The Hidden Infrastructure Challenge Behind Every AI-Generated Avatar

Toward More Controllable AI Video Editing: An Early Research Exploration at Netflix

Netflix giới thiệu hai mô hình chỉnh sửa video AI giai đoạn đầu là Vera và VOID. Vera sử dụng mô hình diffusion phân lớp, chỉ tái tạo vùng chỉnh sửa (kèm alpha matte) thay vì toàn bộ clip, bảo toàn nội dung chưa chỉnh sửa. VOID chuyên xóa vật thể trong video với kỹ thuật inpainting hợp lý vật lý, tái tạo cảnh thực tế khi vật thể bị loại bỏ. Cả hai mô hình đều vượt trội so với các phương pháp hiện có trong nghiên cứu.

Lập trình viên muốn phát triển các giải pháp AI tiên tiến trong xử lý video nên tham khảo để hiểu cách thiết kế mô hình hiệu quả như Vera và VOID, từ kiến trúc đặc biệt đến kỹ thuật điều khiển chi tiết để nâng cao chất lượng và tính khả thi của các ứng dụng AI video trong tương lai.

New EU rules put a driver-watching camera in every car

From 7 July 2026, the EU's General Safety Regulation phase two mandates advanced driver-assistance systems on all newly manufactured cars and vans. Key requirements include pedestrian- and cyclist-detecting emergency braking and a cabin-facing camera that monitors driver attention and issues distraction warnings. The regulation builds on the 2024 first phase (which introduced intelligent speed assistance) and supports the EU's Vision Zero goal of near-zero road deaths by 2050. The driver-facing camera is the most debated element, valued by safety researchers but raising privacy concerns. The rules also accelerate the broader shift toward software-defined vehicles packed with sensors and compute.

AI brings object-level vision prosthetics closer to reality

Researchers at EPFL's NeuroAI Lab have developed AI-based topographic neural network models that predict optimal brain stimulation patterns to evoke perception of complex visual objects — such as faces and houses — rather than just simple light flashes. The models were validated in live trials on sighted monkeys in Amsterdam, showing that model-guided cortical stimulation can bias visual object perception in predictable ways. While the team cannot yet create object perception from scratch (without any visual input), this is the stated next goal. The approach could also extend to improving cochlear implants for auditory prosthetics.

Building a Multimodal Chatbot with Qwen3-VL Instruct and Thinking Models

Step-by-step guide to building a multimodal chatbot using Qwen3-VL Instruct and Thinking models. Covers the model's architecture (SigLIP2 vision encoder, TM-RoPE positional encoding), training pipeline, and benchmark performance. The implementation uses Hugging Face Transformers pipelines with the 'image-text-to-text' task type and builds a Gradio web interface supporting image uploads alongside text. Both Instruct Mode (fast, direct answers) and Thinking Mode (step-by-step visual reasoning with visible traces) are demonstrated with practical code examples and real chatbot outputs.

Best Object Detection Models for Machine Learning in 2026

A comprehensive comparison of the top object detection models in 2026: RF-DETR (by Roboflow), YOLO12 (Tsinghua University), and YOLO26 (Ultralytics). Covers architectures (CNN-based vs. transformer-based), key metrics (mAP50:95, FPS, parameter count), benchmark results on MS COCO, licensing implications, and use-case guidance. RF-DETR leads in accuracy (60.1 mAP on COCO, a record), YOLO26 excels on edge/mobile hardware with 43% faster CPU inference than YOLO11, and YOLO12 introduces attention-centric design with strong nano-scale performance. Includes quick-start code snippets and fine-tuning guidance for each model.

I tested Gemma 4, Qwen 3.5, and Ministral 3 for vision tasks, and only one understood the assignment

A hands-on comparison of three local vision-capable models — Gemma 4 E4B, Qwen 3.5 9B, and Ministral 3 3B — tested on identical tasks: a Docker error screenshot, a busy UI with a counting task, and a low-quality photo of a medicine bottle. Qwen 3.5 9B came out clearly ahead, correctly reading fine print, counting models, identifying active ingredients, and even suggesting UI interactions. Gemma 4 E4B handled simpler tasks but got lossy on busy screens. Ministral 3 3B performed well on the debugging screenshot but hallucinated on more complex visual tasks. The takeaway: vision capability varies significantly among local models, and only Qwen 3.5 9B consistently understood what it was looking at.

Fleek raises $25m to build the AI infrastructure behind global secondhand fashion

London startup Fleek has raised $25m in Series B funding, bringing total funding to $45m, to scale its AI-powered infrastructure for the global secondhand clothing trade. The round was led by Burda Principal Investments, with participation from eBay, FJ Labs, Andreessen Horowitz, and Y Combinator. At the core is Fleek Sort, a custom vision-language model trained on millions of secondhand transactions that automates garment identification, grading, and pricing. The platform connects over 2,000 wholesale suppliers with 50,000+ retailers across 100+ countries, and has helped keep 12 million items in circulation. The funding will expand the AI-native marketplace, engineering teams, and supplier/buyer networks.