You can try the following two prompts in the Thinking mode (via web/app) to get a better model experience in certain domains like counting (Note: keep a line break after the bracketed titles):: [Think with Grounding] ....... [Think with Pointing] ...... These two prompts encourage the model to adopt bounding boxes or points (which are classic fundamentals in computer vision) in its thought process. Personally, I love the pointing approach for solving abstract topological/reasoning tasks. Using points to represent continuous trajectories makes the MLLM's reasoning process feel much more human-like. Speaking purely from my personal exploration: Getting a multimodal model to accurately represent continuous trajectories with points is still a highly challenging frontier task for the entire industry. The current performance on real-world scenarios still has a long way to go.

Absolutely mind-blowing work. 🤯 This reminds me of when we were exploring the concept of 'thinking with visual primitives'. We noticed the model excelled in synthetic scenarios, but real-world generalization for complex tasks still faced bottlenecks. Back then, I kept thinking: "If only someone could provide manual annotations for these real-world scenarios..." Using point-form visual primitives for reasoning would be powerful for tasks like state tracking, spatial topology, and logical testing, but they may heavily rely on high-fidelity, real-world labeled data. I think it’s the same across other domains too. The journey toward AGI cannot skip the hard, grounded work of human-labelled data. Massive respect to this effort! 🫡

With the recent discussions around Gemini 3 Deep Think, I’m revisiting my logs from the Gemini3-Pro-Preview version a few months back. One thing stood out beyond standard benchmarks: its ability to handle scientific long-tail tasks. Specifically, I fed it an image of a complex molecular structure (shown below), and it nailed the formula correctly. It’s an interesting case study on how scientific vision, massive compute, and top-tier research talent converge. Respect to the team. 🫡 #AI# #Gemini# #Science#

Seeing InternVideo-Next explicitly characterize its architecture as "Encoder–Predictor–Decoder (EPD)" and defining the Predictor as a "latent world model" confirms a clear convergence in the field. 📉 From our early work on Context AutoEncoder (CAE, — where we decoupled the Encoder, Predictor ("Regressor" in the paper), and Decoder — to Meta's I-JEPA/V-JEPA shifting entirely to latent prediction, and now this. It seems we are all validating the same core intuition: Understanding ≠ Pixel Reconstruction. 💡 Decoupling representation learning from high-frequency detail generation to build World Models in latent space is evidently the path forward. Glad to see our early intuition resonating with the latest SOTA. 🫡 #AI# #ComputerVision# #WorldModel# #JEPA# #CAE#

Looking at today's LLM / Multimodal LLM benchmarks is like looking at an elite student's exam scores: a high score doesn't guarantee real-world capability. The focus must shift from leaderboards to usability, reliability, and seamless workflow integration.

Introducing DeepSeek-V3.1: our first step toward the agent era! 🚀 🧠 Hybrid inference: Think & Non-Think — one model, two modes ⚡️ Faster thinking: DeepSeek-V3.1-Think reaches answers in less time vs. DeepSeek-R1-0528 🛠️ Stronger agent skills: Post-training boosts tool use and multi-step agent tasks Try it now — toggle Think/Non-Think via the "DeepThink" button: 1/5