WelcomeWEFreedomWindsorEssexCounty-ActionCommunicationsRalliesResourcesSocialEventsWednesdayNightMeetings-ActivismAgoraAgorismCanadaCommonLawCommunityDataCentersDataCentresDoocracyFarmersMarketIntroductionsNaturalLawNewslettersOntarioOrganizingOutreachProjexProjexWikiRalliesResistanceAnalysisSoireesSoiréesWikiSpooksWindsorElectionsTop Subs
1
videos - - (+0|-0)
The Greatest Lie Ever Told(www.youtube.com)
3 commentsbyLarrySwinger
Add topics

Comment preview

[-]x0x71(+1|0)

The standards of how we successfully train AI really excites me for the future. Not only because of our ability to use it but also because of what it can teach us about learning and how best to do it (like measured with real numbers).

When you train an AI you keep all weights flexible. You don't fix one part and then train the remaining weights to the data. Technically you can always over fit the remaining part if it has enough degrees of freedom and you run the cycles. But it won't give you good results.

Modern academia does the opposite. It assumes things like all races must have the same baseline psychological properties, or that if there are different psychological archetypes in the population that each civilization and race will have them in the same probabilities, instead of potentially an entirely different set of personality types. Then it tries to fit a social model over that. That's like keeping your rear most weights fixed in AI.

I'm hoping what AI will teach us is that as a standard practice you should consider all parts of a model flexible and then perform a regression to the data with all layers together. And that if it really is appropriate for part of a model to be fixed then it won't move even when you let it be the trainable part of a regression.

You only ever fix a part of a model if it seems to be bouncing around the same values after some training and you want to stop wasting training rate on it or if it was trained elsewhere and you don't have the resources to fine-tune it as well.

Soon it will be considered unscientific to have fixed assumed parameters in anything. And when every social model includes flexible weights to factor race those racial weights are going to soak up a lot of gradient.

parentlinkreply
[-]LarrySwinger0(0|0)

That's interesting. Do you see accuracy about social sciences or the humanities appearing in any LLM right now? I'm really unimpressed with both ChatGPT and Claude as I believe I've told you. I've only tried Claude a little to be honest. Fuck Fuck Hoe requires two clicks to get to the AI, I don't know if it even has a direct landing page so that's why I'm not using it that much.

parentlinkreply
[-]x0x70(0|0)

Well LLMs are based on the text they read and the reinforcement they are given. There is nothing scientifically objective about them. I mean that the model training practices are healthy for honest science. No givens, no assumptions, fit all parts of the model to your data.

parentlinkreply