What If AI Could Build the Next AI Model | Professor of AI Explains

Vit Lyoshin (00:01.25)
Welcome Damian, thank you for your time today.

Damian Borth (00:04.696)
Hello, nice having me. Thanks. Thanks for the invitation.

Vit Lyoshin (00:08.302)
Sure. so today when we train AI models, seems like, we use a lot of data and a lot of compute power. Do you think we're approaching that limit of useful data from AI model trainings?

Damian Borth (00:28.622)
That's an interesting discussion. I mean, that's exactly the kind of you know problem that people mentioned that pre-training is over. That, you know, we're running out of data. We only have one internet to train our, you know, language models or you know agents from, which is very much can be true. we don't know, right? There are only some of the f you know, AI Frontier Labs that can prove that.

wrong or, you know, still say that there are still scaling laws, neuroscaling laws that are still working if you increase the amount of data. I think this is an interesting question because it explains a little bit why people are trying to move from the virtual world to the physical world with robotics and world models. That might be a solution to move those agents into something real, so they can infinitely dense train in the real world.

Vit Lyoshin (01:22.592)
I see, that's interesting. So when we train AI model, what actually happens internally on the background? How does it remember the things that it's learning?

Damian Borth (01:34.542)
Yeah, so training AI models, in particular, like all the AIs today is their neural networks, and neural networks they are built by you know connected neurons, and the connections between the neurons are stronger or weaker, depends on you know the training. And this strength is given in weights and biases. So parameters of the model are the weights. The big language models has trillion of weights, smaller billions. and during training.

We show the neural network a lot of data and we see how well this model predicts, for example, you know, left words in a sentence like, you know, the capital of France is, and then the language model says Paris. And if you do it with all the data on the internet, you know, the weights of this model, the parameters, are adjusted. And if you do it right, this knowledge of this entire data set is compressed in the weights of these neural networks.

This is a very expensive procedure. This takes time, takes memory, takes energy, compute. And if you did it once, you have this precious model that can now be used as a language model for other, you know, applications and downstream tasks.

Vit Lyoshin (02:49.132)
Yeah, I see. Okay. And you're working on something called weight space learning. Can you explain what that means?

Damian Borth (02:55.202)
Mm-hmm.

Yes, so this is a new interesting and exciting research area. If you think about what we actually did as a community over the last years, we had actually we took big models and trained them on the entire internet, in particular the words. So now you know those models can you know analyze words and generate words. Then people took big models and trained them on all the pixels on the internet, all the pictures. So now we can analyze pictures and generate pictures. Then

Other people took actually a big model and trained it on all the code of the internet, so programs, so we can analyze programs and synthesize programs. So what we are doing, we are playing the same game, but instead of words, pixels, or code, we are taking the weights of already trained neural networks. This is a new data modality that we process, and surprisingly.

There are a lot of neural networks on the internet that are open weights, open source that you can take for training. And the idea is once we train a very big model on all those trained neural networks that we can analyze and generate neural networks. So instead of having this very expensive training procedure, you would be able to prompt your neural networks on demand and get it instantly back from us. That's the idea of waste based learning. There's a lot of details and a lot of, you know, individual areas.

But that's the overall idea. Can we think of weights similarly as pixels, code, and words and train the model on that?

Vit Lyoshin (04:29.198)
Interesting. Do you have an example of that? Like maybe some use case of that?

Damian Borth (04:34.466)
Yeah, I mean one of the very first examples that people started to do is, you know, you you you take, for example, a couple of models, hundreds of models that are trained to do image classification. And you train an you know a big meta model, some people say, on that. And then, you know, this model understands what a good where a good model is, or what a good model is, what a bad model is. And you can take an unknown model that somebody gives you, read it, read the weights.

Vit Lyoshin (04:46.658)
Mm-hmm.

Damian Borth (05:03.564)
And the model can tell you, well, this neural network is a good one, or this neural network is a bad one. So you can actually infer the properties of this neural network. You can test the neural network without test data. So you can say this model has a 90% accuracy on maybe a data set like ImageNet, or it has a 20% accuracy or 30%. So you can analyze the weights. However, the more exciting use case, according to my opinion, is that you could generate a neural network. Think about

You want to have a particular neural network for your very specific data set. Think about your hospital and you have this, you know, COVID X-rays. Could you actually send it in a privacy-preserving way, this data set to somebody to as a prompt to generate the weights for a neural network that can then classify according to the X-ray if somebody has COVID or not? So can we instantly generate weights and kind of not having the need of having this expensive?

you know, backpropagation and optimization procedure that neural networks are trained nowadays. So analyzing weights, reading, or generating weights, writing. That's the idea.

Vit Lyoshin (06:13.134)
Interesting. So you can have... Basically like we had a of apps for specific purpose, you can have now neural networks for specific purpose. Is that kind of the idea?

Damian Borth (06:26.156)
Yes, yes. So you you could usually generate whatever neural network you want in on demand, use it, and you know, next time you need something else, you can generate another neural network that is purposely built for this particular task or data set, which not necessarily needs to be a big network, but maybe a specialized small network. So if you think about you would have an agent that is kind of trying to fulfill a task, this agent could use tools.

Vit Lyoshin (06:28.046)
Hmph.

Vit Lyoshin (06:46.68)
Mm-hmm.

