How to get started with LLMs today

Key Takeaways

in this space of everything that's happening at the moment I presume people know how to go to open Ai and start interacting with chat GPT um right but I'm wondering for people you know programmers and scientists and data scientists and ml people who maybe haven't had a lot of experience with it um how can they get started with playing around with large language models today yeah yeah of course there are kind of different levels so um obstruction Independence I guess how much code you're willing to write you know much um time you have to put into understanding what's going on what's going on behind the scene I guess because there are also like different methods to interact with a language model there are different kind of pipelines to try in a language model if you have data so I think like the main reserves if you want to understand a bit what's going on without like Drop too deep into uh the kind of entire difficulties that like I felt like machine learning and open is still a new face certain phases like this um kind of chords in which you can go through the fine tuning of uh supervised language model for for example classification tasks uh that's gonna give you like a very good intro on now to actually have your own maybe small large language models on your own laptop doing things and that could be already good enough to do many different tasks it's very controllable you can do whatever you want with it and and it's usually run on consumer AdWords most of the time it might be slow but at least it's it's enough to to start playing with it well right now starting like you I'm just adding a link to I mean I presume a lot of people know about hugging face but I've just included a link in in the chat for those interested once again if any questions come up do feel free to ask them in in the YouTube chat as well sorry please go and fit it yeah sure sure and like but right now with all this like huge development there are kind of a lot of packages that drop apis for you that are coming out for example there is a land chain that is kind of an abstraction that you can use to kind of seamlessly um let me use the word talk with multiple language models uh without kind of like having to drag into different kind of setups different kind of systems building kind of like first layers of I don't know safety so long chain is another thing that I think it's like uh very interesting to look at because it's it's it's not like a a a scientific revolution but it's very useful it's like um empirically like it's a good engineering effort to make uh language models accessible to the broader audience and finally if you kind of want to uh maybe write uh some code that is mostly API I think that like using four years or like open AI apis is still unaffecting way to deal with language models you can prompt for questions ask for specification there is an entire field that is called prompt engineering that is kind of the art of asking questions to language models that is becoming more and more relevant there are courses coming up that are that have been like advertised everywhere on most social media platform and this is the direction of like talking directly with the model using for example apis great and people can check out the open Ai apis and I've just um shared a link to cohere and uh Lang chain as well uh Yuri Gordon has a great question I'm actually we're going to get to this type of thing later actually Yuri but I do want to foreshadow this year is asking about the security measures and tools to handle different risks oh pardon me I've got a bit um a bit of hay fever I'm getting so excited um uh but um and yeah Akio has a has an interesting question around integrating um GPT into Ides and that being useful for programmers he also says is this a cause of fear in programmers that AI will take over um I'll I'll maybe say a few words about that um because I do think talking about code assistance you've mentioned co-pilot but um in all honesty man like if I can get stuff like chat GPT to generate templates and basic you know control flow stuff and you know um skeletons like I want to make sure in all the business logic is make sense and all the domain expertise that I need to inject into my software works but I'm all for you know um having a coding assistant helping me with the stuff especially the code I don't enjoy writing but anything business critical of course like let's actually I'll take a a non-controversial analogy if I'm if I'm building a bridge um I'm a structural engineer and I'm building a bridge and I need to figure out how thick to make a pylon to support a part of the bridge and I need to make sure that it's you know it supports this part of the bridge I'm not going to Outsource the measurements of that to something like chat GPT or a large language model maybe I'll get it to do the basic design and these these types of things but anything that would result in the bridge collapsing I'm going to make sure that I figure out by hand right so anything that's super Mission critical like that um so maybe you can speak about that in in the software world uh briefly for that yeah I think like I wouldn't kind of give anything that is Mission critical to judge PPT to develop or at least if I do I would then like check it manually uh before uh ever deploying it when I say that like I think co-pilot is is useful is in the context of helping you uh writing code that takes long to write but it's not like highly complex or um kind of like very sophisticated domain kind of code I don't think kobalo can can yet like do domain modeling for you uh you still have to come up with like your own uh object-oriented infrastructure and like you need to organize process in a meaningful way that respect your kind of flow but what is very useful as a as as um as what a developed is this kind of like very powering task that takes a lot of boilerplate boilerplate code that everybody has a victim and it's already on GitHub and that's autopilot and that's why copilot example like to efficiently write that uh directly for you yeah definitely like I don't expect like software develop meant to disappear soon as I think I expect um as being much uh um or a pre software developer to be much more productive with this kind of tools and I expect people to work more on verifying the code that this kind of to come up with