ChatGPT: In-context Retrieval-Augmented Learning (IC-RALM) | In-context Learning (ICL) Examples

Key Takeaways

how can we teach chat GPT new information new knowledge since we know that after 2021 it has a very limited knowledge amount so very easy before chat gbt4 comes out you can ask it and it tells you well fine-tuning is done pre-training is done so there's no way you can provide me further information but you know chat GPT is cheating here a little bit so let's have a look how we can teach jet GPT new information last week I showed you here crash course on AI prompt here's the YouTube video and we discussed very shortly that there is now a paradigma shift from our pre-training Like a Bird model and a defined tuning for each specific Downstream task with another data set we have now the option with those huge allenzymes of free training prompt and predict and from prompting I showed it a minimum 20 methods available so what we are then discussed very shortly fine-tuning or the alternative is prefix tuning where you have only a small amount of parameter with prefix tuning that you really train on and as I told you if we'll want to learn how we can do this this was the example I've given you you will have a source sequence and you have a Target sequence and with the source sequence we have question then we have some context to the question and an answer to the question it consists of sentences the target sequence we wanted to predict the answer give me our answer here to the question giving here the question here given the context the overall context and I told you this is for example a chapter in your biotechnology book is yes so this is for example how Microsoft trained bio GPT and we are gonna Focus today on a very simple Source sequence just question an answer and if you want to see the research paper for this here from January 2023 from AI 21 Labs there's a beautiful research paper uh it's about in context retrieval augmented language model or large language model and they come up with a new alternative and they call it in context retrieval augmented language model what it is they have an llm and they add some outside information set for example from the internet some up-to-date information and they discuss the mechanism how they can do this in particular they leave the pre-trained language model without any modification or even without an API access and this is exactly what we're gonna do we leave our chat GPT as it is and we will apply this in context retrieval augmented language modeling methodology January 2023 uh you can have different uh architectures you they work with an off-the-shelf Frozen broad retriever I do it manually you can do whatever argumentation or whatever architecture we would like to just want to show you how easy it is so they give you the framework for in context ROM they show you here the aggress the auto regressive model this is whatever you know and then they just say add some documents from an external Corpus and the external Corpus is simply the information I copied personally from the internet and I provide to them and I would just want to show you how easy it is look this is here where they say this is an example of in context one we simply prepend the retrieve document so this here is a sentence that they retrieve from the internet so you can do it automatically from your browser from an AI system or you just type it in like I did and then you you put we simply prepared to retrieve document before the input prefix then you have your prefix the World Cup 2022 was the last Rift 32 teams before the increase too and this is now the open Word and as you can see with the retrieved evidence now pre-pended before the prefix and your feet is in the language model the language model is able to give you an answer that before the increase to 48 in the 2026 tournaments so here we go now with our free chat GPT and you see I have exactly here this this structure of my prompt let's have a look at this oh yes just the second gives me an answer what I want to show you the prompt design so here you see I have a question and I have an answer and this question is simply what new information about Quantum error correction is available this is a question chat CPT will know but it will not know the answer because the answer I have taken here as you can see from Google here from February 2022 2023 this is an article about the latest development in Quantum Corrections and what I'm particularly looking for I'm looking for an uh a clear event and here we have it here we have a logical qubit that were made from 49 fizzle qubit was able to outperform one that we made from 17 Cubit so we have very clearly two integer numbers that we can now ask the system for and this is exactly what I do I provide it with the information in here now as you can see some few short learning yes we will have further questions and then I just ask my question and you see chat GPT answers yes beautiful in the research mentioned in the previous answer in Google's Quantum I demonstrated a logical qubit made from 49 physical qubits important to note that the number may change continue to develop in an improved Quantum Computing correction technology so we achieved our goal our 49 physical qubits we learned this to the chat GPT without having access to the sandbox or to an API where you have to pay for this this is now a very simple and efficient model and this is just one piece of information you can go on and more and more and more so you can ask in a different way and it still gives you the correct information as you can see here it is the particular research has 49 instead of the Formerly Known 17 qubits so this information is now part of it now as we are scientists researcher we do another experiment and here take something from February eight I am looking here from some unique information and here we have it lenses now used more than 10 billion times per month this is an information about 2023 so you can imagine what we do now I just exchange here the block of new information we have the same prompt engineering the same prompt design and this is the way we go so here we have now the answer and we just have to modify the question a little bit the question was what new information about Google lens is available yes this is what I inserted in information this is the retrieved document if you want from the internet and now we can learn that GPT here lenses now used 10 million times and the answer by GPT as you can see is now 10 billion times per month so you see this is the method if you have external knowledge if you write it or you absorb it from the internet with your web browser this is how you can integrate new knowledge into chat GPT without fine tuning it which is very expensive so just to be clear in the free chat GPT version this works only as long as you have your session if you want to save all these new information that you provided to in context learning you have to have or buy from opmei and API so you have your model that you then can it's not fine-tuning it is tune within context learning for new data and this is the beauty so you can experience this on the free chat GPT version and when you feel comfortable you go and you take a license from the new chat gbt4 or whatever comes up and then you know that your knowledge that you insert in the system is available so thank you for watching I hope you enjoyed it a little bit see you in my next video