1,116 karma · joined April 2, 2023
Why does everyone need a full operating-system tier browser to render any website with javascript?
Also it's a lot better at coding. GPT has become exceptionally lazy recently, but i consistently can get 500+ lines of code out of claude (it even has to spawn multiple output windows)
Perhaps the top end 4 might wrong slightly more clever code, but you're hard pressed to get it to do more than a dozen or two lines.
I'm not 100% sure it's related but if you ask it to draw bounding boxes around things it's always off by quite a bit.
Just like you might imagine a leech could evolve some attribute to help defend its host against other predators.
I can't speak on the elastic vs open search topic though - does being part of the fiefdom of amazon work as a survival strategy?
'delete techbros' also generates some great banter. you get some very dated and incorrect references though "Mt Gox fade into fever dream of capital burnt for naught"
Also surprisingly there, grammatical errors[!]
10/10, i can see why everyone has been using claude for their, uh, creative writing use cases.
Odd they call us `humans`, and also a bit that it doesn't say who it's named after.
They're also pretty fantastic for discoverability if you do a bit of filtering to remove hallucinations. Only downside is you end up finding that project is abandoned for years or so.
I guess you benchmarked via API? I've heard even the datestamped models have been nerfed from time to time..
In terms of the actual output it can be annoying if the convo goes on too long and it hits the window, but for technical things this hardly matters as you can just restart it - more of an RP issue as i understand.
It also seems to line up with random paragraph location doing the worst in that paper.
LLMs are that 'blurry jpg of the internet' but now also more finetunable/specialized.
I'm going to look at siccing some LLMs on old Jenkins errors today, ideally it will give me more time to spend on building things (and hopefully less fires later).
I obviously don't expect it to magically fix everything but I have a good idea of what they're capable of, what they need help with, and what their pitfalls are
I can now have some scripts and some "agents" scan through 250GB of logfiles and do various sanity test things and either report back or make a branch and try to fix it, before i even need to look at it. It costs almost nothing and if it finishes early I can have it "go do the same thing but in rolling 3 month windows to see wtf happened when"
FWIW I think LLMs are more powerful for the language side of things. I get a lot of the programming improvement from python/metaprogramming/etc - but the combination is like having a group of interns you can make do a bunch of tedious/shit work that nobody wants to do.
1. things manifest a lot less
2. you don't have to worry about maintenance
I could see a proper structured LLM setup some day even resuscitating 'dead' software.
If you give it some documentation, background info, code and traceback info (and know to avoid context window poisoning) you can find it's capable of even helping solve rather tedious CUDA/etc issues.
Consistency - This is what GPUs were made for, putting a bunch of 'effort' almost instantly into investigating any regression that any real SW dev would ignore because, life's way too damn busy.