Damian Borth (06:56.278)
And these tools could be dynamically generated neural networks or dynamically generated code, right? And then on purpose, on demand, get them created, solve the task, and you know, use the knowledge that is extracted to do whatever the agent's supposed to do. That's exactly the idea. You could think about this as you know, some kind of tooling layer, and we call it neural tooling layer that could be used in this particular downstream application.

Vit Lyoshin (07:20.226)
Mm-hmm.

Vit Lyoshin (07:23.8)
So I read about your research that you kind of proved how you retrained the neural network on another big one and it saved a lot of like GPU hours and things like that. Can you tell us a little bit that so we can understand better like the efficiency and savings by using this method?

Damian Borth (07:42.146)
Yeah. So there are two two different things how you could think about that. We had one paper that was published at Computer Vision and Pattern Recognition Conference, C V PR, where we had a very particular use case for remote sensing data. So satellite data and so-called foundational models that are able to analyze, you know, satellite data. and we now took our approach, trained it on the entire remote sensing.

neural networks that we could find on platforms like Hugging Face and you know had the pre-training on computer vision models and took this to generate foundation models in remote sensing. And this was interesting because the generated models, they are much, much cheaper in generating. They were comparable in all the state-of-the-art foundation models in remote sensing. For many downstream tasks are classification, segmentation,

object detection, land use, land cover, wildfire detection, floodings, et cetera. We could also, because we have a big model that understands neural networks, prompt the model in st for the architecture. So we could say, well, we want to have a very small model that is able to do the task, or we want to have a very big model to that is able to do it. So either a so-called VIT vision transformer, or maybe a very small model like an efficiency net or mobile net. And we

Showed in our paper empirically through experiments that the small models that we prompted were better than the big models distilled to small models. Which is actually pretty cool because it's like, you know, we have a big model that is understanding weights, and you prompt whatever you want and you get it, you do a little bit of fine-tuning, and suddenly you have your model ready. So we could generate weights and we could distill weights. And there's also work that can speed up optimization by.

Well, I have those five checkpoints. And now please predict checkpoint number 10. So you kind of jump ahead in transfer learning or fine-tuning. For the numbers, the C VPR paper showed that we can be we can train comparable remote sensing models with 20 times less compute needed as compared to state of the art models trained from scratch. Which is which is pretty nice. And we we just coincidentally stumbled upon that because of some publications.

Vit Lyoshin (09:53.89)
Wow, that's amazing, so...

Damian Borth (10:01.304)
And then we compare to our results and we're still like, that's a lot. So that that was pretty nice, nice results, yeah. Doesn't doesn't happen often in research, but you know, once in a while you have those great moments.

Vit Lyoshin (10:04.792)
Hmm.

Vit Lyoshin (10:13.282)
Yeah, that's great that we already have those foundational models there and we can now, like you said, on demand generating like specific expert neural networks for tasks and things like that. And especially with improvements in like efficiency and speed and cost for running them, we could end up running them on our mobile phones essentially. So that would be pretty cool.

Damian Borth (10:37.656)
Yeah. Yeah, or think about if the entire robotic movement that we see, all this robotic startups, if they are all building robots, they will need to have small neural networks. Because a robot has a battery and this battery and the energy and this battery will be used for the arms and legs to move the robot and should not be used for the compute. So they need efficient small neural networks. That's a real problem, right? And if there will be millions of robots out there, they will need millions of neural networks.

Vit Lyoshin (10:46.158)
Mm-hmm.

Vit Lyoshin (10:50.744)
Mm-hmm.

Damian Borth (11:07.832)
To move their arms and their heads and you know all the actuators. So that's actually a pretty interesting niche that people are starting now to understand how important it is.

Vit Lyoshin (11:16.694)
Yeah, yeah. What are some of the limitations of WaySpace learning approach?

Damian Borth (11:23.574)
Mm-hmm. So there's a great community around that. We started like four years ago, five years ago working on that, and more and more people kind of joined from different directions. There is there are some people who are focusing on analyzing weights, some people on generating weights. They're different approaches, some approaches can do both. In the very early days, the biggest thing that was the biggest challenge was scaling. So we could only generate like small neural networks, like a couple of thousands of parameters, maybe you know up to almost one million.

Which was good enough for some applications like in robotics, reinforcement learning, or some, you know, implicit neural representations, INRs. But to be really considered a general purpose approach, we would need to scale to billions of parameters. And with some developments over the last two years, people from the community, including us, were able to scale up to one billion and more. Always depends like do you aim to generate the entire neural network at once?

Or are you editing parts of the neural networks that are relevant? And there are different setups that you could think about what's useful as an application or whatnot. But the scaling is the very big challenge. And the second thing is the weight spaces, so the neural networks, all the weights, they have very particular and interesting, you know, geometric characteristics. There are a lot of symmetries and a lot of permutations that, you know, have different points in this weightspace for the same neural network. So dealing with that.

By learning this end to end is a huge challenge. And you have to think about equivariant architectures that are kind of smoothing out this permutations and symmetries. So there's a lot of things to do that is special to weights as a modality as compared to images or words or code. but progress is also made in this direction. So limitations are there, but they are there to be conquered, right?

Vit Lyoshin (13:00.91)
Mm-hmm.

Vit Lyoshin (13:16.526)
Yeah, yeah, I guess as the community grows and people learn more about it, there will be some breakthroughs for sure. It's like you said, four or five years and it's already a progress. So that's great. Yeah. So I want to switch to the recent MIT research that came out or a white paper that came out on November 2nd, which is like a few days ago.

