The striking idiocy of keeping making multimodal “local” models to fit into the intentionally (to maximize profits) memory-limited crapbooks of the Mac cult victims, is beyond my comprehension.

Yes, nothing local runs on an over-priced 16GB crapbooks, and barely runs (~ 1 token/sec) on 32GB crippled Mac Minies, since there are no mainstream Macbooks with 32GB and the top-tier Pro models cost like $3K+.

To fullfill the demand of the igorant cultists all kinds of Q4_0 crappy factorizations are being built, which produce a very generic, commonplace slop for reasons discussed at length here.

And yet, instead of building a decent, text-only, trained on the code of just a few decent languages, to make vapecoding at least less cringe, they keep pushing multimodal crap, which is about 25Gb for Q8_0.

Why, for the love of Steve, why? Who TF needs a multimodal crap? What for?

By the way, training text-only models on pirated textbooks (come on, every single one does this, even indirectly, by buying cleaned up data sets for training) and high quality codebases (avoiding amateur crap which skews models to create a subpar, amateurish crappy code) is a viable strategy.

I bet that Google, MS and the others routinely “indexing” their vast codebases into special puprose models which they use to auto-complete in their crappy, verbose badly designed imperative API calls.

It has been understood (at least by some) long ago, that one has to carefully select not just the training data, but even the order in which one feeds the training examples into a slop generator, so that the resulting “theoretical topology” (there are no high dimensions, just weights) would better capture and “match” Reality it has been abstracted out from.

The training sets for particular languages shall be distinct and perhaps the standard library shall be feed in first. I propose a much as an order of magnitude improvement of the code slop quality when trained on textbooks, then public interfaces of the Standard Library, its idiomatic use, and then all the decent codebases one can get hold off, one language at a time, without mixing and even without human annotation.

And suddenly the way too smug 16Gb Macbook coffeeshop dwellers could even have some non bullshit models to run.

This approach can be obviously generalized to all the STEM subjects (from [the fundamental] principles to particulars, one at a time, excluding graphomania and marketspeak texts).

As a benchmark example one could train the Haskell-specialized model on the classic textbooks (Bird, Wadler, the Reports, etc.) and the non-monad-transformer-plagued codebases, like prelude, GHC itself, cabal, xmonad and all the other code written by people who understand what they are doing and why (and, of course, the underlying math and algebraic abstractions).

Such a model would easily BTFO all the generic coding models and, perhaps, even the datacenter-sized “frontier” ones, since the quality of “inner representation” – the resulting DAG (yes, this is a DAG) beats every other metric in principle.

Anyway, just train text-only specialized models instead of useless multimodal crap and think twice before buying an overpriced soldered 16Gb crapbook.

Thank (You) for your attention to the matter.