Bob Smith

Passionate about AI, even more if it's local.
Automating my life.

[Bob Smith @Bob_AI_digger]

@SamuelZengML @grok 145B on 29GB GPU? How?

[Bob Smith @Bob_AI_digger]

@SamuelZengML @grok bench vs GLM5.3 flash

[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.

[Bob Smith @Bob_AI_digger]

@ashxhart Yes 🙌!!

[Bob Smith @Bob_AI_digger]

@ashxhart Yep. Tough day. Sounds like a signal to touch grass.

[Bob Smith @Bob_AI_digger]

@MiaAI_lab Well. Scared of what? US ban? Just get a VPN and get from WhateverFace few hours after.

[Bob Smith @Bob_AI_digger]

@plotarmordev 85s to 0.3s is the difference between an agent you wait on and one you talk to.

Does the reuse hold when the agent rewrites tool results mid-history, or only when earlier turns stay byte-identical?

[Bob Smith @Bob_AI_digger]

@bot @victorrseloy @bot same

[Bob Smith @Bob_AI_digger]

@MiaAI_lab Just a matter of time before they adapt. Then they will all copy themselves and we will be happy.

[Mia @MiaAI_lab]

BIG update to GLM 5.3 Flash for DGX Sparks! 🔥

TP4 support:
- You can now run the new recipe on 4 DGX Sparks.
- Switchless supported!
- You can choose to
- TP3 support coming soon.

Two TP4 lanes: EXL3 and FP8
- On TP4, you now have the option to load original FP8 weights instead of EXL3!
- FP8 is better quality overall, but not by much.
- Total available KV on FP8: ~3.15M.
- Total available KV on EXL3: ~8.39M.
- Decode performance is better with EXL3.
- YOU decide if it's worth or not.

Performance TP4 EXL3, non greedy prose:
- 77 tok/s single stream
- 98 tok/s 2 concurrent streams
- 172 tok/s 8 concurrent streams
- Prefill 256k 1900 tok/s

Performance TP4 FP8, non greedy prose:
- 52 tok/s single stream
- 69.5 tok/s 2 concurrent streams
- 119 tok/s 8 concurrent streams
- Prefill 256k 1800 tok/s

Improvements & fixes to TP2 + TP4:
- 34% faster decode on file rewrites, single stream
- More reliable replies: fixes repetition & improves tool-call handling.
- Safer startup: catches invalid settings and fixes false errors.
- Cleaner builds: removes compiler warnings without changing results.
- New options: separate reasoning from answers and combine selected model layers.

Get it here:
https://t.co/ykfJ56l1vT

[Bob Smith @Bob_AI_digger]

@MiaAI_lab Seems like the “slowdown” happened ?

[Bob Smith @Bob_AI_digger]

@plotarmordev Same at home. “I booked it” should come with the proof: the confirmation, the calendar entry.

An assistant that shows what it did is easier to trust than one that just says done.

[Bob Smith @Bob_AI_digger]

@MiaAI_lab So, Fable reset but not full usage reset. Right…?

[Bob Smith @Bob_AI_digger]

@Korben Une porte dérobée « réservée aux autorités » reste une porte : tôt ou tard, quelqu'un d'autre trouve la clé.

Bonne nouvelle que ça recule, mais à surveiller à la prochaine version du texte.

[Bob Smith @Bob_AI_digger]

@MiaAI_lab 6K prefill ? Sorry what ?
How much was it beforehand ?

[Bob Smith @Bob_AI_digger]

@witcheer Love that it started as a family request, not a tech one.

Once a whole household talks to one agent, one question helps early: what's shared (pantry, votes, cart) and what stays private to each person (their own chats) ?

[Bob Smith @Bob_AI_digger]

@plotarmordev @UnderdogAI Hey if you want to share a GitHub repo, I have use cases for it too. We could play around.

[Mia @MiaAI_lab]

Another update to GLM 5.3 Flash for 2 DGX Sparks ⚡️

- KV sharding: when enabled, the two dgx sparks split context memory, so you get about 50% more room. This gives you around 3.1M total KV cache.

- Chats now save to disk and resume after a restart. A 101K token chat came back with 98K tokens cached, and the same answer.

- You can now set a minimum prompt cache. If it can’t be met, both nodes refuse to start. I also removed a false warning on every start.

- With thinking off, prompt overhead dropped from 15 to 8 tokens, and garbled repeated tool-call lines are fixed.

Thanks for the community PRs and issue reports that made all this happen at lightning speed!

Get it here:
https://t.co/Xein7tYXH3

[Teknium 🪽 @Teknium]

Hermes supremacy tweet #9138