Post

So here's the problem with AI training. If you do not do any curation of the data, your training data looks like this: a standard Gaussian bell curve, familiar to anyone who took statistics in school.

Caliban Darklock

@darklock.com

This isn't what you WANT the data to look like. You want it to look like this: that huge peak in the middle is culled and compressed, so there's less "average" result, and the entire bottom third of the data -- the low-quality outliers -- gets chopped off. Now your responses skew above-average.

26/10/26, 7:43 pm0 reposts1 like

Caliban Darklock

@darklock.com

21h

But what the tech dumbasses have done instead is curate indirectly, by training on social media: the space where MINIMAL THOUGHT AND EFFORT is put into the data, so it is overwhelmingly in that central peak (which you wanted to REDUCE) and the top end simply isn't there.

Caliban Darklock

@darklock.com

21h

Basically, we have begun a feedback loop that will inexorably drive the machine to give dumber and dumber responses. Furthermore, by letting people post machine results ON social media, you dramatically raise the chance of feeding results back into the machine and hastening model collapse.

Caliban Darklock

@darklock.com

21h

And we've done this because there's an inescapable truth about how we train the machine: we need too much data. The training mechanisms are too slow and make too small a difference. But because our data source is hideously polluted, faster training just makes the problem worse.