Hugging Face podcast mentions

Quote

are provisioning that environment because being the creator. and early adopters of this utility, I'm wondering how you're addressing some of those challenges. Git somewhere? Are you digging into some of the GitHub code spaces? I know that Um there's the ONA project that used to be Gitpod that's leaning heavily into that agent. coding environments. Hugging face is investing in agentic environments. you're thinking about near term, what is the quick hack that you got running? And as you continue to invest in this and plan forward, what are some of the ways that you're thinking about how to actually make that shared developer environment something that is sustainable and maintainable over time. ways like at the bottom the seventh too because the project I started I wrote the first line code like six weeks ago and it has gotten so good so quickly right and I was as like I would kind of rate my productivity developing Agor with Agor. As like maybe 50x pre-AI type of workflow. So we're pretty early. on when it became something shareable that we could you know bring if a handful of people on

Quote

By overwhelming audience demand, we have settled on best papers of twenty four. twenty twenty four overview talks and Oxford style debates for the following category. Computer vision, open models, transformers killers, synthetic Data, agents, and LLM scaling, featuring friends and past podcasts. Podcast guests from Roboflow AI two slash meta recursal together. Hugging face, open hands and semi-analysis, and a special startup landscapes keynote. Note from Sarah Guo of Conviction. Head to HTTPS colon slash slash Lou dot MA slash L S live to sign up and apply to speed or sponsor. If you want to take part in the QA limited in-person Tickets are available now. Lastly, we are making a last call for listener questions. questions for our twenty twenty four recap head to speakpipe.com slash latent space. to submit questions. Watch out and take care. Hey everyone, welcome to the Laden Space Podcast. Alessio, partner and CTO at Decibel Partners, and I'm joined by my co-host Swix, founder Osmalia. Hey and today we're still in our sort of makeshift in between studio but we're very delighted to have uh a former returning guest host. Uh Itamar, welcome back.

Quote

Okay, so basically let's let's take uh Koh here. So they have command. R. Let's say the model type is command r. In mlxlm, if If you go to MXLM and then you look at the models folder, you should be able to see command R.Py. That file contains all the code necessary So, I think that's a good idea That's what I mean by by model type. Huggingface uses model type to load the Tokenizer to load your I think model model uh uh weights as well. Um and we just do it a m a more direct approach by calling the uh, the model type as the model file. And what I mean by that is like usually if it already exists you just run The model, you just download the model from Maggie Face and run it uh through the CLI or through your own Python script. Usually it's just MLX, name of the project, dot generate. And that should be enough uh to generate uh uh an initial response. Um if that does not exist, then I go look at the transformers in implementation which they usually have a put a pi a transformers in t implementation. I look at the code and I just convert one by one.

Quote

is the firewall. So the firewall has two or three layers actually. One is about analyzing giving me visibility and control what your employees are doing. AI services are they using what are they using it for? What is the level of the sensitivity? We have now evolving the agent control, so which agents are used, what type of tools they are using, these agents, what are they trying to achieve and real-time control. And lastly we have a model risk file. firewall. So for large companies that have a team of developers, they tend to like to download models from Hugging Face or other sources that potentially can bring all sorts of threats. And we have a firewall that analyzes the model. So it checks, for example, who is behind the model, it runs like a run time red teaming to find vulnerabilities whether it's protecting against jailbreaking or pront injection. So wrapping back a high level. We enable AI for companies that have problems or restrictions from using AI. We have a new interesting direction by the way, which is a side result of the visa. Visibility, we can give you like intelligence into adoption. So it's only not only about security. Some companies want to know if AI is used effectively. in their company. So we have dashboards that's showing you who's using AI, what type of usage, and giving like tips and tricks, encouraging users.

Quote

It's metadata and deliberate. That's evolved really kind of like in three chapters for us. So chapter one started in the summer of twenty twenty two when we got our initial round of funding. Um off of a slide deck. We had zero lines of code. And at the time hugging Teamface was really the center of the universe for machine learning. And so we said look We're gonna build a set of open source capabilities so that any op any Hugging Face user can get their data ready for models that are hosted in Hugging Face. And so we embarked on that in July 2022. Um, we're working throughout the fall. Um and then in November 2022, all hell broke. Loose with the release of Chat GPT and then a month later with folks Dusting off the rag paper that was written by a bunch of researchers at Meta. Yeah. And then folks were like, okay, we've gotten a lot of the easy data into vector databases. How do we get hard? stuff and we were really well positioned at the time. Open source like a real really rich set of um open source um ed AI infrastructure tooling companies that were emerging at the time. So ourselves, Lang Chang. Lamin DAX, Weviate, Chroma, etc. And then a really And so that open source project really was like chapter one. And that was just a

Quote

started a couple of years ago. It's it's even more than that, like three, four years ago. Gibrier and I we come from this machine learning data science background and uh We come from a time where models were emerging as a new super powerful thing, right? So we started in two thousand uh sixteen. We had uh TensorFlow becoming open source. So machine learning models became a thing. Then we had Pytor. and and everything around. And Hugging Face started to put those sp specialized models. uh available to everyone with uh Transformers Hub, the Hugging Face Hub. Right. And uh we were sort of looking at uh the idea of having these is is specialized models and thinking about about how this is going to behave in in a future where you wanna connect them, right? Or have them becoming some sort of a brain behind the scenes. that maybe you're gonna have uh let's say you have uh models for image recognition recognition, models for image detection, models for entity recognition, time series uh transformers, whatever, right? And how could we have a place where you're gonna connect all of these specialists, right? experts or specialists by the time uh and one how we can connect them

