Padding extends shorter sequences with special placeholder values so examples of different lengths can share one tensor and be processed as a batch. Text tokenizers usually add a dedicated padding token on the left or right; image and audio pipelines may pad spatial or time dimensions with zeros or another defined value. An attention mask or length record tells the model which positions are real so padding does not influence predictions or loss. Excessive padding wastes memory and computation, making length-based batching useful. The padding side and value must match the model’s training assumptions, particularly for autoregressive generation, positional encodings, convolutional boundaries, and exported inference runtimes.
Voice AI startup Wispr has secured another $25M in funding as its Flow dictation app surges in enterprise adoption and user retention, bringing total capital raised to $81M.
Google has released Gemini 3, its most powerful AI model yet, offering advanced multimodal reasoning, coding capabilities, and agentic planning tools across multiple platforms.
OpenAI has announced a strategic partnership with Broadcom to design and deploy 10 gigawatts of custom AI accelerators, expanding its hardware ecosystem beyond Nvidia and AMD.
Google Cloud and Figma have launched an expanded partnership to embed Google’s generative AI models – including Gemini 2.5 Flash and Imagen 4 – directly into Figma’s design tools, reducing latency and accelerating creative workflows.