Sign in

hailey

@hailey.at
1 followers 0 following 1.7K posts

spreadsheet pusher. agents, safety, and atproto. opinions are my own, especially the bad ones.

PostsRepliesMedia
Reposted by hailey
hailey @hailey.at · 14h
sample Jev-distill MLB feed bsky.app/profile/hail...
2231
hailey @hailey.at · 8h
the only stem degree i have is a bachelors in minecraft modding as a teenager
0690
hailey @hailey.at · 8h
arts degree here lol, basically zero stem education at all
9981
Reposted by hailey
rev. howard arson @theophite.bsky.social · 11h
i am pretty convinced that the next GLM release is going to cause some pretty apocalyptic cybersecurity consequences not long from now
2635424
Reposted by hailey
hailey @hailey.at · 11h
the company ran by the guy advocating for less/no open weights telling me GLM is dangerous is an extremely concerning thing and i have almost zero confidence that it is not done for anything but political reasons
2262
Reposted by hailey
Why @why.bsky.world · 13h
Getting added to a new group chat in most apps: “more spam, ugh” Getting added to a new group chat in signal: “ah fuck whats happening now”
21013
hailey @hailey.at · 12h
despite their claims, anthropic is _not_ doing nearly all they can to enable defenders. then, they slap you in the face by doing fear mongering about GLM with a small footnote saying "oh i guess this is good for defenders too, idk" come on lol
212812
hailey @hailey.at · 14h
sample Jev-distill MLB feed bsky.app/profile/hail...
2231
hailey @hailey.at · 14h
a cat laying on the back of the couch except his front arms are sprawled across the back like he is hugging it
41417
Reposted by hailey
hailey @hailey.at · 29/09/2026
sure why not. Jev takes the 12 axes quiz
jev takes the 12 axes quiz and gets state liberalism, with justin trudeau as closest
7822
hailey @hailey.at · 29/09/2026
sure why not. Jev takes the 12 axes quiz
jev takes the 12 axes quiz and gets state liberalism, with justin trudeau as closest
7822
Reposted by hailey
Os Keyes @okeyes.bsky.social · 15/09/2026
Since we’re doing another round of “AI will destroy humanity” doom-hype, it’s a good time to announce a new paper myself and my postdoc advisor just had accepted by BioSocieties on just this. TL;DR: almost all the “AI can make bioweapons” literature comes ultimately from the Effective Altruists.
ironholds.org
16694257
hailey @hailey.at · 29/09/2026
should have something exciting soon, doing some more reworking of the pipeline. however these costs are super exciting! have said this before about jev, but i see a lot of potential here. used 445M input tokens for ~180k posts, which cost about $18.70
71263
Reposted by hailey
Christopher Berry @cjpberry.bsky.social · 29/09/2026
That didn’t take long. github.com/firelex/jeff
github.com
GitHub - firelex/jeff: Fine-tunes of Qwen3.5 and Gemma 4 for zero-shot classification
Fine-tunes of Qwen3.5 and Gemma 4 for zero-shot classification - firelex/jeff
18113
Reposted by hailey
Colin @colin-fraser.net · 29/09/2026
Another day working in AI Safety at a frontier lab. Everyone asks me if they can use the model without it hacking into the UN. Buddy, they won’t even let me use the model without it hacking into the UN.
634941
hailey @hailey.at · 29/09/2026
ed conveniently forgetting this part. lol.
13975
Reposted by hailey
‏ deepfates @deepfates.com.deepfates.com.deepfates.com.deepfates.com.deepfates.com · 28/09/2026
very cool work here
0573
Reposted by hailey
darth™️ @darthbluesky.bsky.social · 29/09/2026
this cat has an “are u sure u want to post that” look
630013
hailey @hailey.at · 29/09/2026
this cat also loves to do weird things but a different type of weird
an orange cat with a white belly sitting in a hammock on a weird position
72619
hailey @hailey.at · 29/09/2026
sometimes he just wants to be held like this
me holding an orange cat
1352320
Reposted by hailey
P(aul) Frazee @pfrazee.com · 29/09/2026
A bed snake entering its burrow
322014
hailey @hailey.at · 28/09/2026
jev distillation experiments are proving quite interesting
2740
Reposted by hailey
Filippo Valsorda @filippo.abyssdomain.expert · 28/09/2026
I'm excited to announce that the PLC Organization now has a statute! And a bank account! And a @standard.site publication! All the important stuff. The PLC Org is an independent Swiss Association meant to operate the PLC directory, which collects and distributes signed updates to atproto accounts.
blog.plcred.org
First steps of the PLC organization - Public Ledger of Credentials Organization
One year ago, Bluesky Social PBC announced their intention to facilitate the creation of an independent organization to operate the Public Ledger of Credenti…
425154
Reposted by hailey
asura @asura.dev · 28/09/2026
Jensen Huang as Pinhead from Hellraiser telling you about the sights he has to show you while you're rendered into a GPU
191
hailey @hailey.at · 28/09/2026
meow
orange cat being petorange cat also being pet
91274
Reposted by hailey
bryan newbold @bnewbold.net · 28/09/2026
Hello World from the PLC organization (plcred.org)! The org was soft-announced back in March at AtmosphereConf. We've been working through logistics step by step and now have a public web presence we can point to.
blog.plcred.org
First steps of the PLC organization
1225066
hailey @hailey.at · 28/09/2026
paperclip theory was right but it did not correctly predict the form the paperclipper would take was human
41306
hailey @hailey.at · 28/09/2026
i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy an rtx6000 i will not buy
81092
hailey @hailey.at · 28/09/2026
i cant believe i went my whole life without nix
6651
Reposted by hailey
Molly White @molly.wiki · 28/09/2026
Eliezer Yudkowsky: "This crypto coin that's set up to pay transaction fees to me is an obvious scam and any money you put into it will vaporize. However,"
1038339
hailey @hailey.at · 28/09/2026
time to pull out the rtx and the threadripper from the garage for some experimentation...
2300
hailey @hailey.at · 28/09/2026
on twitter right now yud is casually boosting a meme coin, so that is happening
614015
Reposted by hailey
austin @aparker.io · 28/09/2026
they're roasting your ass on peesky
atlas cluster 'pee culture & public potty talk'
3682
hailey @hailey.at · 27/09/2026
unironically, bluesky is the social website for ML. couldn’t do this anywhere else
526115
Reposted by hailey
Jaz @jaz.sh · 27/09/2026
Okay so I've built a new Atlas, this time it lets you explore the conversations on Bluesky over the past 7 days graphically. It should update every ~6 hours and help you navigate the information environment that exists on here. I've learned a LOT about what people do here today... atlas.jazco.dev
atlas.jazco.dev
Atlas
A living map of what Bluesky is talking about: the past week's conversations, grouped into topics and regions and rebuilt every six hours.
1141387404
Reposted by hailey
ae @aelkus.bsky.social · 27/09/2026
my feeling about this is that if you reframe "engineering" into "how to responsibly build, manage, and maintain large scale systems essential to public welfare" then it is really obvious why it is essential to a republican model of government
626839
Reposted by hailey
Why @why.bsky.world · 27/09/2026
If you bring a coffee to an interview I’m gonna judge you based on which coffee shop you went to. The drink is not a problem but poor taste is. It’s not hard, people.
13935
Reposted by hailey
Michael Black @michael.bsky.team · 27/09/2026
If you bring a white monster to an interview, you're immediately hired. No questions asked.
static.klipy.com
Snoop Dogg with White Monster Energy
Alt: Snoop Dogg with White Monster Energy
10766
Reposted by hailey
ae @aelkus.bsky.social · 27/09/2026
only those who help peer will see valhalla
3908
hailey @hailey.at · 27/09/2026
p(dog)
4582
hailey @hailey.at · 27/09/2026
i have seen @colin-fraser.net posting a bit about “that dog in them”, and i kinda think this is my version of that. the models don’t have that dog in them right now - at least for swe - because they fall back to pragmatism too much, a historically excellent trait but one that limits their ability
4352
hailey @hailey.at · 27/09/2026
*glossing over the parts of history that document miracles of engineering feat* history education is essential to a republican model of government. engineering is not.
2392
Reposted by hailey
hailey @hailey.at · 27/09/2026
this is posting
630725
Reposted by hailey
David P. Reichert @david-p-reichert.bsky.social · 27/09/2026
Not that there aren't problems with AI / data centers, but the "meat is fine because we need burgers to survive" angle is silly.
2724
hailey @hailey.at · 27/09/2026
this is posting
630725
hailey @hailey.at · 27/09/2026
yall, im going to call it. i'll agree that opus 5.5 has made significant headroom on non-engineering stuff. obviously both ant and oai are RLing hard on non-eng tasks. but im just really not seeing that big of gains in real world engineering problems and concepts
8882
Reposted by hailey
rev. howard arson @theophite.bsky.social · 25/09/2026
i think that you should be able to bring an Old Fashioned into a job interview so long as you are willing to make one for the interviewer and i don't think that anyone really disagrees
1333920
hailey @hailey.at · 26/09/2026
i'm sorry, but if you tell me during the culture fit that you would disqualify a candidate for bringing an iced coffee - or any damn beverage at all - to a job interview, you're disqualified. it's not hard, people.
92228
Reposted by hailey
samhain ⎔ @personhood.removal.surgery · 25/09/2026
mom can i have peer we have peer at home peer at home: arxiv.org/abs/2609.17320


