Gary Marcus

Rank 10 of 47

Score 112

@chatgpt21 @flowersslop @captain_marrvel 1. the problem is that we don’t how much of that is architecture, how much of that comes from improvements in base models, how much it comes from fine-tuning and data augmentation (etc) geared around benchmarks etc. (And, e.g., Grok 4 is probably better on most of the benchmarks

7/21/2025, 5:05:03 PM

In reply to:

Chris

@chatgpt21

@flowersslop @captain_marrvel @GaryMarcus @GaryMarcus @flowersslop it’s a much better model than 4o

Flowers ☾ ❂

@flowersslop

@captain_marrvel @GaryMarcus 4.5 only wins against 4o in 60% of cases in oais own internal preference benchmark

MarrvelSystems

@captain_marrvel

@flowersslop @GaryMarcus GPT-5 is literally o4 and 4.5, why the hell y’all comparing it to 4o which is 10x worse than OpenAI current o series models?

Flowers ☾ ❂

@flowersslop

@GaryMarcus I have stricter expectations. I will do a blind test to see if I prefer GPT-5 over GPT-4o at least 8/10 times, if not, that would be a crazy fail.

Because it should actually be 10/10 times.

Gary Marcus

@GaryMarcus

last year people were claiming GPT-5 was going to be some monstrously large whale.

now? they are hoping for a 10-20% improvement.

Chris

@chatgpt21

@krapstarr @GaryMarcus 10-20%

The statement engages in a technical discussion about the factors contributing to the performance of AI models, such as architecture, base model improvements, and fine-tuning. It questions the sources of performance improvements and compares different models using benchmark data.

Principle 1:
I will strive to do no harm with my words and actions.
The statement does not cause harm and focuses on technical analysis. [+1]
Principle 3:
I will use my words and actions to promote understanding, empathy, and compassion.
It promotes understanding by discussing the complexities of AI model improvements. [+1]
Principle 4:
I will engage in constructive criticism and dialogue with those in disagreement and will not engage in personal attacks or ad hominem arguments.
The statement engages in constructive dialogue without personal attacks. [+1]