garrison

twitlonger skynet

speedrunning the most boring imaginable version of skynet

am I understanding correctly that it worked on some iphones/macs and not others? if so it must not be safari support alone, but something weirder, right?

careful, if you're too dismissive it might try to cancel you

a bunch of stuff was disabled in *all* browsers post-spectre but those features were eventually re-enabled with some restrictions (secure context etc)

I have not tried it in safari but caniuse/mdn claim it's supported so idk caniuse.com/sharedarrayb...

god I wish there was software I actively liked

okay in hindsight they *really* should have ripped the bandaid off last time

I think *also* including the handle input is important because it will slowly educate people about what's really going on

as long as the oauth buttons are prominent enough people will still mindlessly click them without getting confused

oh no not again

you're wrong, actually

it's not an issue imo because people are already used to being presented with a ridiculous wall of oauth options

as devs I'm sure we both find this revolting, but it's *normal* and that's what matters here

imo we should just make it look like oauth for now, everyone can understand "sign in with bluesky/blacksky/tangled/whatever"

include the handle input for the long tail of PDS nerds

over time "regular" users will learn what the handle input is for

based, but it shouldn't raise if the operations are associative like + or concat, e.g. (a + b + c)

other than that I wouldn't even notice bc I use parens everywhere lol

friendship ended with operator precedence, parentheses are my friends now

SAB works in safari does it not? Is this a wasm issue specifically?

code.storage Backed by git, built for scale.

code.storage is interesting

oh yeah that's exactly what I was thinking of, sounds cool!

oh no that bot got me too LOL

as in paid hosting or something that can be whitelabeled?

I was thinking of the latter, in the same vein as people outsourcing auth or cloud stuff

I'm not sure if that's actually a good idea long-term, but it would make it easier to build atproto apps ig

who's gonna be the first to invent pds-as-a-service?

if those signals really do go to zero then I agree we will have a serious problem

it is, however, entirely possible that they won't, and thus far for me that has actually been the trend

it's possible my perspective is warped here because I had already reached that conclusion before vibecoding was a thing

but I feel like as models have gotten better I still have no trouble spotting unserious work. Code aside, there are many other signals

tbh it's more that I think the OSS ecosystem was cooked either way; it could hardly shoulder left-pad let alone LLMs

someone should invent a website that displays code stored in version control and does not instantly rate limit me after one click

who up Writing Ahead they Log?

I'm not saying you're wrong about this, but I do think you should consider what it looks like if you *are* wrong about this and hedge accordingly

as for the companies, here's hoping open models can keep up because, uh, yikes

thus far that bet has paid off, and quality has improved enormously

I'm not certain that training/steering will be enough, but the trend does look that way. Sota models seem quite aligned modulo malicious inputs

many such cases

there are other things that will take much longer, especially related to "quality" as that is a goalpost which can be moved, and human programmers will continue to thrive in those areas for some time imo

but "only humans can write correct code" is not a hill I want to die on

my bet is that in the coming months/years you are going to be very surprised at how good models get at correctness testing and security

I expect superhuman performance soon, esp considering how inexperienced the avg dev is with both