i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark@social.treehouse.systems I didn't know the ideal number for code to behave differently was over 30% of the time!
Then again, I like and don't mind working with legacy code and systems so I personally tend to wonder "why even redo a working thing" -
@whitequark @porglezomp I'm spitting out my drink at j++ → j--. Holy shit.
@xgranade
I think the right is the output from running the model on the right code (center being the "desired output"). So it's not changing the semantics of the loop, just not not changing the loop order to match their desired outcome.
Given that loop order can have behavioral impact (and I would never trust an LLM to be able to tell if it did), that seems like the correct behavior to me though
@whitequark @porglezomp -
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark The Code Randomizer (TM)
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark well two out of three ain’t bad. No, wait…
-
@lu_leipzig I view code as art and so any tool that puts determinism strictly above aesthetics is a net negative to my craft
@whitequark @lu_leipzig Ideally, I think a formatter that learns how I formatted the rest of the buffer would be the goal.
Most of the time I like the deterministic formatting. However, I find deterministic formatting fails me around function headers and long function calls / long boolean statements.
I want it to do the deterministic formatting once, and then if I undo immediately, don't do it again to that area... and preferably learn what I was trying to do.
-
@ireneista TIL that my philosophy is the same as the Extreme Programming philosophy
@krans @whitequark it was a nice name for a movement, it did a good job of conveying that the goal was radical change
at the time, from what we can tell, none of the people saw it as a labor movement specifically, which is too bad... that might have prevented it from being watered down by successive cycles of consulting and renaming
-
@xgranade
I think the right is the output from running the model on the right code (center being the "desired output"). So it's not changing the semantics of the loop, just not not changing the loop order to match their desired outcome.
Given that loop order can have behavioral impact (and I would never trust an LLM to be able to tell if it did), that seems like the correct behavior to me though
@whitequark @porglezomp@robin @xgranade @porglezomp oh you're right
-
@whitequark @lu_leipzig Ideally, I think a formatter that learns how I formatted the rest of the buffer would be the goal.
Most of the time I like the deterministic formatting. However, I find deterministic formatting fails me around function headers and long function calls / long boolean statements.
I want it to do the deterministic formatting once, and then if I undo immediately, don't do it again to that area... and preferably learn what I was trying to do.
@theeclecticdyslexic @lu_leipzig my goal is to be able to run a command on a patch that formats the added lines "more or less like the rest of the file"
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
Figure in question seems to be about "model performing in its ideal conditions"
The author's actual opinion is implied in the Results:
"After inspecting the compilation checking module, we found that DUET CS achieves 55.8% computational accuracy, which is a practical metric for a code generation system. This result shows that more than half of the output code are compilable and implement the same function as the input code. The user can
use this check as an optional layer of the pipeline to guarantee grammar correctness.
...
We found that even the non-compilable outputs display around 60% similarity to the ground truth, which means even if DUET CS cannot always produce grammar-correct code, it can still provide valuable information to help user to transfer code style.
...
Notice, that generally the task of generating the exact same code as ground truth is very hard, especially when the code length is rather long (˜47 lines)." -
Figure in question seems to be about "model performing in its ideal conditions"
The author's actual opinion is implied in the Results:
"After inspecting the compilation checking module, we found that DUET CS achieves 55.8% computational accuracy, which is a practical metric for a code generation system. This result shows that more than half of the output code are compilable and implement the same function as the input code. The user can
use this check as an optional layer of the pipeline to guarantee grammar correctness.
...
We found that even the non-compilable outputs display around 60% similarity to the ground truth, which means even if DUET CS cannot always produce grammar-correct code, it can still provide valuable information to help user to transfer code style.
...
Notice, that generally the task of generating the exact same code as ground truth is very hard, especially when the code length is rather long (˜47 lines)."That last one is a funny statement because it's laughably easy for a human to maintain the execution of a function after a style refactor. You would reprimand a junior if they couldn't do that
-
@theeclecticdyslexic @lu_leipzig my goal is to be able to run a command on a patch that formats the added lines "more or less like the rest of the file"
@whitequark @lu_leipzig that's a pretty reasonable concept I think.
