Numlock Sunday: Matt Lynley on what's actually happening with AI
By Walt Hickey
Welcome to the Numlock Sunday edition.
This week, I was thrilled to talk to my friend Matt Lynley, who wrote “Hugging Face's next act: growing into a $4.5B valuation” for. Here's what I wrote about it:
Hugging Face, the ecosystem that supports large language models and which has repeatedly been compared to GitHub but for the machine learning space, announced that it has raised a $235 million funding round, which would put it at a $4.5 billion valuation. Among the investors are a who’s who of enterprise software and hardware companies trying to buy into the boom, including hardware companies like Nvidia, Intel, AMD and Qualcomm, as well as longtime tech industry titans like Google, Amazon, IBM and Salesforce. Hugging Face currently hosts 1 million repositories and 500,000 models. This more than doubles their $2 billion valuation from last year, which many at the time thought was rather rich.
AI is buzzy, and I like Matt’s view on it. Lots of people allowed themselves to be dazzled at the promise of what social networking could be, only to neglect the real business going on underneath. Matt’s doing that reporting when it comes to AI, looking at the actual business that’s developing and not just the Skynet-style hype.
Matt can be found at.
This interview has been condensed and edited.
Matt, you have started a great newsletter called Supervised. What got you first interested in the field of AI? I know that you've been a tech reporter for a very long time, but when this niche started to emerge, when did you think it was really worth taking a business journalist's eye to?
Yeah. I've been a journalist for about 10 years. I did take a four-year detour through analytics and data science, between the two. But when I decided to come back into journalism, the idea was to cover the stuff that I used when I was working in analytics and data science. So obviously, we're big Snowflake users, we're big DBT users. The idea of coming back to journalism, partly because it's one of the only things I'm good at, sadly, was to essentially write about the things and cover the companies that I used, which I felt was a blind spot in tech reporting. This was before ChatGPT, so when we talked about AI, we were literally talking about statistics, honestly. And the companies that were in this space were focusing on what we called back then neural network and deep learning, which obviously still exist today.
I was covering these startups that were emerging. One of them is called Hugging Face, which essentially is a place where people were uploading and sharing some of these machine learning models, which again was just to a certain extent glorified statistics in some ways. That's a vastly massive oversimplification. I'm sure someone's going to hit me over the head at some point for saying that.
And then, as I was back in journalism and covering this, we started to see some neural network developments happening. The first one probably would have been Stable Diffusion and DALL-E, which I believe was August or July of 2022. That was the first time we started to see, okay, some of these techniques — one is called a generative adversarial network, which is essentially you pit two algorithms against each other until one of them fools the other one — were able to start creating "art," if you want to call figures with seven fingers on their hands art, sure.
But we started to see, okay, some of these techniques that we've been working on for a decade are starting to create some really interesting results. And OpenAI had been working on some tools at the time. GPT-3 was what was available, which was pretty good, and some companies were actually using it for their projects, but obviously it wasn't until ChatGPT came around in November that AI exploded with the launch of ChatGPT. Suddenly it was like, "Oh my God, I'm bored and I can talk to someone. This rules," even if it's a pathological liar.
So, as I was watching this all play out, everyone was focusing on OpenAI and ChatGPT and how's AI going to change the world? Is this SkyNet? Duh-duh-duh. And there's this whole universe of things happening around and underneath it that was going unnoticed.
Honestly, I did what I always do, which is call 100 people and ask for their feedback on it, because I was looking for something that would take one step deeper — not 10 steps deeper, where we're talking about the intricacies of specific loss function — one step deeper where you can throw around the term large language model; you don't have to define it, spend words defining it every single time, right?
And so that was my thesis, which is just everyone's covering OpenAI, everyone's covering Meta and what they're working on, but it's like, go to these AI events and conferences, and what are people talking about? Obviously, they're talking about Llama, the language model from Meta, but oftentimes it's really interesting but super niche stuff that you don't read about. And so I was like, "Okay, well what if I just started a publication that's just, what are people talking about in the industry that no one's writing about?" and so that was the idea of Supervised. What I aim to do is say, "Hey, I called 20 people so you don't have to, and this is what everyone's actually talking about," so that's the way I see it.
