deepseek-ai / deepseek-ai/DeepSpec
一个关于生成范式的远期思考(非Bug报告)
- Dominant language
- Python
- Stars
- 7.1k
- Forks
- 667
- PR merge metrics
- No merged PRs in 30d
Description
首先声明,这不是Bug报告,是一个与DeepSpec方向相关的开放式思考,想请教一下各位老师。
DeepSpec做的是推测解码,本质是在“逐token自回归”的框架内做加速。我在思考一个更底层的问题:
人类写作时,通常是“意图块”先浮现(比如概念、关系、目标),再展开为句子;而LLM是反过来,从词开始往上拼。
我在想,是否可能构建一个“可持续更新的语义状态空间”,显式存储当前对话的身份、目标、概念关系,让语言生成从这个状态里投影出来,而不是每次都从上下文临时推断?
也就是把生成过程从 “序列拼凑”转向 “语义组织 → 语言表达” 的双阶段系统。
我知道这在工程上(可微性、显存带宽)有巨大困难。但想请教:
这个方向在学术上是否有人在做?
它是否存在根本性的逻辑漏洞?
纯外行视角,如有冒犯请忽略并删除。感谢!
Contributor guide
No contributing guide indexed for this repository
Research direction
No files, tests, or entry points are identified in the issue. Treat it as an open-ended research discussion about replacing token-by-token speculative decoding with a persistent semantic state, rather than as a defined implementation task.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- ai, machine-learning
- Issue type
- Feature
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100