[adlpz]

I thought this was going to be yet another post about how AI is ruining Junior devs so we'll have a Senior replacement crysis in a few years.

It sort of is, indirectly, and I agree with pretty much everything.

But the bit about sycophancy was particularly enlightening. I actually thought "plain" ChatGPT-like interfaces could be good for learning. But the Youtube ROAS example is really powerful. If the student can skew the teacher's conclusions so much just by the way they phrase their questions/answers, we're going to mislead new programmers en masse.

I'm not even sure that the extensive prompting they say they use for their "Boots" is good enough.

I guess in the age of AI you still need someone to repeatedly reject your pull requests until you learn. And AI won't be that someone, at least for now.

[A4ET8a8uTh0_v2]

I think I agree. From personal experience, even when I ask for scathing critique, the need to placate and make me feel better seems to bleed through ( both on 4o and 5 as far as I could tell ). I am not sure what to make of it.

We go back to this original prediction that the tool will help those, who both want help and are painfully aware of LLMs peculiar issues.

[Eisenstein]

The incentives align: when the success metric is engagement, it will be optimized for engagement. It makes perfect sense that it is going to agree with you.

[1718627440]

> But the bit about sycophancy was particularly enlightening.

I always try to stay above this by prompting the question twice, with the opposite biases. But I of course don't know, which hidden biases I have that the LLM still reinforces.