谷歌股哥
1.4K posts
e-GPU: An Open-Source and Configurable RISC-V Graphic Processing Unit for TinyAI Applications
arxiv.org/abs/2505.08421
English
谷歌股哥 retweetledi
Gemini 3.1 Pro benchmarks leak @GoogleDeepMind
Take with a grain of salt no guarantees for authenticity 🤷♂️
English
Cool, so waiting for Zuck to come back.
Internal results show Meta’s Avocado model beats leading open-source models even before any specialized training.
as per The Information.
English
谷歌股哥 retweetledi
据爆料，Gemini 3.5 Pro的Snow Bunny(雪兔)版本遭泄露
1、泄露的内部模型Snow Bunny(雪兔)可一键构建完整应用。
2、仅需1个提示词，该模型就能生成3000行可运行的代码。
3、全新的Fierce Falcon模型专攻速度与逻辑运算。
4、全新的Ghost Falcon模型负责处理UI、视觉及音频创作工作。
5、性能超越Claude Opus 4.5。
6、新增Deep Think开关，用于解决复杂逻辑问题。
7、采用System 2思维模式，在作答前先暂停并进行推理分析。
8、在高难度推理基准测试中斩获80%的得分，而竞品平均仅为55%。
9、泄露的代码显示，gemini-for-google-3.5相关变量已准备就绪。
中文
谷歌股哥 retweetledi
谷歌股哥 retweetledi
Gemini 3.5: The "Snow Bunny" Leaks
- Snow Bunny Checkpoint: Leaked internal model "Snow Bunny" builds entire apps in one go.
- 3,000 Lines of Code: It can generate 3,000 lines of working code from a single prompt.
- Fierce Falcon Model: New "Fierce Falcon" model specializes in pure speed and logic.
- Ghost Falcon Model: New "Ghost Falcon" model handles UI, visuals, and audio creation.
- Beats GPT-5.2: It outperforms the unreleased GPT-5.2 (75.40%) and Claude Opus 4.5.
- Deep Think Mode: Features a new "Deep Think" toggle for solving hard logic problems.
- System 2 Reasoning: Uses "System 2" thinking to pause and reason before answering.
- 80% Reasoning Score: Scores 80% on hard reasoning benchmarks vs competitors' 55%.
- API Confirmed: Leaked code reveals gemini-for-google-3.5 variables are ready.
English
哈哈哈，真有人把Anthropic CEO和Deepmind负责人前几天接受访谈的内容对比了。
上周达沃斯论坛上。
Anthropic的CEO Dario Amodei和Google DeepMind的CEO Demis Hassabis都来了。
两个人说的话，听起来既像打架，又像合唱。
Dario说：6-12个月后，AI就能干完程序员的活了
时间线够激进，他的理由很直接：
Anthropic的工程师现在已经不写代码了，他们只是编辑Claude生成的代码。
他们的新产品Claude Cowork，一周半就做出来了，几乎全是AI写的。
是不是有点夸张？
但看看他们的动作就知道了。
Anthropic最近招的岗位，覆盖音频、生物学、网络安全、视觉。
他们还要自己部署100万颗TPU芯片，不租Google Cloud的了。
三年时间，从零做到预计100亿美元营收。
2028年的目标是700亿美元ARR，真的猛啊！
而 Demis说：没那么快，但也快了
Demis的态度稍微保守一点。
他觉得编程确实容易自动化，因为你能立刻验证结果对不对。
但科学发现就难多了，因为验证周期长。
不过他也说：今年我们可能会看到，初级岗位开始受到影响。
时间线上，Demis认为AI达到人类认知水平，可能需要3-5年，甚至10年。
两人的分歧在哪？主要是对"自我改进循环"的判断。
如果AI能自己训练和改进AI，那进度会飞快。
Dario觉得这个循环快要转起来了，Demis觉得还需要时间。
都担心什么？
Dario提了一个很可怕的场景：
GDP增长5-10%，但失业率也是10%。
这在历史上基本没出现过。
高增长和高失业同时存在，意味着财富在极少数人手里快速积累，大部分人被甩在后面。
他还提到一个"零号世界国家"的概念，1000万科技工作者，GDP增长50%，其他人停滞不前。
听起来像科幻小说，但这不就是现在硅谷和其他地方的差距在放大吗？
中国呢？
Dario说了一句很重的话：把美国AI芯片卖给中国，有点像把核武器卖给朝鲜。
Demis的说法温和一些，他觉得中国目前落后美国大概6个月。
两个人都希望能让中国公司放慢进度，但这显然不太现实。
给年轻人的建议
Demis说了一句很实在的话：
如果你今年找不到初级岗位或实习，别慌。
花时间玩这些 AI 工具，玩到比研究人员还熟练。
为啥？
因为研究人员都在忙着造工具，没时间探索工具的"能力溢出"。
你如果能把工具用到极致，甚至比造工具的人还懂怎么用，那你就有价值了。
听起来有点反直觉，但想想也对。
工具的进化速度太快了，会用比会造更稀缺。
我们该怎么办？
如果Dario说得对，我们只有1-2年时间适应AI达到人类认知水平。
如果Demis说得对，时间会长一些，3-5年，甚至10年。
但不管是谁对，有一件事是确定的：这个变化正在发生，且不会停下来。
Demis建议：玩AI工具，玩到比任何人都熟练。
这可能是普通人在这场变革中，唯一能抓住的机会了。
原文见评论区
中文
From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence
arxiv.org/abs/2601.03220
English
谷歌股哥 retweetledi
Paris loves robots! 🫶🏼
... and of course, vice versa!
