Llama-4-Maverick is CRAZY GOOD to power agents It's now the top open model on smolagents LLM leaderboard, beating the much larger DeepSeek-R1!
Congrats @ThomasScialom and team!
RESEARCH
-

Llama-4-Maverick tops smolagents leaderboard, beats DeepSeek-R1
By
–
-

Deep diving into AI: beyond surface-level hype
By
–
Vous scrollez ? Moi je creuse. (Pas pour trouver de l’or. Pour extraire ce qui compte vraiment dans l’IA.) Pendant que certains vous balancent « 10 prompts pour devenir CEO en 24h »… je lis. Je trie. Je relie les points. Et je vous en fais une newsletter lisible par des
-
AI Video Generation Struggles With Natural Physics Simulation
By
–
2. Motion naturalness
— Linus ✦ Ekenstam (@LinusEkenstam) 8 avril 2025
The cheese hovers in mid-air rather than falling naturally to the ground. pic.twitter.com/jexCwbDxDo2. Motion naturalness The cheese hovers in mid-air rather than falling naturally to the ground.
-

AI Scientist-v2: Agentic Tree Search for Automated Scientific Discovery
By
–
Excited to release our technical report: “The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search” https://
pub.sakana.ai/ai-scientist-v
2/paper/
… The AI Scientist-v2 incorporates an “Agentic Tree Search” approach into the workflow, enabling deeper and more -
MedSAM2: AI Model for 3D Medical Image Segmentation
By
–
Harvard and University of Toronto researchers dropped an AI for 3D medical image & video segmentation
— Rowan Cheung (@rowancheung) 8 avril 2025
Built atop Segment Anything Model 2.1, MedSAM2 generalizes across organs, modalities, and pathologies with an 85%+ reduction in annotation costspic.twitter.com/ZW66mTldfjHarvard and University of Toronto researchers dropped an AI for 3D medical image & video segmentation Built atop Segment Anything Model 2.1, MedSAM2 generalizes across organs, modalities, and pathologies with an 85%+ reduction in annotation costs
-

UC Berkeley Stanford researchers use Test Time Training for video generation
By
–
Researchers from UC Berkeley and Stanford tapped Test Time Training to generate one-minute videos
— Rowan Cheung (@rowancheung) 8 avril 2025
They added TTT layers to a pre-trained Transformer and fine-tuned it to generate cartoons with strong temporal consistency
Here's a single-shot example:pic.twitter.com/YCqNTHSvpuResearchers from UC Berkeley and Stanford tapped Test Time Training to generate one-minute videos They added TTT layers to a pre-trained Transformer and fine-tuned it to generate cartoons with strong temporal consistency Here's a single-shot example:
-

Google Researcher Wins IEEE Speech Audio Processing Award
By
–
Congratulations to Google Visiting Faculty Researcher Shrikanth Narayana for winning the 2025 IEEE James L. Flanagan Speech and Audio Processing Award! This award recognizes contributions to speech communication science and technologies for inclusive human-centered engineering.
-
Stanford HAI releases 2025 AI Index annual report
By
–
Drum roll 🥁 @StanfordHAI ‘s annual #AIIndex2025 is out! This is one of the most trusted annual reports on the state of AI globally. Everyone interested in AI should read this! https://t.co/0AOphJBola
— Fei-Fei Li (@drfeifei) 8 avril 2025Drum roll @StanfordHAI ‘s annual #AIIndex2025 is out! This is one of the most trusted annual reports on the state of AI globally. Everyone interested in AI should read this!
-

Audio Algorithms Implementation on Edge Devices
By
–
Stop by the #ICASSP2025 Google booth today at 11:30 AM IST where Arpit Jain will present "Considerations for the Implementation of Audio Algorithms on Edge Devices", covering algorithmic computational complexity, memory use, & DSP/NPU architecture for algorithm implementation.
-

Creator Pauses Llama 4 Analysis Amid Conflicting Reports
By
–
Been a busy few months. Gonna check out for a bit. I made a whole video about Llama 4 and decided not to release it as more details come out around what was announced vs what results people are seeing. Will revisit it after the story’s finished unfolding. For now, gonna jump