AI trends: a Latent Space crossover

Foreign.

Welcome to Practical AI. If you work in artificial intelligence, aspire to, or are curious how AI related technologies are changing the world, this is the show for you. Thank you to our partners at Fastly for shipping all of our pods super fast to wherever you listen. Check them out@fastly.com and to our friends at Fly. Deploy your app servers and database close to your users. No ops required. Learn more@flyeio.

Well, hello. We have a very special episode for you today. I got the chance to sit down with the guys from Latent Space, WIX and Alessio out in San Francisco. They were kind enough to let me into their podcast recording studio and we got a chance to talk about our favorite episodes of both of our shows and some of the overall takeaways we've had from those discussions. We cover some of the trends that we've been seeing in AI and they even get a chance to grill me on my opinions about prompt engineering. So enjoy the show.

Hey everyone, welcome to the Late in Space podcast. This is Alessio, partner and CTO in resident at Decibel Partners. I'm joined by my co host, Swix, writer and editor of Latent Space.

And today we're very excited to welcome Dan Whitenack to the studio. Welcome, Dan.

What's up guys? It's great to be here.

This is a podcast crossover if you recognize this voice. Dan is the host of Practical fm. He's been in my ear on and off for the past five years, covering the latest and greatest in AI. Before it was cool.

Yeah, yeah, yeah. Before, before the AI hype back in these like weird data, science times, whatever that is now.

Yes. Everything is merging and converging. So I'll give a little bit of your background and we can go into a little bit on your personal PhD in math, mathematical and computational physics. And then you spent 10 years as a data scientist, most recently at SIL International, which I actually, I thought it was like an agritech thing. And then I went to the website.

It's actually a nonprofit international ngo. Yeah. So they do language related work all around the world. So I spent the last five years building up a team that's been working on kind of low resource scenarios for AI, if people are familiar with that. So doing machine translation or speech recognition, that sort of thing in languages that aren't yet supported.

Yeah. And we'll talk about this later, but I think episode three on Practical AI was already featuring the global community that AI has and addresses.

Yeah, yeah, yeah. It's been an important theme throughout the whole time. Throughout over 200 episodes.

Yeah. And you recently left Sil to work on Prediction Guard, which we can talk a little bit more about that. You are also interim senior Operations Development Director at NT Candle Company.

Yeah, yeah.

And. Yeah. What else should people know about you?

Yeah, I mean, as probably can be noted from the intro, I love working on various projects and having my hands in a lot of things. But, yeah, I code on the side for fun and that's how I usually get into these side projects and that sort of thing. But outside of that. Yeah, I live in Indiana. I was telling you guys that I'm trying to coin the term cerebral Prairie, so we'll see if that catches on.

Probably not, but you're our second guest in a row from Indiana. Linus from Notion was Indiana. We're talking about how there's a surprising number of international students in there.

Yes, very true.

Purdue's a strong university.

Yeah, yeah, Very strong university. It's a great place to spend time and there's a lot of fun things that happen around that area too. So. I'm also very into music, but not any sort of popular music. I play like, mandolin and banjo and guitar and play folk music.

Low resource.

Yeah, low resource music, low resource languages, all those things. Anything low resource is in my territory for sure.

And maybe we can cover the story of practical AI. How'd you start it? Tell us what the early days were like and just fill everyone in.

Yeah, it was kind of a winding journey. Some people might be familiar with the Changelog podcast, which I think they've been going now for like 11 or 12 years. It's pretty prolific in around. I think originally around more open source now it's kind of software development in general. But they have a network of podcasts now and at a Go conference, actually. So I'm a fan of the GO programming language. That's another fun fact. But at Go for Con, I think it was in 2016, maybe, I met Adam Stakoviak, who is one of the hosts of the Changelog. At the time, I was giving a talk about data science, something I forget, but he kind of pitched me. He's like, we've been thinking a lot about doing like a data or data science podcast, and at the time he had a name, it was like. I think. I think it was hard data or something like that, which. Which never caught on for obvious reasons. But I kind of stored that away. It didn't really do anything with it. But then over the next couple of years, I met Chris Benson who's my co host on practical AI and helped him with a couple of talks at conferences. We met through the GO community as well. And eventually he's a. He was working at a different company at the time. Now he's a strategist with Lockheed Martin, working in AI stuff. But he reached out to me and said, hey, would you ever consider doing kind of like a co host podcast thing? And at that point I remembered my conversation with Adam. So I reached back out to Adam with the changelog and then we kind of started working on the idea. We wanted it to be practical. So at the time, well, there's a lot of people doing things now with AI, like hands on. Back then there were kind of some podcasts that were really hyped AI, like not practical at all. Which is why we kind of came to practical AI, something that would actually benefit people. And that, that's like a great thing to hear from people that when they listen to the show they do actually learn something that's useful for their day to day. That's kind of the goal.

Yeah. Nice. And I think that's one of the things in common with our podcast. You know, there's a lot of content out there that can get a lot of clicks with a fear of AI, you know, and all these different things. And I think we're all focused on more practical and day to day usage. Yeah. Tell us more about Prediction Guard. That kind of fits into making AI practical and usable.