What is the first thing that comes to mind when you think of Paris?
Probably warm croissants and the Eiffel Tower glittering at night.
What if I tell you, that it should be robotics? 🤯
Paris is a great place to launch a robotics startup because it mixes world-class tech talent, big startup energy, and strong public support for deep tech.
First, Paris has excellent research and engineers, coming from places like Inria (a major research institute in robotics and AI) and top schools around Paris/Saclay.
This matters because robotics startups often start from research and need smart people who can build “real” systems.
Second, Paris is one of Europe’s biggest startup hubs thanks to STATION F (one of the world’s largest startup campuses). It gives founders access to mentors, partners, and lots of investors, which makes it easier to meet the right people fast.
Third, France is unusually strong in government-backed deep-tech funding through programs like French Tech 2030 and support from Bpifrance.
Robotics takes time and money (prototypes, hardware, testing), so this kind of support is a big advantage.
Here's the landscape of robotics companies in Paris:
→ Genesis AI is a physical-AI lab building generalist robots + robotics foundation models, and launched with $105M funding.
→ @StanleyRobotics builds robotic “valet parking” systems that autonomously move and park cars.
→ @UMA_Robots is building general-purpose mobile/humanoid robots for industrial and real-world tasks.
→ @LeRobotHF (@huggingface) is an open-source robotics stack with datasets, and models to lower the barrier to robot learning, they also have @pollenrobotics in their portfolio.
→ @EnchantedTools builds friendly humanoid robots (e.g., Mirokaï) for service environments and raised ~$17M seed.
→ Fuzzy Logic Robotics makes a platform to quickly program/deploy industrial robots and raised €2.5M seed.
→ @in_bolt builds real-time 3D vision guidance for industrial robotic arms and raised ~$16–17M Series A.
→ @heex_io builds “smart data” software that captures only the most useful edge data to improve AI systems and raised €6M seed.
→ @WandercraftHQ builds advanced robotic exoskeletons for mobility/rehab (and humanoid robotics R&D) and has $75M total funding.
→ SoftBank Robotics UK Ltd (EMEA HQ) develops and commercializes service robots like Pepper and NAO for customer interaction and education.
→ Galam Robotics builds modular warehouse storage automation robots/systems and raised €10M - Series A.
→ Exwayz builds GPS-free 3D localization / SLAM software for autonomous robots and raised €1M.
→ @phospho_ai (YC W24) builds tools/software to easily connect, control, and train AI models on real robots (“brains for robots”).
→ General Robotics builds industrial robotics solutions for automation and integration.
→ Gobano Robotics develops AI-powered robots to automate dexterous industrial tasks and raised ~€3M pre-seed.
→ Marso Robotics builds brownfield-ready warehouse companion robots to automate hard manual logistics tasks.
→ @DeplaceAI provides data + semantic guidance that turns large-scale demonstrations into training signal for robotics/physical AI.
I'm aware that there might be companies missing, and the map is not completed! More in comments...
~~
♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com
English
谷歌股哥 retweetledi
1.5 million names are flying around the Moon on Artemis II. Is yours one of them?
It's not too late to add your name to the mission—and it's absolutely free: go.nasa.gov/artemisnames
English
谷歌股哥 retweetledi
DeepMind just did the unthinkable.
They built an AI that doesn't need RAG and it has perfect memory of everything it's ever read.
It's called Recursive Language Models, and it might mark the death of traditional context windows forever.
Here's how it works (and why it matters way more than it sounds) ↓
English
回顾一下Anthropic CEO Dario在纽约时报峰会上接受采访说的话——
Scaling Law才是本质的，他已经观察了10年时间。每过一段时间，都会有一点小改动，例如推理模型、Test Time Scale，都是一些小技巧。但背后Scaling Law，更多数据、更多算力才是本质的。
高级分析师@techeconomyana
姚顺雨： Anthropic这个公司很有意思，它不做什么花哨的创新，就是把预训练做大，把RL做好，然后去解决真实世界的任务，模型就会越来越聪明，带来更多价值。做To B，其实所有目标是更一致的：模型智能越高，解决任务越多，带来的收入越大。
中文
