• 0 Posts
  • 294 Comments
Joined 2 years ago
cake
Cake day: February 7th, 2025

help-circle







  • The risk I see with this is that the strategy rests on the assumption of states acting in rational self-interest. While this is generally a reasonable assumption to make, it has become abundantly clear that it’s not always true.

    Hypothetically, if China came under control of a megalomaniac willing to compromise Chinese interests and safety (or unwilling to consider that risk) for the sake of nationalist ambitions to “reclaim” Taiwan, things could get ugly. Would the US actually retaliate? Would it fight open war against a nuclear power headed by someone willing to cut their whole country’s nose to spite their face?

    Conversely, if the US came under control of a foreign asset (or useful idiot) willing to strike a deal with China for cheap access to the chips they need for their American Intelligence, selling out Taiwan in the process and probably getting fucked over in the process, it might allow them to entrench so deeply that even subsequent presidents would have a hard time dislodging them without, again, risking nuclear war.

    I don’t think either scenario is likely, but I also didn’t think the US would let Pedonald run for another term, let alone elect him, so maybe my judgement just sucks.





  • Also generating subtitles for my mother’s foreign films I guess?

    You mean mapping learned patterns (speech) to other patterns (words)? Yes, that’s actually what it’s really good at.

    Of course, if those films invented completely new words, it would be out of its depth. Weird pronunciations may also produce weird text. It might get things wrong either way occasionally.

    But if it’s within the boundaries of its training material, it’ll generally produce reasonable output, and in that context, the occasional aberration doesn’t matter as much (particularly if the human brain can compensate).

    Just don’t ask it to generate anything where factuality matters.







  • Per their product page and personal sources (not secret, just doxxing risk):

    They’re great for construction sites, factory floors, hazard response, because they can move in places humans can’t (or shouldn’t) for inspections, documentation and such, while also more flexible in navigating obstacles than wheeled or tracked robots. They can monitor active operations with greater leeway because they’re less concerned with safety risks.

    If a chain snaps and a human was too close to heavy moving parts, that’s a family’s worth of pain, grief and worries over income, trauma for everyone present, and an expensive cleaning bill if you’re an executive so devoid of empathy things need a price tag to move you.

    If it catches a robodog, that’s less than $100k of damages and maybe some extra footage to help avoid future accidents (and cleaning bills).

    (Granted, Cerberus is a military version, but I assume they thought it’d be the military extension of that principle, detecting mines and such.)


  • I’m not sure AI in this form will ever find a trillion-dollar problem to solve that would justify the sheer amount of money being burned on it. It’s too limited in its capacities and too inefficient for what little it can do reliably.

    If a significant technological breakthrough comes along that permits actual reasoning and evaluation, coupled with some capacity to reliably extract such reasoning tasks from natural language input, its capabilities might expand enough to actually fill some of the roles it has been hyped up for.

    Simultaneously, it would need a massive improvement in efficiency (both in terms of cost and in energy and water demand) to make it financially feasible.

    I think the comparison would work better if, say, some new removable storage medium LargeDisk had been all the rage for transfering large amounts of data, and everyone was building LD factories and founding LD Express services to facilitate fast delivery of large amounts of data. You just call a data distributor, such as a video rental service, order a few movies and a few hours later a courier arrives. You hand them the LD from your last order to return, then take the new back to your home cinema and plug it in.

    Then along comes high-speed internet. Suddenly, transferring LDs is cumbersome and slow, when you can just stream those movies faster. No more ordering in bulk to make the delivery fee worth it, no more waiting for hours. LDs persist, but they become part of your home cinema instead, the place where such movies are stored.

    Sure, the demand for LDs will briefly collapse, then steadily rise as more people build their own home cinema, but it won’t be for the LD’s merit alone.