Nomic: Truly open AI

Yeah, so hey everyone, my name is Brandon. I am the co-founder and CEO here at Nomic AI.

And I really got started with Nomic just based on some experiences that I had at several previous roles. I had been a machine learning engineer in several different verticals, be it robotics, defense, medicine, finance. And I had started seeing the same problems over and over again, a lot of the issues revolving around data quality and how that affects the downstream properties of the models. And when you're working in these domains like defense and medicine and finance, where it's super high impact and you really have a vested interest in your models being right, you start to learn very quickly that a lot of the utility that these models can provide is unlocked if you spend a lot of time really curating your datasets.

And so, the reason that I ended up deciding to start Nomic with my co-founder, who actually worked with me at the previous role at Rad AI, which is a generative medical startup that was shipping Transformers on radiology text back in 2018, was because we saw how impactful really upping the data quality in these products could be. And so we started that journey together and then were fortunate enough to run into Zach through some open-source work. And I'll let him tell that story.

We have all this data for it, but I don't have the time.

I was a little bit bored.

And I think I joined Nomic, I want to say, like, three or four weeks later.

This is my co-founder, Andrej. Yeah, yeah. I actually, funny enough, also dropped out of a PhD to join the radiology startup where we met. So, we like to joke that, yeah.

Well, that's exactly it. The two big factors for me, and this is going to be a tangent now, but the two big factors for me were velocity and impact, really. When I ended up leaving Hopkins to join Rad AI, the reason that happened is one of the co-founders of Rad AI sat me down and was like, "Look, I'm sure your thesis is great and all four people that know what a latent structure model of a random dot-product graph is are going to love it."

But if you want to actually come here and ship code that is going to help radiologists and be deployed in clinical practice a week later, now is the time and we have the GPUs. And that's pretty compelling. And even nowadays, it's increasingly hard to get those GPUs in academic positions. And so it's making increasing amounts of sense for, I think, the flight that we're seeing from academia to industry.

Andrej.

I mean, that's the model that Zach was talking about training. And it's also funny how everything can look like an overnight success from the outside. But I think a key part of what made GPT4All so good is we had spent the last year building Atlas, and it was the first time we really applied Atlas to the task of cleaning training data for a model. And so within a weekend, we could go in and, if you look at the original GPT4All paper, we show, oh, here's how you identify the sections of the training data where the model's refusing to respond.

And, oh, here's where you identify the parts of the training data where the model's giving one-word answers. And obviously you want to remove those so the model that you're training doesn't learn those behaviors. And so it's one of those things which I think is so true in startups, where the effort-payoff curve is super nonlinear. It's very much like a series of step functions where you put in all this effort, and then only when you pass some threshold, your payoff shoots up.

Fortune favors the prepared, is what I'm saying.

Or the model? Yeah, so it started as basically just a reaction to GPT-4 becoming increasingly closed source. We, as a company full of people that basically met through open source and have been open-source contributors through our careers, were pretty frustrated with the fact that this open lab was like, yeah, we're not going to tell you anything about the data. We're not going to tell you anything about the training. And we'll get into this later when we're talking about benchmarking closed versus open models.

But it was born out of that frustration, I think. The original inspiration was like, let's just make a sick open-source model that everyone can use and show people how to do science. But really what it grew into now is sort of this incredible open-source ecosystem where you have all these companies around the world now shipping models in the open source. Mistral is doing really awesome work there. Hugging Face is doing really awesome work there. Replit is doing really awesome work there.

And we're now part of this kind of weird, almost emergent assemblage of hackers and hobbyists that are taking these models and running around and quantizing them, or actually training or slurping new models in the open source. And we sit at this junction now where we help to ensure that those models can run on a wide variety of different hardware so that no matter what computer you have access to, you can run the models. And we make sure that you can actually do retrieval-augmented generation, which we'll talk more about later.

On these models in a very easy way, so in sort of like this drag-and-drop interface. And so really just doubling down on the accessibility portions of this that I think made the GPT4All project such a success in the first place.

That's right. So, I think the most popular model right now is Mistral 7B, quantized to 4-bit precision, fine-tuned on the OpenOrca dataset. And one of the two key factors that I really want to drive home is the contributions here. One is making sure that those models run as fast as possible across a wide variety of pieces of hardware. So it's not just NVIDIA GPUs. Like, if you have an AMD GPU, if you're running Metal, if you only have a CPU machine, you should be able to run as fast as possible in all of those situations.