I like the idea at least.
One thing I will say of deterministic formatters is they have changed my habits over time in order to get it to format the way I want. You can take that as both good and bad, but I think most (maybe 60%) of the things they have forced on me have been good.
Edit: I also get stun locked trying to decide how to format 15 lines of code far less often.
-
@whitequark @lu_leipzig that's a pretty reasonable concept I think.
I like the idea at least.
One thing I will say of deterministic formatters is they have changed my habits over time in order to get it to format the way I want. You can take that as both good and bad, but I think most (maybe 60%) of the things they have forced on me have been good.
Edit: I also get stun locked trying to decide how to format 15 lines of code far less often.
@theeclecticdyslexic @lu_leipzig yeah if a formatter requires me to do things I don't want I simply quit using the formatter (and sometimes the codebase)
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark what the what.
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark Just like, use one of the tools that already exists? It'll be:
- Fast
- Cheap
- Efficient
- AccurateI don't understand any of this "industry" outside of being a massive destructive boondoggle.
-
R relay@relay.an.exchange shared this topic
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
Why is AI as a consumer product being rushed out the door so fast? It's obviously not ready for prime time. It's unreliable, inaccurate, and fragile.
It's like a car being sent out to car dealerships with only 3 wheels with hasty promises of a future 4th wheel.
Possibly the goal isn't a car with 4 wheels but a plan for something else, similar to a Waymo power outage gridlock
Reliance on AI is a national security risk vulnerable to high fuel prices
1/
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark "92 boosts, 115 favourites" damn
I swear to god sometimes Mastodon is just "old person yells at thing".
Researchers spend tons of time and money trying to solve Sudoku in polynomial time not because Sudoku is such a important problem to humanity, but because it's a NP-Hard problem, and you can thus reduce all other NP-complete problems to Sudoku and solve them all in polynomial time if you can solve Sudoku in polynomial time.
The research challenge is disentangling content from style in a learned embedding space, it's a classic representation learning problem that's genuinely hard: 1) Two functions that do the same thing should have identical content embeddings but different style embeddings, 2) Style must generalise to unseen code patterns, not just pattern-match known rules, 3) It's unsupervised, so there's no labeled (code_A, same_code_in_style_B) training pairs.
Code formatting is actually a very good medium to test this hypothesis, because you have an infinite latent space of code that does the exact same thing but is stylistically different.
-
i'm at a loss of words after reading a paper about reformatting code using an ML model that has a measured statistical quantity A_c which says how often the reformatted code behaves the same as the original
the "ideal" (their choice of words) case is 64.2%
@whitequark feck that
-
@whitequark @deborahh @danlyke ie, the sort of thing a linter does?
-
@whitequark "92 boosts, 115 favourites" damn
I swear to god sometimes Mastodon is just "old person yells at thing".
Researchers spend tons of time and money trying to solve Sudoku in polynomial time not because Sudoku is such a important problem to humanity, but because it's a NP-Hard problem, and you can thus reduce all other NP-complete problems to Sudoku and solve them all in polynomial time if you can solve Sudoku in polynomial time.
The research challenge is disentangling content from style in a learned embedding space, it's a classic representation learning problem that's genuinely hard: 1) Two functions that do the same thing should have identical content embeddings but different style embeddings, 2) Style must generalise to unseen code patterns, not just pattern-match known rules, 3) It's unsupervised, so there's no labeled (code_A, same_code_in_style_B) training pairs.
Code formatting is actually a very good medium to test this hypothesis, because you have an infinite latent space of code that does the exact same thing but is stylistically different.
@budududuroiu the reason I was reading the paper is because I'm working on the same problem and I think the encoding presented in the paper makes no sense at all to use
-
@budududuroiu the reason I was reading the paper is because I'm working on the same problem and I think the encoding presented in the paper makes no sense at all to use
@whitequark to use for what? It's research, it's not meant to create something for industry use. Academia already suffers from the "File-drawer problem". I also did research on using GANs for Outlier Detection, when most of the time Outlier Detection is a classification problem, not a learned representation problem.
