9/ Which is also why many are clamoring for a more open, resilient ecosystem, or even a rebel alliance: a handful of leading providers concentrates too much power — economic and political — in too few hands. xitter.cf/mignano/status/2066535…

It is very important for the entire ai startup ecosystem for key inference vendors to be 1) capable of getting capacity that they can freely allocate, 2) secure, 3) reliable, 4) always driving cost and perf let’s not forget how much of an enabler cpu-era cloud was for innovation! no capex! outsource much of your infra ops (and more every year)! every small-mid customer “owning, building and running their own landed power, cooling qnd chips” completely ignores the lessons of economic history. we had colo 20y ago…public cloud won in an age where closed frontier labs weigh every (limited) GPU served by api vs. internal training and 1st party product use, inference abundance beyond them is a prerequisite for a rich startup ecosystem

Fastest growing startups per Brex data

The Department of Homeland Security is requiring all civilian federal agencies to remove DeepSeek products and 'any third-party services developed using DeepSeek implementations,' and to stop buying them.

a lot of people (rightfully) ask about the long-term value of data companies turns out building efficient, high-quality data pipelines is a science and a dark art unto itself. IMO, that's a big part of where the moat comes from really impressive to see what Proximal is building here

Follow-up: The manipulation is in the marketing. Let me be clear: I don't have anything against Pearl or the team. The cuPOW paper is high-quality research. But there is a massive gap between what the protocol actually does and what the shillers claim. Claim 1: "Proof of Useful Work" What they say: Pearl is a "Proof of Useful Work" network. The mining work is actually useful, it does AI inference. The cuPOW paper by Komargodski and Weinstein is the mathematical foundation for Pearl's mining. Its goal is to make sure that miners can't cheat by doing something too easy. The paper says the protocol should produce a "proof of work" that has "prescribed hardness" — meaning the work must be hard enough to be a real proof, and a miner can't find a shortcut that lets them win without doing the hard work. That part is about hardness. The work must be hard. It doesn't say the work must be useful. This prevents shortcuts like: Submitting all-zero matrices, Reusing pre-computed matrices, Exploiting low-rank or structured matrices that cost less than full n³ computation. The Key Distinction: Hardness vs. Utility Computational Hardness: The miner must do a lot of computation. They can't fake it or skip it. This is what the protocol guarantees. Computational Utility: The computation must produce something valuable like a real AI inference that someone actually needs. This is not guaranteed. The paper itself admits this. It says miners choose their own input matrices. The protocol only checks that the matrix multiplication was done correctly. It does not check that the matrices came from a real AI model or a paying customer. Usefulness is left as an "economic hope," not a cryptographic guarantee. The "usefulness" is supposed to come from economic incentives: miners with real AI customers would earn extra money, so they'd choose to do useful work. But loading real model weights costs more than generating random numbers. So the rational miner chooses the cheaper path: garbage mining. Claim 2: "Pearl has an AI inference lane" What they say: Pearl is a blockchain where mining is done as a byproduct of AI inference. What's actually happening: The AI inference isn't running on Pearl's network. It runs on Together AI's servers. Together AI is a separate company with its own GPUs. Pearl's miners don't run inference. They grind matrix multiplication on random numbers. Pearl's role is to subsidize Together AI's inference costs using PRL emissions. Together AI offers a 25% discount to its customers, and Pearl pays for that discount with newly minted PRL tokens. Claim 3: "Pearl is AI Bitcoin" What they say: Pearl is the "AI-native Bitcoin", a Bitcoin fork that uses AI computation instead of wasted hashing. What's actually happening: Pearl is a fork of Bitcoin's code. It swapped SHA-256 for matrix multiplication. That's it. There's no AI in the protocol. There's no inference. There's no model. There's no verification of useful work. The "AI" is a marketing layer. It's borrowed from a third party (Together AI). It's not native to Pearl. If Together AI disappeared tomorrow, Pearl would still mine matrices, but the "AI" narrative would vanish. Claim 4: "Duplex mining" (miners do AI inference + mining at the same time) What they say: The same GPU computation serves an AI customer and mines Pearl blocks. Two rewards for one job. What's actually happening: The protocol cannot tell if a miner is doing inference or just grinding random matrices. It only verifies that a matrix multiplication was done correctly. It doesn't verify that the matrices came from a real model. So a rational miner will skip the inference and submit random matrices. Loading model weights costs memory I/O. Generating random numbers is cheaper. The "duplex" part is optional. The Pass paper (Pearl's own economic defense) admits in its own text that pure mining is 23% cheaper than duplex mining. And here's the part nobody talks about: the so-called duplex mining doesn't secure the chain while doing AI. Even if a miner somehow did both ran real inference and mined a block the protocol still wouldn't verify that the inference happened. The block is secured by the matrix multiplication, not by the inference. The AI work is parallel to the mining, not part of it. The AI happens on Together AI's servers. The mining happens on Pearl's network. The two are linked only by a financial subsidy , PRL emissions paying for a discount on Together AI's inference. There's no cryptographic connection between them. So "duplex mining" isn't securing the chain with AI. It's securing the chain with matrix multiplication while AI happens somewhere else entirely. The "duplex" is a billing arrangement, not a consensus mechanism. pearl-2:native should be called "Proof of Computational Hardness." That's what the paper guarantees, what the network enforces, and what the miners actually produce. "Proof of Useful Work" and "AI native Bitcoin" are marketing terms, not protocol features. The Bubble Is Built on an Assumption The shillers are selling a narrative the protocol doesn't support. They assume miners will choose useful work out of economic incentive. But the protocol doesn't enforce it. And when the cheaper path is always available, rational miners take it. The bubble is this: a token whose value depends on "useful work" that the protocol cannot verify. When the gap becomes obvious and the Cornell study already made it obvious, the narrative bursts. And who's left holding the bag? The shillers who sold the story, and the retail who bought it. "The pigeon saw the cat. Then it closed its eyes. There are none so blind as those who will not see."

I told Claude to make a video about what it's like to be European One day later...

Extropic designed the tasks and the reward, while Prime Intellect handled the RL infrastructure: 1. verifiers to build the environment 2. Hosted Training to run the RL loop 3. Prime Sandboxes to execute the model's code and return the execution score 4. Prime Inference to serve the LLM judge and frontier baselines