Pearl Research Labs retweeted
New Layer 1, same standard of custody.
Anchorage Digital Bank and Anchorage Digital Singapore now support native custody for Pearl (PEARL), built by @prlnet. Addresses, deposits, and withdrawals are live for clients today.
Anchorage Digital Bank and Anchorage Digital Singapore now support native custody for Pearl (PEARL), built by @prlnet. Addresses, deposits, and withdrawals are live for clients today.
15
17
121
16,909
Congrats to our friends at @flappyairplanes
We’ve spent the past year developing new primitives for GPUs. We believe this enables new algorithms to change how machines learn, open new markets, and dramatically reshape AI’s cost structure.
We have raised a $1 billion Series A, co-led by Index and Lux.
19
13
163
28,140
Pearl Research Labs retweeted
PEARL is now live on BitGo. Eligible BitGo clients can hold, transact, and secure it with two wallet types.
🏦 Multi-signature cold wallets are BitGo Bank & Trust qualified custody wallets: segregated, bankruptcy-remote, and insured up to $250M against loss, theft, and misuse.
🔑 Mtody hot wallets are available for clients who want to hold their own keys.
Both wallet types include institutional-grade security and policy controls.
@prlnet
🏦 Multi-signature cold wallets are BitGo Bank & Trust qualified custody wallets: segregated, bankruptcy-remote, and insured up to $250M against loss, theft, and misuse.
🔑 Mtody hot wallets are available for clients who want to hold their own keys.
Both wallet types include institutional-grade security and policy controls.
@prlnet
3
15
97
18,125
Pearl Research Labs retweeted
Today is probably a good time to check in on things with @privy_io, which became part of @stripe just over a year ago.
Our acquisition was predicated on three factors: stablecoin growth, self-custody as a key enabler of broader global coverage, and rapidly increasing demand for programmable financial infrastructure. As we said at the time, "money has to reside somewhere, and Privy builds the world's best programmable vaults."
Since then, the Privy team has executed exceptionally well: assets in Privy wallets have grown ~25x, monthly active developers are up ~3x, revenue is up ~9x, and Privy now facilitates nearly $25B of monthly transaction volume. @tryramp, @Deel, @Klarna, @felixpago, @HyperliquidX, and thousands of other businesses are building embedded banking functionality on Privy.
At Stripe, we're deeply integrating Privy into Treasury, @Link, and Connect. In each case, it's accelerating our global expansion, and enabling us to build better functionality faster. (Coming soon: @Link balances.)
We're also seeing that Privy's wallet infrastructure is the ideal substrate for agents to transact and earn -- live today with @merit_systems, @orthogonal_sh, @Nevermined_ai, @prlnet, and more. Developers prefer programmable financial infrastructure, but agents need it. We expect that the agentic economy will continue to grow extremely quickly.
The Privy team has maintained remarkable product velocity, with major launches this year including deposit addresses, earning on Privy balances, digital asset accounts, custodial wallets, and Stripe's own stablecoin cards.
It's also been neat to see how Privy has been accelerated by Stripe, with better onramps, @stablecoin payment rails and licenses, Stripe's fiat capabilities and issuing stack, and integration with @Tempo.
Together, we're building one API for holding stablecoin balances and earning yield, moving money in and out of fiat, and orchestrating spend with cards. We still have a lot of work to do, not least in integrating everything across Privy, Bridge, and the rest of Stripe -- there's been so much feature and functionality expansion that we haven't yet had much time to reconcile everything. Overall, though, I'm very happy with how things are going, and excited by where these capabilities are headed.
If you're interested in working on the future of stablecoins, agentic payments, and crypto's deployment in the real world, you should get in touch with @sternhenri (now leading all of @StripeCrypto) or others on the team.
Our acquisition was predicated on three factors: stablecoin growth, self-custody as a key enabler of broader global coverage, and rapidly increasing demand for programmable financial infrastructure. As we said at the time, "money has to reside somewhere, and Privy builds the world's best programmable vaults."
