AI is cooked ….

Subq: A Revolutionary AI Model

Introduction to Subq

• A Miami startup has introduced a model named Subq, claiming it could significantly impact the AI landscape.

• Unlike existing models such as ChatGPT, Claude, and Gemini that rely on transformer architecture, Subq claims to be the first frontier model not using transformers.

Understanding Transformers and Their Limitations

• Transformers utilize dense self-attention, comparing every token against all others in a context window simultaneously.

• This process is likened to a meeting where everyone must communicate with each other; manageable with few participants but becomes inefficient with larger groups.

• The computational complexity of this method is O(N²), meaning doubling inputs quadruples costs.

Subq's Innovative Approach

• Subq employs sub-quadratic sparse attention, allowing the model to select only relevant tokens for comparisons rather than engaging all tokens.

• This selective communication reduces computational time closer to linear scale instead of quadratic.

Performance Metrics of Subq

• The result is a native 12 million token context window capable of processing large codebases in one pass.

• It operates approximately 52 times faster than flash attention at 1 million tokens and costs about one-fifth compared to models like GPT 5.5 or Claude Opus for long context tasks.

Current Status and Considerations

• As of now, no independent benchmarks have validated Subq's performance on significant long-context tasks; it remains in private beta.

• Despite its innovative approach, transformers have proven resilient over nine years against architectural challenges. The issue that Subq addresses—the O(N²) limitation—is genuine and warrants attention.