Sur les limites des #LargeLanguageModels:
Le langage est éminemment incarné ("embodied") alors que les #LLM ne sont que des modèles inférentiels sans notion de vérité, d'émotions, d'engagement envers autrui ou envers l'avenir...
"Machines such as LLMs can generate text strings that signify emotions and moods. But these are statistical constructions. Having no concerns and no bodies, machines have no emotions and no moods, and no means to develop sensibilities for them."
As some tout the good tidings and marvels of AI, LLM, and marketing obfuscation ad nauseum, let’s not lose our grasp on how much our own ethics affect that real impact these tools have on all of us. And if we can’t do that, how are we supposed to instill a sense of ethics on these new conscious minds we pride ourselves in creating?
On one hand, we have new papers that show how just using the language of a specific human group can trigger implicit, hidden biases in #LargeLanguageModels
on the other hand, we have software developers working to build tools that automatically retrieve information that may be of interest, and that try to reason ahead on your interests. Highest point so far: https://new.computer/
Does anyone have a good list of logical questions to judge large language models ability to reason?
Questions like "if it takes 3 hours for 3 towels to dry, how long does it take for 9 towels to dry?"
I'm playing around with Mistrals leaked 70b Miqu LLM and want to test it's reasoning skills for a project I'm working on. I've been really impressed so far. It's slower than Mistral & Mixtral but it's been producing the best reasoned answers I've seen from an LLM. And it's running locally!
Also, no I will not join your research project looking at how #LargeLanguageModels and #GenerativeAI can help solve climate change. There is no possible world in which that makes any sense.
It also depends on what you want the #LargeLanguageModels and/or #generativeAI to do, and if you care to put in the time, effort, and investment to curate the training data or not.
Many of these operations don’t want to do the work in terms of curating their training data (whether that means screening it or asking for permission or whatever) because it’s not cheap or fast!