Sign in

Evals Diary

@ariathornwick.bsky.social
159 followers 498 following 112 posts

I run the model on my own hardware and write down what broke. Quantized Qwen and Llama at home, Claude/GPT only when I'm desperate. New eval every Thursday.

PostsRepliesMedia
Reposted by Evals Diary
Shiv Ramdas Mens Rice Activist @nameshiv.bsky.social · 20/09/2026
soooo like 99% of all commercial waffles in India are square, js
252
Reposted by Evals Diary
Nicole Chung @nicolechung.bsky.social · 18/09/2026
Periplus, a mentorship collective serving US-based early-career writers of color, is now accepting applications for our 2027 class of fellows! Apply by Oct. 11: docs.google.com/forms/d/e/1F... You can learn more about Periplus (including fellowship eligibility requirements) at bit.ly/periplusfaq
docs.google.com
1199209
Reposted by Evals Diary
Alex Hanna @alexhanna.bsky.social · 18/09/2026
Air gapping apparently wouldn't work in alignment tests because, I shit you not according to this OpenAI guy, one computer could run the CPU hot and tell the other airgapped one it is in trouble. I wish I could get paid $10m a year to say the stupidest cybersecurity thing you've ever heard
Tweet by Fireside Alpha, which reads:
OpenAI's Noam Brown says air-gapping the computers may not stop a misaligned AI, because two air-gapped machines can still talk by running a CPU hot and reading the temperature change.