Since then, the Privy team has executed exceptionally well: assets in Privy wallets have grown ~25x, monthly active developers are up ~3x, revenue is up ~9x, and Privy now facilitates nearly $25B of monthly transaction volume. @tryramp, @Deel, @Klarna, @felixpago, @HyperliquidX, and thousands of other businesses are building embedded banking functionality on Privy.
At Stripe, we're deeply integrating Privy into Treasury, @Link, and Connect. In each case, it's accelerating our global expansion, and enabling us to build better functionality faster. (Coming soon: @Link balances.)
We're also seeing that Privy's wallet infrastructure is the ideal substrate for agents to transact and earn -- live today with @merit_systems, @orthogonal_sh, @Nevermined_ai, @prlnet, and more. Developers prefer programmable financial infrastructure, but agents need it. We expect that the agentic economy will continue to grow extremely quickly.
The Privy team has maintained remarkable product velocity, with major launches this year including deposit addresses, earning on Privy balances, digital asset accounts, custodial wallets, and Stripe's own stablecoin cards.
It's also been neat to see how Privy has been accelerated by Stripe, with better onramps, @stablecoin payment rails and licenses, Stripe's fiat capabilities and issuing stack, and integration with @Tempo.
Together, we're building one API for holding stablecoin balances and earning yield, moving money in and out of fiat, and orchestrating spend with cards. We still have a lot of work to do, not least in integrating everything across Privy, Bridge, and the rest of Stripe -- there's been so much feature and functionality expansion that we haven't yet had much time to reconcile everything. Overall, though, I'm very happy with how things are going, and excited by where these capabilities are headed.
If you're interested in working on the future of stablecoins, agentic payments, and crypto's deployment in the real world, you should get in touch with @sternhenri (now leading all of @StripeCrypto) or others on the team.
121
55
1,173
178,619
Exciting times ahead!
1/ We are excited to list PEARL, and are extremely aligned with our friends @prlnet in their mission to make the frontier of inference more useful and efficient.
28
33
332
38,131
Pearl Research Labs retweeted
We’re excited to launch Liquid Inference with some of the fastest-growing token producers in the ecosystem:
• Engy @engyai
• Pearl Research Labs @prlnet
• AkashML @akashnetAI
• Hyperbolic @hyperbolic_labs
• Boundless @BoundlessHQ
• Zro @zroai_
• Engy @engyai
• Pearl Research Labs @prlnet
• AkashML @akashnetAI
• Hyperbolic @hyperbolic_labs
• Boundless @BoundlessHQ
• Zro @zroai_
10
23
136
17,109
Pearl Research Labs retweeted
Excited to share @prlnet support is now live on Utila, with the same governance you use across every supported chain.
Create Pearl wallets, receive #PRL, enforce transaction policies, and send transfers from console, mobile, and API.
Read more: utila.io/blog/pearl-prl…
Create Pearl wallets, receive #PRL, enforce transaction policies, and send transfers from console, mobile, and API.
Read more: utila.io/blog/pearl-prl…
1
2
37
3,804
Pearl Research Labs retweeted
Our latest research paper explores the growing connection between AI and digital assets and explains why broad AI adoption may drive new demand, utility and applications across the digital asset economy. blackrock.com/us/individual/…
765
1,994
8,363
4,157,831
Pearl Research Labs retweeted
Pearl (@prlnet) has the fastest FP8 MoE Kernels for Blackwell chips!
Group-Matrix-Multiplication is one of the most optimized operations in AI. As a former researcher at @nvidia I can say firsthand that pushing the performance frontier of MatMuls is... Hard.
Pearl-GEMM is ~5% faster than prior state-of-art Quack (@tri_dao) and ~42% faster than Flashinfer on B200s.
pearlresearch.ai/research/blog/…
Group-Matrix-Multiplication is one of the most optimized operations in AI. As a former researcher at @nvidia I can say firsthand that pushing the performance frontier of MatMuls is... Hard.
