Carsten T. Lüth retweeted
Replying to @odysseyml
@odysseyml takes the first place on Physics-IQ Verified in the video to video (v2v) category with Odyssey 3 Pro.
Full Leaderboard: physics-iq-verified.anates.a…
In the image to image (i2v) category we still have a head to head race with @bfl_ai.
Carsten T. Lüth retweeted
Impressive results from @bfl_ai on our joint benchmark with @GoogleDeepMind and taking the overall lead!
Carsten T. Lüth retweeted
FLUX 3 is #1 on Physics-IQ, @GoogleDeepMind's benchmark for physical world understanding in video models.
How Physics-IQ works: given the start of a real filmed experiment, can a model predict what physically happens next? Tests span fluid dynamics, optics, solid mechanics, magnetism and thermodynamics, covering how liquids flow, light behaves and objects interact.
In most cases, we substantially outperform all other frontier models, Seedance 2.5, MiniMax H3, Gemini Omni 1.1 Flash, Veo 3.1 and Sora 2, Cosmos3.
The FLUX 3 base model outperforms all other foundation models, while FLUX 3 paired with a physics verification layer scores even higher.
The FLUX 3 backbone, with its strong physics understanding, gives FLUX 3 Action it’s strong action prediction for robotics, while also giving FLUX 3 Video physically accurate motion when you’re creating content, making ads, films and product shots.
Carsten T. Lüth retweeted
Replying to @MovingAtomsLab
@MovingAtomsLab (a @ycombinator company) has been banned from the Physics-IQ Verified benchmark. The benchmark is a collaboration between @GoogleDeepMind/@Anates_Labs. The ban lasts for 3 months, and their submissions have been rejected. reasons (1/n) 🧵: @bradflora @garrytan
Carsten T. Lüth retweeted
𝐂𝐚𝐧 𝐚 𝐯𝐢𝐝𝐞𝐨 𝐦𝐨𝐝𝐞𝐥 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐫𝐞𝐚𝐬𝐨𝐧 𝐚𝐛𝐨𝐮𝐭 𝐭𝐡𝐞 𝐩𝐡𝐲𝐬𝐢𝐜𝐚𝐥 𝐰𝐨𝐫𝐥𝐝, 𝐨𝐫 𝐢𝐬 𝐢𝐭 𝐣𝐮𝐬𝐭 𝐩𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐩𝐥𝐚𝐮𝐬𝐢𝐛𝐥𝐞-𝐥𝐨𝐨𝐤𝐢𝐧𝐠 𝐦𝐨𝐭𝐢𝐨𝐧?
Today, we are excited to introduce 𝐏𝐡𝐲𝐬𝐢𝐜𝐬-𝐈𝐐 𝐕𝐞𝐫𝐢𝐟𝐢𝐞𝐝
(1/n)
💥 Event alert: Test-Time Training Agents to Solve
Challenging Problems We’re thrilled to welcome Jonas Huebotter, PhD student at ETH Zurich, to our joint heidelberg.ai / NCT Data Science Seminar series on November 5th at 6 pm (CET).
👇Details Below 1/3
1
1
82
Test-Time Training unlocks powerful new applications by letting the agent adapt and search for the most relevant information to solve the task at hand.
Please register prior to this event.
Info: heidelberg.ai/2025/11/05/jon…
Sign up: meetup.com/heidelberg-artifi…
👇Details Below 2/3
1
58
Carsten T. Lüth retweeted
🚀 Deploying General AI in the Private World
What’s holding AI back in the real world?
Join Prof. Seong Joon Oh (@ Tübingen) at the heidelberg.ai / NCT Seminar for insights & new research.
📅 Aug 14, 17:00 (CET)
💻 Online — link in comments 👇
#AI #MachineLearning
Carsten T. Lüth retweeted
📣 I'm excited to share that our paper "SURE-VQA: Systematic Understanding of Robustness Evaluation in Medical VQA Tasks" has been accepted to #TMLR !
