AI News 2025-12-28
AI Daily Brief
Summary
Alibaba's AntV team launched an AI infographic framework that turns text prompts into structured visuals. Mathematician Terence Tao said AI "solved" old problems by rediscovering forgotten results, highlighting its role as a super librarian. In 2025, the field focused on the AGI-to-ASI transition, with DeepSeek-R1 reaching the cover of Nature and renewed debate on AI tools and human centrality.
Ruby 4.0 shipped with the new ZJIT compiler and Ruby::Box isolation to improve performance and safety. Shanghai AI Lab's SGI-Bench shows top models still fall short of a "qualified scientist" across the research loop. The "vibe coding" hype cooled as studies suggest AI coding has not boosted productivity and can weaken developer skills. Zhipu AI released ZCode, a visual AI coding environment with multi-model support and conversational version control.
Today’s AI News
AntV Infographic enables one-sentence AI visualization. Alibaba’s AntV team released a new infographic generation and rendering framework. Users can input a few sentences and get structured infographics automatically. Highlights include AI-native optimization, a declarative graphics language, high-quality SVG output, 200+ templates, multi-theme styles, streaming rendering, built-in editing, and developer-friendly tooling for education, data display, and reports.
Terence Tao: AI “solved” old math problems by rediscovering forgotten results. Tao noted that some recent AI “solutions” to Erdos problems were already solved long ago but forgotten. He argues this is not cheating but a demonstration of AI’s ability to scan literature and surface “low-hanging fruit” labeled as unsolved. The episode shows AI’s current strength as a super librarian performing knowledge archaeology rather than true from-scratch innovation.
A $2T AI bet: DeepSeek reaches Nature while Meta becomes a 2025 laggard? The industry focused on the AGI-to-ASI transition in 2025, with major gains in reasoning and multimodal benchmarks. Investment surged and competition intensified. China’s open-source DeepSeek-R1 became the first peer-reviewed model on the cover of Nature. Leaders emphasized three trends: AI tool mastery as a career necessity, rising value of soft skills, and the need to keep humans at the center of AI alignment.
Ruby 4.0 released with a new compiler and native isolation. On its 30th anniversary, Ruby shipped 4.0 with ZJIT, a new SSA-based JIT compiler aiming to surpass YJIT through deeper global optimizations. It also introduced Ruby::Box for strict namespace isolation to reduce global pollution and monkey patch conflicts, plus a redesigned Ractor API with Ractor::Port for better concurrency. The release includes broader language and core library upgrades, Unicode 17 support, stricter nil checks, and performance improvements.
SGI-Bench shows top AI models are still far from a “qualified scientist”. Shanghai AI Lab introduced SGI-Bench to evaluate AI scientific general intelligence across four loops: reflection, ideation, action, and perception. The best closed model, Gemini-3-Pro, scored only 33.83/100. Findings show long-chain reasoning failures in reflection, low feasibility in ideation, weak correctness in experiment execution, and limited multimodal comparison reasoning in perception. The study highlights gaps in evidence synthesis, experimental planning, and code-level scientific correctness.
“Vibe coding” faces a 2025 reality check. Despite hype that AI would write most code, real-world use shows limited productivity gains. Surveys report higher error rates for Copilot users and low trust in AI-generated code. Over-reliance can also weaken developer “muscle” and problem-solving drive, a phenomenon described as “machine detours.” The consensus is shifting toward AI as a precise assistive tool, not a replacement, with the biggest value in testing and support tasks.
Zhipu AI launched ZCode, a visual desktop AI coding environment. ZCode addresses the high barrier of CLI-based tools like Claude Code and Gemini CLI. It visualizes command-line operations, integrates GLM, Claude, Gemini, OpenRouter and more, provides conversational version checkpoints, supports MCP for web access and tools, ships with specialized sub-agents for generation and review, and works out of the box without Node.js or Python.