Brian Jalaian

🎥 New 3-min explainer: Visual Reasoning Agent (VRA) = training-free agentic AI for vision systems 💡 Up to +40% accuracy on tough remote sensing tasks — no retraining, just smart reasoning. Think. Critique. Act. 📽️ Watch how VRA boosts robustness👇 #VLM #AgenticAI

@WhiteHouse Thank you President Trump

From coast to coast: 'THANK YOU, President Trump, for decisive action against the Iranian regime.' 🇺🇸

@danvahdat @elonmusk Thank you @elonmusk

Starlink is now free nationwide across Iran. Thank you, @elonmusk. We the people of Iran will never forget your kindness. ❤️🙏🏻

🚀 Want to train ML models with free GPUs in the cloud? In just 2 mins, David Moe from Jalaian AI Lab shows how we teach: ✅ Google Colab + Kaggle ✅ AI-assisted coding (Gemini) ✅ Easy submission.csv for competitions 🎥 Watch full video: youtube.com/watch?v=S5cijA… #AICoding

“Don’t fear uncertainty — that’s where the magic happens.” Shared a 3-min clip from a convo with Arman on books, mindset & the startup journey. ▶️ Watch:youtube.com/watch?v=siwQIE… #entrepreneurship #growthmindset #startuplife #booktalk

Year 3 of our Machine Learning in Data Science course is here—with a new focus on AI tools like Copilot 🤖⚡ Excited to explore how these tools can boost learning and creativity. Big shoutout to the research team pushing boundaries! #MachineLearning #AI #DataScience

We trained ML models to detect cognitive load using wearable biosignal data. 👨‍⚕️ Wristbands + EEG → Features + Spectrograms → Transfer Learning Here’s a 3-min clip breaking it all down: ▶️ youtu.be/z8jsEZ9_x2U #DeepLearning #Neurotech #Biosignals #TechForGood

🧠 Can AI detect your mental effort using wearables? We tested ML, DNNs & transfer learning on biosignal data (EEG, PPG, BVP) to decode cognitive states. 📺 Full video: youtube.com/watch?v=z8jsEZ… #AI #DeepLearning #Wearables #CognitiveLoad #MachineLearning

@sama @Sama 🔥 Practicing o3 3+ hrs/day—what learning path works best? Structured tutorials vs OSS deep dives? 📚 vs 🛠️ balance? Tips on feedback loops or AI/human mentors? How do YOU use AI to level up your #AIresearch game? #skillsmax #o3

if you are not skillsmaxxing with o3 at minimum 3 hours every day, ngmi

🌟 Hydra shows that structured, agentic reasoning can dramatically improve VLM reliability efficiently. We’re excited about future work combining Hydra with even larger LLMs and edge deployment! 📄 arxiv.org/abs/2504.14395 #AI #LLMs #RobustAI #VisionLanguageModels

🚀 Excited to share our new research! We present Hydra, boosting adversarial robustness and reducing hallucinations in VLMs. Hydra enhances reliability through iterative reasoning—no extra training needed. 📄 arxiv.org/abs/2504.14395 #AI #VLM #Robustness #ML #Research

🧪 Results: Hydra outperforms existing plug-in methods across benchmarks like ScienceQA, GQA, and TextVQA. It’s particularly effective against unseen adversarial examples — even without explicit defenses.

🏋️‍♂️ Why is this important? Stronger resilience against adversarial attacks Reduced hallucination rates Better factual accuracy in open-world settings Works without expensive retraining or fine-tuning!

It uses Chain-of-Thought (CoT) and In-Context Learning (ICL) techniques to generate, critique, and refine answers across multiple reasoning passes. Hydra acts like an "agent," adapting based on feedback.

Vision-Language Models are incredibly powerful but vulnerable to adversarial attacks and hallucinations. Hydra tackles both issues by introducing an iterative reasoning framework — no retraining needed.

🚨 Join the AAAI FS 2024 - ATRACC for expert talks on AI Validation, Risk, & Trustworthiness! 🗓️ Nov 7-9, 2024 📍 Arlington, VA 🔑 Hotel Deadline: Oct 17 – Limited rooms! 🏨 More info: buff.ly/4eVwTHY #AAAI #ATRACC2024 #AITrustworthiness #AIConference 🚨

Yes, many of us have always been working on "AGI", whatever you mean by that. Perceptron, General Problem Solver, expert systems, machine learning, backprop, RL, SSL, transformers, LLMs...: For some, these were going to be "The One Weird Trick" that was gonna take us to AGI. 1/

Extraordinary new paper from Google on medicine & AI: When Google tuned a AI chatbot to answer common medical questions, doctors judged 92.6% of its answers right … compared to 92.9% of answers given by other doctors. And look at the pace of improvement! arxiv.org/pdf/2212.13138…

Neural nets that analyze/generate text, speech, image, proteins don't tell us much about the structure thereof. But 1. they are very useful 2. they tell us that SGD is better at structure discovery than humans 3. they upset people who devoted their career to structure discovery

We’re pleased to introduce Make-A-Video, our latest in #GenerativeAI research! With just a few words, this state-of-the-art AI system generates high-quality videos from text prompts. Have an idea you want to see? Reply w/ your prompt using #MetaAI and we’ll share more results.