Actually, really helpful, thank you Google. Wondering how far this technology will go in the next couple of years, if it's this far already in a couple of months.
This was fantastic! While I've been watching The Full Stack LLM Bootcamp, I'm not technically strong enough to start there, and will use these Google Cloud Tech videos as a means to "jumpstart" my knowledge of LLM and Generative AI. This is a great general primer for students and colleagues!
@@ChatGTA345 unlikely. 1 or 2 small companies pursuing this tech with such ambition could be a waste of time. But if all the big players are investing their time and money in this tech, then it has to be something very real and very serious
@@zappy9880 I don't think that necessarily follows. The industry has followed so many hype waves before. The competitive advantage is actually not to do what everyone else does
@@ChatGTA345 well, it is a waste depends on how you use it. but can really be useful in several fields if you know how to use it and how you fine-tune it. just treat it as some sort of assisting tool as of now, and not as something that you actually use as some sort of definitive source of knowledge.
Appreciate the valuable content! Sharing some key takeaways of the video and I hope this can help someone out. 1) 00:50 - Large language models (LLMs) are general purpose language models that can be pre-trained and fine-tuned for specific purposes. LLMs are trained for general purposes to solve common language problems, and then tailored to solve specific problems in different fields. 2) 02:04 - Large language models have enormous size and parameter count. The size of the training data set can be at the petabyte scale, and the parameter count refers to the memories and knowledge learned by the machine during training. 3) 03:01 - Pre-training and fine-tuning are key steps in developing large language models. Pre-training involves training a large language model for general purposes with a large data set, while fine-tuning involves training the model for specific aims with a much smaller data set. 4) 03:15 - Large language models offer several benefits. They can be used for different tasks, require minimal field training data, and their performance improves with more data and parameters. 5) 08:50 - Prompt design and prompt engineering are important in large language models. Prompt design involves creating a clear, concise, and informative prompt for the desired task, while prompt engineering focuses on improving performance. 6) 13:43 - Generative AI Studio and Generative AI App Builder are tools for exploring and customizing generative AI models. Generative AI Studio provides pre-trained models, tools for fine-tuning and deploying models, and a community forum for collaboration. 7) 14:52 - Palm API and Vertex AI provide tools for testing, tuning, and deploying large language models. Palm API allows testing and experimenting with large language models and gen AI tools, while Vertex AI offers task-specific Foundation models and parameter efficient tuning methods. This takeaway note is made with the Notable app (getnotable.ai).
Finding answers to questions has become so much easier now with new tech. I have never been good at writing code, so this is a welcome change as far as I'm concerned! Look forward to more progress in technology.
Minor Correction @ 2:14. "In ML, parameters are often called hyperparameters." In ML, parameters and hyperparameters can exist simultaneously and serve two different purposes. One can think of hyperparameters as the set of knobs that the designer has direct influence to change as they see fit (whether algorithmically or manually). As for the parameters of a model, one can think of it as the set of knobs that are learned directly from the data. For hyperparameters, you specify them prior to the training step; while the training step proceeds, the parameters of the model are being learned.
@@praveenmadduri7181 In the context of neural networks, yes. When parameters are mentioned, one is usually referring to the weights/connections between nods.
Yes, they are different conceptually. Parameters are directly applied/calculated in the hypothesis or model; while, hyperparameters are somewhat heuristically decided based on what works. For example if you were figuring out how to get from home to office, the path details maybe calculated directly by the GPS, but the time at which you leave maybe heuristically decided by you. Another example of a hyperparameter can be how many backup cameras you choose to add should the main camera fail on a robot, there is no 'correct' number, it's more of a cost or design choice. In an ML transformer, choosing the number of encoders or decoders can be a hyperparameter. The parameters would be learned from the language training in the LLM.
1980s or so, there were telephone operator who connects those STD lines. Now they are vanished but their next gen kids are employed in another market. That's how innovation works!!
Anybody who read this comment, you'd want to type this prompt in Chat-GPT or Bard: "I have 15 liter jug, 10 liter jug, and 5 liter jug. How do I measure 5 liters of water?" ---> See what they answer
You first instantiate the model with randomly generated parameters (540B in this case) and use lots and lots and lots of data to make the model "learn" and modify these parameters so they are better. For llms, you need hundreds of powerful gpus and you need weeks or months to train such massive models. Falcon 40B which is a state of the art open source model with 40B parameters was trained for two months.
@@MrAmgadHasan chatgpt was trained for about 2 years , there are 2 seperate models within chatgpt , one to understand context, the other to predict the text 🤪
It can. Once you feed the base of information, it can learn from the questions themselves leveraging possible answers for accuracy. Hallucinations will happen but that's when you start fine tuning it with the correct answers that it could not find on its own or on its data base. A human can't learn everything on their own, we need to study content which is build over time through observation.
Hey there! AI is definitely becoming more prevalent on Facebook. I've noticed more personalized content and ads powered by AI algorithms. It's amazing how AI enhances our social media experience."
At 4:50 I did not understand the third point that the speaker made i.e. "Orchestrated distributes computation for accelerators". Can someone please explain?
you can use a new drive architecture sought via gpu pixels for proximity stream like to not need large.lamguage models, and use multi factor checks to reduce need of a lot of data..thank me now.
Thank you. I understood about half (optimistically) of it. I subscribed to the channel hoping to start from the beginning and understanding more. My ultimate goal: a LLM Librarian, combining the catalog of a library with results from internet search engine, giving the deepest answer possible.
I have an urgent question (school related) -> is LLM part of NLP? Is an LLM always an NLP model? Or can an LLM be another kind of model? "L" for Language in both kinds of models. Both in AI. Both for language. A colleague says LLM is not necessarily an NLP model but then I did not understand LLM and/or NLP and my oral exam is in few days omg
I've been extremely frustrated in my interactions with chatbots, they never seem to tell the truth and it's getting harder and harder to tell what's true from what's not. I honestly like regular Google searches much more!
This is a great overview video thank you. Do you have a reference for how to host open-sourced LLM's in Vertex AI (or other GCP tools)? Overall I'm looking for GCP tools and ways for turning open-source LLM's into API's to be used within our native cloud instance.
@@andrestorres7343 Yes, but it was really pricey since you have to host the underlying infrastructure. Usually large GPU virtual machines and on GCP depending on model size it was $2k - $5k per month to host an open source model. We are sticking with the API version of the big models because of this.
It's a custom built computer chip developed by google to perform matrix operations and train deep learning models. Think of them as gpus specialized for deep learning.
I understand the message of this slide to be not about prompt design, but AI response: that if the app in which the model is embedded first instructs the model to describe the process to get to an answer and THEN feed that back in with the original prompt, that the quality of the final response improves.