• 0 Posts
  • 74 Comments
Joined 2 years ago
cake
Cake day: August 14th, 2023

help-circle
  • The average cost of a hospital stay in a U.S. hospital is about $3,000 per day, but it varies significantly by location. So long stays like yours might cost between $250,000 and $500,000.

    If your insurance covers it (and about 92% of Americans have health insurance), you’d be looking at your annual out of pocket max, which the law caps at $18,000 for family plans or $9,000 for individual plans, but which most people on employer sponsored plans (around 60% of Americans) have out of pocket maxes around $4,000 to $5,000. Source

    So for most Americans, your hospital stay would’ve probably cost the individual patient about $5,000. Insurance would’ve paid another $350,000.

    But for some Americans, they’d be looking at a $360,000 bill and then would just file bankruptcy, start over with close to a net worth of zero, at least for non-exempt assets (people generally get to keep their homes, cars, and retirement accounts in bankruptcy so it won’t actually be starting from zero if you’re well into a middle age in the middle class).

    Or worse, the hospital would realize they’re not getting paid, and then would find a reason to kick you out as soon as you’re stabilized. They have to keep you alive even when you can’t pay, but don’t have to treat you beyond that for free.


  • Let’s take home appliances. Imagine you are a person who knows how to diagnose and repair microwaves. You keep all the most common parts for the most common brands in your warehouse. You bring them with you based on the customer’s description of what is wrong, and you’re prepared to efficiently apply to correct repair as soon as you’re confident in your diagnosis.

    Your typical job looks like this:

    • Get a call, get all the billing information (15 minutes).
    • Drive out to the person’s home (30 minutes).
    • Talk to the customer (15 minutes).
    • Unscrew and disassemble the access panels to the appliance itself (15 minutes).
    • Diagnose and test things to make sure your initial hunch is correct (15 minutes).
    • Remove and replace the faulty part (15 minutes).
    • Put everything back where it belongs (15 minutes).
    • Drive back to your office (30 minutes).

    There, that’s 2.5 hours of your time to do a 15-minute task of installing a part. At the factory, a much less skilled person (who doesn’t need to know how to diagnose different models, or manage a business) could have installed 10 of those in the same amount of time. Maybe more, because they wouldn’t have had to remove an old one.

    Most manufacturing is like this. Assembly is easy. Repair is hard. So repair of heavy/bulky/stationary things is always going to be very expensive. It may be more economical to tow the thing to a central place to be repaired, so that the worker doesn’t have to waste too much time driving from place to place.

    Throw in the need to keep an inventory of dozens of parts for hundreds of models, and you’re also paying for the warehouse space and parts supply chain, and the interest on the money spent up front to stock up, maybe to be recovered later when a job actually needs that part.

    The economics strongly favor assembling new stuff rather than repairing old stuff for anything even remotely simple. It isn’t until you’re up to the $5,000 range that it becomes pretty normal to prefer an all-day repair job over paying for a replacement.

    For $500 devices, it’s gonna be pretty hard to economically repair things.



  • I’m only generally familiar with the big crime podcast/documentaries that spilled into the mainstream about 10 years ago: first season of Serial, Making a Murderer. And both of those were highly critical of the police work and called convictions into question (and actually got the public attention on the wrongful convictions).

    More recently, I’ve seen the HBO series on Karen Read, and it painted a picture of severe police misconduct that at worst tried to frame an innocent person, and at best botched the investigation to make a conviction of a guilty person difficult to impossible.

    So yeah, crime documentaries often do show police misconduct and incompetence. At least the ones that hit my radar.


  • The boring answer: criminal investigative files generally aren’t released, so they’re compiled in a way that mingles information about victims with information about suspects and witnesses and others potentially involved in criminal activity, intentionally or unwittingly, directly or tangentially.

    If you want to export a list of all names in the files, you’ll want to filter out victims for sure, and probably mere witnesses. You definitely don’t want to out informants and make them vulnerable to retaliation.

    So most law enforcement agencies simply will keep everything secret. The idea of releasing names from the file was unusual, and it’s not surprising that Trump’s own people refused to follow through, especially when it’s highly likely that Trump was in that list of names.


  • No, LCOE is an aggregated sum of all the cash flows, with the proper discount rates applied based on when that cash flow happens, complete with the cost of borrowing (that is, interest) and the changes in prices (that is, inflation). The rates charged to the ratepayers (approved by state PUCs) are going to go up over time, with inflation, but the effect of that on the overall economics will also be blunted by the time value of money and the interest paid on the up-front costs in the meantime.

    When you have to pay up front for the construction of a power plant, you have to pay interest on those borrowed funds for the entire life cycle, so that steadily increasing prices over time is part of the overall cost modeling.


  • My problem with nuclear is both the high cost and, somewhat counterintuitively, the very long life cycles to spread that high cost. The economics only make sense if the plant runs for 75 years, which represents an opportunity cost of displacing whatever might be available in 25 or 50 years.

    A solar plant planned in 2025 might be online in 2027, and decommissioned in 2047, replaced with whatever technology/economics are available then. But a new nuclear reactor bakes in the costs for 80+ years, to be paid by ratepayers who haven’t been born yet.

