Ace Eldeib

[Ace Eldeib @alexeldeib]

@PatrickToulme Curious, what’s the intuition on why?

[CoreWeave @CoreWeave]

Kimi K2.6 from @Kimi_Moonshot is purpose-built for coding agents.

As of today, CoreWeave ranked highest in @ArtificialAnlys’s inference benchmark on Speed vs. Price for K2.6.

Speed, scale, and economics. All three at production grade.

[Anne Ouyang @anneouyang]

Excited to share @Standard_Kernel's seed round and some reflections on what we’ve learned about kernel generation and what we believe is next. Grateful to our amazing team, supporters, and the broader community pushing this space forward.

[Aaron @aaronbatilo]

All coding models need to upscale this book like 800x

[Ace Eldeib @alexeldeib]

@bcherny How about no-repo? Or create one on demand like v0/lovable when the user wants to? I’d think you can get away with saving state somewhere else until then. Nice for greenfield projects.

[Ace Eldeib @alexeldeib]

@StasBekman @modal CodeSandbox might be close to what you have in mind, ignoring implementation the UX/SDK side of "sandboxes" also looks to be converging somewhat

https://t.co/eplfnWrgyf
https://t.co/pdcmnWmct6
https://t.co/oKmWSCWiJm
https://t.co/006PZbAnze

[Ace Eldeib @alexeldeib]

@StasBekman @modal If you want persistent filesystem state, you can either snapshot the entire rootfs, or mount a separate blockdev/filesystem. Hard to fit in a tweet but some useful links

https://t.co/cmCeOjT2XH
https://t.co/eaJACvSUM2
https://t.co/SY6vOCZUKa
https://t.co/4KRjmVqFHU

[Ace Eldeib @alexeldeib]

@StasBekman @modal single node simple process trees I think are fine on (4), but if you’re doing multinode or collective comms you’ll probably have a bad time. Check the CUDA checkpoint repo issues around NCCL.

[Ace Eldeib @alexeldeib]

@StasBekman @modal Not modal but my 2c

1. Yes. CUDA offloads from GPU, gvisor (or generically, criu) does the rest.

2. Can be done by mounting code rather than VM snapshotting? But you can make edits and snapshot the filesystem too

4. Multinode not supported afaik.

[Ace Eldeib @alexeldeib]

using this video has me cackling. peak niche gpu crossover meme

[SemiAnalysis @SemiAnalysis_]

Since RCCL is an fork of NCCL, RCCL is basically a copy+paste carbon copy of NCCL except it takes months for new features added to NCCL to reach RCCL. This is clearly not optimal for AMD to be building on their competitors platform as that means that AMD will never be better than NVIDIA or even reach parity (when iso-time) due to the delays/engineering burden of syncing with upstream.

AMD is working on an moonshot project cuz MORI-CCL aims to be an first principles from scratch re-built of the AMD collective library software to not be dependent of their competitor's software. Ironically enough with MORI-CCL, it currently doesn't support AMD's Pensando NICs yet it supports NVIDIA ConnectX-7 NICs with AMD GPUs. Support for AMD's Pensando NICs is coming after support for NVIDIA ConnectX-7 NICs.

[Ace Eldeib @alexeldeib]

@clattner_llvm @value_and_grad useful refs:

https://t.co/P8YuxR05jR
https://t.co/Yk0YjxvvHS
https://t.co/3pYYi6SVb1

Looking at package versions/filenames from https://t.co/uiVSYtl8TG
https://t.co/jqagFHMIpO
https://t.co/SxmYlSGpob

[Ace Eldeib @alexeldeib]

@clattner_llvm @value_and_grad Looks like mojo pushes manylinux 2_34 wheels to PyPi, it will be incompatible with anything < glibc 2.34. jax by comparison publishes manylinux_2_27 on PyPi, and Google pushes manylinux2014 cuda builds (compat back to Ubuntu 13)

[Ace Eldeib @alexeldeib]

@aaronbatilo Skill issue

[Ace Eldeib @alexeldeib]

@CaiGuy64 @TohaSaeto I’ll bite, what event?

[Ace Eldeib @alexeldeib]

@ayaboch Claude code is nice, but seems to suck at frontend for me. ChatGPT for the spec, v0 for the UI, Claude to build the rest. Lovable has superpowers for integrating with supabase for full apps, but once you hit a certain level of complexity, it’s easier to iterate locally.

[Ace Eldeib @alexeldeib]

@VictorTaelin Is this on github or something? Where do we learn more?

[Ace Eldeib @alexeldeib]

@zanehengsperger @joinindustrion How do you find them? Where’d you get this data?

[Ace Eldeib @alexeldeib]

@blelbach @JokerEph Will this one be recorded? Looks like yes? Super interested in cuTile and/or the new CUTLASS Python integrations but leaving early :(

[Ace Eldeib @alexeldeib]

@AdjectiveAlli @AnushElangovan Ah they used the 94GB? https://t.co/KJUXLuhPZ0

[TNG Technology Consulting GmbH @tngtech]

Inference test - @AMD MI300X beats @nvidia H100

We got access to 8x MI300X - 192 GB each and therefore 1.5 terabyte of vRAM in total - provided by @Dell

Running @deepseek_ai R1, we compared inference metrics against 8x H100 - 94 GB each

Here are the results 🚀

1/n

[Ace Eldeib @alexeldeib]

@AdjectiveAlli @AnushElangovan right, how does it fit otherwise? The comparison for H200 still looks favorable but far less extreme https://t.co/hCcK6f6kJa

[Zev Rekhter @zevrekhter]

@casper_hansen_ @lmsysorg @GenAI_is_real @zhyncs42 @deepseek_ai @AnushElangovan @dstackai has a great post https://t.co/q7keOHApgT