大家支持jensen 可以去这个网站签名。standwithjensen.com/
1
32
同一句话,Anthropic 的招股书 15 天里改了两次。10 月 26 日首版没有;11 月 2 日修订版说“可能即将”达到 RSP 里 AI 加速 AI 研发的阈值;11 月 10 日第二次修订改成已经达到。
依据是对未来模型和整体进度的预期,当前模型并没有表现出来。RSP 的条款写的是 observe or expect,预期就算数。
招股书还承认,过去它都是事后才发现研发提速的。
Anthropic again amended its S-1 today.
Amendment No. 2
In its today's filing, Anthropic says it has crossed a line set in its Responsible Scaling Policy: the point where AI is making AI research much faster, which requires stronger safety checks.
A week earlier it said only that it “may imminently” cross it. The call is based on what it expects from future models, not on anything current models have shown. The filing also sets the IPO price at $70 to $78 a share.
2
2
68
当初云居民是为了订阅AI,办理itin,办理美卡,后面顺势就申请美签,下一步估计又会顺势去趟美国,去趟日本。这一切好像也不是主动,只是为了去完成某项事情,接着就进入了另一个局面,越觉得这是时代在推着自己走。
41
Richie.L retweeted
NVIDIA announces CEO transition. Read the release: nvidianews.nvidia.com/news/n…
Meta FAIR 开源了 e-means,拿来替换 k-means。论文说,Lloyd 跑 25 轮的损失,它 1.4 到 2.6 轮就压过,GPU 上比 Flash-KMeans 快 10–45 倍。
质心改用小批量加滑动平均更新,batch 和步长按公式自动算,只需给 k。已进昨天发布的 FAISS 1.16.0。
但对手用 k-means++ 初始化时,5 轮后只领先 0.06%–0.33%,省的主要是时间。
开源地址:github.com/facebookresearch/…
31
给 5 个大模型两件违背道德的事,问哪件更糟,答案直接从第一个 token 的概率里读。
密苏里大学等团队的论文发现,同一组题只把两个选项对调顺序,30 组里有 26 组排出的价值次序变了。可把每个模型的答案平均之后,4 个模型都得出干净、几乎一样的排序,首尾还和多数人类被试一致,却对不上它们自己任何一组的结果。
测的都是非推理模式。看模型的道德评分前,先看它单题排得稳不稳。
原文地址:arxiv.org/abs/2611.10265
This paper finds that LLM moral choices often lack a stable ranking of values, and averaging their answers creates a fake, human-like moral hierarchy.
Many studies conclude that LLMs hold sensible moral values because they do well on morality tests. But good answers do not show that a model weighs competing values against each other.
The researchers gave 5 LLMs, including GPT-5.2, Gemini, and DeepSeek, pairs of moral violations and asked which was worse. They read each answer's probability straight from the model's next-token output, then tested whether the choices fit a consistent ranking.
Reversing the order of the 2 options changed the ranking in 26 of 30 test blocks. After averaging, 4 of 5 models showed a clean, nearly identical ranking that matched no individual test set.
Before trusting a model's moral alignment scores, check that its choices hold a consistent ranking on individual prompts and both answer orders, without averaging.
30
Artificial Analysis 的视频编辑榜上,Gemini Omni Flash 1.2 排第 2;文生视频从 1.1 的第 12 升到第 6,图生视频从第 17 升到第 14。
价格变化更实在:1080p 每分钟 $6.26,1.1 是 $9.12,降了 31%,一条 10 秒片约 1 美元出头(推算)。10 秒 1080p 约 76 秒出片,比文生视频榜上排在它前面、被计时的模型都快。
AA 也写了短板:人体结构、声音和语音编辑偏弱。
Gemini Omni Flash 1.2 takes #2 on AA-Video-Editing v1.1 and improves on Gemini Omni Flash 1.1 on AA-Video-T2V v2.1 (#12 to #6) and AA-Video-I2V v2.1 (#17 to #14)
Gemini Omni Flash 1.2 is Google's latest video model, released on November 5. It supports Text to Video, Image to Video and Video Editing, with native audio, and generates clips of up to 15 seconds. It will replace Gemini Omni Flash 1.1 on the Gemini API on December 3.
