OF LLMS IN SLMS

@gojutechtalk

4.2K followers · usually 486 views a post

Views

4.7K

Likes

124

Comments

Why it worked

The video explains a complex technical concept (specialized SLMs) using a relatable analogy (video games) and a clear, concise explanation of the trade-offs involved. This makes the information accessible and engaging for a broad audience.

Summary

The speaker discusses the concept of specialized SLMs (Specialized Language Models) as a trade-off for efficiency, comparing it to how gamers manage their game libraries. Instead of running all capabilities at once, users will activate specific SLMs for tasks like coding, writing, or translation, making the solution more manageable and effective.

Structure

1Introducing the concept of specialized SLMs
2Explaining the trade-off between generalizability and efficiency
3Comparing SLMs to video games and their management
4Describing the practical application of activating specific SLMs for tasks
5Concluding that this approach makes the solution amendable and workable

On-screen text

OF LLMS IN SLMS Real science has trade-offs: SLMs are only viable if we specialized them. And that's okay. ABSOLUTELY I THINK THAT LIKE BASICALLY TRYING TO DISTILL OUT AS OF THE CAPABILITIES WE CAN ONLY WAY TO SPECIALIZED WE COULD TALK IN GRANDIOSE IDEALISTIC TERMS AND SAY HAVE ALL THE STUFF THAT LLMS HAVE IN SLMS WONDERFUL" AND THEN KICK IN BE ABLE TO DO THAT BECAUSE IT GENERALIZED TRADE-OFFS IN ACTUAL SCIENCE TRADE-OFF IS THEM IN ORDER FOR THEM TO RUN LOCALLY COMPUTE AND LOCAL SPACE HAVE TO REMOVE SOMETHING RIGHT SOMETHING IS GENERALIZABILITY BUT I THINK THAT'S TOTALLY FINE BECAUSE PEOPLE USING AI MODELS I THINK OFTENTIMES FOR TASKS THEY'RE NOT-- WORK DOMAINS AT ONCE PROGRAMMING SLM BE AVAILABLE DOING CODING HAVE THE WRITING SLM KICK IN WRITING EMAILS OR PROPOSALS OR SOMETHING HAVE THE TRANSLATION IN WHEN THEY'RE WRITING ARTICLES THAT GONNA GO TO THEIR FRIENDS WHATEVER IN COUNTRIES THAT LANGUAGES SO YOU WILL ROTATE IN OUT OF THESE SLMS BASED ON WHAT TRYING TO DO THINK LIKE VIDEO SO FOR ALL OF VIDEO GAMERS HAVE 100 DIFFERENT VIDEO GAMES IN YOUR STEAM PLAYING 100 GAMES AT W- YOU'RE NOT OKAY YOU ARE FOR ME I'M MOSTLY 5 GAMES THOSE GAMES CAN SWITCH WHEN I HAVE BUT I DON'T THEM ALL RESIDENT IN MEMORY ALL AT ONCE PHILOSOPHY IS WHAT DO WITH SLMS IS THAT WE'LL HAVE SLMS FOCUSED IN DOMAIN BUT WE'LL ONLY THE ONES FOR TASKS THAT ON MAKES IT AMENDABLE THAT MAKES WORK

Original caption

The idea of distilling LLM capabilities into specialized SLMs for local use is gaining traction. This approach trades generalizability for efficiency, allowing users to switch between models based on specific tasks like coding, creative writing, or translation. It's like having a library of video games – you access the one you need, when you need it. Link to full live stream: https://youtube.com/live/xyBHj2b-oNY #AI #LLM #SLM #MachineLearning #TechnologyTrends #FutureofAI

More from @gojutechtalk

Their other posts in the index, biggest breakout first.

Transcript

OF LLMS IN SLMS. Yes, absolutely. I think that that's like basically trying to distill out as much of the capabilities that we can. And I think the only way to do that is to make them specialized. I mean, we could talk in grandiose idealistic terms and say we're going to have all the stuff that the LMs have in SLMs. It's going to be wonderful and all that and then reality's going to kick in. And we're not going to be able to do that because it it they're generalized and there are tradeoffs. in actual science and the trade-off is in order for them to run locally with local computer and local space you have to remove something right and that something is the generalability but I think that that's totally fine because when you think about people using AI models I think often times for tasks they're not they don't need it to work on a hundred different domains at once have the programming SLM be available when they're doing coding. Have the creative writing SLM kick in when they're writing emails or proposals or something. Have the language translation one kick in when they're uh writing articles that are going to go to their friends or their customers or whatever in other countries that speak different languages, right? So you you will rotate in and out of these SLMs based on what you're trying to do. The way I think about it is almost like video games, right? So for all of you hardcore video gamers, you probably have a hundred different video games in your Steam. But are you playing 100 games at once? You're not. Okay, maybe some of you are, right? But for me, I'm mostly five games. But I have all those games that I can switch over to when I have the desire, but I don't need them all president in memory all at once. The same philosophy is what we're going to do with SLMs is that we'll have all of the SLMs that are focused in the different domain, but we'll only activate the ones for the task that we're working on. And then that makes it amendable. That makes this solution work.