Absolutely amazing video! But I have to ask... LLMs are not explainable/interpretable yet right? It's still a black box. Why would someone deploy such a thing for Education(EdTech)? Is finetuning it after using more data and compute enough to solve hallucinations/make it more reliable?
Great question! LLMs are such huge neural networks, that interpretability is (at least for now) out of the question. Maybe in the future, we can figure out a bit more of why a particular output is a certain way, but for now they're more black boxes. However, finetuning does work well, and if you top it with retrieval (search) augmented generation, they can be pretty trustable, and they "show their work". By "show their work" I don't mean that we can look at the nodes and see what happened internally, but that they would be able to answer a question and show where in the internet (or the database, could be wikipedia or an internal database) they found the information for that answer. So so far that's what we have, but I'm still hopeful that one day we'll understand them better internally, and be able to hold them accountable for their answers. Especially for sensitive things like education, or healthcare.
Thank you, your explanation are easy to understand
Absolutely amazing video! But I have to ask... LLMs are not explainable/interpretable yet right? It's still a black box. Why would someone deploy such a thing for Education(EdTech)? Is finetuning it after using more data and compute enough to solve hallucinations/make it more reliable?
Great question! LLMs are such huge neural networks, that interpretability is (at least for now) out of the question. Maybe in the future, we can figure out a bit more of why a particular output is a certain way, but for now they're more black boxes. However, finetuning does work well, and if you top it with retrieval (search) augmented generation, they can be pretty trustable, and they "show their work". By "show their work" I don't mean that we can look at the nodes and see what happened internally, but that they would be able to answer a question and show where in the internet (or the database, could be wikipedia or an internal database) they found the information for that answer.
So so far that's what we have, but I'm still hopeful that one day we'll understand them better internally, and be able to hold them accountable for their answers. Especially for sensitive things like education, or healthcare.
thanks!
Hi from India
🙌 Promo-SM