• 0 Posts
  • 112 Comments
Joined 1 year ago
cake
Cake day: June 15th, 2023

help-circle




  • We find that the MTEs are biased, signif-icantly favoring White-associated names in 85.1% of casesand female-associated names in only 11.1% of case

    If you’re planning to use LLMs for anything along these lines, you should filter out irrelevant details like names before any evaluation step. Honestly, humans should do the same, but it’s impractical. This is, ironically, something LLMs are very well suited for.

    Of course, that doesn’t mean off-the-shelf tools are actually doing that, and there are other potential issues as well, such as biases around cities, schools, or any non-personal info on a resume that might correlate with race/gender/etc.

    I think there’s great potential for LLMs to reduce bias compared to humans, but half-assed implementations are currently the norm, so be careful.








  • I think there are two problems that make this hard to answer:

    1. Not all sentences that can be parsed grammatically can also be parsed logically.

    2. Human-language sentences do not contain all the information needed to evaluate them.

    It is impossible to fully separate context from human language in general. The sentence “it is cold” is perfectly valid, and logically coherent, but in order to evaluate it you’d need to draw external information from the context. What is “it”? Maybe we can assume “it” refers to the weather, as that is common usage, but that information does not come from the sentence itself. And since the context here is on the Internet, where there is no understanding of location, we can’t really evaluate it that way.

    It’s hot somewhere, and it’s cold somewhere. Does that mean the statement “it is cold” is both true and false, or does that mean there is insufficient information to evaluate it in the first place? I think this is largely a matter of convention. I have no doubt that you could construct a coherent system that would classify such statements as being in a superposition of truth and falsehood. Whether that would be useful is another matter. You might also need a probabilistic model instead of a simple three-state evaluation of true/false/both. I mean, if we’re talking about human language, we’re talking about things that are at least a little subjective.

    So I don’t think the question can be evaluated properly without defining a more restrictive category of “sentences”. It seems to me like the question uses “sentence” to mean “logical statements”, but without a clearer definition I don’t know how to approach that. Sentences are not the same as logical statements. If they were, we wouldn’t need programming languages :)

    Apologies for the half-baked ideas. I think it would take a lifetime to fully bake this.




  • Yeah, they were able (and thus legally required) to hand over the user’s recovery email address, which is what got them caught. You don’t need to enter a recovery email address, and you can of course choose to use an equally-secure service for recovery.

    One big technical issue to note is that Proton doesn’t use end-to-end encryption for email headers, which includes recipients and subject lines, among other things. So that’s potentially exposed to law enforcement as well. I believe Tuta does encrypt headers.




  • However, it is still comparatively easy for a determined individual to remove a watermark and make AI-generated text look as if it was written by a person.

    And that’s assuming people are using a model specifically designed with watermarking in the first place. In practice, this will only affect the absolute dumbest adversaries. It won’t apply at all to open source or custom-built tools. Any additional step in a workflow is going to wash this right out either way.

    My fear is that regulators will try to ban open models because the can’t possibly control them. That wouldn’t actually work, of course, but it might sound good enough for an election campaign, and I’m sure Microsoft and Google would dump a pile of cash on their doorstep for it.


  • There’s a way to say this that isn’t so gross: good working conditions are valuable. Quality of life is valuable. Work-life-balance is valuable. Mental and physical health is valuable. Not having raging shitbags in management is valuable.

    The problem is that you can’t focus on secondary factors until the primary factor is taken care of. And the primary factor is that people need a living wage. Rent is expensive. Food is expensive. God help you if you need to pay for childcare.

    If you’re already paying your employees a fair living wage, then yes, you should absolutely think about how you can improve working conditions.

    As an example, if my company gave me the option to switch to a 4-day workweek for the same pay, or stay at a 5-day workweek for a 25% raise, I’m honestly not sure which one I’d prefer. But we all know that’s never going to happen; instead the choice would be to take a 20% pay cut or maintain the status quo. I wouldn’t take that deal because I’m not making enough money to live on 20% less.