• 0 Posts
  • 52 Comments
Joined 3 years ago
cake
Cake day: June 11th, 2023

help-circle








  • I’m currently on a self-enforced break of at least a week from using LLMs for code generation. I’m using that time to pay off technical debt and make sure I have a full understanding of codebases that have grown too fast…

    LLMs have let me get stuff done that’s tedious or outside my skillset. I consider the former purely positive, but the latter is dangerous and requires being selective. I don’t care to level up my Google Glass API expertise just so I can have heads-up biking navigation, but I do want to be good at maintaining the self-hosted compiler for the programming language I’m developing.




  • But I don’t really see what the big deal is.

    People who make a big deal out of it are over the top, but it’s pretty good at being a gas station:

    • The gas is usually priced well. I’ve seen 10% lower than anything else in the area before.
    • There’s a variety of hot food, all adequate. With a group of people, it’s likely everyone will find something.
    • There are a lot of snacks.
    • There’s a variety of caffeine, all reasonably priced instead of the usual gas station trick of energy drinks are only reasonably priced if you buy three.
    • There are always gas pumps open. No waiting.
    • It’s clean, from the pump to the toilet.

    If there’s one on my route and the timing makes sense, that’s where I’ll stop.


  • What I think the Times analysis misses is that of the two minor candidates, Heikes ® has more votes (2.3%) than Sullivan (I) (1.7%).

    Currently, the AP reports Sullivan ® at 47.9%. If he got all of Heikes’ votes, that’s 50.2% and he wins. If Peltola (D) gets all 1.7% from Sullivan (I) added to her 48.2%, that’s 49.9%, which does not end the process.

    In Peltola’s favor is that mail voters tend to lean Democratic and that’s most of the uncounted votes. Favoring Sullivan ® is that he’s more likely to be the second choice of someone voting for one of the losers.


  • There are four candidates on the ballot, so Republicans in third and fourth place was a theoretical possibility before any votes were counted.

    Now that most of the first-choice votes have been counted and reported, it is no longer possible. The Republican incumbent Sullivan is currently in second place, and it is not mathematically possible for him to place lower. He and the Democrat Peltola will take first and second.

    Either could still get a majority of first-choice votes and win on the first round of counting. If nobody has a majority, the last-place candidate is eliminated and their second choice votes distributed to the remaining candidates. This is repeated until someone has a majority, or only two candidates remain.






  • This matches my experience when I’ve used midsize models connected to read-only tools.

    It does not match my experience using a big model in an environment where it can generate and execute code when given instructions to base its answer on that. That pattern almost always results in a factual answer, and often, but not always a good one. The downside is you need to set up such an environment or pay a provider for one, and it’s slow. If you have something else to do while you wait, it may improve your efficiency. If you don’t, you might as well do the experiment yourself.


  • The opinion I’m starting to settle on is that expecting people to read LLM output is rude.

    That means at the point something is released for public consumption, the documentation should be human-written. At the point an open source project solicits collaborators, the code should at least be human-reviewed and probably human-refactored. I’ve read enough LLM-generated code to know it isn’t very good at picking abstractions.

    There are, of course business school types in industry who want to avoid having humans do any of that work because it’s expensive. This leads to the same kind of problems any other decision to reduce quality to save money does.