All Updates
Page 1249 of 1417
March 2, 2026
Deepfake Attack Hits BSE Boss
The boss of the Bombay Stock Exchange was targeted in a recent deepfake attack. This incident could have cheated many people. It exemplifies a growing global deepfake problem.
He Xiaopeng Pushes L2-to-L4 AV Policy Leap
XPeng founder He Xiaopeng proposed accelerating autonomous driving regulations from L2 to L4 for the 2026 Two Sessions, viewing L3 as transitional. XPeng showcased 2nd Gen VLA Robotaxi tests in unmapped factories, secured Guangzhou L4 road test permits, and nears production. This aligns with XPeng's physical AI strategy for AV leadership.
Xiaomi Robot Meets Auto Factory Screw-Driving Pace
Xiaomi robot starts 'internship' in auto factory die-casting workshop. It autonomously operated for 3 hours at self-tapping nut station with 90.2% dual-side success rate. Performance matches the line's 76-second takt time requirement.
Apple Launches iPhone 17e and M5 Macs
Apple's March event previews iPhone 17e with A19 chip at $599, A18 Pro MacBook outperforming M1, and M5 Pro/Max MacBook Pros. iPad lineup gets A18/M4 upgrades for Apple Intelligence support. Retail prep signals strong demand like iPhone launches.
Lenovo's Desk Robot Arm & AI Scheduler at MWC
Lenovo unveiled two desktop AI concept devices at MWC 2026 for office scenarios. The 'Gongyou' robotic arm and schedule manager all-in-one provide local AI-powered digital companionship. They aim for efficient productivity with a dystopian edge.
Xiaomi Humanoid Runs 3 Hours in Factory
Xiaomiโs humanoid robot completed a 3-hour autonomous operation in a real auto factory. It achieved a 90.2% success rate on the self-tapping nut workstation. This trial highlights advances in industrial humanoid applications.
Project Helix Speeds Zero Trust Onboarding
Project Helix simplifies Cloudflare One onboarding. It leverages automation and Terraform templates for rapid deployment of best-practice configurations in minutes.
Cloudflare Unveils Programmable SASE
Cloudflare offers the only SASE platform with a native developer stack. It enables building custom real-time security logic and integrations at the edge.
Cloudflare One Launches Agile SASE
Cloudflare One positions agile SASE as the key to modernization in 2026. It provides a single programmable connectivity cloud that secures humans, devices, and AI agents.
Kimi K2.5 Hits Yearly Revenue in 20 Days
Moonshot AI's Kimi K2.5 model generates more revenue in 20 days than all of last year, topping OpenClaw benchmarks. Agent economy emerges as AI agents autonomously handle workflows. Vertical industries lag but hold unicorn potential.
Juewu Tech Raises 100M+ RMB Series A
Juewu Tech secured over 100M RMB in Series A funding led by Huachuang Capital. Its modular all-terrain robots have scaled in Chinese agriculture after 1M+ km tests. Expansion targets overseas markets and urban emergency scenarios.
OpenAI Acquires GPT.com Domain
OpenAI has reportedly acquired the GPT.com domain, which now redirects visitors to ChatGPT.com. This follows their acquisition of Chat.com last year, which also points to ChatGPT. GPT stands for Generative Pre-trained Transformer.
Huawei Launches 50Gbps AI Backhaul Tech
Huawei introduced a U6GHz E-band long-distance full-duplex microwave solution to manage explosive traffic from AI wearables and robots. It targets 6G ITU-R KPIs like 50Mbps uplink and 300Mbps downlink edge rates. Current backhaul networks struggle with these demands.
Embodied AI Hurdles at Heterogeneous Computing
Hot money is chasing embodied intelligence amid high industry hype. However, deployment faces major pain points including accuracy, reliability, and data silos. These bottlenecks hinder transition from demos to practical use, with heterogeneous computing as a key challenge.
Qwen 3.5 Needs bf16 KV Cache
Qwen 3.5 models like 35B A3B require bf16 KV cache in llama.cpp for accurate perplexity, not default f16. Tests show bf16 PPL at 6.5497 vs 6.5511 for f16/f32. vLLM defaults to bf16 correctly, but llama.cpp needs manual flags.
SleepLM: Language Models for Sleep
SleepLM introduces foundation models aligning multimodal polysomnography with natural language for sleep analysis. It curates the first large-scale sleep-text dataset from over 100K hours of data across 10K+ individuals. The models outperform SOTA in zero/few-shot tasks, retrieval, captioning, and enable novel capabilities like language-guided event localization.
Silicon Valley Faces AI Anxiety
Interviews with Silicon Valley tech company employees explore AI-induced anxiety amid layoffs and fears of programmers being replaced by AI. Workers discuss continuing to hand-code and their views on future changes. Practical advice is offered to ordinary people navigating the AI era.
RL-CMSA Masters Min-Max mTSP
RL-CMSA is a hybrid reinforcement learning method for the symmetric single-depot min-max multiple traveling salesman problem (mTSP). It constructs diverse solutions via probabilistic clustering with learned q-values, merges them, solves a restricted MILP, and refines via local search. It outperforms state-of-the-art genetic algorithms on random and TSPLIB instances, especially for larger sizes and more salesmen.
PseudoAct: Pseudocode for Smarter LLM Agents
PseudoAct is a new framework that uses LLM-generated pseudocode to create structured plans for agents, encoding control flows like loops, conditionals, and parallelism. It overcomes ReAct's limitations in long-horizon tasks by reducing redundant actions and ensuring coherent execution. Experiments show 20.93% success gain on FEVER and new SOTA on HotpotQA.
ProductResearch Boosts E-Commerce Agents via Synthetic Trajectories
ProductResearch introduces a multi-agent framework to synthesize high-fidelity trajectories for training e-commerce shopping agents. It uses User, Supervisor, and Research Agents to generate comprehensive product reports from nuanced intents. Fine-tuned compact MoE models rival proprietary systems in depth and utility.