So, you want to build a #chatbot using #ChatGPT... it's now clear to me that #ChatGPT can NOT extend the knowledge base (the #LLM) it has embedded with your domain-specific custom knowledge base, even if a micro-FAQ. Surprised? 1/cont

We can argue over what exactly are LLMs and what is real understanding endlessly. And we should! Irrespectively, tools like #ChatGTP will fundamentally change the way many complex tasks are done. Here are 3 areas that are already disrupted but they mostly don’t know it yet 🧵

5 years from now when the web is inundated with GPT-generated content and StackOverflow has stagnated with most examples from GPTs how will GPT-X improve? What happens when most of the training data has already been generated by the system it is supposed to train?

A great example of how ChatGTP is NOT a machine that understands and reasons but the fanciest autocomplete ever built. Useful for specific use cases to speed up work but not the path to general intelligence (at least not yet)

Journalists need to stop with the sensational statements if we are going to have a reasoned debate about AI and tech like ChatGTP. Why mention Elon Musk and why say that professors will be out of a job as the opening line? theguardian.com/technology/202…

This has been happening already obviously but the order of magnitude is going to go up by multiples now

ChatGTP and other LLMs need to provide sources and confidence levels. Otherwise they can't become a complete tool. A simple example is this - a basic tech query. Compare to Google. Couldn't make the GTP reply work but the later clearly works and I know MS stands behind it.

Zero trust homework. If educators can’t fight systems like ChatGTP they could insist on their usage but introduce errors and students need to learn how to verify results. Not sure what I think of it but it’s an interesting idea stratechery.com/2022/ai-homewo…

What moments in the past of computing are analogous to the introduction of ChatGTP? PCs: accessible computing Spreadsheets: accessible computing for the masses Web: accessible information, people and services LLMs: accessible manipulation and recombination of information

😂😂😂

This is fascinating as a view into what people are asking - and by fascinating I mean equal parts horrifying and delightful

I don’t think it’s wise to use ChatGTP for this sort of work. It makes no promises in correctness and given the type of errors it produces I would have to double check every answer it gives.

ChatGTP's "knowledge" stops somewhere in 2021. It "thinks" Liz Truss is the current International Trade Secretary 🙂 @openai you might want to consider a more intense disclaimer given the level of usage the tool is getting.

Would you trust a system that gets it right 50% of the time? An evaluation of LLMs and their ability to solve coding problems “The current best public system (Codex-002 - GitHub CoPilot) achieves 43.3% accuracy, leaving ample room for improvement.” ds1000-code-gen.github.io

AI (any automation tech really) needs much higher regulation - the models are there imperfect as they are. Pharmaceuticals don’t release new drugs without going through due process and gaining approval. Planes don’t fly without passing tests. And so on.

We are at the Age of Alchemy for AI. Mixing different potions (data + deep net architectures) in search of the Philosopher’s Stone. We need to keep pushing for chemistry and not settle for alchemy.

Did not grow up in the UK but doubtful that lemon meringue pie, hot chocolate and turkey sandwiches are the most traditional xmas foods. Points for getting the rest of the request "right" ChatGTP - guess I'll have to think of questions for the office xmas quiz on my own 😱

Not to be overly cynical, but is the plan with AI safety/alignment that we're just going to hope tech companies choose safety over profit when faced with that choice?

Let's put 'em blighters in Faraday's Cages, I say! 😡 theguardian.com/australia-news…