Progressive overload

Economist, consumer advocate

[Progressive overload @pdbarnsley]

@CaponesBones @icpolicy @mattyglesias This has happened once to Gemini, because it was *too progressive* and a number of times to grok, because it was too progressive. Tay, not being a gpt llm isn’t really relevant to this discussion

[Progressive overload @pdbarnsley]

@icpolicy @decentricity @Gmeebus @mattyglesias Do you know that they have AI in other countries, with different overton windows, and that Elon Musk has an AI that I mentioned in the tweet you responded to? Do the textbooks mention those things? I’d be surprised.

[Progressive overload @pdbarnsley]

@decentricity @icpolicy @Gmeebus @mattyglesias But does that tuning of outputs explain the alignment between liberal views and LLM answers across a sweeping range of factual questions? That's completely implausible, even if the training universally leans liberal

[Progressive overload @pdbarnsley]

@decentricity @icpolicy @Gmeebus @mattyglesias His claim wasn't that untuned base models are less progressive, which sounds completely plausible to me, it was that I was voicing specific opinions incompatible with that, but the specific opinions in question were a secret.

[Progressive overload @pdbarnsley]

@jacktymac @mattyglesias I didn't hear the bit about "killing Arabs". Maybe it's near the end? She did say a bunch of other, different stuff though...

[Progressive overload @pdbarnsley]

@shadihamid @mattyglesias Which kinds of seats is the DSA portion of the winning coalition going to be competitive in?

[Progressive overload @pdbarnsley]

@icpolicy @Gmeebus @mattyglesias Again, you are trying to dress up your buddy's "racist beliefs are justified raw data is racist" as some sort of sophisticated analysis of racial differences, when he's right there saying "I am super racist and think AI should be too!". Not a great look.

[Progressive overload @pdbarnsley]

@icpolicy @Gmeebus @mattyglesias I think asking you to identify an example of a thing you've asserted, so I can look into it, is just table stakes for a sensible discussion

[Progressive overload @pdbarnsley]

@icpolicy @Gmeebus @mattyglesias Again, I think you're super confused about what Gmeebus is claiming and therefore what I'm responding to. 'AI naturally becomes racist because raw data is racist' is an assertion that racist beliefs are grounded in fact. No one is arguing about the *existence* of racist beliefs.

[Progressive overload @pdbarnsley]

@icpolicy @Gmeebus @mattyglesias I'd like you to precisely specify an opinion I have expressed, but which I would no longer hold if I tried a base model LLM - at which point I will be eager to improve myself.

[Progressive overload @pdbarnsley]

@icpolicy @Gmeebus @mattyglesias Now, if you want to argue that trainers embed some value judgements in AIs' output, sure. 'can't search for bioweapon recipes' is a value judgement, and there are definitely going to be other, more controversial ones. But that doesn't explain the skew in factual responses above.

[Progressive overload @pdbarnsley]

@jacktymac @mattyglesias I think claiming someone is a "single issue candidate, but they have not publicly mentioned that issue" is kind of a tell.

[Progressive overload @pdbarnsley]

@Gmeebus @icpolicy @mattyglesias But also

[Progressive overload @pdbarnsley]

@Gmeebus @icpolicy @mattyglesias Look! just like you wanted!

[Progressive overload @pdbarnsley]

@helmsson @icpolicy @mattyglesias Have *you* seen what happens when training isn't done? We *have* seen what happens when you want your model to be super based, reality be damned, and then you accidentally turn it into mecha hitler. Making models lie about millions of unrelated facts is super hard...

[Progressive overload @pdbarnsley]

@SamBHodgson @icpolicy @mattyglesias Indeed. And there is also no big shiny button a trainer can press that says "facts, but make them woke". So the vast majority of the *factual* claims you are seeing LLMs make are untrained inference.

[Progressive overload @pdbarnsley]

@Gmeebus @icpolicy @mattyglesias If your mental model is "modern LLMs are just like Tay, but woke" you are not really engaging with the advances that have occurred since "read these posts and then mimic them" was the state of the art.