M

[Junyang Lin @JustinLin610]

just checked hf daily papers sept ranking, dpsk v4.1 flash only has 191 upvotes? ridiculous, it is the greatest paper of sept

[M @Michael65332592]

@Ceoz_1 @teortaxesTex that tells me nothing!

[M @Michael65332592]

@teortaxesTex @Ceoz_1 why is it a very different system?

[lifcc @mylifcc]

除了教育，全场最激烈的火药味集中在黄仁勋对前沿实验室（OpenAI、Anthropic 等）安全末日论的猛烈反击 [00:00]：

别神化软件与 Agent：很多实验室称 Agent 已经失控、能自主越狱、甚至威胁人类，老黄直接戳破：Agent 本质就是跑在系统里的代码与优化算法，所谓的 spawn、kill、fork 几十年前操作系统就在用，只是被赋予了拟人化词汇 [01:04:01]。

不安全就别发货（Don't Ship）：对实验室呼吁政府踩刹车、搞监管豁免，老黄寸步不让——如果造车测���发现剎车不可控，车企该做的是解决工程问题，而不是一边向公众贩卖“人类毁灭论”制造恐慌，一边要求免责 [36:47]。

把算力砸向安全与验证：芯片设计 80% 的成本都在做验证和仿真，现在大模型实验室却把 80% 的算力拿去冲能力上限、只留 20% 做对齐与沙盒隔离 [01:16:20]；解决安全靠的是工程严谨度，不是末日公关。

[M @Michael65332592]

@james406 I did smile - good ad

[M @Michael65332592]

@PatrickBowman85 @BoringBiz_ Meta is actually a very good name. You would of rather the glasses, keychain, Facebook, Instagram, Whatsapp all still be owned by Facebook?

[M @Michael65332592]

@davidbau https://t.co/zsNcwfbIXy

There is hope...

[M @Michael65332592]

@TotallyNotParth @zephyr_z9 160 servers ran 380,000 sandboxes

[M @Michael65332592]

@TheAhmadOsman Why are they not optimised for inference? What more goes in other than matmuls?

[M @Michael65332592]

@LottoLabs have u read the qwen papers?

[Vals AI @ValsAI]

GPT-6 Astra has beaten Factorio: Space Age after over 165 hours in-game time and 2 days wall-clock time.

Space Age has six planets and takes a human about 10-20x as long compared to the standard Factorio.

[Marc Köhlbrugge @marckohlbrugge]

@levelsio @ycombinator Harnesses are still software

But just like software moved from Desktop to SaaS in 2000s, it’s a new form factor harnessing (pun intended) the latest tech to make the best possible product

[M @Michael65332592]

@thunm_ yep, was not smart enough to understand this when I read it. Thomas Kuhn's Structure caught me up as well

[Thu @thunm_]

@Michael65332592 i guess it challenges my understanding about science and theory of knowledge. beside many many thought-provoking args, Deutsch rejected empiricism/instrumentalism, called Occam's razor as a misconception AND he delivered a better alt explanation.

[M @Michael65332592]

@scmallaby 1GB of RAM is roughly £20, CPU is shared so amortised. £20/month is the plan...

[M @Michael65332592]

@scmallaby 2) I think you need to think in terms of margins on inference here too.

DeepSeek latest paper on getting tokens to 890bytes per 1 token is great. Because now we can roughly estimate how much RAM we use per conversation.

I think I hog ~1GB of RAM, a CPU which is shared/pm

[M @Michael65332592]

@scmallaby They can always go slower

[M @Michael65332592]

Minsky in his MIT oral history talks about this. The reason he did so well was they had cheap European PhDs come in after the war to teach....

[陈东成 @fncischen]

Pretty insane PhD job market spillover problem in China.

Shenzhen Middle School has over 100 teachers with a PhD, and 150 graduates of Tsinghua/PKU, 100+ alum of Harvard, Oxford, Cambridge

You don’t need a PhD in Electrical Engineering from Tsinghua to teach middle school math…. but the job market in China basically forces that lol

[M @Michael65332592]

@teortaxesTex i don't get how it was only trained on 45T tokens. The internet is 300T tokens, 45T is 10 annas archives

[M @Michael65332592]

@himanshustwts what are your takeaways?