Good morning. The big one today is a milestone in self-evolving AI: Claude is now training Claude — at $4 an hour, outperforming a $150-an-hour human researcher. Meanwhile, a 32GB Intel GPU is handling full AI animation production, WPS is letting business users build apps from a single sentence, and China's data regulator is wiring the compute grid into the national infrastructure. Here's what matters in AI today.
Today's Top 3
1. Claude Starts Training Claude: $4/hr Beats a $150/hr Human Researcher
Core insight: Anthropic's models are now training their successors — and the loop is getting cheaper and better fast. At $4 an hour of compute, the student model outperforms a highly paid human research assistant, pushing AI "self-evolution" from a slogan toward an actual engineering reality.
Source: qbitai.com
2. Intel's Arc Pro B70 with 32GB VRAM Tackles Full AI Animation Production
Core insight: With 32GB of on-board memory, Intel's new Arc Pro B70 moves AI drama creation from script to finished video on a desktop-class card — a sign that AI content production is no longer gated by cloud-only hardware or datacenter GPUs.
Source: qbitai.com
3. WPS Debuts "Inspiration App": One Sentence Builds a Working Business App
Core insight: WPS's new multi-dimensional-table feature turns plain-language descriptions into usable, continuously iterated business applications — no code required. It's a concrete step toward the promise that every business person, not just engineers, can ship software.
Source: 36Kr
More AI News
- The 734-dependency problem: why local AI deployments lag official versions — tiny differences in the inference software stack can change output tokens, a fresh explanation for the gap between self-hosted and vendor-run models. Source
- OpenClaw: it had its run, and it's winding down — the once-hyped agent framework recedes as Harness takes the spotlight, a small real-world cycle of rise and handover in the agent tooling space. Source
- China's data regulator: the compute grid is now deeply coupled with the "five networks" — at the 2026 Big Data Expo, the National Data Administration framed computing infrastructure as the "brain" woven into water, power, telecom, underground, and logistics networks. Source
- ShanghaiTech team builds embodied world-model infrastructure, closes $10M+ seed — InstAdapt aims to turn real-world robot interactions into reusable training experiences, demoing at NVIDIA GTC 2026 as the only firm co-showcasing physical-AI infrastructure with NVIDIA. Source
Trend Watch
Today's through-line is AI building AI, then putting that capability in everyone's hands. Models training models collapses the cost of research itself, while 32GB desktop GPUs and sentence-to-app tools collapse the cost of creation. At the same time, compute is being treated as national infrastructure — wired into water, power, and telecom grids. The loop is closing: cheaper intelligence, cheaper tools, cheaper compute, all at once.
Disclaimer: This briefing is AI-curated from public RSS sources for informational purposes only. It does not constitute investment advice. Links point to original publishers.