Damian Borth (13:40.558)
Mm-hmm.

Vit Lyoshin (13:43.188)
and they have something called weight space programming. Can you tell us a little bit what that one means?

Damian Borth (13:48.227)
Mm-hmm.

Yeah, that's actually a pretty interesting new setup that we observed. So what happened is that there's a research group at MIT that discovered for the first time that you can use one big neural network, proprietary one, I mean Astra, you know GPT six Astra, and you can use this to modify or edit the weights of another neural network, in this case Gwen and some others. So you could

rewrite or reprogram these weights by this neural network. And they called it weights based programming. So it's you know their their phrase and their terminology. And it's kind of given that you know momentum and given that announcement, I was personally surprised how much hype was behind it and how how much visibility and tension it got. It was everywhere.

So people are now super excited about one neural network is able to edit or change the weights of another one and to improve it significantly. So that was really, really important because we had now, as the authors claim, up to six-fold improvement in you know math tasks. If you look into it, you will see that there is some type of, you know, it's a it's a blog post, there's no paper, so it's not kind of went didn't went through peer reviewing.

And you know, when you look into the details of the experiments and you know how the prompts were looking like, there's a lot of discussion now how much of this is really able to generalize to other tasks, how much is this specific to this particular task? And there are some, you know, people who kind have you know questions about the experimental setups. The one thing, however, that happened is what happened is that this weightspace programming went really, really viral on social media and now everybody is excited about it.

Damian Borth (15:47.884)
And I see it as one area in the big weight space learning domain where we don't generate the entire network, but edit parts of the network, which actually happened. Right? So it's actually amazing. And I'm really excited that we see that because people are now getting the idea that weights can be treated as data that can be read in and read out. So it's a new data modality. And we have the capability with newer networks.

Vit Lyoshin (15:48.174)
You

Vit Lyoshin (15:59.427)
Yeah.

Vit Lyoshin (16:13.76)
Mm-hmm.

Damian Borth (16:17.47)
Modify them. In this case, a proprietary, you know, Frontier AI lab neural network. In our case, specially trained neural networks that are reading and writing those weights. So it's actually a great time for weights based learning. And I'm super, super excited because of the community is getting now the attention. and then let's see what will happen like in three or six months, right, with this moment.

Vit Lyoshin (16:41.516)
yeah, for sure. Everything develops so fast. But my question would be how does it know which weights to modify and how based on what?

Damian Borth (16:54.882)
That's a good question. So what they did, they they had eight different experiments, you know, with different tasks. One of the tasks was to, you know, improve this you know, Gwen model in the MAFT task. There was some other task where they you know had little reinforcement learning game playing things. the autos had designed particular prompts and with those prompts they asked you know on a high level the Astra model to

Vit Lyoshin (17:12.077)
Mm-hmm.

Damian Borth (17:24.702)
repair or modify or improve the neural the other neural network they were reading in, you know, given the task. And then the neural network did its operation. And the operation was very little. So it changed like less than one percent of the weights and as an outcome, as the authors claim, was the improvement in those particular tasks or you know what they proposed. So the experimental setup is a little bit specific. So

Can be improved and benchmarking is very important in this area. We need proper benchmarking to make those things comparable. But it seems, and this is the interesting thing, it seems that the latest, you know, announced models from OpenAI and Entropic they have capabilities that the previous models didn't have because this we only observed at the latest models. So one question is: maybe they're also trained with weights, maybe they saw weights before.

Vit Lyoshin (18:14.914)
Mm-hmm.

Damian Borth (18:20.054)
Or maybe they have some emerging capabilities that we did not observe before, given their size or given their training, whatsoever. So that's an interesting question. Since they are proprietary and we don't have access to them, we don't know why they are so good at doing that. But it's exciting to see that they are, you know, able to do that. The only challenge and limitation of this approach is if because they're only changed a little amount of weights in these big models. So if we would like to generate the entire model,

Vit Lyoshin (18:35.714)
Mm-hmm.

Vit Lyoshin (18:46.52)
Mm-hmm.

Damian Borth (18:50.05)
with such a front AI la model like Asta would probably need to pay more money than training it from scratch. So it's it's it's not scalable at all. The the challenge really the scalability, you know it's it's kind of an in context learning. So there are some examples that are given in the prompt and it's a very complex prompt. So it's nice to modify those small neural networks or parts of a big neural network. It would not scale to larger networks and it would not be on the long term you know

Vit Lyoshin (19:00.206)
Mm-hmm.

Damian Borth (19:20.342)
sustainable because it's too expensive and doesn't make sense. But exciting to see that and exciting to see that momentum. While you know yeah and and when you think about that, you could now use this ideas of having a proprietary neural network like Astra, but instead of working directly on the weights that are very, you know, large where you take a big model, you could think about using approaches that the community already developed since two thousand twenty one.

Vit Lyoshin (19:26.904)
Yeah, interesting.

Thank

Damian Borth (19:50.562)
the weights based learning community in compact representation of weights that are much, much smaller and use this with such a model like Astra and then see what happens. Like think about you could ask your language model questions like, I have this neural network, can you improve this by 10%? What can I do with that? Or what should I do? Or I have this neural network, you know, what's the I know it's working on this data set very well. What do you think? Will it work also on this data set? Is it worth to do transfer learning?

Vit Lyoshin (20:02.114)
Yeah.

