On 08.05.25 12:43, Wouter Verhelst wrote:
Therefore, I conclude that, practically, we cannot include models in Debian if we want them to be reproducible.
Note that there are other obstacles to reproducibility besides "we can't distribute the data 'cause they're non-free".
We might not be able to because they are too damn big.Rebuilding the neural net might require specialized hardware (or just commodity hardware but too much of it) which neither our builders nor our developers have.
Rebuilding might take too long to be useful to anybody.The build process might be nondeterministic (how *do* you coordinate umpteen GPUs that all work to incrementally adjust the weights of a large network? Answer: you don't) and thus not be reproducible in principle.
NB I agree that people-intelligence isn't the same as computer-intellence, but that's 'cause currently there's something missing from the latter; the misspelling is intentional. Call it common sense, grounding in physical reality, learning from adverse as well as exemplary training data, or inheriting a brain structure taht's been refined by a couple million generations via evolution instead of a few ten iterations of neural network engineering by humans
The problem is that all those missing factors are destined to go un-missing — and then what? We can't base our rules on biological exceptionalism.
-- -- regards -- -- Matthias Urlichs
Attachment:
OpenPGP_signature.asc
Description: OpenPGP digital signature