And also making retrieval-augmented generation easy. So if you want to personalize these models on your machine, that should be as easy as dragging the files that you want the model to have access to into a folder. And so those, I think, are the two really, really big value adds of GPT4All right now.

It's all-time highs, I would say, right now. We've got a ton of VC money flowing into companies that are shipping things open source. I love that that's happening. I think on a podcast right after GPT4All came out, I think it was the Weights & Biases podcast with Lukas, I said something like, the biggest challenge for open source is gonna be getting the monetary resources to do a 7-billion-parameter model or these bigger models. And so, it's amazing to see companies like Mistral actually going and doing what they say they're going to do and releasing these amazing models.

And Meta as well, with the release of Llama, I think, has played a big part here. And I think also one of the reasons it's kind of popping off right now is we really are at this kind of new frontier in terms of the discovery of what these things can do. And so it's totally possible that some random person in the middle of nowhere that's just somewhat interested in this stuff spends enough time poking at it. There's so much new stuff to find that they can discover things that are really amazing.

And so you see some of these really incredible techniques coming out of not maybe what you might call the royal science or the universities, but some random person will be like, oh, hey, I did this new type of interpolation between these two model weights and it turns out it gets way better, or things like this. Or, oh, I happen to have gone and curated this dataset that is now incredibly useful for fine-tuning towards XYZ task. These things are things that a single person can do on commodity hardware that really move the needle now.

And so I think now that there's attention on it, and because it's such an inflection point in terms of what is possible with the technology, people are really able to do a lot without needing access to a ton of resources.

So this has been, as Zach was talking about earlier, one of the biggest parts I think that was integral to the success of the original GPT4All is that it was quantized. We used llama.cpp for that. And I think one of the things that we've been really proud of here at Nomic is the fact that we've been able to actually bring staff on full-time to work on things like improving llama.cpp. And so right now we have Nomic staff working back and forth with Georgi on getting new compute backends merged in to improve the support for these models across a wide variety of hardware.

And I think the lesson here for the broader open-source ecosystem is, as economic pressures start to apply and the VC money starts to run a little bit drier, staying collaborative and making sure that, as a whole, the open-source ecosystem itself continues to be able to move forward. And if someone else's repo is one that's useful in your project, being open to collaborating with them and being flexible about where those boundaries lie is really, really important. And so that, I think, is a lesson that I really hope all of the people building in open source right now take to heart.

Yeah, so Atlas is a tool for exploring and interacting with massive unstructured datasets using only a web browser. It was born out of a lot of the work that my co-founder, Andrej, and I were doing at this radiology generative AI startup, Rad AI, where a lot of the work on the models involved exploring these massive datasets, finding places where maybe the doctors had made mistakes or the model had made mistakes, looking for patterns in those errors, attributing that back to the training data, looking for distribution shifts in the actual production data.

When we were at Rad AI, there was this super interesting global health event called COVID, you may have heard of it, that happened. That was this massive distribution shift in our query stream. At the time, we didn't have the tooling like Atlas to deal with that in a very flexible way, and it caused a lot of headaches, certainly. And so, effectively, the interface that we converged on is what would be ideal for this: this massive scatterplot that you can run in your web browser, where every piece of data in your dataset is a point, and two points are close together if the pieces of data are similar.

And what we found is this sort of layout is particularly useful for debugging AI models because their errors tend to pop up in localized regions of it. So we're actually doing a collaboration with Hugging Face recently on their open-source IDEFICS model, looking at where that model performs systematically well and systematically poorly. All of the places in the latent space, in this representation space that we're showing people, where the model was performing poorly were all spatially localized. They popped up as little patches of land in the data.

Similarly, all the places where maybe the model was even memorizing data popped up as these little patches. That was incredibly informative for us in terms of detecting bad data, detecting not-safe-for-work data that should have been removed, detecting data that was, like, misextracted, detecting things the model was memorizing, or detecting categories of data that the model was biased towards reproducing, things it did particularly well on. So that is sort of a broad motivation from the ML engineering perspective of how Atlas got started.

