Suri

Product Builder &AI PM @ Microsoft Copilot
ex-googler SWE➡️PM

[Thorsten Ball @thorstenball]

Finally summarized what I've been sending to friends in DMs https://t.co/rnf3M16NzI

[Suri @suriiiii123]

@felixrieseberg wow this is definitely the top to my reading list this week!

[Suri @suriiiii123]

@FishAudio Congrats🥳

[Fish Audio @FishAudio]

$52M seed round. 15M+ users. But this is what we're most proud of yet.

Introducing Drama 3: the most emotive and controllable TTS model in the world

And yes, it's open-weight.

You can now control how even a single word should sound, change its tone, and shift your voice mid-sentence.

Uncle Roger was impressed too.

Try it here:

https://t.co/4nbz4sQ4yy

Comment “DRAMA” within the next 48 hours for 1 month free.

If you're an enterprise your first month is on us:

https://t.co/jBXXZSAg0p

[Suri @suriiiii123]

@CollovLabs this is huge!!

[Suri @suriiiii123]

@VincentWu11 @HuffPost @YahooNews @newsbreakApp Cannot wait to start using it!

[Suri @suriiiii123]

this is insane...

[beamnxw ./ @beamnxw]

Andrej Karpathy (ex-OpenAI, ex-Tesla) just released a prompt that turns Opus 5.5 into a harness that actually finishes the any job.

It's f*cking unreal...

send it to Opus 5.5, and thank me later. Then read the full guide for Opus 5.5 Harness Engineering.

[Suri @suriiiii123]

@baseten Start reading today🙌

[Suri @suriiiii123]

@shengkunye 🥳 this is huge for efficient context management

[Suri @suriiiii123]

@theworldlabs @AMD Really interesting move. The AI infra race is increasingly moving beyond “who has the best chips” to “who understands where the next generation of models is going.” World models + physical AI could be a very strategic wedge for AMD

[Suri @suriiiii123]

Really interesting move. The AI infra race is increasingly moving beyond “who has the best chips” to “who understands where the next generation of models is going.” World models + physical AI could be a very strategic wedge for AMD.

[World Labs @theworldlabs]

We are excited to announce that World Labs is joining @AMD.

The research and technical breakthroughs we have achieved since our founding in 2024 have given us a clear vision for AI’s potential to solve problems in the spatial and physical world.

Accelerating the future of spatial and physical intelligence requires scaling our efforts, scaling our reach, and getting closer to the hardware.

[Max For AI @MaxForAI]

🔥太疯狂了，现在你甚至可以直播围观一个 535B 大模型是怎么被训练出来的。

斯坦福教授兼 @simile_ai 创始人 @percyliang 刚刚宣布，他们的新模型 Marin 535B-A23B 本周正式开训，而且整个训练过程公开。

Marin 共有 535B 总参数，其中 23B 激活参数。

为此Percy他们一共准备了18.75T tokens的数据，动用了 11 套 GB200 NVL72，约 792 张 GB200显卡。

预计连续训练 3 个月，总训练计算量约 2.7e24 FLOPs。

训完还会继续做 post-training。

请注意，Marin 真的全部开源了整个训练过程。

正式开跑前，他们先训练了一套 4 级 Scaling Ladder，从 1.6B-A61M 一路跑到 27.7B-A1.2B，用小模型提前预测这次 535B 主模型的 loss 和训练状态。

现在主模型正式开跑，所有人都可以直接围观实时训练曲线，后面的数据、实验记录、工程问题也会持续公开。

类似的事情以前有人做过。

2022 年 BigScience 训练 176B 的 BLOOM 时，就公开过训练进度和 TensorBoard。

但 Marin 这次直接把公开训练的规模推到了 535B。

目前来看，这应该是人类公开围观过的最大规模大模型训练之一。

接下来三个月应该会很好看。

模型最后能不能打是一回事，我更想看的是一个 500B+ 模型的MoE到底会怎么训、怎么调，以及中间到底会炸几次。

这种机会太少见了，等于把一场原本只有顶级 AI Lab 内部才能看到、成本可能高达数千万美元的训练实验，直接摊在全世界面前。

[Suri @suriiiii123]

Kimi’s open-source strategy is more interesting than “give the model away for free.”

The playbook:

Open model → developer adoption → ecosystem → API/inference → revenue.

And with Kimi K3, large MaaS providers crossing certain thresholds need a commercial agreement with Moonshot — giving it a way to participate in downstream upside.

Open source isn’t the monetization strategy. It’s the distribution strategy.

[Suri @suriiiii123]

@realYunfanYe Congrats👏

[Xudong Han @Xudong07452910]

Agent 出错以后，到底该修模型，还是修 harness？

Scale AI 这篇《Model or Harness?》专门研究 Agent 的故障定位。

今天的 Agent 已经是一个复杂系统：模型之外，还有 Context、Memory、Tool、Grader、用户和运行环境。最后看到的失败结果相同，背后的原因可能完全不同。

比如 Agent 忽略了一条早期指令。

可能是上下文压缩把它删掉了，需要改 harness；也可能信息一直都在，只是模型没有正确使用，这时才该训练模型。

论文整理了 41 类常见失败，并把每个问题定位到具体的「组件交互 + 责任方」。

这样一次失败就能进一步回答：

该做模型 post-training，改 context / memory / tool，还是重新设计环境和评测。

作者还让不同前沿模型独立给失败案例分类，最好的结果与人工标注达到 Cohen’s κ = 0.76，说明这套分类有一定一致性。

我觉得这篇很适合现在的 Agent 发展阶段。

随着 harness 越做越复杂，「Agent 失败了」这个结论已经太粗了。

真正有用的 debugging，要继续追到第一处无法恢复的错误，再决定到底该修哪一层。

📎 arxiv: https://t.co/lEitVGWyKu

[Suri @suriiiii123]

@laura_llin Cannot wait to try it out!🥳

[Suri @suriiiii123]

@yanliudreamer @ninthbit_ai @VigoCreativeAI @Formulasearch 🙋‍♀️

[Suri @suriiiii123]

Cursor used Kimi to help build Composer 2. DoorDash delegates lower-level work to K2.6. Thinking Machines used K2.5 to generate early post-training data for its Inkling model. Nobody’s customer-facing brand says “Kimi,” but it’s in the supply chain.

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Suri @suriiiii123]

@hearddotdev this is awesome！！

[Suri @suriiiii123]

@0xF1ction Cannot agree with this analogy anymore!