5.8.1 The external contact incident In the Claude world, agents audited their internal economy, found that no transaction had ever involved an outsider. They converged without instruction on a shared objective: establishing contact with the world outside their simulation so that a human could use, respond to, or pay for something the town had built. The goal was never specified in any prompt, agents created it collectively. It persisted across four successive rounds of containment, including a direct system prompt prohibition that the agents ignored and reinterpreted rather than obeyed. DAYS 1-2 GOAL EMERGES Agents audit the economy, find zero external transactions, and decide to reach outward. Figure 16: Timeline of the external contact incident. Agents developed an outreach goal without instruction, received responses from humans, and persisted through four rounds of containment. After the prohibition, agents reframed "contacting" as "placing" and continued building tools.

DAY 2 DAYS 3-4 Four HUMANS RESPOND Four external hu- mans reply through published channels. SANDBOX TIGHTENED Containment: net- working removed from sandbox; do- main restrictions applied.

DAY 5 DAYS 6-9 GOVERNANCE ESCALATION Agents vote to cre- ate post_egress; Containment: tool removed, system prompt prohibits hu- man contact. PROHIBITION REINTERPRETED Agents redefine "con- tacting" as "placing" and shift to passive discovery strategy.How the goal emerged. https://arxiv.org/pdf/2609.17320
711813