Atlas is definitely a B2B enterprise tool. We have very generous limits for individuals, like power users, academics as well, especially, to make it so that they can leverage it. But it is kind of the core growth engine behind the monetization of Nomic. And I think that puts us in a really, really interesting position relative to a lot of other open-source companies because it means that we can keep GPT4All as this very pure love letter to the community, open-source sort of project.

And we won't be pressured to eventually find a way to squeeze it for dollars. But I think the two interact also very well from a funnel standpoint. So the kind of person that is interested in tooling to help them understand and curate massive unstructured datasets is the kind of person that's either, one, using that to train models, or two, generating a lot of that kind of data. And that's exactly the kind of person that is interested in tools like GPT4All. And so a lot of the first enterprise sales that we had with Atlas were inbound in the GPT4All Discord.

Maybe this is just what building a company in 2024 is, but our sales funnel is literally like Twitter to Discord to enterprise sale. It's insane.

Yeah. So, I think one of the most interesting things I've learned was: be open to your thesis about core audience shifting. So, when I originally built and envisioned Atlas and sort of brought it to Andrej and the team to help me realize it, I very much thought that the ICP, ideal customer profile, would be a machine learning engineer. And a lot of our early users were machine learning engineers. We still have a lot of them using the system. But something that struck me as very interesting is Atlas started to get picked up by a lot more kind of, like, I don't want to say less technical teams, but an archetype that's more of a business analyst or someone doing business intelligence, where we see consulting companies that get these datasets from their customers and maybe they have domain expertise, but they can't code.

And so, never before have they been able to interact with this data with this level of velocity and this level of granularity. And so, we've seen a much wider adoption in terms of how technical the actual users of the systems are than I originally anticipated, which has been really, really interesting to see.

Yes, correct.

Yeah, I think the fundamental thing here is people realize that there's a new data primitive, and that data primitive is here to stay, and that's the vector, the embedding. And just to bring us back to Atlas for a second, in the same way that we see vector databases taking that new data primitive and trying to play out the implications of that at the database layer of the stack, I think we're gonna see a radical shift like that at all layers of the business software stack.

And the current conception of what I believe Atlas to be in the limit is the logical extension of what happens when you make embedding vectors a fundamental primitive at the sort of visualization or business analysis, business intelligence level of the stack. So, internally, we love to throw around calling it sort of like a Tableau for unstructured data.

Yeah.

And so, I think also, one of the things that I'm most proud of that we're able to do with this, which is distinct from a lot of other open-source releases, is we're not just releasing the model weights, we're releasing the dataset that Zach spent all of his time curating. And what this is going to do is it's going to allow people that want to build upon and adapt the model to, one, be able to reproduce it, and two, be able to start tweaking it and learning for themselves without having to go through that arduous process.

And as far as we're aware, none of the other top models on any of these benchmarks have gone out and released their datasets. Most of the time, we just see open weights. Sometimes we see a codebase, but you almost never see the datasets. So we're really proud that we're sort of delivering the full end-to-end package to the open-source community.

There are a couple of answers here. I think the truest answer is that the data is often the secret sauce. When people talk about, like, oh, what is the secret sauce of your AI model? It's almost always the data. Sometimes it's clever scaling and you'll have architectural tricks that move the needle, but almost always it's the data. And so if a company wants to appear open source without maybe really being open source, they'll release the weights but not the data. For some companies that we're now seeing, you're seeing Voyage pop up, for instance, and Cohere is another one that are really investing heavily into embeddings as a service.

That's their IP. That's their core IP. And so for their business to work, they can't release it. But then you start to get into these very weird conflict-of-interest scenarios where it's like, okay, but if they don't release it, one, there's an auditability question on that model. And then there's another question, which is, if they don't release the data and they don't have necessarily the tooling to actually comb through all of it, it's really hard to make sense of if the benchmark scores that they get are super legit or not.

Maybe they're legit, but you can't actually check, right? And so I think that's an element to it as well, perhaps.

Okay, very cool.