In our testing, it performs best on dialogue scenes, multi-shot stories and UI and motion design clips in Text to Video, and on complex edits and visual effects in Video Editing. It is weaker at human anatomy and at sound and speech edits.
Gemini Omni Flash 1.2 costs $6.26 per minute of 1080p video on the Gemini API, 31% less than Gemini Omni Flash 1.1 at $9.12 per minute.
It is available in the Gemini API, Google AI Studio, Vertex AI, the Gemini app, Google Flow and YouTube.
Congratulations to @GoogleDeepMind on the release!
See below for our analysis of Gemini Omni Flash 1.2 in Text to Video, Image to Video and Video Editing 🧵
1
27
Wow
Anthropic just killed:
> Google Sheets
> Excel
> TLDraw
> Airtable
> Miro
> Excalidraw
> FigJam
> Whimsical
> Lucidchart
> Lucidspark
> Mural
> Tableau
> Looker
> Power BI
> Metabase
> Hex
> Julius
> Mode
> Sigma
> ThoughtSpot
> Omni
> Lightdash
> Preset
> Redash
> Deepnote
> Observable
> Evidence
> Qlik
> Domo
> Rows
> Smartsheet
> Coda
> Notion databases
> Equals
> Causal
> Pigment
> Quadratic
> Sourcetable
> Grist
> Baserow
> NocoDB
> SeaTable
> Stackby
> Zoho Sheet
> Apple Numbers
> LibreOffice Calc
> Datawrapper
> Flourish
> Polymer
> Akkio
> Eraser
> Napkin
> Draw io
> Visio
> Gliffy
> Creately
> Milanote
> Stormboard
> Conceptboard
> Klaxoon
> Heptabase
> Kinopio
> Obsidian Canvas
> Canva Whiteboards
> Microsoft Whiteboard
> Zoom Whiteboard
> Apple Freeform
Everything's totally fine.
21
10 月 28 日官宣时,Gemini 4 Carbon 只给部分网安防守方用。13 天后,Google 放开了。
Ultra 用户在 Gemini App 和搜索 AI Mode 里能用,开发者走 AI Studio 的 API,科学家可申请 6 个月免费,网安团队走 Fairwind 计划。
官方说多数活仍用 Argon,Carbon 留给最难的。演示里它把鸟类研究做成鸣禽发声器官的 3D 模拟,实时合成 6 种鸟叫。
20
去年 11 月,Demis Hassabis 用 Gemini 3 vibe coding,几个小时重做了他 90 年代写的《Theme Park》测试版,薯条撒多少盐都能调。
11 月 10 日他说,Gemini 4 Carbon 一次就逆向了原版二进制,加了个 DeepMind 主题设施,再移植到网页。
两条帖子隔了 357 天(推算)。一年前是照着记忆重写,现在从老程序本身读起。这款游戏 1994 年发行,他是主程序之一,当时大约 17 岁。
18
7 份公开规划 PDF,206 张工程图,一条提示词,一个晚上。
Google DeepMind 的 Philipp Schmid 用它演示 Antigravity 的 /swarm:一组 Carbon agent 跑 10 轮通宵循环,重建出伦敦 Platform 37 办公楼的 3D 模型,据 Google 博客约 90.6 万个三角面。
他说用了几周,默认 medium 档,任务野心大了 10 倍,还该再大 100 倍。他是 Google 的人,这话听一半。
1
20
$550,8 小时,4.45 亿输入 token。
Felix Rieseberg 想测 Fable 5.2 只靠提示、不给人工反馈能做到哪一步,做了个游戏,模型、音效、贴图全是从零做的,放在 Claude Artifact 里能直接玩。
算一下:270 万输出按 $40 / M 是 $108,其余约 $442 是输入。4.45 亿输入全按原价要 $3,560,全命中缓存只要 $89(推算)。他说的 heavily cached,省下的是大头。
1
1
33
