Quote

That gets pulled inside and executed. And this thing opened a Pandora box because this log for J was everywhere and you get arbitrary code execution on the box. And the basic primitive inside was It's a reference to external code that is loaded inside and just executes inside. This caused a massive havoc uh all across the world in all of our computer system. Honestly if you Try and read around on the number of compromises. We're talking about hundreds of millions of devices. Okay. Now we look at Hugging Face as a library and you look at this wonderful flag. called trust remote code. And what this thing does is that when you load the model, you know like you click use this model, use transformers, inside it gives you like a code snippet to load the Some model inside it has this flag sometimes hard coded. And what this thing does is they say Oh for some models when you load them, you actually want to load the latest the latest representation for an external machine. What this thing does is literally remote code, load it on your machine, execute it. on your machine, load it on top of stuff. So same sort of thing we did back then. We're doing the same again. Today on Hugging Face, I don't know how many users there are, but if you're running your thing running outside of a jail, if you're running your model outside of a sandbox, you are doing a very bad thing to yourself. And the o…

Quote

we're going to be using container use, but it feels like the next level of AI agents is that every tool will have a built-in method for doing multiple AI agents. safely at the same time on the same code base. So that probably is coming. I hope that's coming at least because I need it often when I'm on bigger projects. Next up, I've wondered for months now how to compare the free models, the downloadable ones that you can get from Docker or Olama Hugging Face. And in June, I discovered the Devstril model from Mistral. That's the The French AI company that's competing with Lama from Meta and the other bigger ones, Gemma and Deep Seek and Quinn. And Mistral put out a blog. post showing this devstrl model, which is actually a model small enough you can run locally on a Mac, at least if you have 32 gig of RAM. And they're claiming SWE bench performance numbers that are better than all the other models. Basically the best model that's free. able to run locally for doing software engineering tasking. So if you're not aware of SWE Bench, it's a website. I think it's swebench.com. com and they rank models compared to each other on a percentage of GitHub issues they could resolve through PC.

Quote

They compare to a whole bunch of other open uh code generated. models like code Nama. And as always, with these news stories, it's The best and they do release it under a super open license, Apache two point zero for both research and commercial use. So uh yeah, lots of progress still being made. on LLM specifically for code. And next we got Hugging Face launches Le Robot, open source robotics code library. So Hugging Face uh made some news earlier this year, just a couple of months ago by hiring from Tesla and the first fruit of that package that is used very, very widely for doing work in natural language processing. with the transformer architecture and they want to position this as That type of package but for robotics where it's a toolkit that is a comprehensive platform platform that has you know support for getting data sets for Simulators for different types of robots, for uh training models. Uh, for having pre-trained models, just lots of stuff. And uh

Quote

work with that and build some capabilities I have in mind. And Yeah. If you would put me back a couple of years ago I'm like, ah no, you know what, it's not my time. But now Now with this AI change and I already went through the first time. On Hugging Face, which is these guys are great. But reading it through the documentation I was like Pretty straightforward. So think about how much AI or change the field that I can. I can easily go by a robot like a small robot and I'm planning already ahead of time. this simulator. So you don't need to wait for it to deliver. You built ahead of time the apps. and simulate it that it will work on the robot and then the robot comes with deploy it. So that's That's my go-to like a s what I'm excited for in uh 2026. Yeah, it's kind of crazy. I I I feel like when we started in this field It was like hard enough to get the depression. Dependencies installed for TensorFlow and just be able to run any model. Just like that in and of itself was like are you trying to do it? Trying to give us PTSD? Is that is that the goal here?

Quote

That is just like the working robots that you see like in in warehouses. Like what what's What's happening now? So they're really using a lot of the same model. models but maybe not as heavy focus on research. So I In parallel to all this that's been happening, there's also been a lot more um open source models coming out also the the same way in LLM land there's there's great open source models, uh heavily driven by Hugging Face. Um so I see them being just kind of scrappy. uh using these like VLA uh models That's vision language action models. Uh but maybe just being more more practical about building other systems around them and kind of using teleop where needed. and and kind of just um have a more product And like uh scrappiness uh sort of approach to making things work. Like what are your actually seeing like are are there I mean I've talked to guests on the podcast, I think of sort of every you know, realm of of this, every like spectrum. And certainly like, you know, we talk to folks that have That seems like clearly factories are doing that. A lot of people talk about, you know, like kind of picking up the thing.

Quote

So I have a a purely theoretical question for you. of OpenAI's tech, uh the original GPT two model, basically. Put up for free on Transformers, Hugging Face Hub. You can download Uh, I'm not sure if you're Um spoiler alert, not so great. Uh it it is not particularly sophisticated as compared to the latest state of four you know, four oh one that that's in beta right now. Um So seeing that advancement and this this nonlinear trajectory of Do you feel that it's inevitable that we're gonna hit not this Generalized you know highly capable and intelligence uh in AI that can do this one thing really well. Right.