ITMO Mega School 2024 Igor Kotenkov - State of the LLM Landscape

Interesting. report Generalizing the state of affairs in the LLM (large language modes) from Igor Kotenkova, author of the channel "Sioloshnaya" (@seeallochnaya)Which I read with interest. Specifically in this video, the author 2 It provides a good overview of the latest achievements for ITMO students.

From the presentation (available here) I made the following points. What are large language models, how do they work and what are they currently used for? How LLMs already affect the real world (Especially when they start using the internet to do things.) And what the future holds. How large language models learn – in fact, models predict the next word in context, the model learns from a lot of text (mainstream)Next, the output is a multigigabyte of parameters that set the weights in the model. Next, the model is completed using a person - there are a couple of texts where one of the answers is chosen by a person as the best and it is submitted to the model. As a result, it speeds up the learning process and optimizes human preferences. ChatGPT - to create the model used huge power and high-quality data, and then the product became the fastest growing in the world and reached the highest level of growth. 100 Millions of users approximately 2 month Chatbots can now be used to automate activities, for example, to improve the efficiency of developers (more report) or to improve the effectiveness of consultants (example) LLMs can learn not only to generate text, but also some skills, such as mathematics. Researchers are actively working on the interpertability of LLMs, in the future it allows the model to memorize and apply patterns, rather than remember everything. When using LLMs, there may be problems with leaking confidential information So far, the only guaranteed way to make LLM better is to add more power and toss better data at the learning stage, but researchers are constantly trying to come up with something else, for example, by optimizing algorithms. To the models you can connect tools, for example, a calculator (or Wolfram Alfa, you can read more about it in the book by Stephen Wolfram.What Is ChatGPT Doing ... and Why Does It Work?"that I mean." told earlier) Models can be helped by asking them to build a chain of reasoning. (chain-of-thought prompting) External tools and chain-of-thought can be combined in the ReAct approach. (Reasoning + Actions). ReAct is a method of promting that allows a model to perform several steps of searching for information. Igor gives interesting examples with models To search for information on the Internet and insert text into the model To use the service to cheat the captcha:)

• To synthesize poisons. For the generation of objects in minecraft, where the model created a skill library and learned skills for creating objects LLMs can work not only with text, but also with video, sounds and other modalities. On the basis of LLMs, you can create autonomous browser agents that will perform actions on the site, predicting what needs to be done. Synthetic data generated by other models can be used to train models. This allows you to improve the quality of training and increase model skills. LLMs can generate tens of thousands, hundreds of thousands of answers to problems, which allows you to build a thought tree and find the right answer. In the future, such models can be used to create scientific articles and new knowledge. LLMs can make up facts or accept incorrect facts, but this can be resolved in the future. Next, ML will cover more and more new areas or bring to an ideal those in which technology is already used. As a result, there will be work both to create better models and to integrate existing models to solve real business problems.

#DataScience #ML #AI #Data #PopularScience #Math