oliver910

I think the response is pretty reasonable. Fines/punishments from regulators internalise externalities and disincentivise antisocial behaviour.

If Rio just writes off the fine without improving its processes in tandem then it could make sense for the regulator to intervene further, but in either case Rio should be on the hook for damages rather than WA taxpayers.

I agree that raising awareness that tools like this are possible is important and that sufficiently advanced actors can do this anyway, however I don't think in this case releasing pre-trained weights to the general public is responsible. This could probably be used to help bypass crypto exchange KYC for moneylaundering purposes. I'm not sure what the best access model is - email us with a good reason to get access to the weights perhaps - but what alarms me is there seems to be no consideration for misuse or responsible release at all.

Exactly. Even in the short term, there are some very worrying behaviours in AI systems which have real consequences. For instance:

* Some researchers recently built an AI system to generate low-toxicity molecules as candidates for medicines. They realised if they changed their system to maximise toxicity, it designed thousands of toxic molecule candidates including the VX nerve agent (https://www.nature.com/articles/s42256-022-00465-9)

* When GitHub released their Copilot AI code completion tool, it could autocomplete things like `API_KEY: `

* AI models used for decision-making in hospitals and courts frequently exhibit extreme racial prejudice (eg https://www.nature.com/articles/d41586-019-03228-6)

* and many many, other examples

There's a real sense of utopianism and not much consideration of misuse even on a very short timescale. Even if you don't think AI will become an existential risk in the future, there are enough problems caused by misuse of current systems that it warrants attention.

All of the money I've seen flowing in the AI safety EA space has been put to extremely good use - $30k to a YouTuber making videos on AI safety who introduced me to the field and is my go-to for explaining topics like alignment (what you want vs what you say you want), grants to extremely productive AI safety researchers, grants to fund educational bootcamps and scholarships, etc.

(Disclaimer: I'm an EA aligned person myself so I'm fairly biased!)

> from the last place they worked at

If you're a fresh graduate or self-taught like myself, getting your foot in the door is tough.

And TripleByte's model is also to let you skip stages of the interview process IIRC, which is appealing no matter how much experience you have.