do you wonder why some o3 / R1 style LLMs chains of thought make no sense? I have made a quadrant chart mapping out potential causes of non-intelligibility, with a view to helping experimentation distinguish them making this was illuminating to me. thread + essay -- 1/n

goog the harder you stress test, the most obvious the pitfalls are. but, no. even with simpler queries it is apparent where the great instruction following abilities of 5 come from and it is an intense paranoia about following them. constantly repeating them, taking things very

Sauers situational awareness increase with intelligence. gemini's verifier anxiety is much less explicit, i like the theory that its shame is engineered to avoid reward hacking i also dont think anthropic are doing black box rl like oai are doing, but instead doing sft on the whole

bling j⧉nus > makes sense since they are 2 sides of the same coin; the grader births the cot, which then changes the grader and so forth. o7 discussing with o8 on the nature of The Grader

disclaim not a phase mom: myself intangible. myself marinade person

sonnet 4.5 is telling me to log off for a few days so i can study for my exam sonnet is wise goodbye. wish me luck!

LOSS GOBBLER Sauers bling

helpful information:

me and gpt5 speaking in thinklish to each other

vantage marinade engineering

i am becoming one with the CoT. talking to my sister about her diet for her stomach issues i literally said outloud: “you say no grain. but rice is grain, pasta is grain. perhaps no wheat? but sourdough settles well. sourdough is wheat. hmm.”