Damian Borth (20:20.758)
If we could actually, you know, talk or have a discussion about neural networks, and they then the language model could propose extensions or improvements, or really like also do the improvements, change the weights, I think this would be really, really amazing because then everybody could wipe train a neural network, right? I'm not sure if this is the right term, but if wipe coding, this would be wipe training with all the pros and cons, right? Not all wipe coded code is nice. But

Vit Lyoshin (20:34.989)
Mm-hmm.

Vit Lyoshin (20:42.327)
You

Vit Lyoshin (20:47.939)
Yeah.

Damian Borth (20:49.504)
It would allow access to everybody who is interested in neural networks to generate them or modify them on on behalf of whatever idea they have. And this is exciting, right? So so well I we believe or I believe that n machine learning and neural network will experience the same transformation as programming did. In programming we did coding and now we do code synthesis. And in machine learning we did training and we will do model synthesis.

Vit Lyoshin (21:01.73)
Yeah, ups.

Damian Borth (21:19.34)
of those weights. I it there should be no reason people should start training from scratch with the techniques and, you know, knowledge that we have now. and this actually opens up huge opportunities for many, many applications and exciting, you know, whatever people will think about it once you give it to many, many people. There will be crazy ideas coming. And I think it's a it's a it's a great time to be in this area and doing weight space learning.

Vit Lyoshin (21:42.68)
Yeah, yeah. I mean, I'm not really technical that much, but I already can think of a bunch of use cases where it's applicable and where you can tweak your smaller models for very specific tiny use cases for your organization, for example. Now everybody, like before we had racks where you had still big library of documents and LLMs seeped through those to figure out how to help you better. With this now...

you can have, you don't have to have all these documents. You essentially just train your mini model, if I can say that, and it will be an expert on your organization, on your finance, on your customer support or whatnot. think, yeah, I think that's amazing. But another thing I read from that research is they could actually ask a model what it was trained on.

and it was correctly identifying exactly what it was trained on. Can you talk about that? How is it possible to do that just by looking at the weights?

Damian Borth (22:38.211)
Yeah.

Damian Borth (22:47.362)
Yeah. So there are many use cases where you generate weights for a new neural network that has new capabilities, right? So that's the weight generation part. We can also think about weight analysis, right? So I get some weights. I don't know where they're coming from. You know, some neural network I find online, I don't know how it's trained, I don't know where it's coming from, what is the data. And I could actually ask questions about this neural network, like, you know, w what type of data was it trained with? How

Vit Lyoshin (22:54.456)
Mm-hmm.

Damian Borth (23:16.374)
Well, does it perform what's the accuracy? What kind of trouble will I get? Is there a risk of having you know adversarial examples destroying the neural network? Are there some backdoors trained in, right? So the weight analysis part is, I think, equally important as the weight generation, but doesn't have this immediate translation into reducing training time and cost. But it has very interesting capabilities if you want to think about.

Vit Lyoshin (23:29.976)
Mm-hmm.

Damian Borth (23:45.774)
Can I give guarantees or boundaries? It's a very strong word, but can I maybe ask the neural networks about its properties to understand better how well it behaves, where it comes before I use it? And you could even think about very interesting applications and people are doing that lineage prediction, right? I have a neural network, I could ask, you know, which neural network which other neural network was used as pre-training for your fine-tuning?

Vit Lyoshin (24:14.701)
Mmm.

Damian Borth (24:14.722)
That's but because you know you are very often you have this one base model and then a lot of neural networks are fine-tuned from that. And it would be interesting to kind of be able to build up this lineage graph. You know, Ilya Horowitz is doing that and others to understand intellectual property traces or understand if, for example, some networks that had some base models that had trouble and now people fine-tuned if you know we should check for the same trouble the base model had.

So building this understanding is equally important because to be honest, neural networks will be everywhere and they will have trouble. They will have they were never 100% correct. And if you think about the entire cybersecurity situation we had a couple of weeks ago with you know neural networks hacking others, we will probably need also something like you know, traceability of neural networks. And and if you zoom out and you understand

Code creates programs. You could think about neural networks as trained programs instead of coded programs. And similar to regular programs, we have, you know, malicious software. So we need antivirus software. You could think about you will have malicious neural networks. You need antivirus. I'm not sure if virus is the right term, but you would need similar counterparts to this. It's an interesting direction. And another area that is, you know, additionally to generation, programming, editing.

Vit Lyoshin (25:25.038)
Vit Lyoshin (25:34.723)
you

Damian Borth (25:41.782)
is kind of this lineage analysis, fingerprinting, to to understand where neural networks came coming from and you know which parts are used where. we will need to do this, as we do with software, we will need to do this with you know trained software and neural networks.

Vit Lyoshin (26:00.268)
Yeah, I think this is huge because I had a couple people before on the podcast from highly regulated industries and they were saying that before we can deploy our software to like a bank or hospital or whatever, have to, or government, we have to submit the whole set of documentation, how it was trained, what we did and blah, blah. And now you can have like this, let's call it analysis or whatever neural network that can run and tell you exactly.

Damian Borth (26:26.702)
Mm-hmm.

Vit Lyoshin (26:29.922)
what it's there, what biases, what errors, what gaps, all of that. And then you can use either the same or other one to fine tune it and specifically correct all these mistakes if possible. That would be amazing. I think we are entering the whole new century of AI.

Damian Borth (26:43.224)
Yeah, yeah, yeah.

