3 ms·

The Pentagon’s Ambitious AI Plans Look Less and Less Like ChatGPT

• janalsncm 3y ago
  LLMs aren’t hallucinating because they’re trained on incorrect data. They hallucinate because they’re not generating complete thoughts based on facts, they’re generating word by word until they terminate. There’s no mechanism for truth or verification in statistical language models.
  • sublinear 3y ago
    I sort of agree. They're probabilistic. They do not converge to truth. They converge to the training data, but of course the training data itself is inconsistent and then also it gets worse because it can only asymptotically approach it with, what I imagine, seems to be a very slow rate. In general, LLMs are fine for search relevance and vague summarizing, but unfortunately marketed by con artists. Scaling them up doesn't magically produce "sentience" or "comprehension" or "emergent behavior".
    • afiori 3y ago
      This line of argument feels like it has an implicit assumption that progress for ai technology is to become more human-like. This is certainly a requirement if the objective is to replace humans, but ai could grow to be increasingly more sophisticated, capable of doing more and more tasks by progressing on different paths. This is not to say that the singularity is imminent, but the argument "Human intelligence has X as a necessary piece, AI has little amount of X so it is not intelligent" does not sound convincing to me.
  • PartiallyTyped 3y ago
    In the same vein, many people are hallucinating in their speech, Dunning Kruger and all that; plus every time we recall something, we alter the memory a bit. Adding context and giving more information to models, eg RAG, actually works.