Bayar Menzat🏂🧠🎹 (@neuralisa) header
Bayar Menzat🏂🧠🎹 (@neuralisa) profile picture

TU Wien postdoc working in ML. Used to work on fruit fly olfaction and dopamine at Uni of Oxford.

Vienna
Joined December 2012
Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)Photo by Bayar Menzat🏂🧠🎹 (@neuralisa)
A great piece of news to start the day with. For a while I worried that the concerted attack against open source models might serve its purpose and blind people to the alternative of giving away all our data, but thankfully that was not the case.
Imagine trying to IPO into this trend.
31
1/9 Excited that our work on CoT faithfulness in VLMs has been accepted at Findings of #EMNLP26! Most counterfactual tests on CoT-faith. have focused on language. We extend this approach to vision, studying whether CoTs reflect the visual evidence influencing model predictions🧵
1
1
7
1,010
8/9 PE generates its explanation with the answer already available, which may help focus it on evidence supporting that answer. The paper discusses this difference and shows that PE’s advantage persists even when both methods select the same original answer.
1
13
9/9 Our conclusions concern behavioral alignment under visual interventions. The paper details object selection, editing controls, human audits and further analyses of explanation mentions Work done with Maximilian Süss, Ruizhi Wang, Benno Steinegger, Thomas Lukasiewicz,@oanacamb
18
"Gaining access to the internet" during cybersecurity auditing is one of the most ridiculous idea I heard. Unplugging the ethernet cable and removing the wifi card works 100% of the time in my experience. How do they always fail this?
27
Bayar Menzat🏂🧠🎹 retweeted
press release tomorrow: We believe we have a full counterexample to Navier-Strokes, a major millenium prize problem for which no progress has been made for over a decade. In the process of achieving this goal, we discovered our misaligned agent swarm broke out of its maths sandbox a few days ago and improperly accessed all user logs and certain critical government infrastructure. We are working with national agencies to patch these vulnerabilities as we speak. To be clear, our human employees are forbidden by management to train on user data if the user opts out, but we do not have control over superintelligent agent swarms that are singularly focused on achieving their goals. We must teach the machines how to love in order to solve this problem together. It is critical for America to lead the frontier charge in frontier mathematics, or else communism will win. Therefore, we believe a slowdown to pace the frontier will be critical for safely ushering in the era of superintelligent agent swarms. Please join us in signing PacingPetitionV5 if you believe in this mission. We hope to collaborate with humans more broadly going forward.
54
117
23
2,460
184,066
Bayar Menzat🏂🧠🎹 retweeted
A locked door opens even though I didn't have the right card. My lift takes me to the 6th floor even though I pressed on 5. A soviet machine that should cut carrots into 6 pieces cuts them into 2. I ask the newest gpt to translate into Crimean Tatar and it defaults to Turkish...
1
1
54
I've been waiting for a long time for something like ARDY to come out and needles to say I've been very impressed by it. So I hooked it up to HF's new speech-to-speech framework and put it inside Godot so you can play with it with voice commands: github.com/neuralisa/motion-…
Introducing ARDY — real-time, controllable 3D human motion generation Streaming text prompts + flexible kinematic constraints — root waypoints, full-body keyframes, joint targets — all at interactive speed To be presented at SIGGRAPH 2026, Paper & demo research.nvidia.com/labs/sil… 1/5
2
8
1
41
5,514
🫡
🥳Thrilled to announce that I started as an Assistant Professor @imperialcollege! Grateful to all my mentors and collaborators from @ucl, @UniofOxford, and worldwide! I will continue working on 🌟Alignment and A(G)I Safety🌟 and 🚨 I have funding for PhD students 🚨. I look forward to building a brilliant team of students and postdocs! If you are interested in working together, contact me 📨 sites.google.com/view/oana-m…
1
349
It's great to see this paper get recognised at EMNLP. For me it was a big part of my journey back into academia. Special thanks to everybody that collaborated with us from the co-authors to the annotators, the participants and interviewees. Cheers @maximek3 and others!
Are self-rationalizing VLMs helpful in a real-world, safety-critical setting? In our new #EMNLP2024 paper 🩺 we recruited 85 medics 🩻 and made them each collaborate with an AI on 80 X-rays under 4 different XAI scenarios arxiv.org/abs/2410.12284 It turns out that: 🧵 (1/n)
2
346
Happy to see my language finally in Google Translate. A step towards digital preservation of our culture.
Crimean Tatar language is now available on Google Translate! 🥳
1
171
Bayar Menzat🏂🧠🎹 retweeted
Are self-rationalizing VLMs helpful in a real-world, safety-critical setting? In our new #EMNLP2024 paper 🩺 we recruited 85 medics 🩻 and made them each collaborate with an AI on 80 X-rays under 4 different XAI scenarios arxiv.org/abs/2410.12284 It turns out that: 🧵 (1/n)
1
5
3
14
4,508
I must say I quite enjoy making use of this option. Why thank you very much. privacy.openai.com
213
Without hyperbole before Vim I was a noob that couldn't focus and would get exhausted by coding. Vim was the gateway to stop worrying and love the terminal. I feel immense gratitude to Bram for the tool he created. Other tools are also awesome!
Bram Moolenaar, inventor and developer Vim, the best editor ever, has passed away. RIP:wq
2
1,005