2 operation mode

exploring how humans need to balance between AI-assisted productivity and active learning modes in the post-AGI world

modern ai models have developed 2 modes of using computation power: test time compute (ttc) and gradient based learning, humans in post agi world need 2 operation mode as well.

  1. get things done
  2. learning

with clear goal and things to achieve, ai could help to take over 90% of the work. one has to be mindful to differentiate between the time to intervene and spend time from the time to delegate. the goal is maximize productivity.

learning is very different story. autopilot mode, like simple coding in cursor with tap tap tap provides little to no learning signals.

learning is about change, which is biological version of gradient based learning. most of learnings happen after taking actions, expectation and reality doesn’t match, which means few things:

  • one has to plan to form independent expectation.
  • to take action.
  • to observe and accept the reality check.
  • to kick off the reflection process.

think about human ai collaboration in the context of these breakpoints.

  • if planning phase is delegated, the baseline is gone. reality check wouldn’t provide learning signal.
  • even repeated actions have educational value by building muscle memory and time to sink in.
  • find the delta between expectation and reality is good exercise.
  • delegating reflection process is like reading elon musk autobiography, feeling good but after reading, very few learnings stay. it’s not your life and high level take away are not that different from one great man to another.

learning is change. change is hard. effective change means collecting as many learning signal as possible and let those experience sink in, repeatedly.

i can totally see myself doing tap tap tap and accept all one day, and following tutorial or building something with as little ai as possible the other day for learning purpose.

these modes are complementary as test time compute (ttc) and gradient based learning. with ttc, runtime performance is boosted, valuable experiences are collected. those data would be used in supervised finetuning (sft) and reinforcement finetuning (rft). gradient based learning would build new useful vector functions and learn to combine them more effectively. rinse and repeat.

i’m learning when to use ai and when not to. it’s a delicate balance and never ending process.