Thanh Ha

I'm here for the entertainment.

[Thanh Ha @zxiiro]

Finally got Grok Build and Claude Code to read the same agents.md in a project repo. Grok build seems to not load project level AGENTS.md even when instructed by it's global AGENTS.md. ended up having to write a shell wrapper to pass the rules parameter as a workaround. Hope the @grok team fixes that in the future.

I was able to get grok 4.5 to troubleshoot and submit feedback in the app which was pretty cool.

[PyTorch @PyTorch]

PyTorch 2.13 is here, with 3,328 commits from 526 contributors and updates across FlexAttention, CuTeDSL, nn.LinearCrossEntropyLoss, torchcomms, FSDP2, Python 3.15 wheels, ROCm, Arm, and XPU.

The release blog and notes cover FlexAttention on Apple Silicon with up to ~12x speedup over SDPA on sparse patterns, a deterministic backward path on CUDA, the CuTeDSL "Native DSL" backend for Inductor, nn.LinearCrossEntropyLoss to reduce peak GPU memory by up to 4x, torchcomms for large-cluster training, and FSDP2 communication overlap improvements.

On July 22 at 11 a.m. PT, join @albanDesmaison (@Meta), Andrey Talman (@Meta), Piotr Bialecki (@NVIDIA), and Chris Gottbrath for a live 2.13 Q&A.

🔗 Read the release blog, and register for the live Q&A: https://t.co/FQUCLjjWAI

[Thanh Ha @zxiiro]

Took shelter at IKEA and the water was coming out of the sewer system. #onstorm

[Aravind Sithamparapillai @AravindSitham]

I've had a few experiences where the wife of a dying husband reached out to say they would like to use me as their planner...after the husband passes away.

I have offered to review the plan in advance but they are certain their spouse and/or advisor had it covered.

👇

[Thanh Ha @zxiiro]

@valvesoftware Nice! I signed up.

[PyTorch @PyTorch]

The speed-of-light optimization for Qwen3.5 on the TokenSpeed inference engine is a significant milestone, achieving a record-breaking 580 tokens per second (tps) for agentic workloads on NVIDIA GPUs.

In the PyTorch Foundation's latest community blog post, you can learn all about the complete design, implementation, and optimization of Qwen3.5 models in the TokenSpeed inference framework and see for yourself how this work is improving performance 👉 https://t.co/Qr1PTIhqok

This achievement was a joint effort between the @Alibaba_Qwen inference team, @lightseekorg Foundation TokenSpeed team, @NVIDIAAI , and the Mooncake team, with special contributions from @tri_dao for FlashAttention-4 (FA4) optimization. @KVCache_AI

[PyTorch @PyTorch]

We’re excited to welcome @alibaba_cloud as a Platinum Member of the PyTorch Foundation 🎉

Alibaba Cloud is a global leader in full-stack AI infrastructure and the force behind Qwen—one of the world’s most influential open-weight model families. Having run PyTorch at massive scale across diverse hardware, they bring invaluable, production-hardened engineering expertise to the upstream community.

[Thanh Ha @zxiiro]

@CryptooIndia I ran into this! Glad it's not just me. 😂

[Thanh Ha @zxiiro]

@JJ_McCullough @sethesoup @LionBlogosphere Fwiw I like coffee. I don't like to put sugar/cream or any other things added to it. I even only drink decaf coffee as I prefer to avoid caffeine.

[Thanh Ha @zxiiro]

@FoodProfessor I didn't get mine yet.

[Thanh Ha @zxiiro]

I've been creating markdown files that list all the work I want it to do. I do all the planning ahead of time during the week creating spec files and linking them to the work dock. Instruct it to create separate commits or PRs for each task.

On the weekend before I head out, I ask it to tell me what tools and approvals it needs to work on the list and let it run while I'm out with my family.

Been able to get it to work uninterrupted while I was out a couple of times. Definitely better than sitting at a terminal watching it work.

[Thanh Ha @zxiiro]

@ZainRzv Unfortunately so.

[Thanh Ha @zxiiro]

@ezyang @LaithSakka Good to see this going into a more public space!

[Edward Z. Yang @ezyang]

Thanks to @LaithSakka, we now have a shared developer log at https://t.co/ZlLiuMhFLE on all sorts of PyTorch things. The way to think about it: classic Meta culture is to build things and then post about them in the internal Workplace. Now we ask people to repost them here!

[Thanh Ha @zxiiro]

@_seemethere Sounds like you need a 2nd account. 😬

[Chris Wood @CWood_sdf]

i love how people are saying "if we write a sufficiently detailed specification, the agent can write all our code"

do you know what writing a sufficiently detailed specification that deterministically maps to what a computer's actions is? it's coding

[Mark Collier 柯理怀 @sparkycollier]

The @PyTorch Foundation is leading the way on open source training, inference, and agents

iykyk

[SteamDB @SteamDB]

Steam Controller and Puck CAD files now available! https://t.co/RrvcXQ94EN

[Thanh Ha @zxiiro]

@cra If you use Claude Code in Terminal you can create bash aliases to open Claude with different .claude config directories. It's clunky and I wish they had native support.

[Thanh Ha @zxiiro]

Looking forward to listening to this episode!

[Mark McGrath @MarkMcGrathCFP]

One day my kids will understand how cool it is that I got to hang out with Dave not once, but twice.

So far