Yeah. And you've covered I think a lot of the most interesting elements of it. I think that when social networks were starting out, people were all very starry-eyed about, "Oh, finally the world can come together. Finally, there's a public square for people," and then it was only—
“Facebook is like chairs!”
Exactly. And then it was a decade later, and then we're just realizing, "Oh no, this is a weaponized advertising ecosystem. There's a huge business going on here," and we were so distracted by the promise of some of this stuff that we actually forgot to try to understand how this was going to become just a huge colossal business.
I guess one thing I've enjoyed about your work in particular is that there's a lot of starry-eyed stuff about AI, both from people who support it and look forward to it, and from people who view it as a potential threat. But I think that what I've enjoyed from you is that it is very unvarnished; it is very much looking at it as the industry. Going back to that Hugging Face story, though, what are some of these companies succeeding at and what are they doing business-wise that you think is interesting?
Yeah, I think one of the underrated parts that's getting lost in translation in all this coverage of AI is that they're still figuring out the business models for this stuff. And they're still figuring out what does a business built on AI even look like, and what are the things you need to actually run it. For Hugging Face, for example, you need a place to store things that you store on Hugging Face. The files, it's like you have a JPEG or whatever, the file is called a safetensor. This gets put up on Hugging Face, which you can use to pull down and run a large language model like Meta's Llama and things like that. They're very much focused on the open-source ecosystem, which is like, anyone can pull it down, mess with it, upload it, change it, all that kind of stuff.
So there needed to be a place to store and share all of this stuff, like there needed to be a place to store and share code back in the early software development days like GitHub. These things are too big to run on a laptop, usually. There are exceptions to that rule, but you need to be able to run them on some powerful hardware, on a server somewhere owned by Amazon or Microsoft or Google where they're the cloud providers, but essentially they take whatever it is that you're building and run it on their powerful hardware for cheap. In AI terms, that's called inferencing, so it's like, when I ask a question to ChatGPT and it sends back an answer, that's called an inference. You need a company that can help provide you with a way to run inference on the tool or model that you're building if you're playing around, screwing around with it. That's a business, right?
And you need a place to store all the data that goes into training these models, because what happens is you go through this process where you collect trillions upon trillions of data points, and then you feed it into these black-boxy algorithms — it's not actually a black box, but there are black box elements of it — that teach this model how to talk to you and respond to you and things like that. You need that data, but it needs to be stored somewhere. There are businesses around throwing that data, whether that's a company like Databricks or a company like Snowflake, these sorts of companies that help you host that and access it in a sort of easy way — debatable, your mileage may vary, et cetera. And so, that's a business.
Then there are all these other little individual niches where there are potential businesses. Say you want to do a more advanced AI problem and say you want to tell an AI bot that you want to write a program for it, but that AI bot is not necessarily up to speed with new tools that came out because when these models are trained, they're frozen in time, and so they don't have access to up-to-date, recent information. You need a way to not only tell them what to do, but also what information it should add on top of it. So it's like, copy and paste an FAQ on a new software development, or something like that. You need something to orchestrate that order of operations, so, that manifests in a tool called LangChain and LlamaIndex, and that's also a business.
There are all these individual businesses that sprout up around business models, that sprout up around just firing in and saying like, "Hey, write me a joke in iambic pentameter, but actually it's Carl Jung writing it instead of Shakespeare," and so there's a whole universe of all these companies that can be worth billions of dollars that are just pretty much powering these individual steps. And again, it's a blind spot: I could rattle off 50 company names, and you've probably never heard of them, but they're all worth hundreds of millions if not billions of dollars. They all power individual parts of this underlying mesh that makes all this possible in the first place.
