The only way to avoid being replaced by AI is to make yourself an outlier.
This post is just me talking to myself to reduce my anxiety after seeing my X timeline overflow with massive new harness, agents and models. Are most of us about to be replaced soon?
But when I think back about how the models work, it calms me down.
LLMs and agents can easily make you averagely good, but they won’t make you excellent. Because excellence is, by definition, an outlier.
This is not to undervalue the incredible progress of the field, it’s generating massive value. But we need to remember how the models actually work. The very essence of an LLM is next-token prediction, it is optimised for consensus and regresses to the mean (just a metaphor :-)). Without specific conditioning, they follow the path of highest probability, directed by what is most common in training data. It converges on the familiar, which is inherently anti-exceptional.
True excellence is exceptional, it is an outlier. The uniqueness and exceptions simply are not well represented in the center span of the statistical distribution.
You won’t be replaced by AI, as long as you refuse to be average and strive to be exceptional. You can only be amplified by utilising your AI well.
So, how do we get the most out of the models and push their output toward that exceptional span?
It goes back to the basics of machine learning: how good the model can output depends on the input. To push the model to generate something exceptional, you have to condition its distribution space. You alter the probability distribution of the next token by feeding the model your unique thinking, your personal takes, your unique experiences, your human instincts, your taste, and your judgment.
It’s a bit contradictory. In traditional machine learning, you try to reduce bias. But now you need to inject bias, your own good bias. You are using your unique perspective as a constraint to narrow the model’s choice space down to the exceptional.
Don’t settle for average. Keep your uniqueness, and use AI to amplify it.