Why does it seem to you that writing in a human-like way is an easier problem to solve? It seems harder to me.
Even if we grant that AI solved Navier-Stokes (there's too much nuance to the issue to be so clear-cut about it), it doesn't seem that it did so in a useful way. The point of trying to solve those problems isn't so much to arrive at a final answer as it is that doing so is a means to find deeper mathematical truths. In other words, the end answer isn't what's of value so much as the knowledge gained on the way to the answer.
A whole lot of mathematicians are saying that the genAI solution basically just gave a result and didn't give any valuable insight as a result.
The 'solution' of Navier-Stokes was just for a particular set of conditions, not a general solution. If "AI" gives a general solution that's another matter.
LLMs do write in a human like fashion, it's just the median or average of everything. That's why your CEO is so impressed - it generates better writing than he can.
Take those two things together, and you've got an answer: it didn't and it does.
AI are actively trained to talk that way. They are trained on simulated conversations of helpful robot assistants, to sound like the robot assistant.
It is a neutral role with broad applicability. They could equally be trained for other styles. A model trained to talk as a pirate would get old, especially after it told you to bugger off when you asked for a cupcake recipe. A model trained to write like Hemingway would be frustrating when you want directions.
What would be better? who should it write like?
There is no technical challenge finetuning models to write in a specific style that is indistinguishable from a human
FWIW, I have been pleasantly surprised with the writing quality of GPT 6 Luna. It’s not particularly smart/insightful on its own, but I get good results when having it summarize or rephrase techbrospeak. It also produces coherent and easy-reading drafts from outlines and bullet points, and it generally follows instructions well in these domains.
6 comments
[ 20.3 ms ] story [ 135 ms ] threadEven if we grant that AI solved Navier-Stokes (there's too much nuance to the issue to be so clear-cut about it), it doesn't seem that it did so in a useful way. The point of trying to solve those problems isn't so much to arrive at a final answer as it is that doing so is a means to find deeper mathematical truths. In other words, the end answer isn't what's of value so much as the knowledge gained on the way to the answer.
A whole lot of mathematicians are saying that the genAI solution basically just gave a result and didn't give any valuable insight as a result.
Disclaimer: I am absolutely not a mathematician.
LLMs do write in a human like fashion, it's just the median or average of everything. That's why your CEO is so impressed - it generates better writing than he can.
Take those two things together, and you've got an answer: it didn't and it does.
It is a neutral role with broad applicability. They could equally be trained for other styles. A model trained to talk as a pirate would get old, especially after it told you to bugger off when you asked for a cupcake recipe. A model trained to write like Hemingway would be frustrating when you want directions.
What would be better? who should it write like?
There is no technical challenge finetuning models to write in a specific style that is indistinguishable from a human