AI影像先驅新創Inception推出加速聊天技術

🔑 增強重點摘要

• •Inception Labs applies diffusion models to language modeling, generating text in parallel rather than sequentially like autoregressive transformers, achieving over 1000 tokens/sec on NVIDIA H100 GPUs[1][2][5].
• •Mercury family includes Mercury Coder for code generation and a general-purpose model, both with 128,000-token context windows, priced at 25 cents per million input tokens and $1 per million output tokens[2].
• •Company raised $50M in seed funding led by Menlo Ventures, with investors including Andrew Ng and Andrej Karpathy, to scale diffusion LLMs[2][4][6].

🛠️ 技術深入

• •Uses diffusion process to generate entire blocks of text at once via denoising steps, retaining transformer neural architecture but with different training objective and parallel inference[1][3].
• •Early prototype matched GPT-2 quality at 10x faster speed; scaled up with larger models and better data for commercial Mercury[1].
• •Features animation in chat interface visualizing text sharpening from noise to detail during generation[2].
• •Reduces GPU footprint, enabling larger models at same latency/cost or more users on existing infrastructure[2].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線