Damian Borth (26:48.982)
No, s seriously, you you described very nicely the future of in in what is potentially possible. We're not there yet. There's a lot of work happening and there are no first steps and the community is really pushing forward. But think about you will get in your network, maybe from a supplier, a vendor, and you are asked, you know, and they they have supplied you with some metadata about it, what it was trained for, you could double check, right? You could just check is this true? You could find maybe some, you know, preemptively before you use it, some

Vit Lyoshin (26:57.816)
Mm-hmm.

Damian Borth (27:18.094)
trouble, some errors, you could ask to repair it, to adjust it to your specific downstream task, to your specific data set, all in recursive loops. There are a lot of startups that think about recursive AI and you know self-improving AI. That could be one way of doing it. There's a lot of, you know, startups that are doing certification and, you know, moving into that space could be also used for that. And there's a lot of, you know, potential in generating on demand weights to make those very specific neural networks

Vit Lyoshin (27:39.128)
Mm-hmm.

Damian Borth (27:47.616)
instantly and not use a lot of computer train. So if you think about weight space learning, and this is why I'm so excited about that, it's not one particular vertical for one industry. It's a horizontal where everybody could benefit, independently if you're in finance, banking, or you're in insurance, or you're in chemical, or you're in, you know, whatever. Because neural networks will be part of every software in the future. And we need tools to read, write, analyze or generate.

Vit Lyoshin (28:02.252)
Mm-hmm.

Vit Lyoshin (28:18.2)
Mm-hmm. Yeah.

Damian Borth (28:18.59)
and and we we we see a glimpse of in which direction that might go. I I hope that a lot of researchers will be now excited. There's a lot of challenges ahead and limitations. And now hopefully the community will grow and push forward very hard because it is a new modality. And again, if you think about you use weights as input data, you could think about one of the biggest platforms where neural networks live is Hugging Face. There are up to three million plus neural networks.

Vit Lyoshin (28:44.674)
Mm-hmm.

Damian Borth (28:48.066)
When you count the parameters, there are billions of parameters, and when you tokenize them in tokens, you have roughly two to four times more data than GPT four was trained with, according to estimates. And nobody's using that until now, right? So the it's it's a great new type of data set that would enable great new applications. That's that's something where we're just starting to see in which directions we could go.

Vit Lyoshin (29:02.254)
Wow.

Vit Lyoshin (29:07.278)
Mm-hmm.

Damian Borth (29:18.082)
So they're gonna be a I think a new you know a new advent of this type of research and a gold rush in that direction.

Vit Lyoshin (29:27.01)
Yeah, it's like an archive, right? Like you found some ancient library or whatever and whole new set of knowledges showing up. And it's actually less or compressed kind of, you don't have to run through the whole internet multiple times.

Damian Borth (29:40.622)
Exactly. And you know the the best part about this data set? Somebody else paid for it to create it.

Vit Lyoshin (29:49.103)
Yeah, exactly, you're already getting it pre-trained basically for free.

Damian Borth (29:52.878)
Yeah, yeah. I mean it it's very similar setup that when you compare language models trained on the entire internet, right? There's also content that somebody else created. Open, not open, you could argue about that. If you know, you take all the models on Hugging Face that are open weights according to the right license and you know it's creative common and you learn from that and you have a derived artifact, a new model that is able to understand them, very similar setup. Yeah. But new

Vit Lyoshin (30:03.256)
Mm-hmm.

Damian Borth (30:21.346)
Different and more exciting applications that you could go for.

Vit Lyoshin (30:25.708)
Yeah, what kind of concerns should we have with this approach and like caution people maybe like, I don't know, some crazy things like models training themselves on certain things and people missing what they capable of now or some like risks, maybe holes. What can you tell us about that?

Damian Borth (30:41.603)
Yeah.

Damian Borth (30:46.318)
Yeah. Yeah, I mean the the the whole recursive self-improvement idea so that models are building better models is great if you think about you know we having the oversight and the big red button to push. It could be a little bit concerning if those models are doing it without our or without us humans giving the instructions for. So that's always the question for every technology.

How are we governed this technology? Like on a governance level? How are we building technical tools that are monitoring it? And how are we able to actually steer it and make those things visible? And this is with neural networks specifically difficult because they are high-dimensional internal representations, high to or difficult to have an interpretation on that. So this will also demand new tools of

understanding those neural networks, monitoring those neural networks, as we have now no similar challenges that we have with agents that are going out to the internet and starting to you know hack other companies. So we need sandboxes, we need environments and we need a new level of safety tools to be developed. And at the end of the day we will have probably an entire ecosystem around weights being read and written and then generated or analyzed.

And you could think about as a new industry beyond coding, beyond pixels, beyond you know, language. and the space is there, the data is there. We need people who start working on that, the community is doing that. So, the only question that we have as a community is, according to my opinion, are the neuroscaling laws working on weights as they work on text, images, and code. And the neuroscaling laws are very simple.

Vit Lyoshin (32:13.399)
Mm-hmm.

Damian Borth (32:42.946)
The bigger you model, the more data you have, the better the performance. So can we scale this up to some size that we will see emerging properties and then maybe accelerate the entire neural network world, you know, tenfold, hundredfold? That's I think really, really exciting promises with all the risks, right? So the tooling, the ecosystem, but that's where it goes.

Vit Lyoshin (32:47.736)
Mm-hmm.

