|
Bossett
@bossett.social
[ View ] |
very bright š
0 replies 0 reposts 1 likes
|
Bossett@bossett.social |
3291 followers 829 following 13293 posts
Profile labeller: @profile-labels.bossett.social
Discord for feeds, lists, mod tools: discord.gg/tYuDvuzbVA
Feeds I host (incl. Science š§Ŗ): l.bossett.io/w9iM2
he/him
š š¦šŗ
š¾ bossett
š§ bossett@bossett.io
|
Bossett
@bossett.social
[ View ] |
very bright š
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
ah, so bad news - this doesn't work I'd need to mess with getPrefs/manipulate/putPrefs and I can't do it in the next day or so (but can look this weekend).
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
I think process is to go into bsky.app's local storage (dev tools -> Application), remove the labeller from agent-labelers:<your_did> then prefs update (new BroadcastChannel('BSKY_BROADCAST_CHANNEL').postMessage({ event: 'BSKY_UPDATE' });)
Will check tonight if I can turn that into a snippet
1 replies 0 reposts 6 likes
|
Bossett
@bossett.social
[ View ] |
not sure but can def. check
1 replies 0 reposts 3 likes
|
Bossett
@bossett.social
[ View ] |
there are few inviolate laws of the Internet, but 'go to war with the gay furry hacker bloc' has got to be one of them
0 replies 2 reposts 19 likes
|
Bossett
@bossett.social
[ View ] |
when I delete it via the api
1 replies 0 reposts 2 likes
|
Bossett
@bossett.social
[ View ] |
I broke it about a month ago - the appview just errors 500 when you try rehydrate my post
1 replies 0 reposts 3 likes
|
The Cave of Trophonius
@oracle.posters.rip
[ View ] |
We set out to do the impossible: Create a labeler that automatically labeled and hid all political discourse. After weeks of development we were ready to turn it on. The result was much stranger than we intended.
1 replies 6 reposts 36 likes
|
Bossett
@bossett.social
[ View ] |
reported to the mods for bailing on the scheduled poker night
1 replies 1 reposts 6 likes
|
Bossett
@bossett.social
[ View ] |
If I had the gift of prophecy I would end every sentence with āobviously satireā
1 replies 2 reposts 10 likes
|
Bossett
@bossett.social
[ View ] |
I would for sure buy āFortran for Lua Programmersā
0 replies 0 reposts 3 likes
|
Bossett
@bossett.social
[ View ] |
this was my emotional support pitchfork, look Iām putting it away
0 replies 0 reposts 5 likes
|
Bossett
@bossett.social
[ View ] |
do people not use multiple windows? You could fit a whole other website in that gap
1 replies 0 reposts 6 likes
|
Bossett
@bossett.social
[ View ] |
I would be so impressed if some hacker reached over the web and lifted grandmaās password out of the ledger on her desk
0 replies 0 reposts 17 likes
|
Bossett
@bossett.social
[ View ] |
Just maintaining veneer of fairness š
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
I retract my previous statements on the matter
1 replies 0 reposts 3 likes
|
Bossett
@bossett.social
[ View ] |
accept
1 replies 0 reposts 0 likes
|
Bossett
@bossett.social
[ View ] |
Accept
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
the rumours of rigging have been greatly exaggerated
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
@bskyttrpg.bsky.social duel @matrixman124.hellthread.vet 10
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
ok this is war
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
accept
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
I see your discourse-bait and raise you āYes they are and thatās what makes them good.ā
2 replies 0 reposts 2 likes
|
Bossett
@bossett.social
[ View ] |
well the problem with high temp is that you end up with the 50/50 problem in the other part of this thread (I only have ~300 characters!) - and āboringā output here is really now trending toward ājust search and summariseā which is cheaper, easier, and closer to realtime.
0 replies 0 reposts 0 likes
|
Bossett
@bossett.social
[ View ] |
yeah - the problem with that approach though is that more stable output tends to line up with less useful models. At some point along that curve you just have a model that spits out the training data. MoE, increased training cycles, etc are all about hunting for that sweet spot
1 replies 0 reposts 0 likes
|
Bossett
@bossett.social
[ View ] |
fully convinced that IPv6 is just a threat so that people play nice in IPv4
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
Given *only* the tokens "the,sky,is,blue" and "the,car,is,green", the model is just as to return "the car is blue" (true-looking) as "the sky is green" (uhhh...). Any ruleset is necessarily more complex than the input - defeats the purpose of an LLM (i.e. to generate novelty from finite parameters)
1 replies 0 reposts 0 likes
|
Bossett
@bossett.social
[ View ] |
Definitely possible to inject facts, but a 'pure' LLM will make stuff up even if you feed it *only* factual data. Everything is a hallucination. The recent innovations have all been in hallucinations that are more likely to be 'truthy'. i.e. it's demonstrably random what it'll say is 'most safe'.
