pull down to refresh

Cerebras Gemma 4 Demo

I think we're used to seeing AI output at 20 tokens/second where we can sort of keep up and pretend we can supervise as humans. It's kind of amazing seeing AI code or write an entire short story at ~1800 tokens/second.

Pretty mind blowing... the future of the dead internet will be AI agents talking to each other and it will be impossible for us to keep up.

I stopped looking over a year ago and nowadays I just get results, traces, and so on filed to me for later inspection. I also measure my own performance in how long I keep tickets and PRs open[1].

The bots are imho much better as a background task than as an invasive foreground chat thing.

  1. on a full load during daylight, I average about 20m for tickets and 2h for PRs, but during the night when the backlog is being processed I average 10h - no surprise because I sleep

reply