Deepseek Breaks the Money - burning Curse of AI Training: $12,000, 1/525 of the Cost, MT - Bench Score Comparable to GPT - 4o DeepSeek, jointly with Tsinghua University, collaboratively launches a new AI alignment technology SPCT (Self - Original...

Deepseek Breaks the Money - Burning Curse of AI Training: $12,000, 1/525 of the Cost, MT - Bench Score Comparable to GPT - 4o

DeepSeek, in collaboration with Tsinghua University, has jointly introduced a new AI alignment technology, SPCT (Self - Principled Critique Tuning), which breaks through the traditional model relying on massive training data and dynamically optimizes the output quality during the inference stage.

According to the paper published by the research team on April 4th, this technology enables the model to dynamically correct the output during inference through a recursive architecture of "principle synthesis - response generation - critique filtering - principle optimization".

The report indicates that SPCT significantly reduces the deployment threshold for high - performance models. Taking the DeepSeek - GRM model as an example, the training cost is approximately $12,000, and the MT - Bench score is 8.35.

IT Home｜Arxiv｜Medium

📮Contribute ☘️Channel 🌸Chat

via Tech Circle🎗 Zaihua Channel📮 - Telegram Channel