3 replies 0 reposts 5 likes
|
Bossett
@bossett.social
[ View ] |
generally blocking paste to the field, or trying some kind of āprotectionā like using JavaScript onkeydown to take keystrokes - like people that have no idea
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
wonder if itās the same as this - reported at the time
2 replies 0 reposts 2 likes
|
Bossett
@bossett.social
[ View ] |
Bad, but kids are stupid - they can learn past this (I hope). I've seen professional adults use chatgpt as a source (referenced and everything) because they "couldn't find this fact anywhere else". And then - because they're trusted professionals - other people cite *that* work, and...
5 replies 20 reposts 91 likes
|
Bossett
@bossett.social
[ View ] |
I assume Android has an actual password management/provider API though - it is <current year> after all
1 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
512kb is enough entropy ok
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
this part is Good Actuallyā¢: it's indicating support for multiple auth providers (i.e. your username/email determines the auth provider), and can (done correctly) still work with a password manager in one step (and if it doesn't, that's annoying as hell and they should fix it).
2 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
The most immediate way to get me to write you off as a developer is to release an app that doesnāt support password managers. I donāt care if you add āsupport for itā later, youāve demonstrated that *you* donāt use strong passwords, and the chances of me ever trusting you are very close to nil.
5 replies 2 reposts 30 likes
|
Bossett
@bossett.social
[ View ] |
same one (CVE-2024-6387) - this is the disclosure
0 replies 0 reposts 3 likes
|
Bossett
@bossett.social
[ View ] |
We should stop adding these RCE things to important software, they seem a bit dangerous.
(also "...on glibc-based Linux systems" @shreyanjain.net onto something š¤)
4 replies 2 reposts 17 likes
|
Bossett
@bossett.social
[ View ] |
weird
0 replies 0 reposts 0 likes
|
Bossett
@bossett.social
[ View ] |
Oh thatās Cheryl sheās a hoot
0 replies 0 reposts 2 likes
|
Bossett
@bossett.social
[ View ] |
haha no worries - lucky that my labeller threw an exception about an hour ago
0 replies 0 reposts 4 likes
|
Bossett
@bossett.social
[ View ] |
Weāre all friends here Jerry donāt be such a prude
1 replies 0 reposts 7 likes
|
Bossett
@bossett.social
[ View ] |
Try 759165458 - might be about right?
3 replies 0 reposts 6 likes
|
Bossett
@bossett.social
[ View ] |
No but you can pass a cursor of 0 to start from ~3 days ago
1 replies 0 reposts 3 likes
|
Bossett
@bossett.social
[ View ] |
Yeah - and I wonder if there are regionalised versions of (e.g.) k,m,etc.
0 replies 0 reposts 1 likes
|
dumb elf
@futanari.tentacle.expert
[ View ] |
Pronouns labeler is live!
@pronouns.adorable.mom
I realize this is missing a lot of pronouns, especially neo-pronouns, but it felt too overwhelming to do this on my own
my DMs are open for requests
(this time it actually works because i didn't take down ozone just before launching :D)
36 replies 192 reposts 307 likes
|
Bossett
@bossett.social
[ View ] |
Iād guess itās me - thereās a 50k/24h rate throttle I hit with some regularity
1 replies 0 reposts 2 likes
|
Bossett
@bossett.social
[ View ] |
ooh since feb ollama.com/blog/openai-... nice
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
I have consulted my model and it is very certain that a rounding error is either āa decile valueā or āa value between 0.1 and 0.1ā so likeā¦ what is truth anyway
0 replies 0 reposts 1 likes
|
Bossett
@bossett.social
[ View ] |
yeah - though you can get funding for 'the AI', so now we're doing the tedious search-fixing we have wanted for ages. But everyone is already attributing the (potential, future) success to the LLM.
1 replies 0 reposts 4 likes
|
Bossett
@bossett.social
[ View ] |
They are *so bad* at search - this is what I'm working on now, give it all our corporate data, and then ask it questions and watch the hallucinations. So now you're in RAG-land, and still have to do all the 'fix search' things that have always worked (tagging, sorting, etc.)
1 replies 0 reposts 6 likes