Data Science Agent and Code Transformation | Hacker News Reader

Data Science Agent and Code Transformation

labs.google.com

113 points_endif_12 comments

Wow, their example to "clean up the code" does a bit more than just refactoring to make code more readable, it appears to change the output.

One would have to check the resulting code carefully to see if the meaning is still as originally intended, or replaced with code that is more probable to be correct (but no longer working).

For instance, it replaces this:

  if dataset == 'animals':
    if dataset == 'turtle':
      x_train, y_train, x_test, y_test = datasets.load_turtles(with_bowtie=False)
    elif dataset = 'formal_turtle':
      x_train, y_train, x_test, y_test = datasets.load_turtles(with_bowtie=True)
  else:

with this:

  if dataset == 'turtle':
      x_train, y_train, x_test, y_test = datasets.load_turtles(with_bowtie=False)
  elif dataset == 'formal_turtle':
      x_train, y_train, x_test, y_test = datasets.load_turtles(with_bowtie=True)

The before-code responds to dataset='animals' with `load_turtles(...)` and to dataset='turtle' or 'formal_turtle' with an error; In the after-code this is reversed, although the apparent logic error and the assignment/equals sign error are resolved.

Actually, the code before does nothing if dataset is set to 'animals', 'turtle' or 'formal_turtle', most of the branches are inaccessible. Also, the extra else clause that raises an error and the line

    elif dataset = 'formal_turtle':

are both invalid syntax.

I think 'clean up' here means something closer to 'convert this to what I'm trying to write'.

Agreed, but I have to say data cleaning is actually one of the hardest step, LLMs are simply not there yet.

It's almost impossible to for LLM to tell all the invalid rows at once since the data cannot be fit into the context window. If we prompt the model to thoroughly do data cleaning, there will be many try-and-fail steps. This happens to me as a human, I clean some rows, expect my program to run with the data, only to find there are more malformed data. LLM cannot get it right for now, actually I see many cases that LLM fails because it wants to convert types (e.g. string to date).

Based on my experience, the best way is simply to skip the data cleaning step in the planning stage (you can provide feedback asking the tool to not do any steps).

Unfortunately, this is true with most LLMs.

Very cool. Something similar is Streamline Analyst [1]. It automates preprocessing and model development from an uploaded dataset.

This is great for a more technical user (or student) who want to parse generated code and hack away. Instead [1] generates a dashboard app which is faster to use, but sacrifices interpretability.

[1]: https://github.com/Wilson-ZheLin/Streamline-Analyst

looks great, but it is only available in usa.

That's some privacy notice. I think I'll pass

I feel you, this definitely is a Google-wide issue across products (e.g. https://x.com/levelsio/status/1831840497629065656). These products themselves are worth a try IMHO.

This and the underlying tweet makes me feel seen.

Why does developing with Google have to be so hard?

How is this different from what literally any other hosted services company does? If anything, Google is being comparatively honest and transparent.

100%. The good part is, if you follow the tweet thread, you can see they are trying to improve things. Hopefully the effort can land and the rest products can follow.

At least they put it up front, instead of burying it on page 14 of some bottom link.

Reply on news.ycombinator.com