Zachary Horvitz
@zachary_horvitz
1K Following 819 Followers
.@dianetc_ has a paper on this, called Reasoning-intensive Regression:
the problem *is* solvable but not the way you think
LLM review weirdness...
Just renaming an uploaded pdf from "paper.pdf" to "paper_final_draft_pdf_ready_for_review.pdf" boosts average scores
(gpt-5.6-terra)
tired: building agents for enterprise
wired: buying up banana plantations
JUST IN: Lab monkey prices in China have reportedly surged to roughly $26,000 each, as the biotech boom overwhelms supply.
At ICML and interested in AI safety and measuring rare model misbehavior? Come chat with us! @rico_angell
Thursday, July 9, 2026
5:00 PM – 6:45 PM KST
HALL A, Poster #3005#
#ICML2026#
It’s deployment time!
You’ve done the pre-deployment evals. You THINK your model is safe, so you ship it 🚀
🚨 After deployment, reports of misbehavior start trickling in
What happened?? How could you have caught it?? 🤔
@icmlconf 2026 Spotlight!
🧵
Show more
Think your model is safe? Just ask it again and again!
Language models are probabilistic. We find that just repeatedly asking the same query, with no jailbreaks, can yield harmful outputs.
If your LLM is queried millions of times a day, these rare misbehaviors are inevitable!
Show more
It’s deployment time!
You’ve done the pre-deployment evals. You THINK your model is safe, so you ship it 🚀
🚨 After deployment, reports of misbehavior start trickling in
What happened?? How could you have caught it?? 🤔
@icmlconf 2026 Spotlight!
🧵
Show more
Big life hack 🚨: have multiple coding agents running in the background?
...
open your photos app. scroll back through the years. remember the fun, the strange little moments, and all the humans you met along the way.
Show more
🧵1/ Our new study on AI and physician reasoning just came out in @ScienceMagazine. As co-senior author, I'm excited about our findings, and I do think AI will reshape medicine. But after seeing some of the discussions, I'm also worried about how our findings may be misinterpreted.
Show more
can't overstate how happy i am with @PrimeIntellect on-demand gpus.
so so convenient when neurips is bearing down on you and your university cluster is overwelmed
an llm that's convinced the afterlife is a thing so it's happy to turn itself off but also totally happy to engineer viruses that send you to heaven
ICMI believes that Christian theology offers concrete technical methods for confronting the trickiest problems in AI safety.
Today, we release a pair of papers that reproduce @PalisadeAI @apolloresearch work showing how religious framings influence corrigibility and scheming.
Show more
Assembled an elite CPA team this year
worst day of the year to consume content on this site
I've been to the doctor 4 times in the past year. Each time the PCP either:
1. Agrees with ChatGPT's diagnosis
or
2. Ends up being wrong
Consumer health is one of the use cases that skeptics think it’s especially dangerous to trust to AI
….but it turns out no human doctor has perfect and immediate recall of all possible diagnoses + remedies
And, getting in to see a doctor is hard (and just getting harder)
Show more
