AI 精选动态
智能评分 60
DeepMind内部员工爆料组织问题严重
AI 推荐理由
本次爆料提供了DeepMind内部士气低落的具体证据,但需注意消息来源为匿名员工,可信度待验证。核心解读
DeepMind被爆在Artificial Analysis榜单仅排第5,落后Anthropic、OpenAI和智谱AI;上次显著进步的模型是4个月前的3.5 Flash,而即将发布的Gemini 3.5 Pro被内部认为不是所需的突破性改进,员工士气低落,预计更多顶尖人才将离开。
全文
DeepMind内部现在已经彻底摆烂了
兄弟们,员工自己都这么说。
故事背景:
上周John Jumper刚宣布离开,今天又有内部爆料。
DeepMind目前在Artificial Analysis榜单只排第5
落后Anthropic、OpenAI,甚至被智谱AI超过
上次真正有明显进步的模型
是4个月前的3.5 Flash,Gemini 3.5 Pro预计6月30号发布,但内部共识是“这不是我们需要的step change”。
我感觉这已经不是单纯模型迭代慢的问题了
而是整个组织和决策层出了大问题
资源最多、硬件最强、人才曾经最顶尖,结果连前沿模型都守不住。
核心就一个原因:大厂的流程和官僚,已经把执行力拖垮了。
说真的,当员工自己都说“如果这样还出不了前沿模型,那我们在干什么”。
这比任何外部吐槽都更致命

> **引用原帖 leo 🐾 (@synthwavedd):**
> 🚨 SCOOP: After the release of Fable 5 and with GPT-5.6 looming, the mood behind the scenes at Google DeepMind is increasingly one of frustration and broad discontent over the lab's perceived fall into a distant third—or even fourth—place.
> "I can't blame Noam [Shazeer] for walking. He won't be the last big name to go, either," a well-connected DeepMind employee told me.
> DeepMind's last major model release, 3.5 Flash, was a significant jump over its predecessor; however, it was not meaningfully better in most cases than 3.1 Pro, released back in February. In real-world use, it remains several steps behind the frontier. That was four months ago, and Google's best model now sits in a lowly fifth place on the Artificial Analysis Intelligence Index—lapped by models from Anthropic, OpenAI, and now China's Zhipu AI. Other releases have proven similarly disheartening: the small video generation model Gemini Omni Flash launched to little fanfare and was easily beaten by ByteDance's Seedance 2.
> Gemini 3.5 Pro, slated to launch June 30th, is "not the step change we need to be truly competitive in the race [to AGI]," per another individual at the company. The consensus seems to be that leadership at Google has all but conceded that race to Anthropic and OpenAI, and that "only a big shake-up" will propel them back to the highs of mid-to-late 2025.
> But employees are not hopeful: "We no longer have a frontier model in text, image, video, voice, or even vision... if we can't release a real frontier model after over four months of work with all of these resources, what are we doing?"
> https://x.com/synthwavedd/status/2068000857757741251