Yeah, yeah, sure. Appreciate that. So, yeah, Prediction Guard is what I've been working on since about Christmas time. Ish. Originally I was thinking a lot about large language model evaluation and model selection, but it's kind of morphed into something else. What I've realized is that there's this market pressure, there's internal company pressure for people to implement these kind of generative AI version of models into their workflows. Because enterprises realize the benefits that they could have. But in practice when they go and they go from like ChatGPT, they type in something, it's amazing. And then like how do we do this in our enterprise where we have maybe rules around data privacy or compliance issues and also like we want to automate things maybe or we want to do data extraction. But I just get like text vomit out of these models. Like what do I do with that? I just gave unstructured text. How do I build a robust system out of inconsistent text vomit? So Prediction Guard is really focused on those two things. One is kind of compliance and running kind of state of the Art AI models in a compliant way and then layering on top of that, layers of control for structuring output and validating output. So some people might be familiar with projects like guardrails or guidance or these things. So we've integrated kind of some of the best of those things into the platform, plus some ways to easily do like self consistency checks and factuality checks and other things on top of large language model output.

Nice. We did have a Shreya Rajpal from Gorails as a guest.

Yeah, yeah.

So, yeah, that's another episode that people really like. Yeah. Maybe just to give people a sense of what practical AI is as a podcast, you want to talk about maybe like the two, three favorite episodes that we have and we can go maybe alternate our favorite.

We've done some prep for this episode.

Yes, yes. Yeah.

So this is kind of, I think our conception of this is kind of like a review for listeners who are new to us, either of our podcasts, to go back and revisit the favorites.

Yeah, yeah. I think I can talk about some personal favorites of mine and then maybe like favorites from the audience. I think some of my personal favorites have actually been, we call them like fully connected episodes where Chris and I actually talk through a subject in detail together without a guest. To be honest, those are great episodes just for, like, me to learn something, like have an excuse to learn something. And we've done that recently, like with ChatGPT and instruction tune models. We did it with stable diffusion and diffusion models. We did it with AlphaFold. So all of those are episodes with us two and just talking through, like, how practically can you like form a mental model for how these models were trained and how they work and what they output. Those are some of my favorites. Just because I learn a lot, because I do a little bit of prep. We talk through all the details of those and it helps me form my own sort of intuition around those things. Another personal favorite for us was that we did a series about AI in Africa that was really cool. You mentioned, like, the global AI community. We did actually a series of those. They're all labeled AI for Africa, highlighting things like Masakane. So people don't realize that, like some of the models that we develop here, like in the west coast or wherever, they don't work great for all use cases around the world. And there's a lot of thriving grassroots communities like Masakane and Turkic Interlingua and other communities that are really building models for themselves. Machine translation, speech recognition models that work for their languages around the World or, or agriculture computer vision models that work for their use cases around the world. So those are a couple of highlights on my end.

Do we go with our personal highlights? Yeah, go ahead. I think you already picked one out.

Yeah, I think mine is definitely the episode with Mike Conover from Databricks, who's the person leading the Dall EFRS there. I think obviously the content is great and Mike is extremely smart and prepared, but I think the, the passion that he had about these things. You know, the red pajama data set came out the morning that we recorded and we're all kind of like nerding out, like, yeah, why is that so interesting? Like he was so excited about it and it's great to see people that have so much excitement about things that they work on. You know, it's kind of like an inspiration in a way to do the.

Same for us, I think, personally. So I tend to drive the news driven episode ones like the event driven ones where something will happen in AI and I'll make a snap decision that we'll have an episode recording on Twitter spaces and we'll have just a bunch of people tune in. I think the one that stood out was the ChatGPT App Store, the ChatGPT plugins release where like 4,000 people tuned in to that one.

That's crazy.

And we did like an hour of prep. Right. And I think it's important for me as a quote unquote journalist to be the first to report on something major and to provide a perspective of something major, but also capture an audio history of how people react at the time. Because this is something that we're talking about in the prep. ChatGPT plugins have become a disappointment compared to our expectations then. But we captured it, we captured the excitement back then and we can sort of compare and contrast where we thought things were going and where things have actually ended up. It's a really nice piece of, I guess, audio journalism.

Yeah, I mean it was just last year I mentioned stable diffusion. All that we were talking about this I always had in my mind, oh, everything's going to image generation. Like, should I quit doing NLP and start thinking about image? And now all I do is NLP and language models. But at the time that was, you know, that's what was on our mind.

Same thing. I was working on a web UI for stable diffusion, just like a thousand other front end developers were. And yesterday was the first time I opened stable diffusion in six months.

And a lot has changed and it's still an area that's developing, but it's not. Yeah, it's not driving thought process at the moment.

Yeah, well, especially because I think just it depends on what you think you want to do. And I'm definitely less visual and more of a text driven person so I naturally lean towards LLMs anyway. Like NLP.

