
Co-founder @Calibre_Labs | Applied AI research & consulting | Agents, AI Evals | prev EVP @amplitude_HQ | VC @khoslaventures @sequoia | @stanford @iitbombay
Joined May 2009
- Tweets6.8K
- Following382
- Followers9.5K
- Likes7.9K
He wouldn't be getting fired so quickly if it was just an affair .. or even several.
The immediate letting go of one of the most successful CEOs of all time by a board of close allies implies actual crime. They likely all need to avoid the scrutiny and make him less of a target.
And the timing implies something too, given that this was a known issue in the government for so many months. They have a lot more dirt on him.
Still a great entrepreneur who built a great company. Hope Nvidia continues to support the open source ecosystem as well.
👇
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.
289
so promising! self improving agents are definitely going to be the big theme of 2027 for everyone
Introducing Wave.
The world's first AI that reads your customers' mind.
$1.47 trillion will be spent on software this year, but only a fraction of software features are used by customers.
AI can build anything you want, but it doesn't know what your customers need.
Wave fixes that.
1
370
🤯
Introducing Higgsfield Miracle.
Our latest proprietary post-production technology vectorizes any video into layers for localized, pixel-level editing in real time, with world knowledge built in. Every frame holds the image depth needed for cinema-grade post-production.
Now in private beta with select testers. Coming soon to Higgsfield MCP.
Set up Higgsfield MCP in Claude to be the first to try it.
329
This is a massive incentive alignment problem. Imagine you love AI but you are a white collar worker and your wealth is mostly concentrated in your home equity.
Do you want a noisy data center in your neighborhood that hits your property values? Doesn't create any job opportunity for you? May in fact accelerate the automation of your trade.
That's what the data center referendum showed us last night. We can't hold back what's coming to all knowledge work but we need better ideas for how to have data centers benefit everyone that needs to support them
Woof, data centers on the ballot last night got shellacked.
Maybe not that surprising, but as a proxy for how the general public feels about AI right now it is stark
4
474
Don't want to hear any whinging from right wing tech-bro trolls about wokeness and the fate of data centers.
Labor unions are vocally *PRO* data centers now and it's a solution to a bunch of job and infra problems that just needs some finesse. We can have nice things without bombing random countries and pedophile leaders stealing from the poor.
Thank you for your attention to this matter.
So if i am reading this right Anthropic went from profitable to not in Q3 because SpaceX charged them crazy rates for short term compute?
it’s clear their Q4 growth will be insane thanks to Opus 5.5 (which is dominating openrouter like nothing else exists) but i wonder if that puts even more pressure on their compute cost. Wild times ahead.
Between the political gridlock of the next 2 years and AI adoption finally playing out in enterprise it’s going to be a constant volatile rollercoaster in the public markets
5
6
4,612
This was the first month ever that AI-generated games went viral on Roblox and Tiktok!! Huss Valley apparently had over 174k concurrent players at its peak.
Opus 5.5 inflecting code yet again.
m.youtube.com/watch?v=5IU4Bu…
1
672
Mogging all the big labs way to go Lily 🙌
introducing prism, the world’s first speech-to-speech model you can’t tell from a human in conversation
prism has:
> 300ms response time, 4x faster than OpenAI GPT-Live
> outperforms every other s2s model on intelligence (with NO reasoning)
> the first s2s model customizable through reinforcement learning
we’re a team of 35 with $35M in funding, and we’re unlocking s2s model adoption faster than openai, google, and x.
the first 100 people to sign up here get 1 month free: rime.ai/prism?utm_campaign=p…
2
11
886
Having too many people in a company or team creates huge drag for AI software - both creation and adoption.
We are going to see a lot of this in 2027 as public markets start looking at revenue per employee numbers as well as ROIC.
But before you can take any step like this you need to build your company's AI systems and upskill/acquire agent development teams - most software companies still haven't taken these critical steps, it's going to be forced on them by the market rather than in their own time.
Microsoft offers buyouts to 60,000 employees. This latest exit offer is available to far more employees than the previous retirement one, and comes just months after Microsoft laid off 4,800 employees. Full details here 👇theverge.com/tech/1018666/mi…
I was a big fan of the BYOK model for software but this week has made it very clear that storing LLM API Keys with vendors (Opencode, Railway) or even in your own proxy (LiteLLM) just paints a huge target on their back.
So what are we going to do about this? i can see o-auth style protection working for enterprise customers with token permissions scoped to specific vendors and usage types becoming a thing but that’s not going to scale well to the long tail.
Feels like this strengthens the traditional seat based software model with tokens baked into pricing where the vendor owns security. Building your own internal apps or giving vendors keys is just yet another attack surface now.
3
7
759
AI long form writing with a good skill/examples has now gotten good enough with a few models that i actually like having it to do the final-but-one draft.
in a few months (weeks?) pangram scores will be irrelevant. we will still see slop but wether AI was used will no longer be the right flag.
1
4
580
love seeing a new open source success 🫡🇺🇸
We just served 1 trillion tokens of Beam in 24 hours. To celebrate, we've reset everyone's limits.
Go build something big 👉 platform.reflection.ai
1
12
1,269
congrats Hamel! you’ve been such a big inspiration for me and @justinjbauer as we work with clients on AI ❤️
1
5
1,685















