2027 project.

Yeah. So I've been harassing the team all week sending this research project to a lot of folks. And I'll caveat first right before anyone jumps in on the comments or reaches out that I'm being too pessimistic. I am generally not an AI doomer to explain where I come from. I'm actually quite optimistic about AI, but I am cognizant of of what exponential growth looks like. And to give like a very real use case of what exponential growth looks like, I would like to remind you of what happened in 2020 when a virus that the coronavirus had infected thousands of people in in one month suddenly became a global pandemic. That's what exponential growth looks like. So maybe I'll give some background. AI 2027, it is a research project. I'm not entirely sure what to describe the paper because it's both like an interactive website, but also a highly in depth research paper written by Hope I'm pronouncing the name correctly here, Daniel Coco Taschlo, as well as Scott Alexander, Thomas Larson, Eli Lefland, Romeo Dean, which are some of them are previous OpenAI researchers, some of them are responsible AI researchers, some of them are policymakers, so on and so forth. Definitely one of the more interesting reads I've had in a while in this space. And essentially the premise of this research project research paper is that all of these Frontier Lab CEOs whether Sam Altman, Dario Monday talk about how AGI will come in the next five years and talk about how scaling laws will continue. And we touched upon this earlier in our discussion on O3. So let's make this basic assumption. Let's assume that AGI, that scaling laws will continue and that reasoning models will continue to improve at the pace that they are improving on today. In that scenario, we could see AGI and systems that improve themselves. So systems like the technical term here would be recursive self improvement in the next two years by 2027, which is what AI 2027 era is about. And if that happens, we would be entering real uncharted territory. So imagine the systems that are near human genius level in math and coding being able to improve and build on their existing architecture and build new, more efficient, more intelligent architectures that could lead to what a lot of people call an intelligence explosion, where the rate of progress of AI systems becomes so explosive. Where we go from AGI to artificial superintelligence and the researchers that are behind the scenes, all what they're trying to do here is struggling to keep up with all of the research that these models are putting out. So the paper describes a scenario based forecast where it starts in 2025, I think, or late 2025, describes a hypothetical fictional scenario between a competition between two AI giants. Open brain in the US I wonder who Open Brain is based on. And Deep Sent in China. I also wonder who Deep Sent is based on. Are competing for the first ASI system, the first AGI, the first artificial superintelligence system. And what's important here? A couple of assumptions to make which are reasonable assumptions. Whoever is first to artificial superintelligence arguably wins the future because you have the most intelligent, most capable model at your fingertips. And if you want to really think nefariously in sci fi here, you can potentially do cyber attacks on your competitor that are absolutely insane. Right. And you can't even fathom what they think about. So I won't go into the depth of the paper here. Right, but like the paper ends essentially in 2027 with the arrival of an extremely strong artificial superintelligence, or AGI, essentially that is misaligned, where researchers have uncovered that this system is misaligned. And upon the arrival of the system, the forecast ends with two possibilities. Either we slow down or we accelerate. I'm going to leave the link in the description. We'll let you read what these two scenarios look like here. But both of them are not necessarily the best scenarios and both of them have a lot of difficulties ahead. One of them is definitely worse than the other. Some important context here set by the papers is at that point, if you have developed an AGI that is more intelligent at politics than any politician, or that is more intelligent at tech than any engineer, or that is more intelligent at AI than an AI researcher, then surely it will be so embedded in the daily workflows that politicians, decision makers, leaders use this AI every day, numerous hours a day, and in that sense, like a misaligned system, would have networked access to everyone in power and would be able to influence how things go because you'll be able to connect the dots between politician A, politician B, so on and so forth. And if you think AI will not be used by politicians, I would take a step back here and think one Politicians are people, they need to they need support in managing complexity. Indeed, there is some reporting, albeit not sure how credible. I'm not going to take it at face value that some of the tariff policy may have been set with the help of ChatGPT, so I'll link it in the description below. But again, I'm not going to say that this is 100% real. So with these kind of assumptions and kind of discussions set here, you would think that the natural answer is to slow down and understand better how to build a better, more aligned system. But the dynamics here really resemble a classic Prisoner's dilemma. So let's say first company, company A slows down, then your competitor will be incentivized to go faster because they will be first to artificial superintelligence. But if you refuse to slow down, then you will forego your arrival to artificial superintelligence. So if your competitor slows down, then you have the incentive to keep going the because you have the chance to arrive at ASI before them. So what I really enjoyed about this paper is that it really paints that Prisoner's dilemma and overlays it over current geopolitical dimensions and current geopolitical discussions. And it really for the first time, maybe, maybe I'm not well read on this topic enough. Really created a relatively realistic picture of what the nightmare scenario could look like with AI. So highly recommend everyone reads the paper not to be a Debbie Downer here and ruin your weekend and get a gauge of kind of at least a potential world that we could be headed in. Now that said, I don't want to be an alarmist. There are many critics of this paper. We cannot confidently say that scaling laws will continue. Even if they do, we cannot confidently say that this will lead to recursive self improving systems that will uncover new architectures in AI within days or weeks, or be able to emulate the brain and so on and so forth, let alone artificial superintelligence. Thirdly, as you mentioned early on in our discussion, there will always be a limitation on compute and energy. And fourthly, let's say we do develop AGI. There's also a lot of benefits to it. We're probably going to be able to solve cancer or do a lot of wonderful things. So I'm definitely not a doomer, but it is a very interesting paper I would say. So that's my spiel. Richie, any thoughts about AI27 when you read it? Yeah.