Yeah, I can hit some listener crowd favorites. So we have like one clear favorite, which is actually, I would say it's a surprise to me. Not because the guest wasn't good or anything, but just the. So the topic was MetaFlow. So I don't know if you've heard of MetaFlow. It's a Python package for kind of full stack data science modeling work developed at Netflix and we had Villa Tulos on who was the creator of that package and that has had. So it's like maybe 30% more listens than any other episode and I think the title. So we titled it from, I think From Notebooks to Production or something like that. Yeah, so it's like this idea of From Notebooks to Production. There's all sorts of things that prevent you from getting the value out of these sorts of methodologies. And my guess would be that talking about that is probably like the key feature of that episode. And MetaFlow is like really cool. People should check it out. It is one way to kind of do this, both versioning and orchestration and deployment and all these things that are really important. But I think a takeaway for me was practically bringing into the. Some people might call it like full stack data science or model life cycle things. The model lifecycle things interest people so much. So beyond making a single inference or beyond doing a single fine tuning, what is the life cycle around a machine learning or an AI project? I think that really fascinates people because it's like the struggle of everyday life in actual practical usage of these models. So it's one thing to go to hugging face, try out a hugging face space and create some cool output or even just pull down a model and get output. But how do I handle model versioning and orchestration of that in my own infrastructure? How do I tie in my own data set to that and do it in a way that is fairly robust? How do I, you know, take these data scientists who like use all this weird tooling and like mash them into an organization that deals with like DevOps and like non AI software and all of that? I think those are questions people are just wrestling with all the time.

Yeah, it feels a little bit in conflict with the trends of foundation Models where you, the primary appeal is you train once and then you never touch it again or you release it as a version and people kind of just prompt based off of that. And I feel this evolution moving from essentially the ML Ops era into, for lack of a better word, llmops. How do you feel about that?

No, I think you're completely right. I think there will always be a place for these models in organizations that are task specific models like SciKit Learn models or whatever that solve a particular problem because organizations like finance organizations or whatever will always have like a need for explainability or whatever it might be. But I do think we're moving into a period where like I've had to rebuild a lot of my own intuition as a data scientist from thinking about gather my data, create like my code for training, output my model, serialize it, push it to some hub or something, deploy it, you know, handle orchestration to now thinking about, okay, which of these pre trained models do I select and how do I engineer my prompting and my chain maybe going to fine tuning like that is still like a really relevant topic. But some of these things that like I've been working on with Prediction Guard I think are the things that have a parallel in mlops, but they're slightly like, they're just a slightly different flavor. I think it's how mlops is graduating to something else versus like people are still concerned about ops. It's just like you say, it's a kind of a different kind of ops.

Yeah. And I think that's reflected in our most popular episodes too. So I think all three of our most popular episodes are model based. They're not more like infrastructure based. So I think number one is the one with Reza Shabani where we talked about how to train the replit code model and Amjad Vibes that they used to figure out whether or not the model was good. And I think, you know, that makes sense for our community. It's mostly software engineers and AI engineers, so code models are obviously a hot topic. Yeah, that was really good. And I think like it was one of the first times where we kind of went beyond just listing traditional benchmarks, you know, which is why we did a whole thing about Amjad Eval. It's like a lot of companies are using these models and they're using off the shelf benchmarks to do it. And what other episodes that we'll talk about is like the one with Jonathan Frankel from Mosaic ML and he also mentioned a lot of the benchmarks are multiple choice, but most Production workloads are like open ended text generation questions. So how do you kind of reconcile the two?

Yeah, did you all get into at all, you know, the whole space of LLMs, evaluating LLMs and sort of this was something on a recent episode we talked to Jerry from LLAMA Index about in terms of on the one hand generating questions like you're talking about to evaluate LLMs or using an LLM to look at a context and a response and provide an evaluation. I think that's definitely something that I think is interesting and has come up in a few of our episodes recently where people are struggling to evaluate these things. And so yeah, we've seen a similar trend in one direction thinking about benchmarks and in another direction thinking about this sort of on the fly or model based evaluation which has existed for some time. Like in machine translation, it's very common. So like Unbabel uses a model called Comet that's like one of the most popular, highest performing machine translation evaluators. It's a model, it's not a metric. And that sort of thing like blue. So yeah, that's a trend that we've seen is is evaluation and specifically evaluation for LLMs which can kind of get dicey.

Yeah, we did a Benchmarks 101 episode that is also well liked and we talked about this concept of like a benchmark driven development. You know, like the benchmarks used to evolve every like three, four years and now the models are catching up every like six months. So there's kind of this race between the benchmarks creators and like the models developers to find. Okay, the state of the art benchmarks is here and GPT4 on a lot of them gets like, you know, 90th percentile results. So you know, GPT4 is not a GI. Therefore to get to AGI we need better evals for these models to start pushing the boundaries. And yeah, I think a lot of people are experimenting with using models to generate these things, but I don't think there's a clear answer yet.

Something that I think we were quite surprised to find was specifically in hello Swag where the benchmarks instead of being manually generated, were adversarially generated. And then I was very interested in our. I mean this is kind of like segueing. We're not really going in sequence here. Segueing into our second most popular episode which was on Roboflow, which covered segment anything from Meta. I think you guys had a discussion about that too.

Yeah, it's been mentioned on the show. I don't think We've had a show devoted to it. But.

Well, the most surprising finding when you read the paper is that something like less than 1% of the data of the masks that they released were actually human generated. A lot of it was AI assisted. So you have essentially models, evaluating models and the models are trained themselves, trained on model generated data. We have very many layers in at this point.

Yeah. And I know that there's been a few papers recently about the sort of things that were done with LLAMA and other models around model generated output and data sets. It'll be interesting to see. I think it's still early days for that. So I think at the very minimum what all of these cases show is that models either evaluating models or using simulated data. I think back a few years ago we would probably call this simulated data. Right. I don't think that term is quite as popular.