Pearl-GEMM is ~5% faster than prior state-of-art Quack (@tri_dao) and ~42% faster than Flashinfer on B200s.
pearlresearch.ai/research/blog/…
17
48
273
49,571
Read more on our collaboration with @attestable here: x.com/Yogi_Brn/statu…
The world is debating whether frontier AI needs to slow down. If it comes to it, do we know how?
A lab can say it stopped training. A country can sign a treaty. But, is the compute following the rules? @attestable is building that layer. New blog! 🧵
3
5
63
40,628
Privacy-preserving proof of inference is a key building block for AI safety. @DarioAmodei's call for AI slowdown & independent oversight raises a critical question / How can frontier labs prove their computations follow the rules, without revealing model weights, prompts, or private data?
Two complementary technologies are necessary to make this inference layer possible: Zero-Knowledge proofs (zk-SNARKs) can prove that Y = M(x), while keeping M and x private, and verify compliance using explicit policies. Proof-of-useful-work (PoUW) can
tie that evidence to actual, timestamped GPU computation, with a public immutable record and economic incentives for participation.
@prlnet is partnering with @attestable to build this trust layer for AI. Together, our complementary technologies can make AI objectively accountable, making verification economically sustainable and self-funding.
Today, we are announcing @prlnet's new and ultra-efficient floating-point PoUW scheme for NVIDIA Blackwell chips.
Floating-point arithmetic has been a notorious obstacle to efficient verification of compute for decades. Pearl’s FP scheme brings proof-of-useful-work technology to Blackwell FP computations on frontier LLMs, with NEAR-ZERO additional overhead over end-to-end inference on optimized vLLM.
The FP network upgrade specification is now available (github & whitepaper) and will soon serve frontier models on Pearl x @togethercompute's endpoints in production.
Explore the implementation:
github.com/pearl-research…
Whitepaper: pearlresearch.ai/Pearl_Whitepap…
Run Pearlified models on our endpoint:
pearlresearch.ai
Two complementary technologies are necessary to make this inference layer possible: Zero-Knowledge proofs (zk-SNARKs) can prove that Y = M(x), while keeping M and x private, and verify compliance using explicit policies. Proof-of-useful-work (PoUW) can
tie that evidence to actual, timestamped GPU computation, with a public immutable record and economic incentives for participation.
@prlnet is partnering with @attestable to build this trust layer for AI. Together, our complementary technologies can make AI objectively accountable, making verification economically sustainable and self-funding.
Today, we are announcing @prlnet's new and ultra-efficient floating-point PoUW scheme for NVIDIA Blackwell chips.
Floating-point arithmetic has been a notorious obstacle to efficient verification of compute for decades. Pearl’s FP scheme brings proof-of-useful-work technology to Blackwell FP computations on frontier LLMs, with NEAR-ZERO additional overhead over end-to-end inference on optimized vLLM.
The FP network upgrade specification is now available (github & whitepaper) and will soon serve frontier models on Pearl x @togethercompute's endpoints in production.
Explore the implementation:
github.com/pearl-research…
Whitepaper: pearlresearch.ai/Pearl_Whitepap…
Run Pearlified models on our endpoint:
pearlresearch.ai
24
52
281
90,333
The Pearl founders will be hosting another live AMA with the community!
Tuesday, September 15
6:00 PM CET · 12:00 PM ET · 16:00 UTC
We've been heads-down building, and the questions have been piling up — about the research, the roadmap, and where the network goes from here. On Tuesday, we're answering them.
Here's how it works: head to the ama-questions channel on our Discord and drop in whatever you'd like answered. Upvote the questions you care about most. The founders will open with the five that rise to the top, then it's an open floor for the rest of the session.
Come with good questions. We'll come with real answers.
discord.gg/joinpearl
Tuesday, September 15
6:00 PM CET · 12:00 PM ET · 16:00 UTC
We've been heads-down building, and the questions have been piling up — about the research, the roadmap, and where the network goes from here. On Tuesday, we're answering them.
