@suchenzang The ironic thing is that I still find AI writing to be *worse* in some ineffable way, as compared to human writing

but I do love writing and maybe this is a bias of mine

@chris_j_paxton the premium of human touch! for lovers of writing, readers of smut, but not for contract lawyers, alas

@suchenzang perhaps the most common failure mode of SF's view of the economy is to see knowledge work as a market of intelligence-based services

but it's actually a relational bargaining system layered on top of intelligence

@chris_j_paxton writing as art will remain art

writing as pure utilitarian conduits for synchronizing information flow between biological humans on the other hand...

@suchenzang My issue is people keep sending me Claude generated documents that, after reading 5-6 pages of a 40 page document, end up having two sentences worth of non-trival information.

Surviving the test of time.

@suchenzang I kinda figured it was just that there are limits to LfD but not to reinforcement learning, and you can't really RL your way to a good essay or novel chapter because the success condition is less clear

@suchenzang just look at how large portions of the math community reacted to recent ai progress in their field. that’s part of what he meant.

@suchenzang At least so far, AI writing has the exact same problem as AI coding, which is that it's far too verbose for the limited information content

I think Ilya was making a point that operates on three levels at once personal, social, and civilizational.

He meant that intelligence is a powerful optimizer, but if you make it your highest value, you end up superbly equipped to navigate the world while increasingly blind to why you're living in it at all

@herbiebradley that still sounds quite transactional

@chris_j_paxton more tokens == more exploration space for random-walking* towards a desirable reward

token-based pricing also makes verbosity a very desirable feature to fill the coffers with

win win win

*rl people don't come at me for gratuitously using "random" here

@suchenzang I should really just have claude-offs at work. Whenever someone sends me a Claude-generated document, I only pass it to Claude and ask it to generate a response which I copy-paste back to the "author".

@suchenzang yes, there are definitely better phrasings further away from the transactional end of the spectrum, but this phrasing I've found communicates the vibe of the idea to AI people reasonably

@suchenzang Yes. As intelligence gets cheaper, taste, trust, and emotional precision get more expensive. Human touch stops being a soft bonus and becomes the filter for what people value.

@suchenzang I wish he worked with @Alibaba_Qwen and @NousResearch on a model harness with high agency and "street smarts".

@IntuitMachine No I'm actually continuously disappointed by our lack of intelligence. And AI is only making everything worse.