Augmented.

Yeah, or augmentation data. Augmentation, simulated data. So I think this has been a topic for some time, but the scale at which we're seeing this done is, is kind of shocking now and encouraging that we can do quite flexible things by combining models together both at inference time but also for training purposes.

Well, have you ever come across this term of mode collapse? What I fear is, especially as someone who cares about low resource stuff, is that stacking models on top of models on top of models, you just optimize for the median use case or the modal use case.

Yeah, yeah, I think that one maybe so. Yeah, that is a concern. I would say it's a valid concern. I do think that these sort of larger models and this gets I guess more into like multilingualism and the makeup of various data sets of these LLMs. The more that we can have linguistic diversity represented in these LLMs, which I know, I think cohere for AI just announced like a community driven effort to increase multilingualinguality in LLM data sets. But I think the more we do that, I think it does benefit the downstream lower resource languages and lower resource scenarios more because we can still do fine tuning. I mean we all love to use pre trained models now, but like in my previous work when you were looking at maybe an Arabic vernacular language rather than standard Arabic, there's so much standard Arabic in data sets. Making that leap to an Arabic vernacular is much, much easier if that Arabic is included in LLM data sets because you can fine tune from those. So that is encouraging that that can happen more and more. There's still some major challenges there and especially because most of the content that's being generated out of models is not in Central Siberian Yupik or one of these languages. Right. So we can't purely rely on those. But I think my hope would be that the larger foundation models see more linguistic diversity over time. And then there's these sort of grassroots organizations, grassroots efforts like Masakane and others that rise up kind of on the other end and say, okay, well we'll work with our language community to develop a data set that can fine tune off of these models and hopefully there's benefit both ways in that sense.

Since you mentioned Masakani a couple times, we'll drop the link in the show notes so people can find it. But what exactly do they do? How big of an impact have they had?

Yeah, I would say so if people aren't familiar, if you go to the link, you'll see they talk about themselves as a grassroots organizations of African NLP researchers creating technology for Africa. So we have our own kind of biases as people in an English driven, sort of literate world of what technology would be useful for everyone else like it. It probably makes sense for maybe listeners to say, well, wouldn't it be great if we could translate Wikipedia into all languages? Well, maybe. But actually the reality on the ground is that many language communities don't want Wikipedia translated into their language. That's not how they use their language. Or they're not literate and they're an oral culture. So they need speech. Right. Text won't do them any good. So that's why Masakane has started as this sort of grassroots organization of NLP practitioners who understand the context of the domain that they work in and are able to create models and systems that work in those contexts. There's others, you can hear them on the AI for Africa episodes that we have that talk about agriculture use cases. Agriculture use cases in the US might look like, you know, John Deere tractor with a cam. Like I don't know if people know this, but like John Deere tractors or these big tractors, they literally, they have a Kubernetes cluster on like some of them have a Kubernetes cluster on the tractor. It's like a at the edge kubernetes cluster that runs these models. And like when you're laying down pesticide, there's cameras that will actually identify and spray like individual weeds rather than like spraying the whole field. So that's like at the level that, you know, maybe is useful here in Africa, maybe the more useful thing is around disease or drought identification or disaster relief or other things like that. And so there's people working in those environments or in those domains that know those domains that are producing technology for those cases. And I think that's really important. So, yeah, I'd encourage people to check out Masakane and there's other groups like that. And if you're in the US or Europe or wherever and you want to get involved, there's open arms to say, hey, come help us do these things. So, yeah, get involved too.

What else is in your top three?

Oh, yeah. So one recent one from Raj Shah from Hugging Face. Some people might have seen his really cool videos on LinkedIn or other places. He makes TikTok videos about AI models, which is awesome. And his episode is called the capabilities of LLMs. And I thought it was really a good way to help me understand the landscape of large language models and the various features or axes that they're kind of situated in. So one axis is, for example, closed or open. Can I download the model? But then on top of that there's another axes, which is, is it available for commercial use or is it not? And then there's other axes. Like we already talked about multilinguality, but then there's like task specificity, right? Like there's Codegen models and there's language generation models, and there's, of course, image generation models and all of those as well. So, yeah, I think that episode really helps set a good foundation, no pun intended, for language models to understand where they're situated. So you can kind of when you go to Hugging Face, and there's. What is there like 200,000 models now? Maybe there's. I don't know how many models there are. How do I, like, navigate that space and understand what I could pull down? Or do I fit into one of those use cases where it makes sense for me to just connect to OpenAI or Cohere or Anthropic, helps kind of situate yourself. So I think that's why that episode was so popular, as he kind of lays all that out in an understandable way.

How do you personally stay on top of models? There's leaderboards, there's Twitter, there's LinkedIn.

Yeah, I think it's a little bit spread out for me between the sources that you mentioned as podcasters. I think that's one of the job. Well, it's also a benefit for us. I think if I didn't have. Every week on Wednesday I'm going to talk about this topic, whether I'm planning to think about A certain thing or not, it kind of helps you prompt and look at what's going on. So I think that is an advantage of content creators. It is kind of a responsibility, but it's also an advantage that we can have to have the excuse to have great conversations with people every week. But, yeah, I think Twitter's a little bit weird now, as everybody knows, but it's still a good place to find out that information. And then sometimes too, like, to be honest, I go to Hugging Face and I'll search for models, but I also search and I look at the statistics around the downloads of models, because generally when people find something useful, then they'll download it and download it over and over. So sometimes when I hear about a family of models, I'll go there and then I'll look at some of the statistics on Hugging Face and try a few things.

