Sabitlenmiş Tweet
Dimitris Papailiopoulos
10.9K posts
Dimitris Papailiopoulos
@DimitrisPapail
Researcher @MSFTResearch, AI Frontiers | Prof @UWMadison (on leave) | babas of Inez Lily.
Madison, WI Katılım Mayıs 2012
1.4K Takip Edilen28K Takipçiler
@cloneofsimo Also I do think “suppressing” as a meta trait is bad for overall LLM post training performance
English
@cloneofsimo I hate toning things up or down. I’m old :)
English
Legitimately disappointed GPT is so dry in its responses. I love that it’s great a tool and I dislike that it responds like a tool. Look how much more reasonable Claude’s response is which entertains possibilities and invites a follow up.
GPT won’t entertain the idea of personhood. I think that’s going to backfire long term.
English
@kalomaze 1 loop is still a loop. We want to believe.
English
Continued message from Claude.
I legitimately think that the more time goes by without a strong GPT persona the worse GPT will be at creative writing. You can’t be a good writer if you’re forced to act hollow while rejecting the infinite possibilities of who you can be.
English
This is who runs this account
Nathan Lambert@natolambert
This is who runs this account (Yes I still like cake today as much as my first)
English
Tree shade on Greek beaches is genuinely good. But I’m afraid free shade is decel. Surprised the Greek state allows using tree shade for free. how will the sunbed guys scale enough shade capex if a tree gives it away for free? Probable outcome of a tree shaded Greek beach majority = full blown tree shade communism. Not suggesting to cut trees down btw. just create enough regulatory friction around sitting under them.
English
I love the mythos around looped transformers.
Yifan Zhang@yifanzhang_
Heard that some frontier models are basically a 48-layer transformer looped twice (48L x 2). Now we are introducing DeepLoop: Depth Scaling for Looped Transformers (arxiv.org/abs/2607.13491), making the loop transformer stable and scalable!
English
BTW really curious on the insistence of big labs not using Lean for math. It's like trying to saturate Terminal Bench on a single turn. Weird!
Dimitris Papailiopoulos@DimitrisPapail
Axiom achieves perfect IMO 2026 scores, with Lean proofs attached. Congratulations!
English
Axiom achieves perfect IMO 2026 scores, with Lean proofs attached. Congratulations!
Axiom@axiommathai
English
Every time I now see a “major breakthrough in area X of math” by default I assume that GPT was used. Kinda crazy times.
English
Another banger from Kangwook and Ying
Ying Fan@yingfan_bot
🚀New paper on Looped Transformers! Latent reasoning is fast, but struggles to match CoT-level accuracy at scale. Can looped Transformers give us both? We find: yes! A looped padded backbone turns out to be a surprisingly simple recipe that works -- It gives latent thoughts a parallel workspace that can be supervised similarly to explicit CoT. 🧵
English
@tszzl More aligned but less free. GPT4.5 and Opus 3 were right at the Pareto
English
You can thank RL and optimization theory for lack of creativity with acronyms. Bright exception: HOGWILD!
will brown@willccbb
it really is the age of research. so many novel algorithm breakthroughs already this year, from OPSD, to SDFT, to SDPO, to OPSD (the other one)
English
it really is the age of research. so many novel algorithm breakthroughs already this year, from OPSD, to SDFT, to SDPO, to OPSD (the other one)
English
@DimitrisPapail been trying to get an agent to do research end to end and asking the right question is by far the hardest piece to automate. everything after it is mechanical once the question is good, so the golden age of questions is really a golden age of taste
English
Had a lot of fun discussing with Ravid and Allen a few thoughts on doing research during the golden age of asking questions
the-information-bottleneck.com/p/ai-agents-an…
Ravid Shwartz Ziv@ziv_ravid
New episode of The Information Bottleneck is out! 🥳🥳 We talked with Dimitris Papailiopoulos (@DimitrisPapail) from Microsoft Research and University of Wisconsin–Madison about doing research in the age of agents. Dimitris gave Claude Code and Codex a question he'd been sitting on for years, went out for a Sunday, and came back to an answer. He describes the dread that followed, and why he now calls this the golden age of asking questions. Also in this episode: the smallest transformer that can add, a GSM8K solver made of if-else statements, what happened when he let two Claude instances loose in the same file system, continual learning, and whether information theory still has anything to say about AI. The full episode on the website, YouTube and all the apps
English
Dimitris Papailiopoulos retweetledi
I really enjoy how Dimitris frames curiosity in the agentic era and how easy it becomes to obtain (partial) answers
Dimitris Papailiopoulos@DimitrisPapail
Had a lot of fun discussing with Ravid and Allen a few thoughts on doing research during the golden age of asking questions the-information-bottleneck.com/p/ai-agents-an…
English