 AI chatbots in science. You've heard about them. You probably used them too. I'm talking about chat tbt and other large language models that provide your chat prompt through which you can ask anything and the generative text model will then produce an answer. These tools have turned out to be remarkably versatile thanks to having very large models with more than one billion parameters that have been trained on more than 100 billion tokens or words if you will. The downside to this is that they're extremely expensive to train and have huge hardware requirements to run. I intentionally waited with making this video until the word hype had died down and I had an informed opinion. However it is a rapidly changing topic and some of the stuff I say is likely to be outdated soon. In this video I'll discuss whether chatbots are likely to replace Google, can help you publish, apply for funding, be your programmer or whether they're better used as an interface for other tools. RAI chatbots google 2.0. Large language models are great at summarizing and having been trained on the internet it's tempting to think that you can just ask chat tbt instead of doing a google search. However these models are difficult to update and therefore often have outdated information. Moreover they're prone to hallucination that is they can provide confident wrong answers they simply make up facts which is made worse by the fact that they provide no sources for what they're claiming. So in the end you're going to have to check all the answers you get from asking chat tbt which means that you're going to have to search for sources which means that you're going right back to google so it is no replacement. However they may be useful tools to help you come up with a better query for doing your search in the first place. RAI chatbots are useful for publishing. This is a controversial topic. A chatbot is not a person it cannot take responsibility and as such it cannot be an author and you should not copy paste text straight from chat tbt into your manuscript which people have already been caught doing. As I already mentioned chatbots can invent facts and if you ask for references it will also invent references that do not exist. So you should not use a model to write alone and probably not use a model to write anything from scratch since fact checking it will take more time than just writing it yourself. However it can still be a useful tool. Think about it as a tool for improving writing. Have it be your copy editor that goes through your text identifies problems and even suggests changes to make it better. It can also in the future I think be used to draft certain sections from detailed data. That is all the input is provided in a structured form so it doesn't invent things. This could be for example tool based method sections. Imagine that you have a tool with a graphical user interface like the string database. You do all your query adjust your parameters and get a network and you then have the tool track all your actions and write an accurate description of what you actually did rather than what you think you did. Similarly I could imagine table based results. For example from an enrichment analysis you might choose the rows of particular interest and have the tool write a summary paragraph about these. You may also be tempted to use an AI bot to come up with ideas for your next grant. However since these models are great at summarizing it shouldn't come as a huge surprise that they are terrible at coming up with new ideas. So if you ask it to write a grant proposal for you it will almost certainly be rejected on grounds of lack of novelty. However I think chat bots could be used for reshaping rejected proposals for submission to a different funding agency and for writing certain required sections that are less scientific and just need to be adapted from other sources. Once you have a grant they could also be excellent for reporting filling in long forms writing the most deliverable reports which are often redundant with papers that you've written and may not be read by anyone. What about using your AI chatbot for coding? The underlying language model is typically multilingual and that includes programming languages. However it's not a programmer and if you ask it to write code it will often have many code errors. However new features like code interpreter allows the model to run the code and in an iterative process fix its own syntax errors providing you with runnable code although it may still give you wrong results. I think these AI chatbots are better used as an alternative to copying and pasting from stack overflow, as debugging tools to help you find mistakes, for code review to help improve your code, for code translation to move code from one programming language to another and finally for automatic documentation of your functions. Last but not least I think that AI chatbots are an interesting form of interface that can be used to interface large language models with code. This can happen in two ways. One is that the code can access for example chat GPT via its API and the other is that the model itself can access code like with the GPT4 plugins. Regardless of which way it's done this means that large language models can be used to interpret the input from users translating questions into for example database queries which can then subsequently be answered correctly using either a knowledge graph or by querying the biomedical literature. Large language models can also be used for output as I already talked about with the enrichment analysis. This means that they can be used to translate the results coming out of an analysis into human readable text for the user to interpret. This means that you can effectively wrap a resource with a user-friendly large language model. So what do I think will be the future of AI chatbots? It is famously dangerous to prophecy especially about the future and more so in a field with rapid progress. However I think it's safe to say that we will have even bigger models coming and that we're moving from large language models being standalone to large language models using tools to look up facts, provide references for what they're claiming and execute code instead of doing calculations within the model itself. There will also be a death of many of the AI tools that we currently have popping up because of AI becoming an integral part of other software. This will include your office package and your integrated development environment in which you make code. So in the future we will often not even think about it when we're using AI. That's all I have to say about AI chatbots for now. I hope you found this video interesting and if you did I suggest you watch this next. Thanks for your attention.