|
|
|
|
|
|
by jchw
2 days ago
|
|
|
Actually I disagree, it's pretty apparent that modern frontier LLMs are nearly completely incapable of writing good prose for some reason. I'm not sure if it's the RLHF phase or what, but even when you explicitly tell them to try to avoid the cliches it's never enough. They're geared toward writing heavily punchy, low-substance prose, and it shows up everywhere in their output, even in places like documentation and just normal chat replies. The exact way in which the models are fucked up seems to depend on circumstances, but I think right now one thing I've noticed out of the latest versions of Claude Opus is that it really really likes to use the word "honest" in its summaries. "What Remains (The "Honest" Part)". I figured this was maybe something to do with it just repeating the system prompts but no. It turns out the word "honest" does appear in some fragments of system prompts in Cluade Code, but it doesn't appear to be anywhere where it would've been in the context of my recent runs. I think this is a tuning issue and that eventually, someone will figure out a good way to prevent models from getting skewed this way. Still, the bad prose quality is not really the biggest issue. In fact, it's kind of handy that the prose quality is shit because it makes it easier to tell when someone just doesn't seem to care about what they're writing. If the prose quality was really good, yet the amount of effort put in was the same, we would be having an even worse problem right now. |
|
|
The world is bigger than Claude and an oversimplified worldview.
If you haven't even read good quality AI written articles, you have either been wilfully blind to them or you have a bigger problem, because you surely haven't avoided them.