Here's how it works: head to the ama-questions channel on our Discord and drop in whatever you'd like answered. Upvote the questions you care about most. The founders will open with the five that rise to the top, then it's an open floor for the rest of the session.
Come with good questions. We'll come with real answers.
discord.gg/joinpearl
11
13
119
37,269
Pearl started with a simple question:
Why should AI and blockchain mining compete for the same GPUs when the same computation could do both?
In this talk, Pearl co-founder @komargodski explains how that idea became a live proof-of-useful-work network turning AI inference into h minimal overhead.
Why should AI and blockchain mining compete for the same GPUs when the same computation could do both?
In this talk, Pearl co-founder @komargodski explains how that idea became a live proof-of-useful-work network turning AI inference into h minimal overhead.
19
22
189
34,215
Proof-of-Useful-Work forced us to solve a problem ordinary blockchains do not have.
A Pearl miner performs large matrix multiplications as part of a real AI workload. Naively, when it finds a valid block, it would need to publish those matrices so the network could verify the ces may contain private model data or user information, and a single weight matrix can be tens-to-hundreds of megabytes. Publishing them on-chain would be both prohibitively expensive and a privacy risk. This is why we encapsulated the entire mining computation inside a zk-SNARK.
A zk-SNARK can take an arbitrarily complex computation and compress it into a small cryptographic proof. The proof reveals nothing about the underlying matrices, while still allowing the network to verify that the miner found a valid block.
Instead of placing tens-to-hundreds of megabytes of potentially sensitive data on-chain, Pearl miners publish a zero-knowledge proof of roughly 60 kilobytes. The blockchain verifies the useful computation without ever seeing the model weights, activations or private data behind it.
This is what makes Proof-of-Useful-Work practical at scale: large AI computations happen off-chain, while only a compact, privacy-preserving proof reaches consensus.
A Pearl miner performs large matrix multiplications as part of a real AI workload. Naively, when it finds a valid block, it would need to publish those matrices so the network could verify the ces may contain private model data or user information, and a single weight matrix can be tens-to-hundreds of megabytes. Publishing them on-chain would be both prohibitively expensive and a privacy risk. This is why we encapsulated the entire mining computation inside a zk-SNARK.
A zk-SNARK can take an arbitrarily complex computation and compress it into a small cryptographic proof. The proof reveals nothing about the underlying matrices, while still allowing the network to verify that the miner found a valid block.
Instead of placing tens-to-hundreds of megabytes of potentially sensitive data on-chain, Pearl miners publish a zero-knowledge proof of roughly 60 kilobytes. The blockchain verifies the useful computation without ever seeing the model weights, activations or private data behind it.
This is what makes Proof-of-Useful-Work practical at scale: large AI computations happen off-chain, while only a compact, privacy-preserving proof reaches consensus.
9
10
110
25,571
@nvidia Read the full paper here: arxiv.org/pdf/2603.20421
1
2
17
11,082
GPU non-determinism forced us to reverse-engineer arithmetic that modern AI hardware does not expose.
The same matrix multiplication can produce different results across @NVIDIA Ampere, Hopper and Lovelace GPUs. The reason is not randomness - it is architecture-specific behavior ation, internal precision, rounding, normalization and handling of subnormal values. That makes exact verification difficult: a CPU re-execution may disagree with a perfectly correct GPU execution.
Hawkeye solves this by systematically probing Tensor Cores with carefully constructed matrix multiplications. These tests recover the hidden numerical pipeline behind each architecture, including how products are grouped, when significands are truncated and when intermediate values are normalized. The recovered behavior is then encoded into a CPU simulator that reproduces the original GPU computation bit-for-bit.
The results show meaningful architectural differences. Ampere uses a two-stage accumulation structure with a 24-bit internal significand, while Hopper combines all 16 products in one stage using 25 bits. Both use truncation-based, round-towards-zero behavior rather than conventional round-to-nearest arithmetic.
