#no cats

[lowkeloki]

"Sorry no cats."

[knight-in-sour-armor]

[my-blue-birdie]

I veto cats.

[tilbageidanmark]

Ignore the sign

[bits-of-wit]

[umbrellacorphq]

There are no cats in this facility. Repeat: there are NO cats in this facility. If you have recently seen a cat, lock yourself in a nearby office or storage room immediately.

[hardtruthsfromsoftcats-blog]

AI Alignment

While this hiatus continues, here are (as promised) some things that can (possibly) tangibly improve humans' relation to AI. I thought today was a good day for this post since it's finally caught Congress's attention. 1. Require all training methods of neural networks to be documented and released. Require a precise, anonymized description of the data that is being used; if the data is public, release the datasets.

2. Ditto on test and validation sets, and report accuracy.

3. Report other metrics as well; precision, recall, etc.

4. Develop standards that leverage progress in the Safe ML community to quantify risk of errors, and require an impact report that states the possible risks of those errors along with comparisons to other models.

5. Enforce monopoly laws that are on the books and prevent giant horizontal and vertical companies from existing and having access to huge troves of multimodal data at no additional cost.

6. Require companies above a certain revenue/user count to fund educational programs for the public and schools on how to use AI technology to benefit them. Especially target this at historically underserved communities.

7. Require companies to pay a carbon tax for the footprint of training and deploying large (and small) models.

8. This doesn't actually solve much, but rename OpenAI to ClosedAI so people who are not in the know don't get confused; then, allow a not-for-profit to create a consortium called OpenAI that focuses on releasing software, maybe StabilityAI.

9. Require publicly disclosed risk assessments from corporations pertaining to trained models. I can easily see how we could move into a new era of LLMs trained specifically for generating malware and deploying them at scale. That's just one risk.

10. Last one for now, require people whose data is used in training a model to receive a cut of the profits. If Spotify can do it (poorly), Google, OpenAI, MS, etc. can do it (at least poorly) too.