Very interesting review, we can see some interesting strengths and weaknesses. Also gives a good idea of the different ways to evaluate such models.
LLMs training had a bias from the start... and now we got a feedback loop since people are posting generated content online which is then used for training again. Expect idiosyncrasies to increase with time.
Now this is a very interesting use of generator models. I find this more exciting than the glorified chatbots.
Interesting experiment. It makes for a very large file but there are a few clever tricks in there.
Interesting terminal oriented tool to interacting with LLM. Let you choose to self-host or run locally.
I guess this will end up being a popular plugin for Krita.
The newer text to speech models are really good now.
Excellent piece about everything which is wrong in the current industrialization moment of generative AI.
A balanced view, that's refreshing. Indeed we see too many "let's call the OpenAI APIs and magic will happen". This is very short sighted, much better can be done.
Ever wondered what you can do with vector databases and LLMs? Here is an interesting use case.
I fully agree with this. All those doomsayers about AI are there to sell something willingly or not. Let's look at the current threats we're seeing now, not the imaginary ones being the results of people taking SciFi a bit too literally.
Attacks on machine learning models are getting more accessible. This means even more care will have to be taken to deploy and use those.
Another open source LLM available out there. This one seems to have interesting properties.
A very needed tool unfortunately. This is fascinating research as well. The world is really so cyberpunk now.
The risks of too much automation... this is chilling to see what using machine learning based system all the way can lead to. It's just a ridiculous situation and no way to get out of it.
On specialized and formalized domains like this it might lead to something interesting. That said there's a tension with the fact that it doesn't know when it doesn't know which might be problematic. Also I wonder how it fares compared to computational models like WolframAlpha. In the end very formal domains like this have large knowledge bases already available.
A good way to get some control back if you want to use a LLM. You can host it locally, it's free software. Definitely a step in the right direction.
I agree with this, be precise about what you're using. Blanket terms won't lead you anywhere to creating proper systems.
Looks like it was a very interesting talk. Situation still needs to be monitored in any case, it's uncertain how those cases will be ruled.
