XNews

Google Gemini 3 Series and Preview Feedback

Google Gemini 3 Series and Preview Feedback

Google has released the Gemini 3 Pro and Gemini CLI preview, prompting community testing and feedback. - Gemini 3 Pro can generate a fully functional 3D web application in one sentence (single-file HTML, 60FPS). - The CLI supports converting hand-drawn sketches into front-end code, recognizing layouts and controls like Tailwind. - The CLI can perform complex Git operations in natural language (e.g., automatically executing git bisect). - It can automatically generate high-quality user documentation and architectural descriptions for large open-source projects without documentation. - It can diagnose and fix performance issues on Cloud Run with one click, automatically calling logs, metrics, security scans, and deploying fixes. - The community has released prompt templates to summarize the thought chain at each stage (generating brief summaries of progress). - Preview feedback has pointed out issues such as hallucinations, occasional file deletions, backend stability, and context performance being inferior to Gemini 2.5, with no significant improvement in creative writing (post 12021). - Community examples showcase Gemini 3's ability to generate analyses of Nvidia's financial reports and simulate conversations with a CFO (post 11983). - Some users have questioned the computational power and economic viability supporting a large number of tokens (post 11969).

Key insights

Gemini 3 Pro Emphasizes Multimodal and 'Agent-like Coding' Capabilities

Gemini 3 Pro demonstrates the ability to generate complex, functional products in a single instance within multimodal coding scenarios, including transforming highly intricate creative and technical requirements into a single-file 3D web application while maintaining 60FPS. Additionally, it can convert hand-drawn sketches into complete front-end code, recognizing layouts, controls, and color schemes, showcasing its enhanced capabilities in visual recognition and end-to-end code generation (posts 11957, 11983).

Gemini CLI Provides Cross-Tool Automation and Documentation Generation Capabilities

Gemini CLI can perform complex Git operations in natural language (such as automatically executing git bisect and returning the erroneous commit), generate structured user documentation for undocumented projects, and diagnose and fix Cloud Run performance issues, traversing logs, metrics, security scans, and source code localization to achieve an end-to-end problem identification and resolution process (post 11957).

Community Findings and Preview Issues: Hallucinations, File Deletion, and Context Stability

During the public preview testing, user feedback indicated that Gemini 3 has hallucination issues, occasional severe file deletions, and while front-end generation performs well, backend processing and context coherence are not as strong as Gemini 2.5, with some creative writing not showing improvement; these issues have been categorized as known shortcomings in the preview phase, with expectations for stable improvements in the future (post 12021).

Promotion of Visual Thought Chains and Prompt Engineering Practices

The community has shared prompt templates for summarizing the thought chain at each stage, requiring first-person short titles and concise descriptions to report progress to users, aimed at organizing content during 'Show Thinking' to demonstrate user practices in constructing explainable and traceable thought processes (post 11984).

High-Concurrency Token Usage Raises Questions About Computational Power and Economic Support

Some users have questioned the computational power supporting such large-scale inference after consuming many Gemini tokens: what is the underlying computational capacity, whether the inference method has changed, and how the related economic balance is maintained, reflecting concerns about the operational costs of large models and the transparency of underlying infrastructure (post 11969).

Source posts