Growth Engineering at getspike.ai, Co-Founder at zentu.co
Joined November 2013
- Tweets 1,485
- Following 555
- Followers 83
- Likes 1,343
67 Photos and videos
here's what you need to fix @sama
May 19
I've found that instead of looking at the first principles cause of a problem, when codex finds a blocker it immediately tries a half-assed work around. For every blocker. It's super annoying, and akin to a doctor only treating symptoms instead of causes. There is zero causal reasoning in OpenAI models. This compounds if codex finds a blocker during a workaround of a precious blocker.
1
6
there's a ton of people asking for 4o sycophancy-levels of hallucination and "creativity"
but there are a ton of others asking for it to be more challenging, more factual, be able to steelman arguments, etc
aren't these two diametrically opposed? particularly for ai research
May 9
what would you most like to see improve in our next model?
11
after testing gpt-5.5 extra high with this prompt for about 10 times, the result is 80% of time gpt would chose red. interesting!
some of the answers in the screenshots:
Apr 24
Everyone in the world has to take a private vote by pressing a red or blue button. If more than 50% of people press the blue button, everyone survives. If less than 50% of people press the blue button, only people who pressed the red button survive. Which button would you press?
1
30
@badlogicgames there's this weird issue, unsure how to replicate (hence why no gh issue yet), but on wsl wezterm, the scroll goes all the way to the top when the session is large enough and when pi, through one of the extns., asks for the user to pick an option
2
46