David

[David @PureXView]

@shengtang135754 Google Play那边安好？

[Thinking Machines @thinkymachines]

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor.

Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon.

https://t.co/VR02tw73fJ

[David @PureXView]

@moger2882 路由器开启SSH，然后交给Grok CLI、Codex CLI。

[Google Research @GoogleResearch]

Today on the blog, we present a new system, published in Nature, that leverages standard consumer smartwatch sensors to passively detect signs of high blood pressure — validated against the clinical gold standard of 24-hour ambulatory bp monitoring. More →https://t.co/p5bCtcaHrj

[METR @METR_Evals]

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development.

We are now publishing a report summarizing the results of this risk assessment.

[Amy Hua @wenyanhh]

Frevana will pay your Higgsfield bill

Send us your 2026 receipts from Higgsfield, Runway, fal, or any AI platform

We'll match your spending in @frevana_ai credits

→ 1:1 match any AI sub
→ Every 2026 receipt counts
→ First 10k users only

Comment "BILL" to claim yours 👇

[RoboHub🤖 @XRoboHub]

Apple’s robotics team seems to be preparing humanoids for the home, too.

In their demo, this Unitree G1 crouches under a table and turns sideways to get through a half-open door. It can also push a box, follow a person and sit down when instructed.

Behind it is HALO, an Apple research system that uses vision and language to generate whole-body motion. The team created 300 hours of training data across 1,000 simulated homes.

Across 360 single-task trials in six unseen real-world environments, none of which were pre-scanned, the team reports a 91% success rate. In a separate multi-task test in one real environment, it achieved 70% across 30 trials.

The team has also released code for generating training data, training the motion models and deploying them on a G1.

[向阳乔木 @vista8]

提示词：“从 hackernews 12G数据库找出最被大家推崇认可的人生经验”

1. 不要等有空了，再陪重要的人。

来源于 PG 的一篇文章，很有共鸣。

2. 一种平凡、可靠、有人相伴的生活，本身就值得。

来源于《一个平凡人生的讣告》，一个没有显赫成就、却认真照顾家庭的人。

3. 别把“获得高分”的能力，当成生活的目标。
来源于《The Lesson to Unlearn》

4. 钱很重要的一种用途，是买回时间和选择权。

来源于 Sam Altman 的一篇文章《Tips for a Better Life》

5. 让别人觉得自己很蠢，通常不会让你的正确产生作用。

批评观点，同时保留对人的尊重。

6. 允许自己刚开始做得差，靠持续完成作品追上审美。

7. 遇到拖延，先找阻力，别先���备自己。

8. 少用“我就是这种人”，提前关闭自己的可能性。

“我不擅长社交”“我不是运动型的人”“我只适合技术”，都可能从描述变成限制

---
所有文章和讨论见评论区，没想到山姆·奥特曼的《Tips for a Better Life》帖子上榜两个。

改天得拜读下。

[Alexandr Wang @alexandr_wang]

🚨MUSE FEATURE ALERT 🚨

muse voice is out for everyone in the US!

you can now phone call and video chat with your muse whenever you need.

i personally have a sticky situation today, and asked muse for some advice. sage and calm as always ❤️

[Manus @ManusAI]

You can now export Manus Studio projects to Premiere Pro, DaVinci Resolve, CapCut, or Final Cut Pro.

Create with AI and edit right in Manus, or export your video to the tool of your choice.

More flexibility, more freedom to create.

[Tarush Agarwal @tarush_agarwal_]

Introducing Cekura Bench: the world’s leading Voice AI Benchmarks

We benchmarked 50+ models from openAi, Google, Nvidia, xAI, Inworld, Gradium, Assembly, Fish audio and more

Now you can:
→ Compare naturalness, performance and cost
→ See how the best voice agents are being built.
→ Find the best fit for what you are building.

Our tests cover Speech-to-Speech models, Speech-to-Text models, Text-to-Speech Models, and Voice Agent platforms. We recorded real humans doing tasks in different domains and created the most representative voice benchmark out there!

Explore the benchmarks: https://t.co/eZAFy7P0Nj

We would love to collaborate with more teams on these benchmarks. Reach out if you are interested.

[Perplexity @perplexity_ai]

Introducing teammates.

Teammates work for you. They take on jobs, and work proactively using your Computer tasks, skills, memory, and tools. Teammates collaborate with other teammates and can message coworkers in Slack.

Available now for Max subscribers on the web.

[Will Eastcott @willeastcott]

This isn’t video footage.

It’s a Gaussian splat rendered in real-time, walked by a simulated person filming on a virtual phone: gait, head turns, motion blur, rolling shutter, footsteps.

You can tap to burn the Gaussians it's made of. 🔥

Built with @playcanvas
🧵 👇

[BennyDaBall @BennyDaBall_OG]

Comfy-Org dropped RelayVSR this morning, a 2x video upscaler built for AI video. Put it head to head with RTX Video Super Resolution on my 5090, same MiniMax H3 clips, same 2688x1536 output.

