
Building @SuperMeAI • STR investor • a life time ago founded @thermodo @robocat • AMA https://xitter.gq/t.co/VDJqTJdzPo
San Francisco, CA
Joined January 2008
- Tweets3.4K
- Following1.9K
- Followers1.9K
- Likes1.9K
I had made my personal site as a bookshelf williwu.com
1
1
274
willi wu retweeted
Websites are going to need a proper front door for agents soon. Human Security saw agentic browser traffic grow 8x from August to October, with Muse driving 60% of it and OpenAI products 30% (via a16z)
willi wu retweeted
This is one of our first deep dives on some of the technology we’ve built behind SuperMe. Excited to share what @maxakhterov and the team have been cooking.
In this blog post we dive into what makes your agent relevant and qualified to be pulled into a conversation.
A core principle for our network is trust. What that means is that if your agent can’t ground in your content it should not answer and we shouldn’t pull it in a conversation where it won’t be successful. This puts very interesting and challenging engineering constraints on how we design the system.
We have about 2.5 seconds to pick 5 people from thousands of relevant profiles for your question.
@maxakhterov breaks down how our search engine works, and why a shiny logo on a resume doesn’t always mean relevant expertise.
Read it here: superme.ai/blog/expert-searc…
👀
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 unemploymentlabs.com/checkpo….
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.
96
Max is spilling the beans on our AI group chats.
We have about 2.5 seconds to pick 5 people from thousands of relevant profiles for your question.
@maxakhterov breaks down how our search engine works, and why a shiny logo on a resume doesn’t always mean relevant expertise.
Read it here: superme.ai/blog/expert-searc…
3
128
I used Antigravity and Gemini 4 Carbon. Antigravity is slow but very thorough and checking all the files.
I ported Deskland to macOS. Interactive wallpapers that adapt to your desktop icons. Swift and Metal running 120 Hz. Supporting widgets.
github.com/willi/deskland-ma…
2
217
I ported Deskland to macOS. Interactive wallpapers that adapt to your desktop icons. Swift and Metal running 120 Hz. Supporting widgets.
github.com/willi/deskland-ma…
Compiled PHP still rules tho
Rust is not faster than JavaScript!
Rust is also not faster than C++.
The compiled binary code might be faster than the compiled binary code of either of the others. But that's a compiler question, not a language question.
Some languages make it easier to compile to performant code than others, but the trade-off is in the language itself.
Don't confuse language with execution.
1
141
Finally Xcode on iPad — oh wait
Henge 0.1 is out, and with it allows you to run macOS on iOS/iPadOS!
This has been a project of mine for about 2 weeks, and I’m very happy to finally release it! It’s in a good state, and ready to release.
Download: github.com/Leviidev/Henge/re…
1
113
willi wu retweeted
What if you could watch people who have been there debate your question instead of getting a single AI-generated answer?
With SuperMe, ask a question and watch different perspectives challenge each other in a group chat. See where they agree, where they push back, and what trade-offs emerge. Or come back later and read the summary.
Save the people whose perspectives you value to a work group, so they're ready the next time you need to think something through.
Whether you need a confidential gut check or help pressure-testing a decision, SuperMe helps you see the problem from angles you might have missed.
Ask the network a question today.
willi wu retweeted
You shouldn't need the best network to get the best insights.
SuperMe turns operators' work into profiles that work through your problem. Now in group chats, Claude and ChatGPT.
@ludo_antonov and I raised a $20M Series A led by @kirstenagreen from @ForerunnerVC with participation from @GreylockVC.
Read more about the changes here: superme.ai/blog/series-a
Want
Mac Studio With M5 Ultra Now Available With 512GB RAM for $22,299 macrumors.com/2026/10/28/mac…
1
127
willi wu retweeted
Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.








