Across 100,000 randomly generated matrix multiplications in FP16, BF16 and FP8, Hawkeye reproduced the GPU result with 100% bit-exact accuracy. This creates a foundation for auditors to verify AI training and inference without modifying or slowing down the original GPU workload.
Looking forward, matrix multiplication is only the first step. Extending this approach to attention, fused kernels, distributed training and cryptographic proof systems could make end-to-end, hardware-accurate verification of AI workloads possible.
The same matrix multiplication can produce different results across @NVIDIA Ampere, Hopper and Lovelace GPUs. The reason is not randomness - it is architecture-specific behavior ation, internal precision, rounding, normalization and handling of subnormal values. That makes exact verification difficult: a CPU re-execution may disagree with a perfectly correct GPU execution.
Hawkeye solves this by systematically probing Tensor Cores with carefully constructed matrix multiplications. These tests recover the hidden numerical pipeline behind each architecture, including how products are grouped, when significands are truncated and when intermediate values are normalized. The recovered behavior is then encoded into a CPU simulator that reproduces the original GPU computation bit-for-bit.
The results show meaningful architectural differences. Ampere uses a two-stage accumulation structure with a 24-bit internal significand, while Hopper combines all 16 products in one stage using 25 bits. Both use truncation-based, round-towards-zero behavior rather than conventional round-to-nearest arithmetic.
Across 100,000 randomly generated matrix multiplications in FP16, BF16 and FP8, Hawkeye reproduced the GPU result with 100% bit-exact accuracy. This creates a foundation for auditors to verify AI training and inference without modifying or slowing down the original GPU workload.
Looking forward, matrix multiplication is only the first step. Extending this approach to attention, fused kernels, distributed training and cryptographic proof systems could make end-to-end, hardware-accurate verification of AI workloads possible.
11
27
134
30,953
Proof-of-Useful-Work forced us to invent a quantization format that did not exist.
Pearl Protocol adds noise to both the weights and activations before matrix multiplication. That creates a constraint ordinary inference stacks do not have - the model values and the noise must 8. So we quantize the model to an effective W7A7 range, reserving one bit of headroom for the noise. The perturbed operands still fit in INT8, multiplication uses INT8 inputs with INT32 accumulation, and the low-rank noise can then be peeled off exactly, recovering the exact result of the quantized computation.
Existing quantization for INT7 were unsatisfactory, so we built our own - a non-trivial fusion of GPTQ-style weight reconstruction, SmoothQuant-style activation smoothing and custom inference kernels. The result is an end-to-end W7A7 pipeline that, in our internal benchmarks, even outperforms many FP8 baselines (in quality and in efficiency). We are writing a paper detailing this technique so stay tuned.
Looking forward, W7A7 is only an intermediate step. Our upcoming upgrade is designed to support FP8 workloads directly, eliminating the need for the extra integer-quantization step.
Pearl Protocol adds noise to both the weights and activations before matrix multiplication. That creates a constraint ordinary inference stacks do not have - the model values and the noise must 8. So we quantize the model to an effective W7A7 range, reserving one bit of headroom for the noise. The perturbed operands still fit in INT8, multiplication uses INT8 inputs with INT32 accumulation, and the low-rank noise can then be peeled off exactly, recovering the exact result of the quantized computation.
Existing quantization for INT7 were unsatisfactory, so we built our own - a non-trivial fusion of GPTQ-style weight reconstruction, SmoothQuant-style activation smoothing and custom inference kernels. The result is an end-to-end W7A7 pipeline that, in our internal benchmarks, even outperforms many FP8 baselines (in quality and in efficiency). We are writing a paper detailing this technique so stay tuned.
Looking forward, W7A7 is only an intermediate step. Our upcoming upgrade is designed to support FP8 workloads directly, eliminating the need for the extra integer-quantization step.
9
18
113
22,282