All Updates
Page 370 of 891
March 28, 2026
AI Boom Drives CPU Memory Price Hikes
AI demand surges prices across electronics: AMD Ryzen 9 9950X3D2, Intel/AMD CPUs up 15%, Qualcomm Snapdragon 8 Elite Gen 6 by 30-50%. ARM launches self-made AGI CPU for AI data centers on 3nm. Google TurboQuant algorithm hints at future memory relief but training demand persists.
Augmentations as Invariance Assumptions
This Reddit post argues that data augmentation in machine learning is often used heuristically, proposing instead to frame each transform as an invariance assumption. It discusses challenges like task-specific validity, optimal strength, and risks of corrupting training signals, with examples from computer vision. The author invites community experiences on validation and best practices.
AI Prototypes Kill Lengthy PRDs
AI tools like Claude Code enable 2-hour prototypes, shifting PM work from docs to validated assumptions and user tests. Anthropic and Zapier teams use sprints, demos over roadmaps/docs. Focus now on hypothesis testing, iteration judgment, and quick MVP building.
Fine-Tune LLMs for Storyboard Scripts
General LLMs fail at storyboard scripts due to weak lens rhythm, vague visuals, and unstable formats. Author fine-tunes open-source models like Qwen using custom datasets for better cinematic output suitable for AI image tools. Covers model selection, data creation, and pitfalls.
7 Flaws in US AI Great Divergence Narrative
US CEA report analogies AI revolution to industrial 'Great Divergence,' urging dominance. Critique exposes weak history use, overoptimistic GDP predictions (vs Acemoglu's 1-1.6%), fast tech diffusion like DeepSeek R1. Highlights China's open-source AI resilience despite chip bans.
Caocao Mobility Profits Q4 2025, Accelerates Robotaxi
Caocao Mobility achieved its first adjusted Q4 profitability in 2025. Custom vehicle moats are driving a full-speed Robotaxi transformation. The company is entering a profitability growth trajectory.
Google Paper Tanks Memory Chip Stocks
Google's new research paper triggered a collective plunge in memory chip vendors' stock prices. The exact content and mechanism remain unclear from available details. Likely tied to AI efficiency advances reducing memory demand.
NeurIPS Apologizes After China AI Boycott
Chinese societies like CCS and CAAI urged boycott of NeurIPS over submission ban on US-sanctioned entities. NeurIPS issued a public apology on March 27 via social media for improper guidelines. This follows collective 'blacklisting' by Chinese academia.
Huang Shifts Nvidia: Chips to Tokens
Nvidia CEO Jensen Huang made intensive appearances over three months. He is redefining Nvidia by shifting from chip sales to token sales.
Moonshot AI Pivots Away from OpenAI
Moonshot AI stops waiting for OpenAI and executes a strategic shift now. The change is metaphorically called a rock musician's 'breakup'.
Robot Open-Source Revolution Unfolds
Xiaomi, Ant, Nvidia and others release open-source robot VLA models challenging closed giants like Tesla. Four factions: academia (OpenVLA, Octo), giants (GR00T N1.6, Gemini), China startups, precision ฯโ. OpenVLA's 7B model beats Google's 55B RT-2-X by 16.5%.
China Releases AI Speech and Corpus Standards
Ministry of Education and National Language Commission released two AI-focused language norms: machine-synthesized Mandarin evaluation standard and AI corpus basic terms. Developed to address AI-era language needs. Now published by Yuwen Press.
XAG Ag Robots Profit & File HKEX IPO
XAG, an agricultural robotics firm, filed for HKEX IPO on March 27 after turning profitable. Revenues grew to RMB 11.66B in 2025 with net profit of RMB 124M, driven by scaling drone sales and high-margin overseas markets. Drones comprise 87.6% of revenue, highlighting product concentration risks.
Claude Strongest Model Codenamed Capybara
A leak in a Chinese tech news brief reveals Anthropic's upcoming strongest Claude AI model is codenamed 'Capybara'. This suggests ongoing major advancements in their large language model lineup. No release date or capability details were provided.
Kuaishou's AI Caution vs Douyin's Boldness
ByteDance excels at spinning new AI/video apps like CapCut, Hongguo, Jianmeng from Douyin ecosystem due to content-driven user flow. Kuaishou, relationship-focused, limits to Keling AI video gen as sole new dept amid listing constraints. Analysis argues Kuaishou shouldn't copy, seek Keling-main app synergy instead.
Caocao Chuxing's First Profitable Q4 Fuels Robotaxi Push
Caocao Chuxing achieved 2025 full-year revenue of 20.2 billion CNY, up 38% YoY, with its first-ever adjusted profitable Q4 driven by scale, AI efficiencies, and custom vehicles. The company deployed over 100 Robotaxis, launched Robotaxi 2.0, and plans 2026 expansion including L4 vehicles via Geely synergy. This marks a shift to self-sustaining growth in China's ride-hailing market.
Anthropic Tests Claude Operon for Biology
Anthropic is testing Claude Operon, a new desktop app mode tailored for biology and health research. It includes dedicated project tools and session management features. This aims to enhance scientific workflows using Claude.
DGX Sparks vs Mac Studio: 397B Model Tie
A user spent 96 hours setting up dual DGX Sparks and Mac Studio M3 Ultra for Qwen3.5-397B inference. Generation speeds tied at 27-29 tok/s, but Sparks excelled in prefill (730 tok/s vs 317), while Mac won embeddings (112 vs 76.6 sentences/s). Setup was far easier on Mac, taking 4 hours vs 4 days on Sparks.
PentaNet Beats BitNet with Pentanary Quantization
PentaNet is a new LLM architecture using native pentanary {-2,-1,0,1,2} quantization, achieving 6.4% better perplexity than BitNet on WikiText-103 while preserving zero-multiplier inference via bit-shifts. Weight distributions stabilize without collapsing to ternary, enabling more model capacity. Code, PyTorch layer, and models are open-sourced on GitHub and HuggingFace.
GLM-5.1 Weights Drop April 6-7
GLM-5.1 model weights are expected to release on April 6 or 7, according to a Discord announcement from Zai. The info comes with an image preview shared on Reddit.