Yeah, and some of these forks, I see the download numbers, but I've never heard of them outside of Hugging Face.

Yeah, it's true. It's true. Yeah. And some of them, like, there'll be a fork or like a fine tune or something. And you do have to do a little bit of digging around, like licensing and that sort of thing too. But it is a useful. Like, there's tons of people doing amazing stuff out there that aren't getting recognized at the, like, you know, Falcon or MPT level. But there's a lot of people doing cool stuff that are releasing models on Hugging Face, maybe that they've just found interesting.

Any unusual ones that you recently found?

Well, there's one that I'll highlight, which I thought was cool because I don't know if you all saw the META released this.

I think it was like six modality model.

Yeah, yeah. And it was interesting because we did this work with Masakane. When I was at sil, we did this work with Masakane and Koki, which is speech tech company, to create these language models in like six African languages. And I was like, okay, that's cool. Like, we did that. We form, like, the data sets. It was satisfying, but now I'm like learning that. Then Meta went and found that data on Hugging Face and that's kind of incorporated in these new models that Meta has released. So it's cool to see like the full, full cycle thing happen where there was grassroots organizations seeing a need for models, gathering data, doing baselines, and now there's like extended functionality in kind of like a more influential way, I guess, at like, that higher level.

Yeah, yeah. I think, I mean, talking about open and closed models. When we started a podcast, it kind of looked like a cathedral kind of market where we had cohere anthropic OpenAI stability. And those are like the hottest companies I think now as you mentioned, you go on hugging face. I just opened it right now. There's this Others News research. 13 billion parameters model that just got released, fine tuned on over 300,000 instructions. It's like models are just popping up everywhere, which is great. And yeah, we had an episode, as I mentioned, with Jonathan Frankel and abhinav from Mosaic ML to introduce MPT7B and some of the work that they've done there. And I think like one of their motivation is like keeping the space as open as possible, like making it easy for anybody to go obviously ideally on the almost a MLS platform and train their own models and whatnot. So that's one that people really liked. I thought it was really technical, so I was really a little worried at first. I was like, is it going to fly over most people's head? But it was actually.

We're going more technical.

Exactly. Now that was a leaning in. Exactly.

And Jonathan is super passionate about open source. He had this rant halfway through the episode about why it's so important to keep models open and I actually edited in a crowd applause into the podcast, which I kind of love. I love little audio bonuses for people listening along and I think the Changelog guys do that really well, especially in their newer episodes.

Yeah, we need to. There is a way for us to integrate some of those things. Yeah, like the soundboard thing. And we've never got into it too much. I need to work with Jared from the Changelog and see, it just spices it up. Exactly, exactly.

You can only have so many hour long conversations about ML.

Yeah, I keep thinking that, but then we keep going. Right, right, right, right.

Sorry, I didn't mean like it was like. No, it switches it up and makes it audio interesting to add variety.

Cool.

I don't know if there are any other highlights that we want to do for.

I'll just highlight maybe one more. Kirsten Lum was on. She had an episode about machine learning at small organizations. I think that's a great one. If you're a data scientist or a practitioner or an engineer at either a startup or a mid sized company where I think the thing that she emphasized was these different tasks that we think about, whether it's curating a data set or training a model or fine tuning a model or deploying a model, sometimes at a larger Organization, those are functions in and of themselves. But when you're in this sort of mid range organization, that's like a task you do, right? So to think about those tasks as tasks of your role and time, box them and understand like how to do all of those things well without getting sucked down into any one of those things, that was an insight that I found quite useful in my day to day as well, is to sort of start to get a little bit of like spidey sense around. Hey, I'm spending a lot of time doing this, but which probably means I'm like stuck in too much. Like I'm making my ML ops too complicated, right, to track versions and like tie all this stuff together. Maybe I should just like do a simple thing and like paste a number in a Google sheet and move on or something.

I think that's a good segue into some of the other work that you do. You run the Datadan IO website, which is kind of like a different types of workshop and advising that you do. I think a lot of founders especially are curious about how are companies thinking about using this technology. There's a lot of demos on Twitter, a lot of excitement. But when founders are putting together something that they want to sell, they're like, okay, what are the real problems that enterprises have? What are some of the limitations that they have? We talked about commercial use cases and something like that. Can you maybe talk a bit about 2, 3 high level learnings that you had from these workshops on how these models are actually being brought into companies and how they're being adopted?

