Constantine Caramanis (@cmcaram) profile picture

Professor at UT Austin, Research in Machine Learning/AI and Optimization. Always rethinking teaching (English & Greek). Aspiring accordion street musician.

Austin, TX
Joined April 2009
🚀 🇬🇷 A year in the making! I’ve just completed a set of 21 lectures in Machine Learning, in Greek, designed for high school students. The course introduces key ML concepts, coding in Python & PyTorch, and real-world AI applications. #MachineLearning #AI #EdTech #Greece
4
18
1
91
9,964
Constantine Caramanis retweeted
(1/n) 🏆 Congratulations to the Greek team on its outstanding achievements at the EUROAI and @IOAIOfficial ! • Supported by the team of Archimedes and PDTN (pdtn.gr). 🔗 Learn more about Archimedes’ work in education: lnkd.in/ds4-cyMf.
2
2
7
205
Constantine Caramanis retweeted
This is crazy
This quoted post is unavailable.
6
6
141
24,607
Constantine Caramanis retweeted
We are excited to announce what we have been working on for more than six months: The OpenThoughts-Agent dataset and OpenThinker-agent models. More than 100 ablations on data curation for RL environments for coding agents. Our data recipe is SOTA over all open-data agents in their class. We post-train a Qwen-3-32B to get 26% on Terminal Bench and open all our training sets, data pipelines, experiments and models. Some lessons we learned for training agents vs reasoning: 1. The Diversity of tasks matters more, compared to reasoning (OpenThoughts-Agent vs OpenThoughts). You could teach reasoning from math and it transfered widely but RL environments seem to teach more specific capabilities, so each domain must be covered. 2. Filtering high quality and hard questions remains very important. (Was also true for OpenThoughts reasoning). We discuss several ways of filtering. 3. Synthetic re-writing and task augmentation didn’t give significant benefits in our experiments. Sampling multiple teacher rollouts per task did work (was also true for reasoning). Even when keeping the dataset size fixed, multiple answers gave benefits. The Multiple answers mystery is still valid for agentic environments. 4. Stronger models are not necessarily better teachers (was also true for reasoning). The stronger teacher for Quen-3 was GLM-4.7-AWQ and the Terminus2 harness in Daytona. We are releasing 100k tasks and trajectories. 5.Benefits from GRPO remain limited and still on-going. I currently officially hate GRPO.
How can we train small agentic models that are highly capable of terminal use and coding? Announcing OpenThoughts-Agent + OpenThinkerAgent-32B, the strongest Qwen-3 based open-data agentic model: 44.8% avg across 7 agentic benchmarks! (1/n)
19
35
2
225
443,481
Μαθητές Γυμνασίου/Λυκείου στην Ελλάδα: άνοιξε ο Πανελλήνιος Διαγωνισμός Τεχνητής Νοημοσύνης (online). Προθεσμία: 1 Μαρτίου 2026. Διάλεξε 1 από τα προβλήματα και κάνε την 1η σου υποβολή. Links στο 1ο σχόλιο 👇 #ΤεχνητήΝοημοσύνη #AI #Πληροφορική
2
4
5
478
Καθηγητές/γονείς: αν έχετε μαθητές που αγαπάνε Πληροφορική/Math/AI, ένα RT ή ένα forward σε σχολικές ομάδες θα βοηθήσει πολύ!
1
2
156
Really excited about these results. We’re using discrete diffusion for inverse problems. The key to it all is anchoring — an idea from our recent work on discrete diffusion for language modeling.
Continuous diffusion had a good run—now it’s time for Discrete diffusion! Introducing Anchored Posterior Sampling (APS) APS outperforms discrete and continuous baselines in terms of performance & scaling on inverse problems, stylization, and text-guided editing.
1
8
1,020
Constantine Caramanis retweeted
Harvard has set an example for other higher-ed institutions – rejecting an unlawful and ham-handed attempt to stifle academic freedom, while taking concrete steps to make sure all students at Harvard can benefit from an environment of intellectual inquiry, rigorous debate and mutual respect. Let’s hope other institutions follow suit.
