News for hackers. Stuff that matters.
Re:Asking an LLM to build a simple web tool by didgeoridoo on Wednesday July 10, 2024 @08:59AM (#40926390)
Claude 3.5 is an absolute revelation. Its writing is very humanlike, and the new “artifacts” window is a fantastic UX improvement. It has completely replaced GPT-4o for me in every use case.
Re:Asking an LLM to build a simple web tool by indigodaddy on Wednesday July 10, 2024 @09:48AM (#40926888)
How/where can you use it for free? Code editor plugins etc too?
Re:Asking an LLM to build a simple web tool by didgeoridoo on Wednesday July 10, 2024 @09:49AM (#40926897)
I don’t. I canceled my OpenAI subscription and shifted that $20/mo over to Anthropic.
Re:Asking an LLM to build a simple web tool by brylie on Thursday July 11, 2024 @10:16AM (#40937073)
I would gladly do that too if Claude supported voice-to-text prompting like Whisper.

I use ChatGPT around town and can’t always type easily. That and it is easier for me to speak stream-of-consciousness than slowly typing my thoughts. Typing is a significant bottleneck.

Re:Asking an LLM to build a simple web tool by AJSDfljff on Thursday July 11, 2024 @11:09AM (#40937568)
Whats your usage? The 'around town' triggered me to ask :)
Re:Asking an LLM to build a simple web tool by brylie on Thursday July 11, 2024 @11:20AM (#40937677)
All sorts of things.

I’m often thinking of ideas while walking or riding the bus. For example, I make plans for an English discussion group I’m facilitating while walking to the class. Or I’ll iterate on software project ideas as they come to mind, like creating design documents, GitHub issues, etc.

Another use case is for live translation where I speak English to ChatGPT and it translates the output to various languages. Then, I show the translation to the other person and let them speak their response for ChatGPT to translate. The interesting thing is somehow Whisper translates the foreign language back to English when it returns the transcribed text, so I don’t need to submit that text to ChatGPT. This live translation of spoken conversations is a really important feature particularly since I interact with many people with varying degree of English proficiency speaking many languages (namely, Russian, Suomi, Czech, and Farsi.)

Re:Asking an LLM to build a simple web tool by lgas on Wednesday July 10, 2024 @02:51PM (#40930116)
Re:Asking an LLM to build a simple web tool by genewitch on Wednesday July 10, 2024 @02:04PM (#40929644)
can you ask claude to build a python program that uses multiple processes/threads to check the collatz conjecture for N numbers, where N is however many "vCPU" you want to use; and paste that somewhere?

It doesn't even need to do the timestamping/benchmarking that my singlethreaded does, but i couldn't make 4o nor any local LLM give code that gave correct output when run. This is by far the most trivial and easily checked thing i can personally verify.

Re:Asking an LLM to build a simple web tool by lgas on Wednesday July 10, 2024 @02:50PM (#40930105)
https://claude.site/artifacts/e38182c3-840a-44ae-bc5a-8c2bb7...

Output:

    % python /tmp/collatz.py
    Number of vCPUs available: 12
    Checked Collatz conjecture for numbers 1 to 1000000
    Number with longest sequence: 837799
    Length of longest sequence: 525
    Time taken: 2.08 seconds
    python /tmp/collatz.py  19.47s user 0.41s system 728% cpu 2.731 total
Re:Asking an LLM to build a simple web tool by genewitch on Wednesday July 10, 2024 @03:12PM (#40930326)
that's good, i can consider paying for that service...

I'll have to check the actual results, of course, but that's much cleaner than what all the other LLMs output

Re:Asking an LLM to build a simple web tool by GistNoesis on Thursday July 11, 2024 @05:26AM (#40934914)
There is a bug :

N = 1000000 # Total numbers to check

chunk_size = N // num_vcpus

tasks = [(i * chunk_size + 1, (i + 1) * chunk_size) for i in range(num_vcpus)]

When num_vcpus = 12, chunksize = 83333, and 12*chunk_size = 999996, therefore we don't check the last four numbers.

Re:Asking an LLM to build a simple web tool by genewitch on Friday July 19, 2024 @03:17AM (#41002797)
that's okay because that N=1000000 line wasn't part of the spec but at least allowed me to run it as is without loading the rest of the stuff i test.
Re:Asking an LLM to build a simple web tool by reducesuffering on Wednesday July 10, 2024 @04:57PM (#40931402)
Now consider that 99.9% of humans can't do this, and combine that with the model trajectory and you have a very concerning next 5 years...
Re:Asking an LLM to build a simple web tool by mr_toad on Wednesday July 10, 2024 @09:59PM (#40933157)
I’m not worried. You still need a someone to specify the problem in a way the LLM will understand and check the code and the results. For now, that still needs to be a programmer.
Re:Asking an LLM to build a simple web tool by AJSDfljff on Thursday July 11, 2024 @11:13AM (#40937615)
It will be interesting how long business people overestimate what they can do until the tools get good enough that it is okay again.

Take small companies with eveyrone being business people besides the nerds in the IT department. If you don't have one security aware person, this IT department will write unsecure code without anyone in the company every questiong this every.

You know like 'hidden' api keys in an mobile app everyone can extract, still using plaintext for passwords or md5, having ports open, old operating systems etc. etc.

Re:Asking an LLM to build a simple web tool by latexr on Thursday July 11, 2024 @07:15AM (#40935434)
> Now consider that 99.9% of humans can't do this

OK. Now consider that those same humans have no need to do it or even attempt it.

Re:Asking an LLM to build a simple web tool by agmater on Wednesday July 10, 2024 @12:20PM (#40928540)
> The results are still not perfect in terms of alignment and space, but they're good enough.

CSS frameworks come in very useful for this, ask it to generate Bulma or Bootstrap valid markup.