Yeah, I think maybe one higher level comment on this is even though we see all these demos happening, everybody's using ChatGPT. The reality in enterprise is most enterprises still don't have LLMs integrated across their technology stack. Right. So that might be a bummer for some people. Like, oh, it's not quite as pervasive, but I actually find it as refreshing maybe because some of us feel like stuff happens every week, it's exhausting to keep up. Like, oh, if I don't keep up with this stuff then I'm getting left behind. But it takes time for these things to trickle down. And not everything like we were talking about the stable diffusion use case and others, not everything that's hyped at the moment will be a part of your day to day life forever. Right. So you can kind of take some comfort in that. I think it's really important for people to, if they're interested in these models, to really dig into more than Just kind of a single prompt into these models. The practical side of using generative text models or LLMs really comes around either what some people might call prompt engineering, but you know, understanding things like giving examples or demonstrations in your prompt, using things like guardrails or, or regex statements or prediction guard to structure output, doing fine tuning for your company's data. These things go. There's kind of a hierarchy of these things. I think you all know Travis Fisher. He was a guest on Practical AI and talked about this hierarchy from prompt engineering through data augmentation to fine tuning to eventually like training your own generative model. I've really tried to encourage enterprise users and those that I do workshops with to think something like that hierarchy with these models, like get hands on, do your prompting. But then like, if you don't get the answer that you want immediately, I think there's a tendency for people to say, oh well, it doesn't work for my use case. But there's so much of a rich environment underneath that with things like LangChain and Llama index and you know, data augmentation, chaining, customization, fine tuning, like all this stuff that can be combined together. It's a fun new experience. But I find that enterprise users just haven't explored past that very most shallow level. So I think, yeah, in terms of the trends that I've seen with the workshop, I think people have gone to ChatGPT or one of these models. They've seen the value that's there, but they have a hard time connecting these models to a workflow that they can use to solve problems. Before we all had intuition. I'm going to gather my data. It's going to have these five features. I'm going to train my scikit learn model or whatever. I'm going to deploy it with Flask. And now I have a cool thing. Now all of that intuition has sort of been shattered a little bit. So we need to develop a new workflow around these things. And I think that's really the focus of the workshops is kind of rebuilding that intuition into a practical workflow that you can think through and solve problems with.

Practically you have a live prompt engineering class. Prompt engineering overrated or underrated?

Yeah, I think prompt engineering as like a term is probably too hyped. I think engineering and ops around large language models though, is, is a real thing and it is sort of what we're transitioning to now. How much you want to say is like that term gets used in all sorts of different contexts. It could mean just like, oh, I wrote a good prompt and I'm going to like sell it on Twitter or something.

Prop base the marketplace of prompts. I wonder how they're doing it to be hon because they get quoted in almost every article about prompt engineering. They got really, really good pr.

Yeah, yeah. I mean, if people can sell their prompts, I mean all for that, that's cool.

I got prompts right here.

But I think it goes like some people might just mean that and I think that's maybe overhyped in my view. But I do think there's this whole level of engineering and operations around prompts and chaining and data augmentation that is a real workflow that people can use to solve their problems. And that's more what I mean when I'm referring to however you want to combine the word engineering with prompting and language models.

Yeah, I've just been calling it AI engineering.

AI engineering. That's good.

Wrangle with the AI APIs, know what to do with them. That is a skill set that is developing, that is a subspecialty of software engineering. Yeah, it is what it is. And I think part of something I'm really trying to explore is this spillover of AI from the traditional ML space, like where you need a machine learning researcher or machine learning engineer. It's spilling over into the software engineering space and there's this rising class of what I'm calling AI engineer that is specialized in conversant in the research, the tooling, the conversations and themes.

What do you think are the unique challenges that someone coming from that latter group, engineers that are advancing into this AI engineer position versus probably more like my background where I was in data science for some time and now I'm kind of like transitioning into this world. What do you think are the unique challenges for both groups of people?

Oh, I mean, so I can speak to the software side and you can speak about the data science side. It's simply that we are for many of us dealing with a non deterministic system for the first time. That by the way, we don't fully control because there's this conversation about did GPT4 regress in its quality? And we don't know because model drift is not within our control because it's a black box API from OpenAI. But beyond that there's this sense that the latent space of capabilities is not fully explored yet. There's 175 billion or 1 trillion parameters in the model. We're maybe using 200 of them literally. Is that meme where we're using 10% of our brain, we are probably using 10% of what is capable in the model and it takes some ingeniousness to unlock that.

Yeah. I think from the data science perspective there's probably a desire too quickly to jump to these other things around fine tuning or training your own models, where if you really do take this prompting chaining, data augmentation seriously, you can do a lot with models sort of off the shelf and don't need to jump immediately into training. So I think that is like a knee jerk reaction on our end. And fine tuning is going to be around for the foreseeable future as far as I can tell. But data scientists have maybe a different because we've been dealing with the uncertainty or non deterministic output for some time and have developed some intuition around that. But that's mostly when we've been controlling the data sets, when we've been controlling the model training and that sort of thing. So to throw some of that out but still deal with that, it's a separate kind of challenge for us.

I just remembered another thing that we've been developing on the latent space community, which is this concept of AI ux. Right. That the last mile of showing something on the screen and making it consumable, easily usable by people is perhaps as valuable as the actual training of the model itself.

Yes.

So I don't know if that's an overstatement to be honest. Obviously you're spending hundreds of millions of dollars training models and putting it in some kind of react app. It's not the biggest innovation in the world, but a lot of people from OpenAI say ChatGPT was mostly a UX innovation.

