zlacker

>>modele+(OP)
"In March 2023, OpenAI released its most powerful language model yet, GPT-4. GPT-4 is not just capable of reasoning. It is better at reasoning than average humans. It scored in the 90th percentile on the Uniform Bar Exam for lawyers. It scored in the 99th percentile on the GRE Verbal Assessment. It even scored a 77% on the Advanced Sommelier examination."

One could argue a common characteristic of the above exams is that they each test memory, and, as such, one could argue that GPT-4's above-average performance is not necessarily evidence of "reasoning". That is, GPT-4 has no "understanding" but it has formidable reading speed and retention (memory).

While preparation for the above exams depends heavily on memorisation, other exams may focus more on reasoning and understanding.

Surely GPT-4 would fail some exams. But when it comes to GPT-4's exam performance, only the positive results are reported.

https://freeman.vc/notes/reasoning-vs-memorization-in-llms

>>1vuio0+vn2
> One could argue a common characteristic of the above exams is that they each test memory, and, as such, one could argue that GPT-4's above-average performance is not necessarily evidence of "reasoning". That is, GPT-4 has no "understanding" but it has formidable reading speed and retention (memory).

I don’t think they will make this argument since it will heavily negate their (expected) argument that they’re not open-sourcing the model because of safety concerns.

>>giggle+WG2
I’m not so sure of this. The fear generally isn’t that the model itself is dangerous, more so that how it is used by someone might be.