Shamar

The documents provide smoking-gun evidence that Meta, Google, Snap, and TikTok all purposefully designed their social media products to addict children and teens with no regard for known harms to their wellbeing, and how that mass youth addiction was core to the companies’ business models. The documents contain internal discussions among company employees, presentations from internal meetings, expert testimony, and evidence of Big Tech coordination with tech-funded groups, including the National Parent Teachers Association (PTA) and Family Online Safety Institute (FOSI), in attempts to control the narrative in response to concerned parents.

“These unsealed documents prove Big Tech has been gaslighting and lying to the public for years

# Declaration

We, the undersigned members of the Open Source community, assert that Open Source is defined solely by the Open Source Definition (OSD) version 1.9.

Any amendments or new definitions shall only be recognized if declared by clear community consensus through a transparent process to be determined.

The problem with hype around LLM is that people without much experience in the field can't think of anything else.

So much they forget the basics of the discipline.

What do you think cross validation is for?

To compare different weights obtained from different initializations, different topologies, different hyper-parameters... all trained from the same training dataset.

Even for LLM, have you ever tried to reduce the size of the vocabulary of, say, Llama?

No?

Yet it's a totally reasonable modification.

What's the preferred form to make modifications like this?

Can you do it fine tuning llama weights?

No.

You need training data.

That's why training data are the preferred form to make modification, because whatever the AI (hyped or not) it's the only form that let you make all modifications you want.

Compiling Emacs on a PDP-11 wasn't cheap either.

And by today standards, a PDP-11 is quite comparable with the cost of the server farms used in training.

And yet Emacs was released under GPL.

So the economic argument is pretty miopic.

It's not just marketing: European AI Act impose several compliance obligations to corporations building AI system, including serious scientific scrutiny on the whole training process.

Such obligations are designed to mitigate the inherent risks that AI can pose to individuals and society.

The AI Act exempts open source from such scientific scrutiny because it's already transparent.

BUT if OSI defines black boxes as "open source", they open a loophole that will be exploited to harm people without being held accountable.

So it's not just marketing, but dangerous corporate capture.

The term "absorbed" was not for the people in the field, but for people who don't know what folding means.

IMHO it's a better metaphor then "learning", because learning is a _subjective_ experience that everyone does and using that term lead inevitably to anthropomorphisation.

"Absorb" match the insight of filters and pipelines, that can be easily understood from any CS student, any "ML expert", any lawyer and any other citizen.

____

As for the network, my argument is simple: if I get back the source dataset from the executable, I think we can agree that such dataset is projected on the numerical matrices that such executable record.

Now where is the dataset?

You might argue that it is recorded _only_ into the gradients logged there (the gradients applied to one single "neuron" for each "layer"), but if so you could reconstruct the source dataset from the logs alone, and in fact, you cannot. You need both the "model" and those gradients in the correct order (and the encodings of inputs and outputs, obviously).

You might ask: "fine, but how much of the source dataset is projected into the gradients and how much is projected into the model?"

To answer, we need to consider that

- the vector space that constitutes the executable is non-linear (the "model" part) and hierarchical (the vectors of the gradients are not independent neither between layers nor between samples)

- (initialization apart) all the values (and the operative value) that the "model" contains comes from the source dataset

Thus I argue that a substantial portion of the source dataset is contained in the "model".

This does not exclude that another substantial portion of the source dataset is also contained into the few logged gradients!

And in fact I've never stated that the "model" contained the whole source dataset.

But if the portion contained into the "model" was negligible, you would be able to get back the sources from those logged gradients alone with negligible errors.

AFAIK, it is not possible, but if you can, please teach me how! I'm always more than happy to be proven wrong if I can learn how to do something that I previously thought impossible!

I strongly encourage you to prove your statement with a script that use the logs saved and a random "model" and get back the exact source dataset.

Beyond the global output and error of each sample from the last epoch, the log also includes the weight update of one single (fully connected) node for each layer.

During the compilation phase, the training dataset is projected on a complex vector space that is constituted by both the "model" of the "neural network" and these logs.

It's just like projecting a shadow over a bidimensional surface: if you discard the data pertaining to one dimension you have no hope to guess what projected it: you need both dimensions.

The logs that are preserved in the compilation process is the part of the vector space that is usually discarded during the "training".

But discarding the "model" would have exactly the same effect: you cannot get back the source dataset from those logs alone. That's why this does not "smuggle the training dataset back".

Indeed the fact that the source dataset is obtainable from the couple "these logs" + "final model", but neither from "these logs" alone nor by the final model alone, proves that a substantial portion of the source dataset is always embedded in the "model", that becomes a derivative work of the sources.