Yeah. I think leading up to ChatGPT, when I saw the output of ChatGPT, I don't think I had the same earth shattering experience that other people had in believing, oh, this output is coming from a model. Sure it came from a model, but the reception to that interface and the human element of the dialogue. So maybe it's both. And right, you're not going to get that experience if you don't have the innovation under the hood in the modeling and the data set curation and all of that, but it can totally be ruined by the ux. I typically give the example one day in Gmail I logged in and I was typing my email and then had the gray autocomplete. Right. I did not get like a pop up that said like do you want us to start writing your emails with AI? It just was so smooth and it happened and it made like it created value for me instantly. Right. So I think that there is really a sense to that, especially in this area where people have a lot of like misgivings or fear around the technology itself.

And we're going to have Alex Gravely on in a Future episode. But GitHub, when they had the initial Codex model from OpenAI, they spent six months tuning the UX just to get copilot to a point where it's not a separate pane, it's not a separate text box, it's kind of in your code as you write the code. And to me that's traditional. That's more the domain of traditional software engineering rather than ML engineers or research engineers.

Yeah, yeah. I would say that is probably yes. To circle back to what we were talking about, like challenges that are unique to engineers coming into this versus data scientists coming into this. That's something data scientists I think have not thought about very much at all. At the very most it's data visualization that they've thought about. Right. Whereas engineers generally there's some human. I mean unless you're just a very pure backend systems engineer thinking about UI UX is maybe a little bit more natural to that group.

You mentioned one thing which is about dataset curation. We're in the middle of preparing this long overdue episodes on datasets 101. Any reflections on the evolutions in natural and NLP data sets that have been happening?

Yeah, great question. I definitely, like, I think are you all familiar with Label Studio and that it's one of the most popular kind of open source frameworks for data labeling and they've been. I think they've been on. We have them on the show like we try to have them on the show every year as like data labeling expert. Maybe it's time for that. It's just reminding me they just released.

So Aaron McHale is in the latent space Discord. I think you had her on at odsc.

Yeah, she was at odsc.

So they just release new tools for fine tuning generative AI models.

Exactly. Yeah.

It's location.

I think maybe that being an example of this is maybe a trend that we're seeing there is around augmented tooling or tooling that's really geared towards an approachable way to fine tune these models with human feedback or with customized data. So like I know with Label Studio a lot of the recent releases had somewhat to do with like putting LLMs in the loop with humans during the Label Process similar to like I think Prodigy has been doing this for some time, which is from spacy. So this sort of human in the loop labeling and update of a model, they brought some of that in. But now this new kind of set of tooling around, specifically instruction tuning of models, I think before maybe people and I've seen actually this misconception. I was in an advising call with a client and they're really struggling to understand like, okay, our company has been training or fine tuning models. Now we want to create our own like instruction tuned model. How is that different from what we've been doing in the past? And kind of what I tried to help them see is yes, some of the workflow that happened around reinforcement learning from human feedback is unique. But reinforcement learning is not unique. There's an element of training in that, there's data set curation in that, there's pre training that happens like before that whole process happened. So the elements that you're familiar with are part of that. They're just not packaged in the same way that you saw them before. Now there's this clear pre training stage and then the human feedback stage and then this reinforcement learning happens. So I think the more that we can bring that concept and that workflow into tooling, like what Label Studio is doing to make it more approachable for people to where it's not like this weird reinforcement learning from human feedback sounds very confusing to people like PPO and helping people understand how reinforcement learning works, it's very difficult. So the more the tooling can just have its own good UI UX around that process, I think the better. And probably Label Studio and others are leading the front on leading the way on that front.

I was thinking. So labels are one thing and by the way, okay, I'll take this side tangent on labels and I'll come back to the main point. I actually presume that scale would win everything.

Yeah.

And it seems like they haven't. Yeah, sorry. There's scale, there's snorkel, there's this generation.

Of labeling companies that's data centric AI companies.

Right. How come there's still new companies coming out? There's Label Box, there's Label Studio. I don't have a sense of how to think about these companies. Obviously labels are important.

Yeah. I think also even before that there was at least features, even from cloud providers or whatever. Like AutoML came before that, like upload your own data, create your own custom model. So I think that maybe it's that companies that want to create this sort of custom models. And this is just my own opinion, I'll preface that. Maybe they don't want like when they're thinking about that problem, they're not thinking about, oh, I need a whole platform to create custom models using our data. They're more thinking about like, how do I use these state of the art models with my data? And so it's still if those statements are very similar, but if you notice, like one is more model centric and one is more data centric. So I think enterprises are still thinking like model centric and augmenting that with their data. Whether that be just through augmentation or through fine tuning or training. They're not necessarily thinking about a Data platform for AI, they're thinking about bringing their AI or their data to the AI system. Which is why I think APIs like Cohere OpenAI that offers fine tuning as part of their API. It's sort of like people love that it makes sense, like, okay, I can just upload some examples and it makes the model better. But it's still like model centric, right?

Yeah, I get the sense that OpenAI doesn't want to encourage that anymore because they don't have fine tuning for 3.5 and 4. And then so the last thing I'll do about data sets and we can go into the lightning round is I was actually thinking about unlabeled data sets for unsupervised learning or self supervised learning. Right. That is something that we are trying to wrap our heads around, like common crawl, stack overflow, archive the books. I don't know if you have any perspectives on that, like the trends that are arising here, the best practices. As far as I can tell, nobody has a straight answer as to what the data mix is and everyone just kind of experiments.