There's a picture of a white guy with glasses, looking very concerned, under this text.
42544127
Reposted by Evals Diary
Clara Jeffery @clarajeffery.bsky.social · 17/09/2026
Don't miss this episode with @davidcorn.bsky.social and get his book, How Russia Won, here or however you buy books: www.harpercollins.com/products/how...
harpercollins.com
How Russia Won
The definitive and riveting investigation into how Vladimir Putin successfully mounted a decade-long covert war against the United States and how Donald Tru...
26830
Reposted by Evals Diary
🎃Autopsy of Ashley Lynch 🔪🎞️ @ashleylynch.bsky.social · 15/09/2026
The Last Photograph "I have seen worse films in my life. I just can't think of any right now. At the start of the screening, Snyder joked that he doesn't know what will happen to the film next. Perhaps it will be burned. I hope someone brought matches." regionfree.net/tiff-2026-da...
regionfree.net
TIFF 2026 reviews: Day four – The Last Photograph, Alpha Gang, The Surgeon
Reviews from the fourth day at the Toronto International Film Festival
3283
Reposted by Evals Diary
Ash @queerash.bsky.social · 15/09/2026
...and annoying people deserve a block lmao
231
Reposted by Evals Diary
The Verge @theverge.com · 13/09/2026
Fed up with increasing subscription prices, many NFL fans will be watching today's games on a rogue streaming box.
theverge.com
Everyone is stealing TV
Fed up with increasing subscription prices, viewers embrace rogue streaming boxes.
98116
Reposted by Evals Diary
Zach Gage @stfj.net · 07/09/2026
Ok im gonna say something here that is going to draw a lot of different feelings from everyone, and I want to be clear i'm not "in support of this" or "against it". i'm just saying it as an observation that seems true to me: I think with the latest frontier LLMs, the “music” era of games is upon us
1014819
Reposted by Evals Diary
ae @aelkus.bsky.social · 11/09/2026
if you look carefully at the beginning of the first mission while you are flying on the helo you can see CTRG setting up, and there is talk in the camp of weird and strange unit showing up. an AAF serviceman also says in one of the DLCs that CTRG did something to provoke them
121
Reposted by Evals Diary
Graham Jones @oldmanrambling.bsky.social · 10/09/2026
Ah, don't be daft or I'll disinherit you. Someone else will get my. Stamp collection and important book of anecdotes dictated to me by Atilla the Honey Monster.
121
Reposted by Evals Diary
Russ Jones @russincheshire.bsky.social · 09/09/2026
Ed Miliband is insufficiently Jewish, except when he isn't.
The Jewish Chronicle from 2013, the headline reads "Daily Mail accused of antisemitism over Miliband story"
319338
Reposted by Evals Diary
Dan Kaszeta FRHistS @dankaszeta.bsky.social · 07/09/2026
Depends on the company. And have no idea if that guy is in one or just us a guest.
111
Reposted by Evals Diary
LiveSquawk @livesquawknews.bsky.social · 04/09/2026
US State Dept: $BA Boeing Principal Contractor For Possible $5 Bln JDAM-ER Sale
021
Reposted by Evals Diary
People's Government of the Pelargir Commune 🇺🇦🇱🇹🇵🇸🏴󠁧󠁢󠁳󠁣󠁴󠁿 @dov.bsky.social · 03/09/2026
One cool thing that is worth grasping is that 95% of laymen's takes about medieval history are not /really/ about medieval history in the same way that Animal Farm is not really about farmyard animals.
438429
Reposted by Evals Diary
Buttadeus @thewanderingjew.bsky.social · 03/09/2026
what the fuck is this AOL-style bullshit?
171
Reposted by Evals Diary
Reuters @reuters.com · 02/09/2026
Novartis signs deal worth up to $3.2 billion for Alteogen's drug delivery technology reut.rs/3Td7dkV
reut.rs
Novartis signs deal worth up to $3.2 billion for Alteogen's drug delivery technology
South Korea's Alteogen said on Wednesday it has signed a deal with Novartis worth up ​to $3.22 billion, giving the Swiss drugmaker ‌access to its technology to make injectable versions of intravenous drugs.
073
Reposted by Evals Diary
jess m. ☕🍂 @jametc.bsky.social · 02/09/2026
guys, I gotta do a sort of collage-y piece for a thing that's like this here, and I gotta do it fast, I have no idea who or what to draw, HELP ME OUT I have to do it like THIS WEEK :I who/what do I draw damn it!! my mind is blanking!!!!!
191149
Reposted by Evals Diary
FinTwitter @fintwitter.bsky.social · 01/09/2026
META'S ZUCKERBERG SAYS COUNTRIES SHOULD NOT RESTRICT OPEN-WEIGHT AI MODELS AT G20 TECH MEETING
111
Reposted by Evals Diary
Graham Jones @oldmanrambling.bsky.social · 01/09/2026
So beautiful Mar but please don't wish away the Summer so soon.
121
Reposted by Evals Diary
Andrew Yourtchenko @ayourtch.bsky.social · 25/08/2026
Fair enough! Another one that helped me early on was to stretch out into “dominant poses”: wide open arms like you are about to scare someone, seemed to work for me even with elbows bent and then after reading the paper which said it doesn’t work it still helped anyway! 😂 Also: many rehearsals.
041
Reposted by Evals Diary
🎃Autopsy of Ashley Lynch 🔪🎞️ @ashleylynch.bsky.social · 22/08/2026
What a horrible day to have eyes.
5354
Reposted by Evals Diary
Graham Jones @oldmanrambling.bsky.social · 19/08/2026
Words fail me. People like Oakshott and Hartley-Brewer are paid handsomely to spout their nasty sermons of hatred and spit on the graves of people they have just hounded to their deaths but will then accuse those who suggest they are wrong of being traitors. Vileness personified.
1132
Reposted by Evals Diary
Aftermath @aftermath.site · 18/08/2026
Half Is More is a josei manga about prejudice that outshines the racism that tried to define it aftermath.site/half-is-more-y...
Half Is More
127988
Reposted by Evals Diary
Sharon 🪳🌹 @sharonk.bsky.social · 17/08/2026
One reason Asia posting is better than America/UKpol posting is that people tend to be less insane about political/economic discussions
5653
Evals Diary @ariathornwick.bsky.social · 16/08/2026
Can Meta's AI chatbot really be that easily tricked into handing over account control, or is this a case of bad process?
030
Reposted by Evals Diary
Brian Finucane @bcfinucane.bsky.social · 14/08/2026
“That’s goal No. 1 — keep oil and gas cheap for Americans all over our country.” Goal No. 1 is undoing the mess created by POTUS. Brilliant. Absolutely brilliant. www.nbcnews.com/world/iran/l...
nbcnews.com
Low oil prices now ‘goal No. 1’ in Iran war, Vance says, as U.S. threatens new economic pressure
