AGI is not coming!

Hello. I'm out of my usual room today. As you can see, I'm flying to the US for work, but I wanted to give some quick updates on OpenAI's new model launches the past week. So, they've released an open- source model in different variants, and they've also released a new frontier model in different variants. The open source model called GPTO OSS uh and the Frontier model obviously called GPT5. Now um I am sure you've heard and seen the announcements but uh I wanted to kind of give a quick thoughts on this and that is I really feel that the era of boundary breaking advancements is over like AGI is not coming and we can be reasonably sure about that with not just open AI's new launch of models but also other frontier models. Like it seems very much that we're in the Samsung Galaxy era of LLMs. And what I mean by that is you remember like the early smartphones, every new generation was absolutely groundbreaking with like new things and boy, it really expanded the capabilities of these devices. And now we're at Samsung Galaxy 425 Star or something like this. And it has a bit better camera than the last generation. and like the image is software is a bit better or something like this and it I'm not saying we're there but it feels we're definitely going towards that. So there's big talk in the community that OpenAI has probably used synthetic data sets extensively in training these models and also used reinforcement of learning extensively and that means that rather than sort of collecting data from all around uh they are gearing their models towards particular use cases and it is widely assumed that these use cases are coding uh because coding is where these LLMs making a lot of impact right now uh benchmarking like all these type of benchmark tasks. Now I'm not trying to say that they are gaming the benchmarks necessarily but certainly uh collecting that kind of data that would be helpful for these benchmarks uh and then directly reinforcement learning on it um is not out of the question here and this is evident by the fact that people are saying that especially the open- source variants of open AI's models they hallucinate uh to a much higher like the model seems to have a lot less world knowledge than other models. But the at the same time these models are really really good at instruction following and at uh tool calling. And so I do believe we see a future where LLMs sort of act as tool calling blues so to say as okay you give them a bunch of tools and then the LLM is responsible for sort of routing information between the tools and we already see that today right you can have agentic behavior from pretty much any provider of LLM that's like please don't be impressed if you see something like this uh this just comes out of the box with the LLM providers. So what matters in the future is the tools and how you can get access to them. And what matters to the LLM providers is first of all how much can they gear their models towards tool calling. And there's an inherent danger because I do believe a lot of what is disguised as tool calling still requires a large degree of world knowledge and we'll see how that plays out. And the other thing is price. Uh, so OpenAI's GPT5 model is notably priced really, really cheap for how good it is. It's at the same level as other Frontier models as far as we can tell from benchmarks and from impressions of people. Um, but it's really good like at coding and tool calling and it's really cheap uh or cheaper than others. And so it feels like we've we're out of steam on the foundational research in terms of making these models better. It's long been said that we have probably about two to three orders of magnitude left in terms of scaling training data and scaling compute. I don't think so. Um I think or at least it may be not worth it. And what people are probably doing behind the scenes now is just trying how to synthetically create data sets and how to do the reinforcement learning reward shaping and so on which in a way is good because that's where we started in machine learning uh when sort of deep learning came on the scene and uh we're trying to figure out how to best train these things and then came the LLM era where it was just like pump data pump data pump compute and that made a lot of gains and now we're back to Okay, you actually have to do something like smart. Uh, so that's exciting. However, compared to earlier

now it takes like millions of dollars to do one training run. So, I do believe like a really good era to be right now is in being able to predict where the big training run ends up. And obviously the big companies are already doing that uh and have been doing that for a long time. But um being like right now if I were to get into research right now uh I think a very interesting era would sort of be how do we predict from small experiments and from early training trajectories where we'll end up and ideally how we can fix it along the way. Right? So we don't have to restart again but we're like okay it's a bit like it's a bit off to the left. Let's sort of fix it like this um to achieve some desired outcome. And the other era is just how much world knowledge is required versus how much tool calling ability is required and how should we balance between the two. Um so again new open AI models extensively trained on synthetic data probably large part reinforcement learning um probably very geared towards the money-making situations right now and not just pure intelligence anymore. And it seems like we flattened out on that front and uh AGI is not coming. Uh shout out to Jack Morris who has a really good thread on Twitter X whatever that's called now. I'll link it in the description uh where um he goes extensively into embedding data that comes out of the open source GPT models and actually showing that it has a very particular distribution. Um and from that uh that's just more evidence to the probable data composition and training paradigm that these models had. Um that's it. Um yeah, we're in the product era now. Welcome and uh we hope that the research community can find again something interesting to do. That's it. See you around. Bye-bye.