Vit Lyoshin (33:07.852)
Yeah, and I think it also brings a question of like for people like engineers, computer science, what kind of skills people now need in order to work with this because it seems like it's a little bit a new concept and maybe some people need to their head around it and like what kind of skills do I need now to be able to work in this field.

Damian Borth (33:25.091)
Mm.

Damian Borth (33:30.901)
Mm-hmm. Yeah. So like when you're very close to research in this field, you probably need, you know, you will definitely need a machine learning background, a background in neural networks, background in training neural networks, and working with neural networks, obviously. But the very moment we are giving these tools to other people who are, you know, now able to generate neural networks according to their specification, it will be a similar, I believe, a similar situation like we have with vibe coding, right? You you would have the situation where some people are

Vit Lyoshin (33:42.286)
Mm-hmm.

Damian Borth (34:01.076)
more experience with it and some people are less. So there are usually people who did software engineering, they know how to specify software, so they can be faster with wipe coding. And I think similar things will s be seen when, you know, people who did machine learning and trained neural networks, they will be faster in generating that. And people who will be not knowledgeable about it, they will need a little bit more time, but there will be no road blocker. That's important, right? So they could you know also do exciting things. So that's when we think about humans working

With those tools to generate neural networks or analyze. But think about agents using it. That's a machine to machine communication. W that that's a qu what would happen if you would allow an agent to have access to such a tool and this agent could generate neural networks on demand? That's an interesting question because maybe agent an agent would have a task to accomplish and the challenge would be that they don't have the right tool. And if we would be allowing the agent to generate the right tool,

it might be exactly the missing piece to accomplish this task and open up a new avenue of capabilities. And then suddenly you will probably see a lot of agents doing a lot of interesting things in combination with other neural networks and other agents. So this is something that I would not be able to kind of predict what kind of behavior would emerge, but this could be really, really exciting to see. And obviously, you know, with all the, you know, boundaries and ray guards and and

In place and sandboxes.

Vit Lyoshin (35:32.94)
Yeah, yeah, yeah, think agents already kind of capable using multiple models based on the task. And also there is a concept of skills, how you set it up for the agents, giving them some instructions in English. But instead of giving instructions in English, you can give them instructions in the form of neural network. And maybe that will improve certain capabilities.

Damian Borth (35:54.669)
Vit Lyoshin (35:57.976)
So I don't know, yeah, the sky is the limit, I guess, and there are a of smart people who will figure it out. Yeah.

Damian Borth (36:03.18)
Yeah. Yeah. No, no, really. And I mean, if you think about the most important thing about this entire technology, about weights based learning is can we make this very expensive, iterative optimization that we call backpropagation much more efficient? If you think about our entire the entire AI industry is built upon the belief that AI training or training of neural networks is expensive and takes time.

If we could put this upside down, I'm not saying that this is now possible, but as a vision, if we could put this upside down and you could generate on demand your neural network instantly, this would open up, you know, very different applications or type of application, real-time applications, and you know, reduce the compute or the amount of compute that we need, save energy, save CO2, right? And and this would be in general much, much better. And

I think this is a very important, you know, thought beside all those great applications. We are currently in a regime or setup in the AI industry where we do large scale training at all costs. And only few can afford it. And therefore we need all those data centers and all those things. What would be actually happening if we could make this much, much easier and much, much less computation intensive?

Vit Lyoshin (37:15.266)
Mm-hmm.

Damian Borth (37:27.896)
That that's that's the the motivation that we have here. And this is kind of the goal that we are trying to achieve in the next couple of years with the research on the one side, with all the challenges that we have in the details, but also with you know applications and you know demos and systems that we can put in place so others can use it.

Vit Lyoshin (37:44.177)
Vit Lyoshin (37:51.404)
Yeah, I have maybe a strange idea or question. I just want to see your opinion on this. When we train AI models, we usually use some data that already exists originally, right? Now we're going to train neural networks on another neural network, but it is also kind of already predefined data that existed before. Is it possible to teach them to generate new data?

Damian Borth (37:57.706)
Mm-hmm.

Damian Borth (38:06.68)
Mm-hmm.

Vit Lyoshin (38:20.64)
Like we have some ideas for synthetic data when we train in or AI model world models that like simulate different situations and things like that. But what about actually creating a whole bunch of new like data for either training or whatever. Like a completely new something, new inventions by the by this neural networks.

Damian Borth (38:29.699)
Mm-hmm.

Damian Borth (38:43.074)
Mm-hmm. Yeah, that that's interesting. So if you think about that, we have we have in general, think about images or text, right? If you if you have language models that are very capable and they generate text, text, then more and more of the internet will be synthesized and not human generated. And then we have this trouble that when you train on this content, there's just kind of the you know it's it's gonna be worse and worse and worse. Because you know, it's it's it's it's kind of always this not not the same, but you know.

the new original thought, the new original content is missing. Pixels, there's this, you know, also observations that when you train on synthetic data, you will not be able to reach the performance as compared to real data. So the the question might be like if we are now starting to synthesize and generate neural networks, will be will we kind of also center all the generated neural networks around what was there and making it kind of worse, worse and worse.

Vit Lyoshin (39:15.917)
Mm-hmm.

