danpalmer
I'm 20 hours in and starting to see examples of mechanics being generalised in a way I wasn't expecting, i.e. an early games thing turns out works in a way more general way that elevates the mechanic in complexity and feels like a very satisfying moment.
I have a suspicion about another that would completely change my perception of the game if it turns out to be true, so... meta puzzles? Not sure, but certainly there's something more than just the 4 initial rulesets colliding.
However, so far I can't point to anything substantial that should have been skipped. Almost every level is teaching me something, and those things keep coming up later.
Ultimately, Order of the Sinking Star applies the same formula as The Witness (although hopefully with a little more story, but probably not a lot more), and dials it up to 11. I loved The Witness, so really it's exactly what I asked for, and I think the feeling of the distance left to go is an artifact of a desire for instant gratification.
I think the game is deeply rewarding and fantastic, as long as you understand you're signing up for a slow burn. That's not for everyone, and that's ok.
My setup is optimised for being quiet and low power, and part of that means bulk HDD storage that is mostly spun down. Seeding a large long tail of data would probably mean a lot more disk activity.
The 1000s of applications folks are doing on the SWE subreddits are mostly juniors (or over titled seniors) with not much experience.
Maybe this guy is in that latter category, but it seems unlikely. I think this is just an issue of remote jobs being scarce.
> I know I can work remotely effectively because I did it for the last 6 years.
6 years means that Ryan started right around COVID. Looking back I don't think COVID was the remote renaissance that it appeared at the time, I think it was a temporary blip. That sucks, I'd much rather it was permanent, and some things have changed for the long run, but unfortunately I think this is likely a much bigger factor. Remote jobs have always been few and far between, and super competitive.
Claude is the author on the commits.
Reinforcement learning selects for models that appear to complete tasks, on the basis that appearing to complete a task is correlated with completing a task.
Both have problems. The latter is probably a better correlation, but by no means flawless. If we just take at face value that RL is training models to complete tasks, we're not engaging deeply enough with the process enough to see where it falls short.
Concrete example, "coding is free" (TBC, but let's assume), and now we're code review constrained. Even if code review becomes free we probably just become product discovery constrained, or something else. Making all software free will just cause us to hit on hardware constraints, and so on. It's still progress, and more progress will bring more benefits, but there's a limit to how much better the world as a whole can get from any one change like this.
I think AI/ML systems are unique compared to other software in this regard, because the "how it works" is not inherent in the software itself, and not working once doesn't mean that it doesn't work at all, it could have been luck.
There's functionally no alternative for consumer logins, and certainly not one that doesn't require learning what PKI is. "like SSH with PKI" is a bit "draw the rest of the owl".
But as I said, there are plenty of UX issues with passkeys, those UX issues however don't include getting your accounts hacked, phished, passwords lost in data breaches, etc, which I'd suggest are a lot worse.
Passkeys aren't solving for security nerds who like portability and have 30 character long passwords, they're solving for grandpa who will happily read their bank's 2FA SMS to the "bank" over the phone. The FIDO alliance has literally billions more of these users than they do users who know what hardware attestation is and how it impacts their auth workflows.
> Addendum: I feel I have made a compelling case for passkeys in this article. And I haven’t even pointed out one of their best security features: they are (in most use cases) domain-bound, meaning they are resistant to what are called Man-In-The-Middle attacks.
"I made a compelling case... but didn't point out the most important security feature that motivated passkeys in the first place"
Honestly I don't think this post did make a compelling case. There's a lot of assuming bad faith on the part of the FIDO alliance, while completely missing that preventing MITM attacks was a key design feature. The FIDO alliance members deal with stolen accounts all the time, and passkeys largely solve that. Stealing accounts is almost always done with phishing or social engineering, and passkeys solve the former and reduce the surface for the latter.
2020: If we solved currency it would transform the world completely
2021: If we solved delivering small items within 30 minutes it would transform the world completely
2022: If we solved being able to sell jpegs it would transform the world completely
2024: If we solved the programming part of software engineer's jobs it would transform the world completely
Each part of this is very easy to implement, even if writing the code by hand. A chat interface, plus memories, plus semantic search, a little database, some embeddings, cron triggers. It's all trivial to implement... unless you want it to work, and more than that, if you want to know for sure that it works.
Because the harness is key, the prompting is key, the exact wording of the skills matters, the way the memories work, the embeddings used, all of these have influence over how well the system actually performs at being a personal assistant.
I've plugged together the dozen libraries necessary to do this sort of thing, and then it sorta works and sometimes doesn't. I realised that the whole problem is actually one of evaluation. If you don't have evals that say that it recalls memories with a certain performance, if you don't have evals that say that it actually completes tasks some percentage of the time, then you don't know if it actually works.
I've learnt not to bother investing my time in anything like this unless there's evidence that it actually works beyond "on my machine" from the devs.
I don't think someone is an idiot for thinking that the information they type into their private Claude account is private. I also don't think people are idiots for thinking their phone is listening to them and giving them targeted advertising based on that. Both are reasonable deductions from their lived experiences. Both are wrong.