QuBlock retweeted
its time to update the mainnet and release it in the form it was originally planned, the very first vision is probably the only correct one
special thanks to all the validators and participants who supported the devnet and continue to do so - now it's time for the mainnet, permission less access is planned for tomorrow/the day after (the exact time is unknown, any second now)
everything needed to run a mainnet node will be available in a separate release, thanks everyone
special thanks to all the validators and participants who supported the devnet and continue to do so - now it's time for the mainnet, permission less access is planned for tomorrow/the day after (the exact time is unknown, any second now)
everything needed to run a mainnet node will be available in a separate release, thanks everyone
30
30
141
7,167
🚀 #Pearl FP8 update
PR #467 (Speed up V4 prover) merged into fp8-scheme on Nov 5.
Another step toward production FP8 mining, following recent merges of #417 and #471.
PR #467 (Speed up V4 prover) merged into fp8-scheme on Nov 5.
Another step toward production FP8 mining, following recent merges of #417 and #471.
2
1
28
3,156
⚡ Faster proofs
PR #467 includes 57 commits across 28 files, optimizing V4 prover performance, #Blake3, and STARK components.
Reducing proof overhead could improve mining efficiency and make FP8 more practical.
PR #467 includes 57 commits across 28 files, optimizing V4 prover performance, #Blake3, and STARK components.
Reducing proof overhead could improve mining efficiency and make FP8 more practical.
1
0
2
561
🎯 Lottery improvements
PR #471 simplified the mining lottery using contiguous h × w tiles and 16 subtiles.
PR #417 optimized Blake3 verification.
Together, these changes improve the technical foundation for FP8 mining.
PR #471 simplified the mining lottery using contiguous h × w tiles and 16 subtiles.
PR #417 optimized Blake3 verification.
Together, these changes improve the technical foundation for FP8 mining.
0
0
2
471
Proving an AI computation was performed and proving that the correct AI model produced the answer are two different problems.
@AnthropicAI addresses the second problem. @prlnet primarily addresses the first.
@AnthropicAI addresses the second problem. @prlnet primarily addresses the first.
Big news from @AnthropicAI on verifiable LLM inference, with an exciting connection to Pearl’s proof of useful work.
Two orthogonal questions:
• Did the model you paid for actually produce this response?
• Was computational work performed?
Anthropic reports that generating inference proofs currently takes ~70–95× the GPU time of ordinary inference. Pearl’s PoUW, on the contrary, adds negligible overhead. These costs buy different guarantees. Proving correct inference doesn’t automatically establish proof of work and the other way around. But the two are deeply complementary. Verifiable inference could certify which model produced an answer, while Pearl׳s PoUW only certifies that a computation has been done.
As inference proofs get more practical, that combination could make AI easier to trust and more economical to provide.
0
0
8
360
QuBlock retweeted
Big news from @AnthropicAI on verifiable LLM inference, with an exciting connection to Pearl’s proof of useful work.
Two orthogonal questions:
• Did the model you paid for actually produce this response?
• Was computational work performed?
Anthropic reports that generating inference proofs currently takes ~70–95× the GPU time of ordinary inference. Pearl’s PoUW, on the contrary, adds negligible overhead. These costs buy different guarantees. Proving correct inference doesn’t automatically establish proof of work and the other way around. But the two are deeply complementary. Verifiable inference could certify which model produced an answer, while Pearl׳s PoUW only certifies that a computation has been done.
As inference proofs get more practical, that combination could make AI easier to trust and more economical to provide.
Two orthogonal questions:
• Did the model you paid for actually produce this response?
• Was computational work performed?
Anthropic reports that generating inference proofs currently takes ~70–95× the GPU time of ordinary inference. Pearl’s PoUW, on the contrary, adds negligible overhead. These costs buy different guarantees. Proving correct inference doesn’t automatically establish proof of work and the other way around. But the two are deeply complementary. Verifiable inference could certify which model produced an answer, while Pearl׳s PoUW only certifies that a computation has been done.
As inference proofs get more practical, that combination could make AI easier to trust and more economical to provide.
11
17
101
10,553
🧵 I dug into Pearl’s new FP8 mining mechanism and its fp8-scheme code.
This isn’t simply replacing INT8 with FP8.
The design combines AI inference, deterministic noise, matrix multiplication, a tile-based mining lottery, and cryptographic verification.
Let’s break it down. 👇
This isn’t simply replacing INT8 with FP8.
The design combines AI inference, deterministic noise, matrix multiplication, a tile-based mining lottery, and cryptographic verification.
Let’s break it down. 👇
1
0
1
102
15/16
For GPU miners, FP8 FLOPS alone won’t determine profitability.
What matters:
• Eligible lottery tiles/sec
• Valid proofs/sec
• Proof overhead
• GPU memory bandwidth
• Inference throughput
• Power consumption
Hopper & Blackwell are the target architectures.
For GPU miners, FP8 FLOPS alone won’t determine profitability.
What matters:
• Eligible lottery tiles/sec
• Valid proofs/sec
• Proof overhead
• GPU memory bandwidth
• Inference throughput
• Power consumption
Hopper & Blackwell are the target architectures.
1
0
0
32
16/16
Pearl’s FP8 design is ambitious:
Useful AI computation + a verifiable mining lottery.
The 4-row minimum and 16-lane extractor are confirmed in code. Commercial inference demand and production deployment still need proof.
PIP-3:
github.com/pearl-research…
Pearl’s FP8 design is ambitious:
Useful AI computation + a verifiable mining lottery.
The 4-row minimum and 16-lane extractor are confirmed in code. Commercial inference demand and production deployment still need proof.
PIP-3:
github.com/pearl-research…
0
0
1
40