Comment by akst

I know "next-generation" is just SEO slop, but I'm going to hyper fixate on this for a moment (so feel free to ignore if you're actually interested in Positron).

I think the future of data science will likely be something else, with the advent of WebGPU[1] (which isn't just a web technology) and the current quality/availability of GPUs in end user devices, and a lot of data computation clearly standing to benefit from this.

The real next generation of data science tools will likely involve tools that are GPU first and try to keep as much work in the GPU as possible. I definitely think we'll see some new languages eventually emerge to abstract much of the overhead of batching work but also forces people to explicitly consider when they write code that simply won't run on the GPU, like sequential operations that are nonlinear, nonassociative/noncommutative (like highly sequential operations like processing an ordered block of text).

I think WebGPU is going to make this a lot easier.

That said I'd imagine for larger compute workloads people are going to continue to stick with large CUDA clusters as they have more functionality and handle a larger variety of workloads. But on end user devices there's an opportunity to create tools that allow data scientists to more trivially do this kind of work when they compute their models, process their datasets.

[1] Other compute APIs existed in the past, but WebGPU might be one of the most successful attempt to provide a portable (and more accessible) way to write general GPU compute code. I've seen people say WebGPU is hard, but having given it ago (without libraries) I don't think this is all that true, compared to OpenGL there are no longer specialised APIs to load data into uniforms everything is just a buffer. I wonder if this has more to do with non JS bindings for use outside the browser/node or the fact you're forced to consider memory layout of anything your loading into the GPU from the start (something that can be abstracted and generalised), just in my experience after my first attempt at writing a compute shader it's fairly simple IMO. Like stuff that always complicated in rendering like text is still complicated, but at least its not a state based API like web/opengl.