I thought it was really interesting when you were reporting on Hugging Face in particular, not to just keep coming back to them, but they're just, I think, a really interesting part of this economy. And you've written quite a bit about them because they are one of the big, public faces of AI. A lot of the companies that threw money at them work in hardware. And you remarked on that, and I thought that was really interesting. What do you make of that?
To take a quick step back, one of the bigger developments in AI is this conversation around open source versus closed source. If you're familiar with software development, it's decades old, and it is as vicious of an internet fight as you'll possibly get within software development, in some parts of the internet. And that notion extends itself to AI in the sense that people creating AI models — and when I say AI model, it's basically like what you would use to power a chatbot, or what you would use to power a writing application, or a programming copilot or something along those lines, making those available so people can download them and do whatever they want with them, or modify them and then do whatever they want with that, or reupload that. And so, that dynamic of open source versus closed source is very much playing out in machine learning as well. And Hugging Face is basically at the center of that: They're the ones that host all these open-source models, including the one from Meta, the company formerly known as Facebook.
They pretty much are at the direct center of that, I wouldn't call it a clash, but it's a complicated part of it. Where you have a company like OpenAI that wants to keep all of their stuff secret, keep their secret sauce in-house, and you access it through the web only, you don't get to download it or blah, blah, blah, or you have Meta releasing one that they're like, "Please use it, do whatever you want with it, and modify it however you see fit." And so, every single time someone uses a model, someone inferences a model, it requires compute from a NVIDIA or from Microsoft or from whomever. Every time someone trains a model, it requires compute from an NVIDIA or an AWS or whomever. Every time someone modifies it, it requires compute from an AWS or a NVIDIA or whomever.
The joke in AI is the house always wins, which is that for everything that's happening with OpenAI and all these other companies, at the end of the day, NVIDIA is just making a ton of money off of it, and Microsoft with its cloud service is making a ton of money off of it, and all this kind of stuff. They all have a vested interest in promoting the growth of this because it's good for their core businesses; it's good for them to have people using their cloud services to power whatever AI tools are out there. Basically it's in everyone's best interest that Hugging Face not only exists, but is successful.
But it's complicated because a lot of people take the viewpoint that developing a closed-source model through the web is actually the best way of doing it, and a lot of people take the viewpoint that all this development should be happening in the open because that's safer. It highlights a lot of the deep nuance in the industry right now, which is actually evolving in a very similar way to how software development evolved originally, in the closed versus open source debate or conversation or whatever you want to call it. Basically, they're one of the most important companies, and they sit directly at this point where an entire field of AI is being built off of the products that they offer.
Wild. It's interesting because again, it does seem like there's going to be a lot of money at stake here, and these decisions are going to have ramifications that will last with us for quite some time.
Yeah, totally. Some of the larger, more complicated AI tools are very expensive to run. NVIDIA hardware, they make a graphics processing unit, a GPU, which you would use to render Counter-Strike originally, but they have custom versions of it now that are specifically toward training machine learning models that are incredibly expensive and hard to get ahold of, and companies are raising literally hundreds of billions of dollars just to get ahold of this hardware in order to compete with each other. I think there's just much more money flowing in the industry just for the pure development of it, and not even promoting ChatGPT or something along those lines, just actually getting the stuff off the ground.
Got it. The newsletter is Supervised, and I really like it. It's a good thing. It's a really cool and extensive look into an industry that, again, gets talked about in a lot of generalities, but I think that you get into the specificities that actually make me learn things about this. So where can folks find you? And I guess if you had to summarize the pitch in a sentence, what are you bringing to the table that you think the bulk of other coverage doesn’t?
I went to a conference so you didn't have to.
No, seriously, I want to be able to surface up what everyone is talking about a month before it shows up in some more public domain. And unfortunately with AI, a month is about as good as I can get. I want to be a year ahead, but things are moving a little too fast, so I try to be reasonable. But hopefully, in Supervised you read about things a month before they start hitting mainstream and AI.
If you have anything you’d like to see in this Sunday special, shoot me an email. Comment below! Thanks for reading, and thanks so much for supporting Numlock.