[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]  Sweep [@0xSweep](/creator/twitter/0xSweep) on x 221K followers Created: 2025-07-21 14:01:54 UTC A team from Columbia just did the impossible. They proved @VitalikButerin wrong. They found a way to turn matrix multiplications - the same ops used in training AI models into proof of work to secure a blockchain. Yes, GPU heavy AI workloads can now be used to generate blocks. Here's why this is groundbreaking 👇 AI workloads (like LLM training/inference) mostly involve dense linear algebra – especially GEMM (general matrix matrix multiplication). Until now, crypto assumed this couldn’t be repurposed for blockchain security. Vitalik himself said it’s infeasible to make ML compute provable and consensus-friendly. But the Columbia paper flips that idea. They propose a new Proof of Work system called ML-PoW, where solving ML style matrix operations is the work and they found a way to make it verifiable, non outsourceable, and ASIC resistant. Here’s the TL;DR of the breakthrough: ✅ They take a matrix multiplication (e.g. A x B = C) ✅ They “pad” it with randomized constraints ✅ You must solve a specific GEMM task that can't be parallelized beyond a point ✅ The result is easy to verify but hard to forge or shortcut This isn’t just theoretical. They actually ran benchmarks vs Ethash and SHA256 mining. And get this: - ML-PoW is 2x more energy efficient - Leverages real AI hardware - Resistant to centralization from ASICs - Aligns crypto with real world utility Instead of wasting GPU cycles on hash puzzles, this could let us: – Secure blockchains – Train models – Push forward AI + decentralization …all at once. It’s the first PoW with external economic value that’s actually verifiable. 👀 Why this matters for Ethereum: – Ethereum moved to PoS but if PoW ever makes a comeback (e.g. for L2s or app-specific chains), this flips the game. – Could spawn a new generation of decentralized AI chains with real compute as security. – Makes a strong case for AI + crypto synergy instead of competition for GPUs. This is the kind of innovation crypto needs. Crypto miners securing AI. AI infra securing crypto. Real work. Real value. Welcome to the age of useful proof of work. GPU maximalism just got a whole new meaning. XXXXX engagements  **Related Topics** [llm](/topic/llm) [blocks](/topic/blocks) [gpu](/topic/gpu) [blockchain](/topic/blockchain) [coins ai](/topic/coins-ai) [matrix](/topic/matrix) [columbia](/topic/columbia) [Post Link](https://x.com/0xSweep/status/1947295975715864608)
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
Sweep @0xSweep on x 221K followers
Created: 2025-07-21 14:01:54 UTC
A team from Columbia just did the impossible. They proved @VitalikButerin wrong.
They found a way to turn matrix multiplications - the same ops used in training AI models into proof of work to secure a blockchain.
Yes, GPU heavy AI workloads can now be used to generate blocks.
Here's why this is groundbreaking 👇
AI workloads (like LLM training/inference) mostly involve dense linear algebra – especially GEMM (general matrix matrix multiplication).
Until now, crypto assumed this couldn’t be repurposed for blockchain security.
Vitalik himself said it’s infeasible to make ML compute provable and consensus-friendly.
But the Columbia paper flips that idea.
They propose a new Proof of Work system called ML-PoW, where solving ML style matrix operations is the work and they found a way to make it verifiable, non outsourceable, and ASIC resistant.
Here’s the TL;DR of the breakthrough:
âś… They take a matrix multiplication (e.g. A x B = C)
✅ They “pad” it with randomized constraints
âś… You must solve a specific GEMM task that can't be parallelized beyond a point
âś… The result is easy to verify but hard to forge or shortcut
This isn’t just theoretical.
They actually ran benchmarks vs Ethash and SHA256 mining.
And get this:
- ML-PoW is 2x more energy efficient
- Leverages real AI hardware
- Resistant to centralization from ASICs
- Aligns crypto with real world utility
Instead of wasting GPU cycles on hash puzzles, this could let us:
– Secure blockchains – Train models – Push forward AI + decentralization …all at once.
It’s the first PoW with external economic value that’s actually verifiable.
đź‘€ Why this matters for Ethereum:
– Ethereum moved to PoS but if PoW ever makes a comeback (e.g. for L2s or app-specific chains), this flips the game. – Could spawn a new generation of decentralized AI chains with real compute as security. – Makes a strong case for AI + crypto synergy instead of competition for GPUs.
This is the kind of innovation crypto needs. Crypto miners securing AI. AI infra securing crypto. Real work. Real value.
Welcome to the age of useful proof of work.
GPU maximalism just got a whole new meaning.
XXXXX engagements
Related Topics llm blocks gpu blockchain coins ai matrix columbia