A weekly curated collection of robotics videos from IEEE Spectrum, featuring highlights such as a disembodied robotic hand crawling on three fingers from Tangent Robotics, MIT CSAIL's SoftMimic approach for safer human-robot interaction, Agility Robotics' Digit navigating obstacles, an aggressive firefighting robot from DEEP Robotics, Unitree R1 humanoid performance, event cameras handling motion, and research on using AI-driven robots to understand whale communication. Upcoming robotics events including RSS 2026, IROS 2026, and Actuate 2026 are also listed.
Nguồn: https://spectrum.ieee.org/video-friday-robot-grippers. 8sync News chỉ tóm tắt và dẫn link; bản quyền nội dung thuộc tác giả và nguồn gốc.
Google Cloud vừa giới thiệu TPU Developer Hub, một nền tảng giáo dục tập trung dành cho nhà phát triển ML sử dụng TPU, bao gồm kiến trúc phần cứng, stack phần mềm (XLA, Pallas kernels), công cụ gỡ lỗi XProf, chiến lược tối ưu hóa (như offloading KV cache) cùng networking và bảo mật. Nội dung đa dạng từ Colabs tương tác, mã nguồn mở đến tài liệu chuyên sâu, hỗ trợ tích hợp AI-assisted development.
Lập trình viên ML nên đọc để hiểu cách tối ưu hóa hiệu suất và chi phí của mô hình trên TPU với các công cụ mới như XLA, Pallas và các chiến lược parallelism, từ đó tiết kiệm thời gian và nguồn lực trong triển khai sản phẩm AI.
Firefly has launched the AIBOX-9075, an industrial Edge AI box powered by the Qualcomm DragonWing IQ-9075 SoC. It delivers up to 200 TOPS of AI performance via a Hexagon Tensor Processor NPU, paired with 36GB LPDDR5 ECC memory and 128GB UFS 2.2 storage. The device targets edge AI workloads including private LLM deployment, robotics, and computer vision. It supports popular AI frameworks (TensorFlow, PyTorch, ONNX) and LLMs such as DeepSeek-R1, Llama, and Gemma, achieving up to 22 tokens/s with Llama2-7B. Connectivity includes dual 2.5GbE with TSN, optional Wi-Fi 6, 4G/5G, and 8x GMSL2 camera inputs. The rugged aluminum enclosure operates from -40°C to 85°C. It runs Ubuntu and Yocto Linux and is priced at $1,239 from the Firefly store.
Researchers at Princeton have developed AI-driven methods to automate the design of radio-frequency integrated circuits (RFICs), traditionally a highly manual and expert-driven process. Using reinforcement learning and inverse design with a convolutional neural network emulator, the system can generate novel circuit architectures from scratch—without relying on human-designed templates—achieving record performance in 5G millimeter-wave power amplifiers. A diffusion model was also introduced to make AI-generated electromagnetic structures more interpretable for engineers. The approach reduces design time from months to minutes, though challenges remain around hallucinations, verification, and the need for open datasets to train more generalizable foundational models.
A practical framework for choosing between TPUs and GPUs for AI/ML workloads, covering silicon architecture differences, use-case fit, and total cost of ownership. TPUs excel at large-scale JAX-based pretraining (100B+ params) on GCP with committed-use discounts, but their static shape requirements, GCP-only availability, and smaller ecosystem make GPUs the default for most teams. GPUs dominate due to PyTorch/CUDA ecosystem maturity, dynamic shape support, multi-cloud portability, and viable spot automation. The post also covers GPU cost optimization strategies including rightsizing via DCGM, spot instance automation, MIG partitioning, and inference density improvements, with Cast AI promoted as a solution for automating these optimizations.
The University of Maryland is funding a research project combining quantum computing and machine learning to accelerate the discovery of single-atom catalysts for cancer detection and treatment. Part of the university's Grand Challenges Grants Program, the project brings together engineers and computer scientists to build a predictive framework that models complex atomic and chemical behaviors — tasks difficult for classical computers. Quantum simulations would generate reliable databases of electronic structures and catalytic pathways, which machine learning models would then search to identify promising catalyst configurations. The team also plans to release benchmark datasets and reproducible computational tools to support open science. The research is preclinical and focused on discovery, not immediate clinical application.
Behavioural interviews are often the deciding factor for levelling in data science and ML roles, yet most candidates neglect them. The post outlines a preparation strategy built around three pillars: building a 'story vault' of 2–3 deep, reusable project examples; researching the target company's culture and value principles and mapping stories to each; and using the R-STAR-L framework — an extension of the classic STAR method that adds a Repeat step (to confirm understanding and buy thinking time) and a Link Back step (to explicitly tie your answer to the company's stated values). A detailed worked example for a DoorDash data scientist interview illustrates how to naturally weave in a culture-fit signal without sounding scripted.

When a RAG evaluation set is repeatedly used to identify failures and tune the system, it quietly becomes a training set — a form of overfitting. The post explains how this happens through prompt tuning on the same test questions, cherry-picking easy examples, and writing questions derived from already-indexed documents. The fix mirrors classical ML discipline: maintain a genuinely held-out test set, build questions independently of system behavior, and treat suspiciously high scores with skepticism. The broader pattern is framed through Goodhart's Law — when a measure becomes a target, it stops being a good measure.
Meta's engineering team describes a hybrid pattern for privacy-aware infrastructure (PAI) asset classification at scale. The core approach combines LLMs for handling ambiguous or novel data assets with deterministic, versioned rules for routine enforcement. Key principles include: building structured 'evidence briefs' from distributed context (code lineage, ownership, semantic annotations) rather than raw fields; keeping human-reviewed labels separate from model-generated recommendations; using a multi-panel LLM judge with Cohen's kappa for quality control; and progressively distilling stable LLM-discovered patterns into auditable deterministic rules. In production, ~85% of traffic is resolved by fast deterministic rules, with LLM fallback for the remaining ~15%. The system includes self-regulation mechanisms (tuning controller states) to prevent runaway optimization loops. Lessons learned emphasize that context quality beats prompt quality, accuracy alone is insufficient for imbalanced taxonomies, and distillation into deterministic rules is the sustainable production model.