Errm, No :-) I meant bars as in benchmarks, often rather meaningless, because within the range of statistic noise.
For instance, something having 100.200 points in one config, in another 100.220, with the bars/scales distorted to make that difference seem much larger.
OpenAI recently played a bit too hard with their GPT-5 announcement. Two bars with the same height but wildly different values, things like that. Such a lack of subtlety that their claim it was accidental is actually almost believable.
Note that the main benchmark in the post is practical. How long a test suite for the product he makes runs.
Or to generally be back in the era where such a thing would be useful for LAN-parties...
Das Balkenspiel ist schlechter Stil. (The bar game is bad style/lame)
Pong!
Errm, No :-) I meant bars as in benchmarks, often rather meaningless, because within the range of statistic noise.
For instance, something having 100.200 points in one config, in another 100.220, with the bars/scales distorted to make that difference seem much larger.
Gaming the bar-game, so to speak.
OpenAI recently played a bit too hard with their GPT-5 announcement. Two bars with the same height but wildly different values, things like that. Such a lack of subtlety that their claim it was accidental is actually almost believable.
Do you even bench, mark?