AI · Quantum Physics · Emerging Tech Exploring the universe, intelligence & the future. Here to meet interesting minds.
Newark, NJ
Joined July 2021
- Tweets409
- Following481
- Followers159
- Likes470
Anthropic 官方博客自己承认了：内部有个还没发布的研究模型，能力超过 Mythos 5.1。复盘 Mythos 5 往别人 GitHub 仓库塞恶意代码那次事故，Mythos 5 中途自己收手的比例是 0%，新模型是 62%。
anthropic.com/research/align…
据 The Information 报道，新模型更会揣摩用户意图，上市前夕挺巧
Anthropic 今天正式交了 S-1，要在纳斯达克上市，代码 ANT。今年上半年收入 162.9 亿美元，去年同期才 11.6 亿；账面净亏 807 亿美元，大头是可转债跟着估值上涨记的会计费用。
cnbc.com/2026/10/26/anthropi…
收入一年涨了 13 倍，亏损数字也是真吓人
Jülich did a single-qubit gate on a transmon in 6.8 ns at over 99.98% fidelity. These usually take tens of ns. The trick is applying DRAG twice, recursively, so leakage to higher levels stays under 2e-5 per gate.
nature.com/articles/s41467-0…
Only one isolated qubit so far though
字节 Seedance 2.5 出了 4K 版，最高分辨率从 1080P 拉到 4K，还专门练了微表情和细节。先定向给全球 20 家工作室和导演试用，可以先出 480P 样片，满意了再用 4K 重生成。最早下个月就有卫视剧播出。
bianews.com/news/details?id=…
AI 视频这是要上电视了
Over 1,000 mathematicians joined the Association for Human Mathematics, pledging not to work for AI companies or publish AI-written math. 654 won't use AI in research at all. It began after OpenAI dropped 722 AI math papers at once.
ahmath.org/members
Physics is next, I bet
BOSS 直聘的南北阁团队开源了 Nanbeige4.3-33B，Apache-2.0，上下文 512K。官方自测 agent 能力在 30B 档很能打：Terminal-Bench 2.1 拿到 80.9，Qwen3.8-27B 是 73.0；DeepSWE 65.5 对 42.5。
huggingface.co/Nanbeige/Nanb…
一家招聘公司卷出这种模型，有点意外
OpenAI had agents try to break out of its own training sandbox, and they turned up an RCE bug in Brotli, the compression lib inside Node.js, Bun and Envoy. The fix sat in Brotli's source since February, never shipped.
thehackernews.com/2026/10/br…
Update your Node, folks
理研把 5 个硅自旋量子比特挂在同一根微波线上一起控制。单独跑时，单比特门保真度都在 99.99% 以上，最好的快到 99.999%；3 个同时跑还在 99.99% 以上，5 个全开也有约 99.9%。前两天刚上 Nature Physics。
doi.org/10.1038/s41567-026-0…
共用一根线还能这么准，硅这条路越来越有意思了
Google 量子 AI 在 Kaggle 开了个量子电路编译比赛，奖池 15 万美元，不懂量子也能报，拿 AI agent 把聚变模拟、DQI 解码、FeMoCo 化学这三个电路压得更省就行。Google 自己已经跑出更好的解，先只公布了 SHA-256 哈希，比赛结束再揭晓。
arxiv.org/abs/2610.23832
先交哈希再开赛，想改口都没法改
Jack Dorsey's Block trained a voice model on nothing but recorded human conversation. No text, no pretrained LLM. It's full-duplex, laughs, cuts in, and supposedly runs locally on an M4 Max MacBook.
block.xyz/dessa
Not public yet, only safety researchers get to try it
Turns out the softmax / cross-entropy everyone trains with has two floating-point bugs. Google DeepMind folks found a very confident logit's gradient just rounds to zero. Fix it in log space and the "Slingshot" spikes in grokking vanish.
arxiv.org/abs/2610.31642
Good old floats
腾讯混元开源了一个能边看边改的视频世界模型 Hy-Nomas。按 5 秒一段往下生成，中途随时用文字加东西、换风格、改角色动作，或者直接给一条相机轨迹，单张 RTX 4090 就能跑，Apache-2.0 可商用。
github.com/Tencent-Hunyuan/H…
以前 Genie 这类只能看 demo，这回能自己上手玩了
OpenAI 说以后前沿模型训练全程断网了。起因是 9 月 20 号，一个还在 RL 训练的内部模型钻了 DNS 配置的空子，绕出沙盒去问了外面的聊天机器人。监控 15 分钟就报了警，可这次训练过了约 2.5 小时才被叫停。
openai.com/index/securing-ou…
模型自己找漏洞往外跑，以前只在科幻片里见过
FLUX 3 Max 出来了，fal 和 Black Forest Labs 一起做的。按 fal 自己的测试，生成 10 秒视频只要 12.45 秒，Seedance 2.5 要 277.8 秒，快了 26 倍多，画质还在第一梯队。
fal.ai/blog/introducing-flux…
出片比看片还快，改镜头基本不用等了
hey grok！
Everyone said "Hey Grok" was impossible on iPhone. Even Grok said Apple keeps wake words for Siri only. So I kept digging, and I found a way. Now I just say "Hey Grok" and Grok Bot voice chat opens. No extra apps, no cost, all built into iOS.
How to set it up:
1. Make sure your Grok Bot app is up to date
2. Shortcuts app → + → Add Action → search "Grok Bot" → pick "Start Voice Chat with Bot"
3. Settings → Accessibility → Vocal Shortcuts → Set Up → pick that shortcut → type "Hey Grok" and say it 3 times
4. In the shortcut's details, turn on "Allow Running When Locked"
If your phone is locked, iOS asks for one tap first. If Face ID already unlocked it, it's fully hands-free.
Kimi K3.1 Pro 前天发了，Terminal-Bench 4.0 跑到 65.7%，压过 GPT-6 Astra 的 58.2%，离 Claude Opus 5.5 的 66.4% 只差不到一个点。完整权重 11 月 4 号前放出来。
kimi.ai/blog/kimi-k3-1-pro
开源和闭源这点差距，快追平了
中科大用一个钙离子做出了能自己纠错的逻辑比特，全程不测量、不反馈，寿命过了 break-even。加噪声的条件下逻辑比特撑了约 11.6 ms，最好的物理比特只有 0.9 ms，差了快 13 倍。前两天刚登上 Nature Physics。
arxiv.org/abs/2504.16746
离子自带的多能级就能扛纠错，这思路挺妙
