It’s tool use (PIL/cv2, BFS) debugged with multimodal perception of tool output within the CoT; it’s not 4o-style multimodal output. See the share link in the thread for more info on the tool calls.
MACHINE LEARNING
-
TPUs excel at inference for mixture of experts transformer models
By
–
In general, we've found that TPUs are quite effective at inference, especially for mixture of expert models and handling attention for transformer models with long context. The high speed multidimensional torus interconnects are quite fast and effective. I'm not sure I have a
-
InstructPipe: Visual Programming Framework Simplifies ML Pipeline Design
By
–
Visual programming frameworks like Visual Blocks can speed up ML pipeline building, but novice users can still struggle with setting them up. Read how InstructPipe, a research prototype for ML pipeline design, helps users build workflows more easily →https://t.co/Lbxv98G7Qy pic.twitter.com/3abJEv7AN7
— Google AI (@GoogleAI) 18 avril 2025Visual programming frameworks like Visual Blocks can speed up ML pipeline building, but novice users can still struggle with setting them up. Read how InstructPipe, a research prototype for ML pipeline design, helps users build workflows more easily →
https://
goo.gle/4lB4pav -
DeepSeek MoE Multi-Node Deployment Improves Throughput Latency
By
–
Our AI team just published a blog detailing how we've achieved improved throughput and latency with multi-node deployment of DeepSeek MoE models. Read more:
-
Test-Time Scaling in Large Language Models: Survey
By
–
What, How, Where, and How Well? A Survey on Test-Time Scaling in Large Language Models
Paper: https://
arxiv.org/pdf/2503.24235
Repo: https://
github.com/testtimescalin
g/testtimescaling.github.io/
… -
Creating Self-Healing Code Agents with Reflection Steps
By
–
How to Create a Self-Healing Code Agent Code generation is one of the most common use cases we see with LLMs, but how do you improve the accuracy and performance of the code generated by your code agent? In this video, we demonstrate how you can create a reflection step in
-
Dynamical World Models and JEPA for AI Planning Since 2016
By
–
Exactly. I've been arguing for dynamical world models (for planning) since at least 2016 (e.g. in my NIPS 2016 keynote) and doing so using JEPA since 2021/2022 (before ChatGPT), e.g. in this 2022 paper: https://
openreview.net/forum?id=BZ5a1
r-kVsf
…
This has nothing to do with Llama4. -
When Do Random Forests Work? Research Analysis
By
–
When do Random Forests work? C. Revelas, O. Boldea, B. J. M. Werker: https://
arxiv.org/abs/2504.12860 #ArtificialIntelligence #DeepLearning #MachineLearning -
Debate: Can Scaling LLMs Achieve AGI?
By
–
On est toujours divisés en deux camps sur la définition de l’AGI, malgré toutes les sorties de modèles SOTA (State Of The Art) : Groupe 1. Les LLMs ne mèneront pas à l’AGI
Groupe 2. Faire monter les LLMs en puissance permettra d’atteindre l’AGI Alors, de quel côté êtes-vous ? -
InternVL3 dethrones Qwen 2.5 as new king of open VLMs
By
–
New king of open VLMs: InternVL3 takes Qwen 2.5's crown! -> InternVL have been a wildly successful series of model : and the latest iteration has just taken back their crown thanks to their superior, natively multimodal vision training pipeline. Most of the vision language