China Unveils Five-Year Plan to Mandate AI Education Across All Levels of Schooling
On June 29, China's State Council unveiled a five-year blueprint for incorporating AI education into every level of schooling, from kindergarten through university. According to Bloomberg, the plan positions AI as a core competency for all students and forms part of President Xi Jinping's strategy to secure dominance in frontier technologies.
The plan calls for systematic curriculum development, teacher training, and preparation of teaching materials, all aimed at cultivating AI talent early. As the world's second-largest economy searches for new growth engines, the decision to AI-ify its entire education system signals that state-led AI investment is expanding from infrastructure into talent development.
With the US-China AI race intensifying, mass talent cultivation through education policy will be a source of long-term competitiveness — a move that could also shape how other countries, including Japan, respond.
OpenAI Publishes a Map of the EU's AI Labor Market — Which Jobs Are Affected
OpenAI has published a report, "Mapping Europe's AI Workforce Opportunity," that maps how AI will affect labor markets across the EU.
The report analyzes which occupations are exposed to automation, which fields will grow, and how business processes will change. While some occupations will shrink with AI adoption, new jobs and fields of surging demand will also emerge — throwing into sharp relief the challenge of supporting workforce transition.
OpenAI has long published research on AI's impact on employment, and this EU-focused analysis appears aimed in part at building trust in Europe's tightly regulated market.
DeepSeek V4 Stable Release Coming in Mid-July
Members of Reddit's LocalLLaMA community have shared the contents of an email DeepSeek sent to Chinese-speaking users, revealing that the stable release of DeepSeek V4 is scheduled for mid-July.
A pre-release version of DeepSeek V4 is already available and runs locally on llama.cpp. The stable release is expected to deliver further improvements in reasoning quality and tool-use capabilities. In anticipation of the mid-July launch, the community is already working on quantization support and local-runtime optimizations.
DeepSeek has drawn attention with its cost-effective models, and depending on the stable release's performance, the map of the open-source LLM race could shift once again.
"Thinking While Speaking" — Chasing Frontier-Level Performance in Real-Time Conversational AI
A paper attracting attention on Hugging Face's Daily Papers, "Thinking While Speaking," proposes a method that combines a lightweight on-device model with a high-performance frontier LLM to achieve both real-time responsiveness and high answer quality.
The mechanism is simple: a small "Talker" model starts responding within milliseconds, and as knowledge arrives from the "Reasoner" — the frontier LLM running behind the scenes — it is woven into the conversation. This speeds up time-to-first-response by 7–19x while targeting near-frontier accuracy. Notably, it can run on an Apple M2 (16GB) laptop.
The approach is highly practical for domains where responsiveness is critical, such as voice assistants and customer support. Resolving the speed-versus-quality trade-off through model collaboration points to where conversational AI is heading.
Google Cloud Begins Offering SandboxAQ's Science-Specialized AI Models
Google, a subsidiary of Alphabet, will start offering software company SandboxAQ's "Large Quantitative Models (LQMs)" on the Google Cloud Marketplace, Bloomberg reports.
Unlike LLMs, which specialize in text generation, LQMs are trained on scientific equations and experimental data. They can be used alongside Google's Gemini models in fields such as drug discovery, materials science, and semiconductor manufacturing. Where LLMs process "words," LQMs process "numbers and the laws of physics."
AI applications in scientific research have expanded rapidly since Google DeepMind's AlphaFold delivered breakthroughs in protein structure prediction. The SandboxAQ partnership marks a step toward making science-specialized AI more accessible to research institutions and companies.
Skipping Transformer Layers at Load Time — A New Approach to Making LLMs Lighter
In Reddit's LocalLLaMA community, a fork of llama.cpp has implemented a "--skip-layers" feature that skips specific transformer blocks at runtime.
Based on recent research papers, this technique leaves unneeded layers uninitialized at model load time, reducing VRAM usage and inference cost. Unlike bake-time (pre-merged) pruning, it lets you select layers dynamically at runtime.
Crucially, performance varies by orders of magnitude depending on which layers are skipped, so a selector mechanism for picking the right layers has been implemented. Benchmark results are published on Substack, and community feedback is pouring in.
For local LLM runtimes, techniques that shrink model size while preserving accuracy are in high demand, and integration into mainline llama.cpp is eagerly anticipated.