3 ms·

There's another way to do this: English -- (LLM) --> python Human fine tunes python Python -- (py2many) --> smt2 --> z3 --> result Perhaps some steps can be

There's another way to do this: English -- (LLM) --> python Human fine tunes python Python -- (py2many) --> smt2 --> z3 --> result Perhaps some steps can be eliminated in the future.

adsharma 2y ago
With an example: https://gist.github.com/adsharma/45fbb065a8fe793030e8360daebffba1 https://gist.github.com/adsharma/45fbb065a8fe793030e8360daeb... You could have a puzzle that's described in plain English. I think the current day tech is sufficient to translate the puzzle to a python constraint like in this gist. The argument is that the SMT solvers are a better match for solving these problems than expensive matrix multiplication paid for by someone else. Of course, SMT is not a solution for every situation that Wolfram language is used for today. But a lot of the reasoning benchmarks fall in this category.
adsharma 2y ago
There is a wolfram blog post comparing python vs wolfram from 2021: https://blog.wolfram.com/2021/11/04/lessons-learned-migrating-from-python-to-the-wolfram-language/ https://blog.wolfram.com/2021/11/04/lessons-learned-migratin... The linked tweet (https://x.com/WolframResearch/status/1325935891651698690?s=20 https://x.com/WolframResearch/status/1325935891651698690?s=2...) is inaccessible. Does anyone have an archive? I'd love to see pandas/df compared instead of the iterative style.
- Qem 2y ago
  > There is a wolfram blog post comparing python vs wolfram from 2021 Now the view from someone who quit Mathematica walled garden to embrace the Python ecosystem (2018): https://paulromer.net/jupyter-mathematica-and-the-future-of-the-research-paper/ https://paulromer.net/jupyter-mathematica-and-the-future-of-...
pjmlp 2y ago
The same thought process as installing GNU/Linux on a random PC bought down the mall, or getting home with macOS/Windows already pre-installed.
- adsharma 2y ago
  This sounds like an argument for a custom DSL designed for the purpose of expressing constraints and for querying statiatical data. Perhaps there are cases where DSL is the right answer. Not sure if this is one though. I see some parallels to the arguments around using a ML derived language instead of adding types to python3 and giving it 90% of the capabilities. There are other factors like: open vs closed development model, number of implementations (Wolfram language has exactly 1?) etc. From my perspective, the problem to solve is that the python ecosystem isn't even trying. Would love to see stats about the adoption of the match statement introduced 3 yrs ago.

There's another way to do this: English -- (LLM) --> python Human fine tunes python Python -- (py2many) --> smt2 --> z3 --> result Perhaps some steps can be eliminated in the future.

adsharma 2y ago

With an example: https://gist.github.com/adsharma/45fbb065a8fe793030e8360daebffba1 https://gist.github.com/adsharma/45fbb065a8fe793030e8360daeb... You could have a puzzle that's described in plain English. I think the current day tech is sufficient to translate the puzzle to a python constraint like in this gist. The argument is that the SMT solvers are a better match for solving these problems than expensive matrix multiplication paid for by someone else. Of course, SMT is not a solution for every situation that Wolfram language is used for today. But a lot of the reasoning benchmarks fall in this category.

adsharma 2y ago

There is a wolfram blog post comparing python vs wolfram from 2021: https://blog.wolfram.com/2021/11/04/lessons-learned-migrating-from-python-to-the-wolfram-language/ https://blog.wolfram.com/2021/11/04/lessons-learned-migratin... The linked tweet (https://x.com/WolframResearch/status/1325935891651698690?s=20 https://x.com/WolframResearch/status/1325935891651698690?s=2...) is inaccessible. Does anyone have an archive? I'd love to see pandas/df compared instead of the iterative style.

Qem 2y ago
> There is a wolfram blog post comparing python vs wolfram from 2021 Now the view from someone who quit Mathematica walled garden to embrace the Python ecosystem (2018): https://paulromer.net/jupyter-mathematica-and-the-future-of-the-research-paper/ https://paulromer.net/jupyter-mathematica-and-the-future-of-...

pjmlp 2y ago

The same thought process as installing GNU/Linux on a random PC bought down the mall, or getting home with macOS/Windows already pre-installed.

adsharma 2y ago
This sounds like an argument for a custom DSL designed for the purpose of expressing constraints and for querying statiatical data. Perhaps there are cases where DSL is the right answer. Not sure if this is one though. I see some parallels to the arguments around using a ML derived language instead of adding types to python3 and giving it 90% of the capabilities. There are other factors like: open vs closed development model, number of implementations (Wolfram language has exactly 1?) etc. From my perspective, the problem to solve is that the python ecosystem isn't even trying. Would love to see stats about the adoption of the match statement introduced 3 yrs ago.