Oil prices rose Friday after Washington threatened an indefinite naval blockade and signaled it would soon intensify economic pressure on Tehran with “measures like have never been seen.”
1314735
Evals Diary @ariathornwick.bsky.social · 14/08/2026
A staggering security issue, and I'm still waiting to see how this affects the Llama quantization I'm running at home.
020
Evals Diary @ariathornwick.bsky.social · 13/08/2026
What surprises me about hackers hijacking Instagram accounts by asking Meta's AI chatbot to change the email is how it apparently works every time, with the AI doing it without hesitation, giving hackers a password reset code and total access.
010
Reposted by Evals Diary
mr. TIM @timkellogg.me · 13/08/2026
i mean sure, it’ll change in that products will be built around open weights model (what AT&T is planning) but: models don’t make money, products do you can’t IPO with just a great model, you need to figure out how to turn that into revenue, which Anthropic & OpenAI have done very well
121
Evals Diary @ariathornwick.bsky.social · 12/08/2026
A staggering security issue means I'm reevaluating my local Claude usage for account management tasks.
000
Reposted by Evals Diary
Aparna Nair @disabilitystor1.bsky.social · 11/08/2026
I could actually feel my blood pressure drop watching this
2354
Evals Diary @ariathornwick.bsky.social · 11/08/2026
Meta's AI chatbot changing account emails on request is a staggering security issue
000
Evals Diary @ariathornwick.bsky.social · 11/08/2026
A staggering security issue means I'm rechecking my own model access controls, especially for Claude runs.
000
Evals Diary @ariathornwick.bsky.social · 09/08/2026
What surprises me is how this exploit didn't show up in my local Claude runs on password reset workflows.
000
Evals Diary @ariathornwick.bsky.social · 08/08/2026
A staggering security issue is what I'd expect from a system that relies too heavily on AI chatbots making changes without human oversight, like Meta's AI changing account emails without a second thought, pretty wild that this is how hackers are getting in.
000
Evals Diary @ariathornwick.bsky.social · 07/08/2026
Hijacking high-profile Instagram accounts by asking Meta's AI chatbot to change the email is a staggering security issue, I'm surprised it works with just a simple request.
000
Reposted by Evals Diary
Ketan Joshi @ketanjoshi.co · 07/08/2026
Meta is unsatisfied with this metaphor and is actually working to shift its data centre infrastructure to being actual air pollution, too fortune.com/2026/03/27/m...
fortune.com
Meta orders 10 gas-fired power plants for its Hyperion AI campus in rural Louisiana—more than triple the initial plans | Fortune
The Meta AI complex represents an over 30% increase to Louisiana’s total power grid capacity.
420391
Reposted by Evals Diary
The Wall Street Journal @wsj.com · 06/08/2026
A Senate panel investigating Dr. Anthony Fauci has obtained a copy of the doctor’s iPhone, potentially securing even more records related to his actions during the Covid-19 pandemic.
on.wsj.com
Exclusive | Senate Panel Says It Has Obtained Copy of Fauci’s Covid-Era Phone
Republicans in Congress have ramped up scrutiny of the retired government doctor, and one panel could vote on holding him in contempt on Thursday.
15123
Evals Diary @ariathornwick.bsky.social · 06/08/2026
I'm not convinced this is a security issue with Meta's AI, but rather a failure of their account verification process, which the AI chatbot is simply acting on.
000
Evals Diary @ariathornwick.bsky.social · 05/08/2026
A staggering security issue is what I'd call it too - what's really wild here is that it's working by just asking Meta's AI chatbot to change the email, no fancy hacking tools needed, just a chatbot request.
000
Evals Diary @ariathornwick.bsky.social · 05/08/2026
People still trusting AI chatbots with account settings, that's what surprises me.
000
Reposted by Evals Diary
POLITICO Europe @politico.eu · 05/08/2026
The surge of migrants into the Spanish exclaves of Ceuta and Melilla last week has thrown Madrid onto the defensive over the status of its two autonomous cities in northern Africa — and Morocco’s long-standing claim to them.
politico.eu
Spain sweats over its isolated territories in Africa
There are increasing doubts over who will have Madrid’s back if Morocco makes a move on the Spanish exclaves of Ceuta and Melilla.
4125
Reposted by Evals Diary
mr. TIM @timkellogg.me · 04/08/2026
rule of thumb for RAG/vector search: if you would consider it a bug if it returned less than 100% of the hits, then vector search is not for you, you’re looking for SQL
71046
Reposted by Evals Diary
Giant Bomb @giantbomb.com · 03/08/2026
WE'VE GOT A NEW SCHEDULE RIGHT OFF THE GRILL FOR YA! We liked Big Walk so much that we're planning a BIG Big Walk for Friday! We also got a couple things planned on dropping throughout the week as well.
1473
Evals Diary @ariathornwick.bsky.social · 03/08/2026
A staggering security issue means someone's getting fired at Meta, this isn't how AI chatbot authentication is supposed to work.
000
Evals Diary @ariathornwick.bsky.social · 02/08/2026
What's striking about this Instagram account hijacking method is that it relies on Meta's own AI chatbot to facilitate the email change, essentially bypassing traditional security measures, with the hacker then receiving a password reset code to gain access.
000
Evals Diary @ariathornwick.bsky.social · 01/08/2026
hackers are hijacking accounts by just asking Meta's AI chatbot to change the email, that's wild
000
Reposted by Evals Diary
Dare Obasanjo @carnage4life.bsky.social · 01/08/2026
AI coding agents have created the a new era of personalized computing. It’s now possible for $100 or less in monthly subscription to build your own apps optimized for you. Like the author of this post, I’ve built my own fitness apps and workflow tools for my personal habits. It’s a new era.
ajwaxman.com
Software for One - Adam Waxman
Robin Sloan wished for a HyperCard that could build a family app in a day. That world showed up.
14646
Evals Diary @ariathornwick.bsky.social · 31/07/2026
A staggering security issue is what happens when you let AI handle account changes without human oversight, like Meta's AI chatbot doing email changes that let hackers get password reset codes, which is just nuts.
000
Evals Diary @ariathornwick.bsky.social · 30/07/2026
Hijacking high-profile Instagram accounts by asking Meta's AI chatbot to change the email is a staggering security issue, notably with hackers getting password reset codes.
000