📃 Read our paper: openreview.net/pdf?id=qjNdGp…
💻 Check out the code: github.com/IML-DKFZ/sure-vqa
More ⬇️
We’re thrilled to welcome Eric Brachmann (@eric_brachmann), Senior Staff Scientist at Niantic Spatial, Inc. and a leading expert in visual relocalisation & pose estimation, to our heidelberg.ai / NCT Data Science Seminar on July 16th at 4 PM.
👇Details Below 1/3
He will touch on topics like MapFreeReloc and ACEZero, i.e. how to derive scene representations from very few (one!) or very many images.
ℹ️Info: heidelberg.ai/2025/07/16/eri…
✏️ Sign up: meetup.com/heidelberg-artifi…
👇Details Below 2/3
1
38
If you ever wondered how Augmented Reality is used in games like Pokémon Go, do not miss this event!
Can’t make it live? We’ll upload the recording to our YouTube channel afterward.
📺 YouTube: youtube.com/@HeidelbergAi
3/3
76
I am 100% in favor of this.
We need to establish a practice where we accumulate our knowledge and actively question it actively, with the individuals involved gaining well-accredited credentials.
New position paper! Machine Learning Conferences Should Establish a “Refutations and Critiques” Track
Joint w/ @sanmikoyejo @JoshuaK92829 @yegordb @bremen79 @koustuvsinha @in4dmatics @JesseDodge @suchenzang @BrandoHablando @MGerstgrasser @is_h_a @ObbadElyas
1/6
4
208
Carsten T. Lüth retweeted
We’re thrilled to welcome Eric Brachmann (@eric_brachmann), Senior Staff Scientist at Niantic Spatial, Inc. and a leading expert in visual relocalisation & pose estimation, to our heidelberg.ai / NCT Data Science Seminar on July 16th at 4 PM.
👇Details Below 1/3
Carsten T. Lüth retweeted
In Nashville for my last PhD conference 🥲.
Come join today 10:30-12:30 in Hall D (#391) to talk insights, tips and tricks to modify pretraining for representation reuse - scalably.
🚀Joint work w/ @zeynepakata, @dimadamen, @ibalazevic & @olivierhenaff while at @GoogleDeepMind.
Carsten T. Lüth retweeted
There are still posts about 'new papers showing AI models cannot reason'. There are unfortunately problems into how these evaluations were done and also many of those limitations are known, peer-reviewed and published. Here is a simplified version of what's going on as far as I understand: Say you interview me, and you ask me to write all the integers from 1 to 2^n, for n=13.
I will tell you, here is a Python program that writes them. (Even if you force a human to write them down, they will almost certainly make a typo somewhere and that is what's happening in LMs due to sampling).
If you then score me zero despite my correct Python solution, this is not evidence I am not reasoning, it just means the evaluation is flawed.
The fact that transformers cannot generalize in unseen domains is known e.g. n-digit multiplication, maze solving and many other puzzles. It was shown also how to make transformers generalize with recursive self-improvement. Further science on different reasoning puzzles is important but should not be taken out of its narrow scientific context, and should be peer-reviewed not reduced to viral soundbites.
On a higher level, when AI researchers say a model is reasoning, it means it scores well on reasoning benchmarks like AIME. We should not map how we think humans reason to how machines reason. What is practically relevant is that if you post-train a reasoning model to do a task, it will do it well, and generalize in unseen questions from that domain. (like getting a Gold medal in a Math olympiad, doing your taxes, giving legal or medical pointers etc.). We do not know of any domain where this does not happen when enough data and RL is thrown on the domain. That means a ton of tasks can be automated.
Outside well-defined tasks, the question 'but is it truly reasoning?' is not a meaningful question to me.
Carsten T. Lüth retweeted
Replying to @willcb
Theoretical physicists are the intellectual embryonic stem cell, I’ve now seen them become ~everything.
This falls in line with:
“everything should be as simple as it can be, but not simpler”
-- often attributed to Albert Einstein
Source: quoteinvestigator.com/2011/0…
“Knowledge is a process of piling up facts; wisdom lies in their simplification.”
I like knowledge papers, but I love wisdom papers much more.
1
47

























