Large Language Models: How Large is Large Enough?
Vložit
- čas přidán 14. 12. 2023
- Explore IBM watsonx → ibm.biz/IBM-watsonx
When it comes to large language models, one of the first things you may read about is the enormous size of the model data or the number of parameters. But is "bigger" always better? In this video, Kip Yego, Program Marketing Manager explains that the truthful answer is "It depends". So what are the factors that will drive this decision? Kip breaks them down and explains one-by-one how to decide.
Get started for free on IBM Cloud → ibm.biz/sign-up-now
Subscribe to see more videos like this in the future → ibm.biz/subscribe-now
#llm #ai #generativeai #ml
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
These videos are fantastic!
Thank you so much for making them available :D
Thank you, this is the information I was searching for. I was explaining the concept in theory to someone. The idea was to use smaller models that are trained for specific domains. By eliminating or reduce all the other domains, the model should perform better and reduce messy results.
Well done video, Get kip to more of these please.
Great video, super informative!
Great video Kip!
At the moment it seems that bigger equals better. Time to change that perception accordingly
already the trend, see mixture of experts concept
0:16 🤣 You tell me.
INTERESTING. 😀
Thank you, that was informative. One question I have is how you determine domain specificity, and perhaps potential lost opportunity?
For example, using financial services tasks as in your example. If you ask someone working in finance about what insights they'd be looking for, tax or perhaps transfer pricing may not be what they consider as part of their domain. However, transfer pricing and tax could have a huge impact on what finance should consider when taking decisions. How do you ensure the domain specificity is not too narrow?
I believe anything remotely related to the domains should be included in the training data. He didn't talk about the training data size, but I believe it should still be very big (but smaller than a general one)
The MBA in me says, beyond some point, the trade-off isn't worth it. Then again, that's probably what they said about the Apollo mission.
It really depends
How can we find a domain specific models or how to train them?
I think that's the next business idea, lol... At this point, pick an industry, and create specific domain model! It's a race! Also, specific domain Vector Databases will be needed!
My llm's so large, it reaches almost every 1 and 0 it can write on; you can literally call it a "wipe"
5:38 THANK YOU BRO. Definitely feel more confident after hearing that.
The Bloke
Lets go phi. chroot chroot chroot
I see no point whatsoever in comparing a domain specific finetuned model to a non finetuned model to draw conclusions or suggest any insights doing this.
bro you got worse handwriting than me!!! Good info though. lol
Kip, that’s a very poor analogy.