Aryan Karamtoth wrote > But LLMs very often go into hallucinating loops.
there is not LLM hallucination. It is the normal function of it. It is not deterministic so if you have two statements A and B, the same LLM can tell you that A is true and B is false; in another thread can say A and B are true; in another one that A is not a valid statement. It cannot understand what is true, ... Second when we speak about IA we assemble a lot of different think in a single basket. One of AI father (sorry I don't remember his name) say that LLM is a dead end, you cannot improve it to have not such "hallucination" things. It say that a true AI is the one build to play go. First version of this AI can beat each human player at each game. Second version can beat the first one at each played game. And least what we are doing to use LLM everywhere? We are building mega data center and selling only AI device (from PC to electric toothbrush), but: * they use a lot of electricity so there is place where: - electricity cannot be always available for residential house - you cannot have a electricity in your new build house - you cannot have build house permit on your field - ... * they use a lot of water, so there is place where: - water cannot be always available for residential house - water cannot be available for agricultural and industry - ... * electricity and water price rise * the actual data-center cannot hold the new heavy LLM's machine so they need to build new one * lack of RAM, hard-disk, ... and price rise * some regions/states are banning the construction of new data-center * ... * they emit a lot of greenhouse gases and other pollutants because they need a lot of electricity and for have it they are: - reactivating or prolonging life of coal power plant - using jet engine to produce electricity - burning diesel in diesel generators to produce electricity - building or reactivate nuclear power plants - ... * ... I think that Debian and free software in general must look also in making a better life for everyone and not to help to increase problems. Ciao Davide -- La mia privacy non รจ affar tuo https://noyb.eu/it - You do not have my permission to use this email to train an AI - If you use this to train your AI than you accept to distribute under AGPL license >= 3.0 all the model trained, all the source you have used to training your model and all the source of the program that use that model
