Tomer Ullman

new preprint: "Directing large language models to follow the letter or spirit of the law" arxiv.org/pdf/2609.23083 (by Qian , Li, Chen, Murthy @soniakmurthy.bsky.social , Belinkov, and me) this is particularly cool/important, and I'm allowed to say it because it was headed by …

arxiv.org

I gave comments for a news piece in Science on a recent preprint on AI finding loopholes ("Large Language Models Hack Rewards, and Society", Liu et al.) since these things are always cut short, I wanted to expand here: www.science.org/content/arti... arxiv.org/abs/2606.04075

science.org

I gave comments for a news piece in Science on a recent preprint on AI finding loopholes ("Large Language Models Hack Rewards, and Society", Liu et al.) since these things are always cut short, I wanted to expand here: www.science.org/content/arti... arxiv.org/abs/2606.04075

Large Language Models Hack Rewards, and Society arxiv.org

Janelle Shane @janelleshane.com in particular has explored many ways in which AI models go wrong along these lines, like these color pallettes or Victorian Christmas cards (www.aiweirdness.com/victorian-ch...)

Victorian Holiday cards by AI aiweirdness.com

"I used an LLM to help me write, but it was just to polish the style and grammar" is something you've probably come across, or done yourself. Many conferences and journals and teachers are ok with this. But does it...? Just focus on the style and grammar...?

from afternoon conversation with colleague in Math, talk turns to AI: "we're having faculty meetings and they're telling us all 'it's urgent that...'. I've spent my professional life in the math department, do you have any idea how unused I am to the phrase 'it's urgent'?"

i get many emails every day from students about various things, but one that made me particularly sad recently was asking about "what we're actually going to be good at when AI can do most of what we're training to do" it was the phrase "what we're training to do" that made me real sad

I'm chaperoning my 4th grade daughter on a school trip and one boy next to me just taunted the other by yelling "HUGO YOU ARE LITERALLY AI GENERATED!!"

I appreciate that some places have negative consequences for AI-slop-submissions (desk reject, ban from submitting for X time) but what do we do about reviews that were clearly LLM-written? Ban those people from being reviewers? Oh no! Do we threaten them with some delicious ice-cream too?

I just came back from a Dagstuhl Seminar on "social intelligence in AI". There's much to consider, but I worry a current path we're on is:

ML's Lament (to the tune of 'Part of Your World')

(i'm reviewing papers for a few ML conferences atm) we should give standard intros standard numbers, so people have the option of skipping them in both writing and reading e.g. a paragraph saying "LLMs are amazing [cite], BUT they still can't do many things [cite]" will be designated OPENING #7

looking forward to the CogSci symposium on "CogSci-Inspired Evaluation of Large Language Models" at the #cogsci2026 conference I'll be talking about evaluating cog-things in AI-stuff

I wonder if anyone's done meta-stylometry to recover interactions based on an non-vanilla LLMs output(?) I mean to say: repeated use of current language models can shift their output to suit the user; So, take the output of a modified model and work backwards to infer prior interactions