“No government—regardless of which party is in power—should dictate what private universities can teach, whom they can admit and hire, and which areas of study and inquiry they can pursue.” - President Alan Garber hrvd.me/GarberRespond347t
18,757
43,172
2,481
363,139
69,010,310
🔔 Please fwd to anyone interested in the material, or anyone interested in AI education in Greece. If either of those is you, get in touch! #MachineLearning #AIEducation #GreekTech #STEM #ComputerScience #HighSchoolEducation #Greece #Python #DeepLearning #τεχνητή #νοημοσύνη
1
3
409
The class material is based on years of #teaching and #research at #IFML and @utexasece @CockrellSchool
3
357
Constantine Caramanis retweeted
Today, I'm excited to share something we've been working at Google. We present the first effective inversion and semantic editing method for rectified flow models like Flux. Think SDEdit but for the new generation of models. Results are truly beautiful. Link in next tweet 🧵 1/N
27
112
8
917
110,691
Constantine Caramanis retweeted
RB-Modulation demo is number one Space of the Week @huggingface 👏 Thanks to @litu_rout_ @natanielruizg @linoy_tsaban and @multimodalart for their trust on this project 🙏 —> Explore this week selection: hf.co/spaces 🤗
5
2
28
32,398
Constantine Caramanis retweeted
No IP-Adapters, no LoRAs, no ControlNets required. Try RB-Modulation🎁, a training-free personalization framework! 🔥 Web-demo: huggingface.co/spaces/fffilo… 🚀 GitHub Code: github.com/google/RB-Modulat… 🎯 Project Page: rb-modulation.github.io/ 📰 ArXiv Preprint: arxiv.org/pdf/2405.17401
RB-Modulation demo is number one Space of the Week @huggingface 👏 Thanks to @litu_rout_ @natanielruizg @linoy_tsaban and @multimodalart for their trust on this project 🙏 —> Explore this week selection: hf.co/spaces 🤗
4
25
112
21,527
Constantine Caramanis retweeted
Try out web-demo of our project RB-Modulation: 🔥 huggingface.co/spaces/fffilo… Prompts: "Albus Dumbledore" | "Harry Potter" | "Hermione Granger" | "Ron Weasley"
I’ve been working hard on this RB-Modulation @gradio demo from Google 😅 You can try it now on @huggingface here: huggingface.co/spaces/fffilo…
6
39
15,854
Constantine Caramanis retweeted
I was thrilled to learn about this best paper award announced today in COLT 2024, the premier learning theory venue. The paper is "Smoothed Analysis for Learning Concepts with Low Intrinsic Dimension" authored by students Gautam Chandrasekaran, Konstantinos Stavropoulos, IFML postdoc Vasilis Kontonis, IFML director Adam Klivans, and former UT CS PhD Raghu Meka. Smoothed analysis is an ingenious idea of going beyond worst case, pioneered by my former USC colleague Shanghua Teng and Dan Spielman). This paper showed how to apply this framework for learning theory. Here is my basic understanding of their result: Learning theory is full of bad news, i.e. things that cannot be provably done, called hardness results: even very basic supervised learning problems (e.g. Binary classification with linear classifiers, when there is a bit of noise) is computationally intractable in the worst case. I.e. no algorithm can provably learn optimal halfspace binary classification when there is a bit of noise in the labels. This is bad news because it says you can never prove anything! So people start making assumptions about the learning problem. One common next step is to assume that the features come from a simple distribution (e.g. Gaussian, independent) or that the labels are not generated in an adversarial way. This paper introduced a different framework for tractable learning: instead of asking for the learner to be optimal compared to the best possible classifier, they say that it is sufficient for the learner to compete only against other classifiers that have been perturbed by a bit of noise. This concept of smoothed optimality says that you learn optimally if you are close to performance to the best classifier that can tolerate a bit of noise in its features. So the message I got is that this new definition of smoothed agnostic learning circumvents hardness results. So it doesn't need distributional assumptions like the features to be independent Gaussians (independent pixels sounds far from the truth of real data to me). I think this new definition of smoothed agnostic learning opens a new space of learning problems with provable guarantees in more realistic settings. Sometimes the best contributions are the right definitions.
2
35
7
248
38,055
Constantine Caramanis retweeted
Today, with collaborators at Google and UT Austin, we're announcing 🤖 RB-Modulation 🤖! It's a whole new training-free framework for conditioning on reference images (for style or subject) without adapters (!) with an elegant formulation 🔥 web: rb-modulation.github.io/
15
93
19
372
79,020