I have a few years of limited experience in LLM development, where I was also one of the first to publish work on Retrieval Augmented Generation (RAG). I've been conducting research on domain adaptation of models since 2016.
Latest ads
-
Colombo Red Hat Certified System Administrator (RHCSA) - RHEL 10- Sanjeewani95
- Updated:
-
NURSING , CAREGIVER , HOTEL & BEAUTY COURSES- IVA Para Medical Campus
- Updated:
-
Handmade Character Soft Toys Peppa Pig Family- anil1961
- Updated:
-
Video Content Creator
- pramukag
- Updated:
-
QA Engineer Intern
- pramukag
- Updated:
Ask anything related to LLMs and how they work. I will try to answer based on my knowledge.
- Thread starter AiLankan
- Start date
I barely Know about LLMs. But have some experiences using those. I have particalarly used them for creative writings such as songs and audio productions such as creating audible songs. Asumming those are some instances of LLMs. Or are they different kind of foundation models? Already the outputs are really natural. I want to know the process of outputting audio using language models or video or even image. do they use other songs, videos, images to train the models and are there any middle ground . I mean converting audio/video/image to some kind of text and train and convert back it again to desired format?
Machan, sorry, I will type in English, not because of Luku sceen ok
Artificial General Intelligence is the highest point in AI. Such an AI system can understand, learn, and perform any intellectual task that a human can do. Imagine an AI system that can do its own research and break all the barriers.
Then LLMS - Just think of LLMs as one way to achieve AGI (there could be many. But as researchers this is the best we have seen so far). Because these models can really understand the dynamics of the world via the vast amount of available knees.
Well, there are many foundation models, such as images, audio, and text. So, they are not exactly LLMs. But again, we can now have models that generate text, images, and voice. Chatgpt pretty much does it. Now researchers like to use the term "Omni Models".
As you correctly understand when it comes to videos and audio , they do need similar data sources to learn. It could also be a mix of data. That is what we call multimodal learning.
For example: you can take a YouTube video and feed it to an algorithm; with its transcript, you can get textual info, then, of course, the audio and video.
* We do not convert other modalities to text. When it comes to these foundation models, we convert all text, images, video, and audio into vectors, which means a stream of numbers that computers can read. Then we convert them back to text, audio, and video,
------ Post added on Jan 7, 2025 at 3:37 AM
Not at all. Same principals. Backpropagation, gradient descent, normalization, probability, and non linearly. Pretty much the basic machine learning concepts.
------ Post added on Jan 7, 2025 at 3:38 AM
Can a current domestic computer what we have at the moment host local LLM that can be used for personal use?
Thank you brother. I got it.Machan, sorry, I will type in English, not because of Luku sceen ok
Artificial General Intelligence is the highest point in AI. Such an AI system can understand, learn, and perform any intellectual task that a human can do. Imagine an AI system that can do its own research and break all the barriers.
Then LLMS - Just think of LLMs as one way to achieve AGI (there could be many. But as researchers this is the best we have seen so far). Because these models can really understand the dynamics of the world via the vast amount of available knees.
Well, there are many foundation models, such as images, audio, and text. So, they are not exactly LLMs. But again, we can now have models that generate text, images, and voice. Chatgpt pretty much does it. Now researchers like to use the term "Omni Models".
As you correctly understand when it comes to videos and audio , they do need similar data sources to learn. It could also be a mix of data. That is what we call multimodal learning.
For example: you can take a YouTube video and feed it to an algorithm; with its transcript, you can get textual info, then, of course, the audio and video.
* We do not convert other modalities to text. When it comes to these foundation models, we convert all text, images, video, and audio into vectors, which means a stream of numbers that computers can read. Then we convert them back to text, audio, and video,
------ Post added on Jan 7, 2025 at 3:37 AM
Not at all. Same principals. Backpropagation, gradient descent, normalization, probability, and non linearly. Pretty much the basic machine learning concepts.
------ Post added on Jan 7, 2025 at 3:38 AM
I thought both of them were the same thing
Yeah I know under the hood it's the same principles. But what I see is that LLMs or Generative AI generally has become to a place where those fundamental principles can be abstracted. Apart to the researchers, a considerable number of people work in GenAI haven't even heard of the word "Gradient Decent".
So is it possible to train a LLM to cater our needs and maybe even provide services to outsiders using it? How practical is it ay this moment?
Yes, it's possible to train an LLM for specific needs, but it's somewhat challenging due to hardware limitations. While you can work with smaller base models to build customized LLMs, it requires significant investment to finalize and deploy a model for your requirements. There are various services available to help with this process, including OpenAI's fine-tuning options.
The practical feasibility depends on your specific use case, budget, and technical requirements.
https://platform.openai.com/docs/guides/fine-tuning
https://www.anthropic.com/news/fine-tune-claude-3-haiku
Yeah surely.. Most of small LLMs are catching up with bigger models. I myself working in a small language model startup in the valley.
Tbh adaptation is the key engineering skill in the near future.
@Scarface
------ Post added on Jan 7, 2025 at 7:30 AM
Yeah mate kinda smaller versions or CPU-friendly checkpoints versions. You can check apple’s MLX library. They are doing some dope work on the on-device.
------ Post added on Jan 7, 2025 at 7:35 AM
Similar threads
