Passionate of ML, NLP and algorithms. Principal Scientist @ NVIDIA, Assoc Prof @ UPB. Growing the ML community in Romania.
Bucuresti
Joined August 2008
- Tweets89
- Following190
- Followers197
- Likes751
Traian Rebedea retweeted
1/4 We trained an 8B safety classifier on just 1,275 carefully annotated prompts.
It beat GPT-5.4, Claude Sonnet 4.6, and leading open-weight guardrails trained on far more data, on average across 5 safety benchmarks.
Quality > quantity.
Really great to be part of @CoLLAs_Conf this here and host the conference in Bucharest! Looking forward for the next editions and papers in continual learning, lifelong learning agents, and unlearning.
📢 As #CoLLAs2026 comes to a close in Bucharest 🇷🇴, we want to send a massive THANK YOU to the team who made it happen! 🙌
To our incredible Local Organizers, Elena Burceanu (Bitdefender, Politehnica University of Bucharest) & Traian Rebedea (NVIDIA, Politehnica University of Bucharest), and our dedicated General Chairs, Sarath Chandar (Mila - Quebec AI Institute, Polytechnique Montréal) & Razvan Pascanu (Google DeepMind, Mila), thank you all so much for an unforgettable event!
See you at CoLLAs 2027 in Montreal 🇨🇦! 🌍❤️
Best Romanian AI Thesis 2026 is now open! Submit your thesis until June 1st.
🚀 Romanian AI Days is organizing the third edition of the Best Romanian AI Thesis Awards 🏆, recognizing and rewarding the best Bachelor's, Master's and PhD theses in the field of AI 🤖. Submissions are open until July 1, 2026 ⏳brait.airomania.eu/
Traian Rebedea retweeted
as chair of ACL publication ethics committee i will push for:
* ban on prompt injections & other adversarial content hidden from humans in papers
* scan of anthology, openreview, softconf for cheating manuscripts
* sanctioning cases as abuses of fair review
@ReviewAcl
Traian Rebedea retweeted
Romanian AI Days și Asociația Română pentru Inteligență Artificială lansează Best Romanian AI Thesis Awards 🏆🎉, un concurs destinat celor mai bune lucrări de licență, disertație și doctorat 🎓 din domeniul AI 🧠🤖. Pentru detalii, intră aici: days.airomania.eu/competitio…
Point 2 there is so not obvious even to very techy people.
2. Accelerated computing is mostly not hardware. I’m surprised people still see NVIDIA’s business as hardware, after all these years.
Great to be part of HuggingFace Daily Papers.
In this work, we show that topic-following is an essential part of LLM alignment and can improve both TODS systems, but also safety alignment. Our CantTalkAboutThis dataset will be on HF soon, together with topic-aligned models.
Traian Rebedea retweeted
CantTalkAboutThis
Aligning Language Models to Stay on Topic in Dialogues
Recent advancements in instruction-tuning datasets have predominantly focused on specific tasks like mathematical or logical reasoning. There has been a notable gap in data designed for aligning
Traian Rebedea retweeted
LLMs are finally catching up to deep RL - we have been training on test from long before it was cool.
Traian Rebedea retweeted
If you're seeking an advisor in NLP, consider exploring options with recently appointed faculty. While established senior folks are well-known, junior folks shine as rising stars. In fact, many successful PhDs are the first few students of their advisor. Don't miss the chance!
Traian Rebedea retweeted
I've said it and will say it again:
#1 Smaller, cheaper, faster, more customized models will cover 99% of use-cases. You don't need a million dollar formula 1 to get to work everyday and you don't need a banking customer success chatbot to tell you the meaning of life!
#2 All companies will ultimately want to build AI themselves (based on open-source AI) versus outsource to third-party APIs and there will be as many models as code repos today. AI is a foundational technology to build tech and the same way you don't outsource or have the same code-base as your competitors, you won't want to outsource your AI development or use the same models as your competitors.
What's going on here?
Microsoft wants to build its own advanced AI to reduce costs and dependence on OpenAI.
What does this mean?
Microsoft is directing its researchers to create conversational AI models that perform nearly as well as OpenAI's but are much smaller and cheaper to run. This in-house AI is already being tested by Microsoft's Bing team for features similar to ChatGPT.
The goal is to save on the ballooning compute costs of large AI models. Even though Microsoft has invested over $10 billion in OpenAI for exclusive access to its tech, unchecked costs from widespread use could get out of hand fast. Smaller "distilled" models that mimic the capabilities of behemoths like GPT-4 can help control expenses while providing customers with powerful AI features.
Why should I care?
This gives Microsoft options beyond OpenAI to deliver performant and affordable AI products. It's a smart business move to avoid overreliance on external partners, however fruitful the relationship may be currently. Plus having in-house alternatives puts Microsoft in a better negotiating position with OpenAI down the road.
100% agree
I get asked a lot: why stay in academia, all the excitement in AI is happening in industry with massive compute. And I am seeing some profs leaving academia, but also seeing lots of researchers in industry looking to go back to academia, especially those who don’t work on LLMs.
I have spent some time and actively working with industry, and it has been a great experience, but for me the answer is simple:
1. Students: Nothing can replace working with really smart & amazing students.
2. Academic freedom: In general, industry will one way or another dictate what you should work on. Today it is LLMs, tomorrow it is something else. The good old days of "here is a cool idea, let me investigate and publish" are pretty much over. But as an academic, I can work on whatever I want: I can start working on "black holes" tomorrow if I choose to.
3. And my favourite one: No reorgs -- big tech really loves reorgs that happen every few months.
Finally, as an academic, I can always take some time off, work in industry or start-up, and come back if I want to. Going into industry is usually a one-way street: It is far easier to go to industry from academia and much harder the other way around.
Traian Rebedea retweeted
Many people are more capable than their boss.
AI systems may become more capable than you, but you'll still be their boss.
If you feel threatened by having a staff -- of humans or machines -- that is smarter than you, you are not a good boss.
If you want to build controllable and trustworthy dialogue systems with LLMs, check out the NeMo Guardrails toolkit we have just lauched!
NeMo Guardrails uses Colang, a simple English-like formal conversation modelling language, to add rails to a LLM.
NeMo Guardrails, a toolkit for developing trustworthy LLM-based conversational applications. It works with all #LLMs and natively supports @langchain, adding a layer of safety, security, and topical guardrails to existing applications. - @NVIDIAAI nvda.ws/3NfoDrC
ALT Available Now: NVIDIA NeMo Guardrails for LLMs
Traian Rebedea retweeted
NeMo Guardrails, a toolkit for developing trustworthy LLM-based conversational applications. It works with all #LLMs and natively supports @langchain, adding a layer of safety, security, and topical guardrails to existing applications. - @NVIDIAAI nvda.ws/3NfoDrC
ALT Available Now: NVIDIA NeMo Guardrails for LLMs
Traian Rebedea retweeted
One sign that determination matters more than talent: there are lots of talented people who never achieve anything, but not that many determined people who don't.
Traian Rebedea retweeted
🎉🇷🇴📢The first *in-person* edition of Romanian AI Days will happen in Oradea, 12-13 October 2022 days.airomania.eu/. Amazing speakers, panels, poster session. Get your ticket!
@andreeadeac22 Cornelia Caragea David Crater @dianaborsa Lucian Busoniu @PetarV_93 Tudor Badea
Traian Rebedea retweeted
Diffusion models like #DALLE and #StableDiffusion are state of the art for image generation, yet our understanding of them is in its infancy. This thread introduces the basics of how diffusion models work, how we understand them, and why I think this understanding is broken.🧵
