抛弃自回归,Celeris用扩散模型造最快LLM
原创
当所有人都在卷自回归模型的token生成速度时,Celeris抛出了一个反问:为什么LLM必须一个字一个字往外蹦?他们把赌注压在了扩散模型上。自回归天然串行——第N个token得等第N-1个出来才能算。扩散模型可以同时去噪所有位置,理论上速度能质变。这不是概念验证,是工程落地。问题是,扩散模型做语言生成的质量能不能跟上?Celeris显然认为可以,而且足够快能赢。我觉得这是个值得跟踪的方向,不管最终结果如何。
原文:Building the fastest LLMs: why we're starting with diffusion · 来源:Hacker News
版权声明
所有资源都来源于爬虫采集,如有侵权请联系我们,我们将立即删除
相关文章
- Nvidia's Seven Billionaires Have Lost a Third of Their Wealth
- Stabyourself.net Games
- 【IT 早报】2026年11月19日 - 本周要闻与趋势
- 从均匀平铺到自适应路由:RINs 如何让扩散生成提速 10 倍?
- Claude Code makes a finished subagent send its report twice (70-second repro)
- Becerra faces a $2B budget deficit as potential AI bust looms
- Show HN: Agent Hive – cute Grokbot like agentic teams with queen and many bees
- Show HN: Lazycow Studio – turns footage you never edited into reels, locally
- Experiments in Weight-Space Programming
- Lindy Effect
最新文章
- Nvidia's Seven Billionaires Have Lost a Third of Their Wealth 44秒前
- Stabyourself.net Games 1小时前
- 【IT 早报】2026年11月19日 - 本周要闻与趋势 1小时前
- 从均匀平铺到自适应路由:RINs 如何让扩散生成提速 10 倍? 2小时前
- Claude Code makes a finished subagent send its report twice (70-second repro) 2小时前
- Becerra faces a $2B budget deficit as potential AI bust looms 2小时前
- Show HN: Agent Hive – cute Grokbot like agentic teams with queen and many bees 3小时前
- Show HN: Lazycow Studio – turns footage you never edited into reels, locally 3小时前
- Experiments in Weight-Space Programming 3小时前
- Lindy Effect 4小时前
itfan123




