True for first generations of LLMs using statistical learning, transformers, and self-attention over 100s billions parameters.
New generation has chain of thought reasoning models that retrain LLMs with multi-stage reinforcement learning not to hallucinate plausible nonsense but correct responses.
New generation has chain of thought reasoning models that retrain LLMs with multi-stage reinforcement learning not to hallucinate plausible nonsense but correct responses.
Comments
DeepSeek is better at resoning/maths/coding/logic than other models - and shows its step by step reasoning.
It's not as good at creative tasks - writing, conversation, general knowledge etc.