[rvieira]

I really think that Python is not a good language for ML, it just got "there" first.

The ecosystem is the real plus, of course. But the language is a headache for this. I agree with the "false economy" angle. I would happily trade the "agility" of dynamic "gluing" with some kind of real type safety, human-readable error messages and performance[0].

[0] - hiding C in Python's clothes doesn't count :)

[kaba0]

Python is the de facto glue language with one of the biggest ecosystems out there, that makes it possible to use any kind of over-the-top library that does 1093 things after a single `import antigravity`. Also, ML absolutely makes sense for python, it’s not like most PLs have actual support for video cards — ML is very specifically about manipulating data (a fundamentally dynamic task) and calling out to specific libraries for training, a very glue-task. Give me any language better than python for that.

Quoting Brooks (butchered): “the only significant productivity improvement comes from relying on code that is already written”. Your fancy “better” language has not even 1/10th of what python has, it won’t replace it.

[rvieira]

> Python is the de facto glue language with one of the biggest ecosystems out there

I never contested that.

> ML is very specifically about manipulating data (a fundamentally dynamic task)

I disagree strongly with using dynamic languages for data. Data has dimensions, units, types. You need to know that you're not adding coats to horses or USD to EUR. You need to know that you didn't silently sliced by the wrong axis. You may want formal verifications. You may to transform data without worrying about silent errors.

All the "metadata" and wrapper classes ML in Python are just trying to give you what the language can't.

> a very glue-task. Give me any language better than python for that

That's my point exactly. ML has evolved beyond glueing a few C libraries. It needs complex, big programs, which is an area where Python is terrible. Also the different nature of the "glued" components (each with its own data formats, protocols and calling conventions) makes the glue a mish-mash of untyped mixed magic idioms.

[kaba0]

Training and using ML is different. It has been relatively common to bundle the trained weights with a different programming language system - but training can be a more exploratory phase, so python is not a bad fit for that.