Hacker News Reader

Comment by kevmo314 | Hacker News Reader

Peter Norvig has written a good analysis with some interesting takeaways: https://github.com/norvig/pytudes/blob/main/ipynb/AlphaCode....

View on HN

Kevin Wang, competitive coder: "I also try to minimize the amount of code I write: each line of code is just another chance for a typo."

Peter Norvig: This is why AlphaCode learned to write code with one-letter variable names, and with no comments or docstrings.

To be fair, unless you're doing something absurd the length of variable names shouldn't significantly affect the number of lines. Also comments with typos aren't really a problem. I don't like hard-to-read code either, but the more years I work the more I think that the shortest (within reason) code is the best code.

That is to say, golfing is bad unless for fun, shorter code is usually better, the shitty hard to read code is probably a reflection of human laziness rather than misguided tersity.

"If I had more time, I would have written shorter code"

Heh definitely an eyebrow-raising comment! On a second read of it though, I took it more as "because minimizing the code written is a pattern observed in the wild, AlphaCode mimics it as that's what it learned from."

I might be terribly wrong though. :-D

This is why neural networks need to have pain buttons that humans can repeatedly press. Of course that's how we get killbots.

I tried throwing below programming problems at ChatGPT

1. Write a static file server in Go

2. Write Go code to convert Color image to B/W

For both I got results. I know both are simple but still it's fascinating that AIs can write code. I have written more about it here https://rohanrd.xyz/posts/surprising-capability-of-ai-code-g...

I wrote a similar prompt but for a discord bot that would take a youtube url and some timestamps and return a gif.

Afterwards I asked it to add an option to "deep fry" the gif. Not only did it produce the correct code it also understood what I meant when referring to deep fried gifs. I was definitely impressed.

I threw a leetcode easy description into ChatGPT and submitted the solution without any editing. It passed all the test cases and got in the 95th percentile for efficiency of C++ submissions.

I haven't tried anything more advanced, but to go from simple requirements to solution without any clarification or even method signatures (It guessed the correct method signature down to the name and input params) was pretty dang impressive to me.

Chances are, you gave it a problem that was actually in its training set.

It's still impressive to me even if it was in its training set. The data used for training is not stored verbatim in the OpenAI model. It would still need to parse the context anew and solve the problem given its understanding of the boundaries of what I've written.

On contrast, I did it for projecteuler and told it needs to be extremely efficient or math. Didnt give me an appropriate result.

I have tried given a simple test suite and asking for an implementation that fulfills it, then asking to modify its impl and the test suite with a new requirement and it did it

There’s something to be said about publishing blog posts as Jupyter notebooks straight to your GitHub. A strong argument won’t need decoration.

Doesn’t display correctly on my iPhone though. :/

Almost half the screen real estate on my iPhone SE was white space. I had zero interest in reading something, especially something that long and in depth on a page that wastes so much space for literally nothing.

Typography, layout, and presentation counts for more than coders such as myself will ever admit publicly :)

This is incidental. The current Jupyter CSS needs to be improved - so what? That's not a bug unique to Jupyter, I've seen dozens of other websites with similar problems (some with extremely bare HTML and CSS that was minimal but bad - so the problem is unrelated to complexity, too).

This has nothing to do with the parent comment's topic, which was about using computational notebooks as blog posts, and is in effect "complaining about tangential annoyances—e.g. article or website formats, name collisions, or back-button breakage" - which the guidelines explicitly forbid.

I've been following the progress of AI code generation and I have to say, the results are impressive. There are still some limitations to these models, such as reproducing poor quality training data and difficulty maintaining focus throughout a problem.

I think using more diverse and high-quality training data could help address these issues, and incorporating additional constraints and regularization techniques into the model's training could prevent hallucinations and improve its overall reasoning abilities.

While there is room for improvement, the progress in this field is exciting and I can't wait to see where it will lead.

NB: This comment was written by GPT-3 after reading the article. The last few months of AI have been frankly mind-boggling.

Reply on news.ycombinator.com