So I have to say when we start thinking about, okay, what's the future of artificial superintelligence going to be? This is where I feel like a lifetime of reading sci fi books is finally paying off.

So.

It'S good because yeah, we don't know what the future is going to be, but think about these things in a storified manner works really well. So I think my personal take is if we don't have this hard takeoff where we have like a really, really just rapid advancement in AI, the most likely sort of long term implications of artificial intelligence, they're going to be somewhere close to the implications of flight. So on the whole having airplanes is beneficial for humanity, makes communication easier, makes you more productive and you have joyful holidays around the world. But also there are some side effects like obviously there's the environmental cost of flight, but then you also have specific things like military use cases where you've got bomber planes and you know, flights result in people dying in ways that weren't possible beforehand. So AI probably going to long term land somewhere like that, net benefit to humanity. But there are going to be some new kinds of problems that are caused.

Honestly after reading AI 2027, I'll take that any day of the week over these scenarios.

So all this AI safety stuff is actually really, really fascinating and it's good because there are some very well defined scenarios in what can go wrong. Like there are researchers, you know, the smart people have been thinking about this sort of stuff for decades and we know like specific things that we shouldn't do in order to have a really bad outcome. So there are obvious things like you don't want AI that is going to help people do cybercrime, but at the same time also pretty much every spy agency, every cybercriminal organization is trying to find exploits and so software making use of AI to do that. So it's almost Certainly happening just behind the scenes. You're just not very well aware of it. There are other things like using AI to find novel viruses and things like that for bioterrorism where it's like, well, you want to have like smart reasoning AI for medical researchers to find cures to existing diseases and chemists to do cool things. But also at the same time, if you are a bioterrorist, then you probably want to use the same tools to murder people. So there is a double edged sword and it's going to be some very thorny sort of edge cases in terms of like, it's going to be subtle differences between we have a really great future and we have a really terrible future. The stakes are pretty high in getting this right.

Yeah. And this comes back to one of the episodes that we had with Ian Bremmer and Ximena Riveros on the need for a global AI governance framework, at least to hinder some of the potential harmful consequences of AI. But given the race dynamics that we're in today, I think we need a much more peaceful world to be able to achieve these types of frameworks and treaties. Final thing, Richie, any final note on AI 2027 you'd like to share with our audience today?

Yeah, I mean, just to follow up on that last point. So one of the recurring themes throughout this episode has been rivalry between the US and China. And over the last decade we've seen a kind of unravelling of globalization and the US China competition is the biggest thing. So having the two organizations from the two countries compete in terms of AI, it's good in that it speeds up progress with AI and it gives different viewpoints. But at the same time, if we do end up with a situation where the United States and China really fall out, then that's going to make existing conflicts like Russia, Ukraine or Israel versus most of the rest of the Middle east. They're going to be small for I compared to US China. So that's the big geopolitical worry, I suppose.

Yeah, let's. Let's hope that it does not reach that point. Now, Richie, anything else before we wrap up? Any plans for the weekend?

Well, I was going to just eat chocolate for Easter, but actually now I think I'm going to go buy a bunker somewhere very remote.

Well, for those who are not aware, I have been on the prepper train for a long time, so. So it's not, it's not a bad idea. But I would still put a low probability on it as of today. Okay, cool. Richie, always a Great chat, Levy. Thank you so much.

Pleasure chatting with you.

D. Yeah, talk.