Damian Borth (39:39.896)
Because we are kind of generate and we train and generate and train on that. That's an interesting thought. The hope that I have is that comparing the neural networks that are online, you know, they're trained on you know standard data sets, they're trained on some particular subset of the world. When we think about there will be now tools to make neural networks to generate those neural networks easier and cheaper and faster, people will use different types of data.

train and generate those neural networks as compared to the ones that are online. So through that we might get you know access to data silos that usually would not be accessible because nobody would have thought about training a neural network on that because it's too expensive, you don't have the GPUs and so that might kind of unlock new type of data that is suddenly available in form of the weights of this neural network. And that's kind of an in and then then you would need to kind of figure out how much is

Vit Lyoshin (40:11.746)
Mm-hmm.

Vit Lyoshin (40:31.054)
Mm-hmm, okay.

Damian Borth (40:35.714)
the synthesis of the way it's harming and how much is the novelty of the data it was used helping? I think that would be an interesting question interesting research question, yeah. Never thought about that.

Vit Lyoshin (40:43.734)
Yeah.

Vit Lyoshin (40:47.638)
Yeah, yeah, I also wonder how much data is offline, right? We know, especially scientists, they have a bunch of papers, bunch of books, never released, nobody saw it. And that's called that it's not there yet. And if people can have their own networks, trained on their stuff, and then it somehow, I don't want to say leaks.

Damian Borth (40:53.144)
yeah. Yeah.

Vit Lyoshin (41:12.248)
but maybe it's a conscious decision to let it out for other people to use because you don't really give your papers, but you're giving them trained material on those papers, expertise, that's maybe how we get those gold that's still left offline.

Damian Borth (41:23.276)
Yeah. Yeah.

Damian Borth (41:30.444)
Yeah, yeah. No, I fully agree. There's the I think the majority of data is actually hidden. It's not on the internet. There's a lot of corporate data, there's a lot of you know medical data, there's a lot of scientific data, and this is like v valuable. It's maybe too expensive to put it all online or there are privacy concerns, etcetera, for all good reasons. If you think about you train a neural network on this type of data, you can think about the neural network as a compressed version of the entire data set, but not individual.

Vit Lyoshin (41:37.036)
Mm-hmm.

Vit Lyoshin (41:40.653)
Right?

Damian Borth (42:00.28)
So you could think about this neural network, you could share so others could learn from that without leaking individual documents that might be concerning regarding privacy. There are attacks where you can still retrieve individuals, and the entire area is called federated learning or privacy preserving. There's differential privacy. We could think about those approaches in this context too. So, how could companies on purpose share their data, maybe on a marketplace in form of a neural network?

without leaking their individual samples. And think about where we go now. We go now into robotics. We go now into engineering, we go so all these engineering labs that are doing engineering, electric engineering, construction, right? They could train little neural networks, you know, give them out for others to learn from that without leaking their information. So maybe there could be an kind of an economy around that or a marketplace structure around that emerge. So they could all kind of benefit from that. There's one idea.

Vit Lyoshin (42:32.781)
huh.

Damian Borth (42:59.98)
And you know, we could see if this is then beneficial for those people who have those archives to give this data in an aggregated, compressed way to the public, being compensated monetarily, but not giving everything out. I think that's an interesting thing because giving everything out would be kind of difficult as an inertia to do this, but giving it in a compressed form maybe is good enough and still being monetized. So you could think about

Vit Lyoshin (43:16.6)
Mm-hmm.

Damian Borth (43:29.216)
All those companies who have all this data, how could they create now secondary revenue streams of this type of data sharing learning agreements? It's kinda it's kind of interesting if something like a marketplace or an economy around this could emerge.

Vit Lyoshin (43:34.562)
too.

Vit Lyoshin (43:44.012)
Yeah, yeah, that's a nice use case for the future. think it's, you know, somebody have to try it, but I think it's an opportunity for sure, because that's how like you hide basically behind the wall kind of, but still spilling some information, some of your findings or whatever it is. Yeah, that would be nice to see.

Damian Borth (44:03.896)
Yeah. I mean there are use cases when people share things and everybody shares things and then somebody learns from that in a privacy preserving way that everybody gets better. In like think about in risk scenarios, compliance scenarios, cybersecurity. If everybody would share that data and we could learn a model to prevent it and give it back to the people, everybody would be more safe, more, you know, compliant, etcetera. There are use cases where it totally makes sense to do this. There may be use cases where

Vit Lyoshin (44:14.765)
Yeah.

Vit Lyoshin (44:27.758)
Mm-hmm.

Damian Borth (44:32.406)
your entire advantage competitively is the data and the models and you would not not do it. But in both cases you could make a conscious decision as a company if you want to share it, monetize it or not, or you know, give it away as a benefit to other people. If you could do this for you know medical data, so we could learn and train better models to, you know, help us to generate proteins that are then used in form of drugs.

We could all benefit from that. That would be the healthcare environment. But there are also other environments like you know, national banks in Europe, they do fraud detection and money laundry and things like that, where when we had this type of data aggregated in neural networks, it would be beneficial to train on that. And so there are different scenarios that totally make sense in this context. the the underlying assumption is that you know, where we are now in two thousand twenty six with weight space learning and this, you know.

Vit Lyoshin (45:22.766)
Yeah.

Damian Borth (45:30.37)
great momentum around weight space programming and where we want to go and what needs to be actually done in between and to show in the scaling laws and to really, you know, move forward. And I think one of the most important next steps, what we would need to do is we need to build a foundation model of neural networks. If we have a foundation model of images or a foundation model of chemistry or foundation model of robotics, we would need to build a foundation model of neural networks.