Yeah, well, I think that's partly driven by the fact that the most popular models, you don't really have a clear picture of what the data mix is. Right. So the people that are trying to recreate that and they're not achieving that level of performance. Right. Then one of the things they're thinking about is, well, what are all the different data mix options that I can try and try to replicate some of what's going on. Right. So I think it's partly driven by that is like we don't totally know what the data mix is like sitting behind the, the curtain of OpenAI or others. But I think there's a couple of trends, I guess, which you've already sort of highlighted. One is like, how can I Mix up all of these public data sets and filter them in unique ways to make my model better. So some, I listened to a talk, I believe it was at last year's ACL and they did this study of Common Crawl and they found that actually a significant portion of Common Crawl was mislabeled all over the place. Right, like trash. Yeah. So like I think it was 100% of the data that was labeled as Latin character Arabic. So Arabic written in Latin characters was not Arabic, like 100% of it. And there was like all sorts of other problems and that sort of thing. So I think there's one side, one group of people or set of experiments that you could think about as like, how do I take these existing data sets which I know have data quality issues or maybe other data biases or problems that I would like to filter out, like not fit for work data, that sort of thing. So how do I create my own special filtered mix of these and train a model? So that's one kind of genre and then there's the other genre which is like maybe taking those but augmenting them with this like simulated or augmented data. Right. This, that's out of a model, like a GPT model or something like that. So I think you could combine those in all sorts of unique ways. And I think it is a little bit of like the Wild west because we don't totally have a good grip on what is the winning strategy there. And so I think that's where I would also encourage people to try a variety of models. So this is maybe a problem with benchmarks in general, right? Like you can see like a, the open large language model benchmark on hugging face and these models are at the top and you could come away with that and say, well, anything below the top three I'm not even going to use. But the reality is that each of those had a unique flavor of this data under the hood that might actually work quite well for your use case. One example that I've used recently in some work is the Camel 5 billion model from Rider. You know, it doesn't work great for a lot of things, but there's certain things around like marketing copy and others that it does a really good job at. And it's a bit smaller model that I can host and run and I can get good output out of it if I put in some of that workflow and structuring around it. But I wouldn't use it for other cases. But that has a lot to do with the data. And you know, I'M guessing writers focus on that copy generation and such. So yeah, I would encourage people specifically on this topic to maybe think about what's going on under the hood and also give some models a try for different. Like gain your own intuition about how a model behavior might change based on like how it was trained and the mix of data that went in.

Awesome. Let's jump into the lightning round. We have three questions for you. It's lightning, but you can take 30 seconds to answer.

All right, cool.

So the first question is around acceleration. What's something that already happened in AI that you thought would take much longer?

Yeah, I think the thing that I was thinking about here was how general purpose, these large language models are beyond traditional NLP tasks. So it doesn't surprise me that maybe they could do sentiment analysis or even NLI or something like that. These are things that have been studied for a long time. But the fact that I can, like at odsc, I was in like a workshop on fraud detection and they were using like some, I forget the models. They were using some statistical models to do fraud detection. I was like, I wonder if I just like do a bit of chaining and like insert some of the examples of these insurance transactions into my prompts if I can get the large language model to detect a fraudulent insurance client. And it seemed to like, I got pretty far doing that. So that fact of you can do something like that with these models are that generalizable beyond traditional NLP techniques I think is surprising to me.

Awesome exploration. What are the most interesting unsolved questions in AI?

Yeah, I think there is still such a focus on English and Mandarin. It's like that, like you're kind of large language model wise. If you look at the drop off in performance after you get past like English, Mandarin, German, it's Spanish to some degree, but German is actually better than Spanish because of how much it's been studied in nlp. And of course Mandarin has a lot of data. Spanish still does good. But like there's languages even in the top 100 languages of the world that are spoken by millions and millions and millions of people around the world that don't perform well in these models. So that's like thing one. But even modality wise, I know there's a lot of work going on in the research community around sign language, but there's all of these different modalities of language. Written text does not equal communication. Right. Written text is a synthesis of communication into a written form that some people consume. But the combination of all of these modalities along with all of these languages. There's just so much room to explore there and so many challenges left to explore that will eventually, I think help us learn a lot about communication in general and the limitations of these models. But is an exciting area. It's definitely a challenge, but an exciting area.

Awesome, man. So one last takeaway. What's something or a message that you want everyone to remember today?

Yeah, similar to when you were asking about my workshops, I think I would just encourage people to get hands on with these models and really dig into the new sets of tooling that are out there. There's so much good tooling out there to go from like a simple prompt to inject your own data to form like a query index to create like a chain of of processing, even like trying agents and all those things like get hands on and try it. That's the only way that you're going to build out this intuition. So yeah, that would be my encouragement.

Excellent. Well, thanks for coming on.

Yeah, thank you guys so much. This is awesome.

Thank you for listening to Practical AI. Your next step is to subscribe now if you haven't already. And if you're a longtime listener of the show. Help us reach more people by sharing Practical AI with your friends and colleagues. Thanks once again to Fastly and Fly for partnering with us to bring you all Changelog podcasts. Check out what they're up to@fastly.com and fly IO. And to our Beat Freakin residents, Breakmaster Cylinder, for continuously cranking out the best beats in the biz. That's all for now. We'll talk to you again next time.