    So if in 2050 a 2030-constructed nuclear plant is still imposing costs of $66/MWh on ratepayers, to finance the interest and construction costs from 25 years earlier, will that be competitive with the state of solar/wind/batteries/hydrothermal at that time? Given the past trend lines, it seems economically foolish to lock in today’s prices for the next 80 years.






  • But the other misleading part is they looked at 20 years which is close to the life cycle for solar/batteries and not even half the life of nuclear

    I think Lazard’s LCOE methodology looks at the entire life cycle of the power plant, specific to that power plant. So they amortize solar startup/decommissioning costs across the 20 year life cycle of solar, but when calculating LCOE for nuclear, they spread the costs across the 80 year life cycle of a nuclear plant.

    Nuclear is just really, really expensive. Even if plants required no operating costs, the up front costs are so high that it represents a significant portion of the overall operating costs for any given year.

    The Vogtle debacle in Georgia cost $35 billion to add 2 MW 2GW (edit to fix error) of capacity. They’re now projecting that over the entire 75 year lifespan the cost of the electricity will come out to be about $0.17 to $0.18 per kilowatt hour.



  • just spitting the information back out, without paying the copyright source

    The court made its ruling under the factual assumption that it isn’t possible for a user to retrieve copyrighted text from that LLM, and explained that if a copyright holder does develop evidence that it is possible to get entire significant chunks of their copyrighted text out of that LLM, then they’d be able to sue then under those facts and that evidence.

    It relies heavily on the analogy to Google Books, which scans in entire copyrighted books to build the database, but where users of the service simply cannot retrieve more than a few snippets from any given book. That way, Google cannot be said to be redistributing entire books to its users without the publisher’s permission.


  • What does an LLM application (or training processes associated with an LLM application) have to do with the concept of learning?

    No, you’re framing the issue incorrectly.

    The law concerns itself with copying. When humans learn, they inevitably copy things. They may memorize portions of copyrighted material, and then retrieve those memories in doing something new with them, or just by recreating it.

    If the argument is that the mere act of copying for training an LLM is illegal copying, then what would we say about the use of copyrighted text for teaching children? They will memorize portions of what they read. They will later write some of them down. And if there is a person who memorizes an entire poem (or song) and then writes it down for someone else, that’s actually a copyright violation. But if they memorize that poem or song and reuse it in creating something new and different, but with links and connections to that previous copyrighted work, then that kind of copying and processing is generally allowed.

    The judge here is analyzing what exact types of copying are permitted under the law, and for that, the copyright holders’ argument would sweep too broadly and prohibit all sorts of methods that humans use to learn.


  • specifically about the training itself.

    It’s two issues being ruled on.

    Yes, as you mention, the act of training an LLM was ruled to be fair use, assuming that the digital training data was legally obtained.

    The other part of the ruling, which I think is really, really important for everyone, not just AI/LLM companies or developers, is that it is legal to buy printed books and digitize them into a central library with indexed metadata. Anthropic has to go to trial on the pirated books they just downloaded from the internet, but has fully won the portion of the case about the physical books they bought and digitized.


  • No. The court made its ruling with the explicit understanding that the software was configured not to recite more than a few snippets from any copyrighted work, and would never produce an entire copyrighted work (or even a significant portion of a copyrighted work) in its output.

    And the judge specifically reserved that question, saying if the authors could develop evidence that it was possible for a user to retrieve significant copyrighted material out of the LLM, they’d have a different case and would be able to sue under those facts.


  • The law says this is ok now, right?

    No.

    The judge accepted the fact that Anthropic prevents users from obtaining the underlying copyrighted text through interaction with its LLM, and that there are safeguards in the software that prevent a user from being able to get an entire copyrighted work out of that LLM. It discusses the Google Books arrangement, where the books are scanned in the entirety, but where a user searching in Google Books can’t actually retrieve more than a few snippets from any given book.

    Anthropic get to keep the copy of the entire book. It doesn’t get to transmit the contents of that book to someone else, even through the LLM service.

    The judge also explicitly stated that if the authors can put together evidence that it is possible for a user to retrieve their entire copyrighted work out of the LLM, they’d have a different case and could sue over it at that time.


  • Does buying the book give you license to digitise it?

    Does owning a digital copy of the book give you license to convert it into another format and copy it into a database?

    Yes. That’s what the court ruled here. If you legally obtain a printed copy of a book you are free to digitize it or archive it for yourself. And you’re allowed to keep that digital copy, analyze and index it and search it, in your personal library.

    Anthropic’s practice of buying physical books, removing the bindings, scanning the pages, and digitizing the content while destroying the physical book was found to be legal, so long as Anthropic didn’t distribute that library outside of its own company.


  • The court’s ruling explicitly depended on the fact that Anthropic does not allow users to retrieve significant chunks of copyrighted text. It used the entire copyrighted work to train the weights of the LLMs, but is configured not to actually copy those works out to the public user. The ruling says that if the copyright holders later develop evidence that it is possible to retrieve entire copyrighted works, or significant portions of a work, then they will have the right sue over those facts.

    But the facts before the court were that Anthropic’s LLMs have safeguards against distributing copies of identifiable copyrighted works to its users.