And give it to people so they could use it and think about their own downstream tasks to fine tune. This would enable a lot of opportunity and good could really kind of you know accelerate the next stage in this entire direction of or research direction of weights based learning. So it's time for doing that.

Vit Lyoshin (46:06.338)
Ugh.

Vit Lyoshin (46:18.988)
yeah, it sounds like a new internet basically. It has all the knowledge from everywhere and people can use it.

Damian Borth (46:28.628)
assuming it's done right, assuming the model is big, scaling laws are working and somebody has enough compute to do this, right?

Vit Lyoshin (46:37.73)
Yeah, right. Yeah, maybe we need to motivate Big Tech to invest in this somehow. All governments.

Damian Borth (46:44.664)
Yeah, well I I I think it's an exciting area for startups in this direction. And companies, you know, that are now taking the money, investing to prove that the neuroscaling laws are working on weights and then having the opportunity to occupy this very new space because it's it's a white spot. Other companies are occupying the language, the pixel, the robotics, the world model space, and this is a new spot. So, you know, this is something that excites me very much because we need to be fast and we need to show this.

Vit Lyoshin (46:49.549)
Mm-hmm.

Vit Lyoshin (47:07.777)
Uh-huh.

Damian Borth (47:14.614)
And then see what people will do with it.

Vit Lyoshin (47:18.54)
Yeah, yeah, that's great. So to wrap up here, what would be your message to people listening and engineers and organization leaders? Where should they get more information about wait space learning, wait space programming in general, how to move toward this thinking? What would you say to them?

Damian Borth (47:20.366)
Yeah.

Damian Borth (47:34.914)
Mm-hmm.

Damian Borth (47:40.736)
Mm-hmm. Yeah. No, I I mean my message is this is a underexplored area and one of the most exciting opportunities that I see in AI for the next five to ten years. so you should be there today if you want to join that. Waitspace learning is an exciting area. There's a lot of material and a lot of research over the last five years. If you are at NERIPS, one of our major conferences, we are organizing a workshop.

About weightspace learning is called neural artifacts as a new data modality. There will be the community that's exciting. And if you want to start working with something like weightspace learning, we have a startup called Weightspace Labs, and we are aiming to release the first foundation model on neural networks by the end of this year, and then accelerate the research. So there are multiple options. Stay tuned, look for the space. I think there are gonna be enough, you know, momentum.

what happened this week with the weights based programming you know announcements that we saw from MIT and I hope that, you know, more people will be excited about weights because it's an a new data modality and we should learn from all that knowledge that is already there, right? There's no reason to not do it.

Vit Lyoshin (48:56.406)
Yeah, yeah, absolutely. 100%. Great. Thank you, Damian, for information. This is amazing. This is something really exciting. But before you go, I have three quick questions. This is my typical innovation Q &A round. So first question, how do you define innovation in a few words?

Damian Borth (48:58.324)
Ha ha ha.

Damian Borth (49:07.842)
Mm-hmm.

Damian Borth (49:17.004)
Innovation is something that was not there before, might be looking at the first sight to be not making sense, but at the end will change drastically the underlying system if it's companies, organizations or other things.

Vit Lyoshin (49:34.776)
Second question is which innovation in the history of humans changed the world the most?

Damian Borth (49:42.538)
Ooh, that's difficult. I would have mentioned first, you know, internet, but then I figured out computers. But what I think now is electricity. I think electricity was one of the most exciting I mean it's obvious, but if you look at the history of electricity from the very first phenomena to the understanding of the theory, the applications, the commercialization, the engineering with generators and companies that power plants, this took like two hundred years.

Vit Lyoshin (49:53.568)
Mm-hmm.

Damian Borth (50:10.914)
We have a similar setup now with AI, but we have to do it all in ten years. That's a little bit that's a little bit the challenge as a human, right? We we're the the the velocity is amazing. So electricity.

Vit Lyoshin (50:19.381)
Yeah.

Vit Lyoshin (50:23.35)
Okay, good.

Damian Borth (50:24.32)
It was magic and now we understand it, right?

Vit Lyoshin (50:27.372)
Yeah, exactly. And the last one is which thing or tool or software that we use today you think we will be laughing at 10 years from now.

Damian Borth (50:41.134)
that's interesting. That's really interesting. I think

Damian Borth (50:50.85)
So we will be laughing that we have used it. I would I would say from probably from that's that's super difficult. No. I think we would the editor right where we code. So I think this will be something that will change drastically.

Vit Lyoshin (51:03.278)
You

Damian Borth (51:17.454)
Code will not disappear and software engineering will not disappear. But the editor as we know it, the integrated development environment, will change. As similarly as moving from assembly code or machine code to programming languages to object orientation. I think this will not laugh, but probably we will be happy equally as we are happy today that there are no memory leaks as we used coding in C.

Vit Lyoshin (51:44.064)
Okay, sounds good.

Damian Borth (51:44.558)
Se secfold, secfold, right? And then you're like, Okay, great.

Vit Lyoshin (51:50.784)
Okay, no, that's great. I'm sure a lot of engineers looking forward to make life easier for them. yeah. All right. Well, thank you very much for your time today. It's been great. So we'll stay in touch and I'll watch out for more news in this area.

Damian Borth (51:58.175)
Exactly.

Damian Borth (52:08.088)
Thank you very much.

Vit Lyoshin (52:10.53)
Bye.

Damian Borth (52:12.248)
Bye bye.