Re:Grok 4.20 beats all other AI models in Alpha Arena test
by ben_w on Saturday December 06, 2025 @05:22AM
(#46172173)
> I assumed all of the models were doing that, using at least Web Search tools.
Sometimes. The other week I was asking ChatGPT about the UK PM, and had to stop the generation early because it started ~"Prime Minister Rishi Sunak…"
The unreliability is also why techniques as simple as "ask 5 times and have it take a vote of its own answers" boost performance. Or "thinking" modes which are approximately just replacing the end token with "Wait." and continuing for ten rounds.
Sometimes. The other week I was asking ChatGPT about the UK PM, and had to stop the generation early because it started ~"Prime Minister Rishi Sunak…"
The unreliability is also why techniques as simple as "ask 5 times and have it take a vote of its own answers" boost performance. Or "thinking" modes which are approximately just replacing the end token with "Wait." and continuing for ten rounds.