I think in general the model makers and to some extent their clients want a slightly higher pass rate over efficiency. This makes sense: for critical first week impressions clients notice pass rate much more, and only later start to grapple with cost. For example this is why High is the default reasoning for Fable, not Medium, and that choice of priorities propagates throughout the stack.
I really never thought it would come up in a HN thread but I’m actually working on a modified version of Anki as a personal project (not quite ready yet though, but will open source probably in a few weeks) where improving the editing/curating/creating experience is a big focus. I’m trying to make it markdown-based too.
Just to pick your brain real fast, when creating new cards (from scratch?) what would a better experience look like for you?
The tricky question at the forefront of education research is, at least in my mind, trying to thread the needle between effective techniques that students don’t like, ineffective techniques that students do like, and poorly defined techniques that administrators like. And on top of all that student self-reports aren’t actually very reliable indicators of learning progress at all!
I think the book Superlearning just gives the simple rule that you should spend 10% of your budgeted time on meta concerns (what to learn, what order, what strategies to use, etc) and that’s always seemed a sensible, good-enough default that takes away the doubt. Knowing how deep to go is, as you point out, only half the battle unfortunately, but the rule helped me at least.
Except the same companies have shown (see: closure of the DS online storefront) it's still an issue, just 20 years delayed instead of 10. Sure, it's better, but it's not good. I'll still take it of course, but we need to pressure lawmakers to set up something more fair and stable.
There are different grades of vendor lock-in. There's mechanical lock-in (which is a thing, like .claude folders) and economic lock-in but then we don't pay enough attention to behavioral lock-in. Habit is powerful, and if you can habituate users into a certain flow, change feels bad and they are more likely to stay.
It didn't stop all of Facebook's behavior, far from it, but we did get to see Zuckerberg hauled in front of Senate committees multiple times (who we do vote for).
This has never happened in China, and will never happen, nor anything like it. Some open oversight is almost always better than possible secret oversight (and do you think that the Chinese government has user privacy on even its top 10 priorities?)
Has anyone created a tool for easily doing this? Imagine if all the Virginia Guiffres of the world created crypto-provable diaries they could later reveal.
Bombs are a bit of a non sequitur here. The point is that Chinese companies are demonstrably hostile to American ones historically (and threatening in some specific structural ways to the American consumer). The presentation may be similar but to attribute American ethics to a Chinese decision is dubious.
Yes you can! Well, mostly, depends on how pedantic you are with definitions: you can transplant layers but not weights, which in common parlance are conceptually similar. But usually it isn’t a good idea for a few reasons.
There’s a really fascinating example[1] where a guy identifies a particular set of layers and transplants them. Overgeneralizing, early layers are encoders and the later layers are decoders and in the middle some blocks seem to do specific things or tasks related “reasoning”. So you can actually create a FrankenLLM and it sometimes works.
This needs architectures to be roughly similar however and internal representations to be consistent-ish so for “stealing” it’s not really a thing (other practical concerns aside)
I’m not normally one for doom and gloom, everything is getting worse kind of thinking. However, one observation from the past few years in particular is exactly that thing, a specific kind of faulty media literacy.
Someone will read Lord of the Flies, for example, and come to the conclusion that it says something real about human nature. No dude, it’s just a story. Someone made it up. It’s not real. It can be interesting and insightful, even a useful mental framework or thought experiment, but it’s not proof or evidence at all! But I increasingly see people treat fiction as if it’s real things that actually happened, and that worries me.
It might be irrational but ticks terrify me (grew up in the Northwest where they aren’t so much a thing) and make me want to avoid moving to the East entirely
If they have a year sub, then yes I agree (even if it’s implicitly always part of the risk of buying so far in advance) but if they are month to month this position is absolutely nonsensical.
But I’m hopefully optimistic that we’ll see a renewed emphasis on speed and responsiveness, since users do notice that despite most products ignoring it.