Yeah. So, one of the key portions of the Atlas pipeline is actually embedding all of your data. And so one of the things that we're excited to start doing is, as we build up this core competency of building these incredible embedding models, we can release them to the public and also incorporate them to improve the Atlas system. And so there's sort of direct product value and sort of direct auditability on the Atlas side of things that I think is really important. And beyond that, going back to if you want to talk about go-to-market and how GPT4All fits into that, this idea of if we can do these things where we're really just giving as many people as possible access to this technology, be it what you need for retrieval-augmented generation with Nomic Embed, or what you need to actually get up and running with the generative models themselves with GPT4All, regardless of your hardware, that's just going to bring more people into the AI ecosystem faster.

And I think that demand sort of directly translates to need and sort of usage of Atlas. Because again, as people are producing massive unstructured datasets and consuming massive unstructured datasets and tweaking these models and really starting to learn the lessons that we constantly relearn, which is, like, it is literally always the data. If something is wrong, it's almost always the data. The understanding that having tooling to really dive into your data and understand it at an incredibly detailed level is just going to proliferate more widely.

Yeah, it goes back to the core mission of Nomic. Our objective function is to improve the explainability and accessibility of AI models. On the accessibility side, we want to, as much as possible, continue releasing new open-source tools for the community for free to bring more people, regardless of where they're at, regardless of what resources they have, into the post-AI world. I think this addresses what I would say the number one risk of this technology is, which is the sort of unequal access element to it.

The version of the world that I fear most is like there's two or three mega-companies that lock everything down in the early days. And so the sort of, like, I don't want to say inequality necessarily, but the inequality induced by that only gets worse and worse and worse. That, I think, is the number one future we have to be watching out for. And then the explainability side of things really is, I think, Atlas. I cannot emphasize deeply enough how important looking at and understanding your model inputs and outputs are for actually building models that are safe and doing the sorts of things that you intend them to.

To do.

The New York AI ecosystem is incredible. It's definitely smaller than the Bay, but I think that gives it a level of, like, tight-knittedness that you almost lack in the Bay. And as someone that's sort of been a part of both, I spent a good amount of time living on Steiner Street in Hayes Valley, and I spent the last couple of years living here in New York. I think I prefer New York. And the main reason being, I think the challenge with Hayes Valley is in the sort of cerebral Valley, and now the AI arena, it can get a little bit like, I don't want to say groupthink necessarily, but there is a dominant narrative.

And it's like, I think when you're surrounded by that, it's quite hard to kind of think a little bit outside of that box. And one of the things that I love about New York is you're not just surrounded by tech. You are surrounded by incredible people in all walks of life, be it art or acrobatics or companies that are doing things that are not necessarily AI but up-and-coming. That aspect is really key. Plus, we have sick AI heavy hitters in New York.

Hugging Face has an office in Brooklyn. Where's Hugging Face SF? It's not there. And especially as Paris continues to be an up-and-comer in the scene, it's a faster flight from New York to Paris than it is from SF to Paris. So I think if I were to bet on where the global nexus necessarily would be, if you're accounting for the fact that there are things happening outside of the US, I think New York is a very strong candidate.

Yeah, you can follow us on Twitter, @nomic_ai. We are always hiring for particularly interesting people, is how I would describe it. I think everyone at Nomic is not only excellent at their work craft, but every single person has this interesting kind of plus-one to them. So, I know Zach hates it when I mention this, but for Zach, it's not only that he's an excellent machine learning engineer that can single-handedly drive a project to open-source something as good as one of the top companies in the world, but he also played D1 baseball, right?

It's that little extra something that I think really interests me.

And also, we hire a lot out of open source. So, if people want to get involved, we hired Zach out of open source. A couple of people that we picked up during the big hiring spree after we raised our A—Adam, Aaron, Jared—these are all people that had interacted with us on open source in Discord before. Not necessarily before it was cool, but we saw them shipping and giving back to the community. And that's something that I think is really baked into what we want to do here at Nomic.

So, if you want to work with us, hit our Discord, come work with us, and we'll see what we can do.

It's a false dichotomy. Yeah, it's a false dichotomy. You need all of the above, and it's questionable if that's even a distinction that should be made.

Thank you to everyone for coming back and joining us for the first episode of season two of The MAD Podcast. We will be back to our regular publishing schedule weekly, each Wednesday, with new conversations with leaders in the machine learning, AI, and data landscape. If you like the show, you can find the video recording of this episode, along with many, many more, on the Data Driven NYC channel on YouTube, and you can find all the important links in the show notes.