[colesantiago]

> LLMs have no intelligence at all, they don't reason.

So the benchmarks (1) of LLM's like Claude 3.5 are categorically flawed at reasoning and they will never get better?

Yet their outputs are heuristically good enough to reduce jobs from engineers by driving the cost of code generation down to almost 0.

If we take the idea that you believe that it cannot reason at all in any capacity, as whole it would be regretful for you to bet against these systems not getting better or that another breakthrough would change the perspective of a new generation of LLMs.

(1) https://www.anthropic.com/news/claude-3-5-sonnet

[cinntaile]

Anthropic is an AI company, it's evident that they want their system to seem intelligent. That's what they're selling after all.

> it would be regretful for you to bet against these systems not getting better or that another breakthrough would change the perspective of a new generation of LLMs.

Now you're putting words in my mouth. The results are impressive for sure, although I don't think we have seen any major breakthroughs since their original release. Just small, incremental steps. I'm not convinced that'll get us to reasoning capabilities but I don't mind being proven wrong. Predicting the future is hard.