tushar

making compilers grin, building distributed systems, previous @ihc__official, research @iiit_hyderabad

[tushar @gpu004]

@ptr_to_joel @TigerBeetleDB lately i have doing small implementation of casandra paxos lwts, tigerbeetle viewstamped replication, cockroachdb raft per replica group

playing around with different consensus ideas

[tushar @gpu004]

I was trying it on different way I quoted websites I created

I think I haven't touched bombardil in 2026 so it was all in 2025 only

Let's say I want to test or run bombardil on top of t3 code, to find different bugs or issues in there react native app or the desktop GUI thing

How should I approach it

[tushar @gpu004]

Hi Oskar, I'm a really big fan of antithesis and have been using Hegel property testing a lot. I love using it. It's an awesome solution.

But the idea of Bombadil is yet to click for me. I've tried using it but I wasn't able to do anything good with it. Maybe I'm not using it correctly or something like that. Please if you can, give a small intro, video, or explanation on how to use Bombadil or what exact problem it solves. I'm yet to figure it out. I really want to but I'm not able to.

I'm a really big fan of Antithesis blogs and Hegel and all the different solutions and talks there but this idea, this specific project, has yet to click for me. Can you please help and explain or click it for me?
@owickstrom

[tushar @gpu004]

@danellisona hell yeah

[tushar @gpu004]

i understand hegel decently, i think hegel as a framework for testing the property of some piece of code or function, which feels way better than writing unit tests for that piece of code, kind of like the fuzzzing idea passing a distribution of inputs to our function to test if this function is following that property properly or not, passing the hypothesis created distribution to our function

i tried to approach bombadil, thinking it's kind of same thing but from frontend, that's it,

i think my mental model is wrong about bombadil (btw, super cool name)

[tushar @gpu004]

@yakkomajuri there blogs and ai engineer talks?

[Jelena @Feralfairy13]

I think the secret to happiness is just becoming a morning person

[Aram Hăvărneanu @aramh]

Language design starts with semantics, not syntax.

[Horace He @cHHillee]

I'll be giving a talk tomorrow at the SGLang Summit at 10am!

Oftentimes, I find explanations in ML infra to be unsatisfying. A common failure mode is that they'll explain the *literal* tradeoffs but fail to provide a mental model for how to reason about the choice.

This talk won't explain everything about inference (only 15 minutes!), but I'll provide two mental models I've found useful when reasoning about inference, and I hope you find them useful too :)

[tushar @gpu004]

they can rent but us will have full control over those rents and they can cancel of that rented capacity instantly if they want to

us will have full control over that kill switch whenever they want

plus so much money willl follow to us companies who will be renting this to chinese companies,

oracle and nvidia ig are lobbying for this alot lately

[tushar @gpu004]

@joeksluis great work man

[tushar @gpu004]

@sarthak811kohli @anieasyy

[tushar @gpu004]

@chaaai It reminds me of this article

[tushar @gpu004]

Sounds pretty plausible

[tushar @gpu004]

SpaceX and Tesla don't function without Elon, Nvidia doesn't function without Jensen.

[tushar @gpu004]

He was at the helm for 33 years, and even though the board pressured him out (or he “resigned”) Nvidia would still be materially better off with him there instead of gone. The company was uniquely built to suit the man himself. The company without the man might be an entirely different beast.

[will depue @willdepue]

I have a major result to share.

I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.

This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.

Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.

I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.

I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.

It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.

These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.

I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at https://t.co/YOtzND0Jqk.

There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

[tushar @gpu004]

but after years of burning my head on aws console, the developer experience have improved such magically, everyone is improving alot, competing with each other

i am enjoying this competition and new innovation world

currently modal sits at the top of developer experience accoridng to me

[zack @zack_overflow]

It actually will be studied in history how well Sam played the coup

Literally not only was able to come back

But able to fire literally everyone on the board who sided against him

Was able to gain even more power