My numbers (warm, median of 3, 124 frames 1344x768 -> 2688x1536):
- Upscale time: RelayVSR 5.8 s vs RTX VSR 1.3 s
- GPU memory: 11.6 GB vs 2.9 GB
- Detail added over a plain 2x resize: RelayVSR 4.8x vs RTX 2.7x
- Flicker vs the draft: RelayVSR 0.98 (calmer than the source) vs RTX 1.18 (sharper AND shakier)
- Face match to the draft: RelayVSR 0.87 vs RTX 0.97
- Spoken line survives both, 9 of 9 clips

Honest take: these are two different answers to the same job. RTX VSR is NVIDIA's feed-forward CNN upscaler: a second per clip, keeps the exact face, but it passes every bit of H3's flicker through and adds some. RelayVSR runs a 1-step Wan 2.2 5B diffusion model on the keyframes and a small decoder on the rest: nearly 2x the detail, and it's the only one that comes out steadier than the draft. It costs 4x the time and 4x the VRAM, and it drifts a little further from the source.

All local. ComfyUI, one RTX 5090. RelayVSR via kijai's PR (not merged yet) + the Comfy-Org int8 repack. RTX VSR via the comfyui-rtx-simple node.

Weights: https://t.co/G0iwWEaHpz
Credits: @ComfyUI, kijai (PR #17511), the RelayVSR authors (Xijun Wang et al.), lightx2v for the native 2x model.

[Kiro @kirodotdev]

New in kiro crew: Kiro bots 👻

Self-improving agents for distinct roles. channel watch, issue triage, deploy guard. each builds expertise with every run, with its own memory and parallel sessions, so you delegate processes instead of tasks.

Keep the why with the work.

Full breakdown 👉 https://t.co/gyV7TCqcZq

[Niko Veit @coldniko]

Strata 0.2.11 is now out! 👀

- Supports new Unsloth Qwen3.8-Flash-Next quants
- Layer splitting is now allowed for all configurations
- Performance increase
- App & Terminal improvements

More at: https://t.co/fQVJHq1SfZ

[Gabe @fishjar]

有人排查家里宽带网络异常，抓包发现运营商正在 53 端口上明目张胆地劫持所有未加密的 DNS 解析。

原理非常粗暴：运营商在骨干链路直接挂旁路分光设备，只要检测到发往公共 DNS（比如 8.8.8.8 或 1.1.1.1）的标准 UDP 53 请求，立刻伪造一个带污染 IP 的应答包抢先返回。因为 UDP 无握手验证，本地客户端直接采纳了最快到达的伪造响应，把正规响应静默丢弃。

这种把戏只对明文 UDP 有效。一旦路由网关强行开启基于 TLS 的 DoH（DNS-over-HTTPS）或者 DoT，旁路设备看不懂 SNI 报文就只能原样放行，虚假抢答当场失效。

只要出口协议还是��奔的明文，你的本地网关就不是你自己的。

完整分析：https://t.co/y78SYtjnC1

[Congresswoman Yassamin Ansari @RepYassAnsari]

Top of the Democratic agenda on day 1 needs to be a pause on advanced AI development, regulation of this out of control industry, and a ban on artificial superintelligence.

[Michael Guo @Michaelzsguo]

Jeff Dean 最近在斯坦福大学举办的 “Confronting Our AI Future” 会议上发表闭幕主旨演讲，并与李飞飞展开了一场精彩对谈。

报告的核心是 Jeff Dean 的全新创业项目 Discovery Loop，旨在利用 AI 打造���自动化的科研与工程发现闭环，让 AI 系统自主提出、执行和分析海量实验，大幅缩短原本漫长的科研迭代周期。

两人还深入探讨了机器学习的发展历程，也对最近被热炒的“递归自我改进（RSI）”做了一番深入浅出的解读，厘清了概念和不少常见误解，并分享了不少前沿科研背后鲜为人知的故事。

两位 AI 顶流聊起创业、技术和往事，既有深度又有幽默，贡献了超多好玩的梗：

下属从 4400 暴减到 3 个人
Jeff Dean 曾管理 4,400 人的 Google Research 团队，天天被繁琐的预算会折磨。于是他发了封邮件，把直属汇报砍到了 4 个人。结果下周 Geoffrey Hinton 宣布退休，直接变成了 3 个人。

煎蛋卷与“递归自改进”
李飞飞用做蛋卷解释 AI 概念：从第一次烧焦到第十次能吃，叫自改进；但如果在煎蛋过程中，大脑突触重构、甚至长出第六根手指且越学越快，才叫递归自改进（RSI）。Jeff 补充：人类之所以特别，是因为我们拥有“学会如何去学习”的能力。

咖啡馆里 3 周面试 400 人
创立新公司 Discovery Loop 初期，他们为了抢人，短短 3 周就在大学街的一家小咖啡馆（Cafe Venetia）狂刷了 400 场面试；后来把办公室直接搬进了一家旧餐馆。

被 NeurIPS 拒稿的 4 万引神作
Jeff Dean 和 Hinton ��人写的“知识蒸馏”经典论文，2016 年曾被顶会 NeurIPS 拒稿，理由是“不太可能有影响力”。如今该论文引用已突破 40,000 次，Jeff 调侃应该建个网站专门收录被顶会拒掉的神作。

台上无预告突击小测验
李飞飞台上突然考 Jeff：“2017 年 6 月 12 日是什么日子？” Jeff 秒答：“Transformer 论文上线 arXiv 的日子。” 李飞飞调侃：“你估计 6 月 10 号就看了，毕竟需要你审批。”

[OJI3 @OJI3_NIKKE]

とにかくRGBMocap @toyxyz3 が優秀すぎる
動画からなかなかのクオリティのモーションが短時間で作成できる。
楽しい😀
minimax H3 (ローカル)
H3-Character-Swap-LoRA
#生成AI