> - Worse at logic (e.g. it contradicts itself in a single sentence, and is unable to figure it out)
That takes most of my use cases. "logic" is what makes GPT often feel like AGI.
Use cases like translation seem less impressive in comparison to logical reasoning because it feels like it's just something where you can throw a lot of data and it'll do better. While with logical reasoning it still feels like model "learned" something more than pure pattern matching
Exactly. That’s also why I find low parameter LLMs to be useless for me personally. I simply cannot trust anything that is so very illogical. GPT-4 is the first LLM that crossed into usable territory for me. Even GPT-3.5 was a fun toy and maybe good for summarization, but that’s it. It will be revolutionary when GPT-4 is cheap enough that thousands of calls don’t cost much. To imagine an LLM much smarter GPT-4…the future is bright.
> - Worse at logic (e.g. it contradicts itself in a single sentence, and is unable to figure it out)
That takes most of my use cases. "logic" is what makes GPT often feel like AGI.
Use cases like translation seem less impressive in comparison to logical reasoning because it feels like it's just something where you can throw a lot of data and it'll do better. While with logical reasoning it still feels like model "learned" something more than pure pattern matching