What does it cost to prove an AI model ran as claimed? A new study: one layer of a 7B model at 1,024 tokens took 33.6s to prove, 341x plain inference. All 32 layers at 2,048 tokens: ~25 min. It works. Making it cheap is the open problem.
arxiv.org/abs/2611.17042
Honest status of onchain AI proofs: a FlashProve verifier now runs in the SP1 zkVM with an ERC-8274 contract on top. Still missing: one real proof, est. ~4-15 hours on an H100 for a small example. Works on paper; proving cost is the open problem.
github.com/anthropics/FlashP…
The math that secures blockchains is moving into AI. FlashProve uses a hash-based zkSNARK to prove which model wrote an answer. Still a research prototype: ~70-95x GPU overhead, 1,024-token limit. Early, but the right direction.
anthropic.com/research/prova…
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
The Binance Alpha Box is now live!
This event uses the updated Alpha Box model, with the airdrop pool consisting of tokens from two projects: Yei Finance (CLO) and Bitlayer (BTR). Users holding at least 245 Binance Alpha points can claim one token airdrop on the Alpha Events page.
Upon claiming, users will be placed into different reward tiers and receive one of the following rewards: 467, 583, or 1666 CLO tokens; 560, 700, or 2000 BTR tokens. Rewards are distributed on a first-come, first-served basis.
If rewards are not fully distributed, the points threshold will automatically decrease by 5 points every 5 minutes.
Please note that claiming the airdrop will consume 15 Binance Alpha Points. Users must confirm their claim on the Alpha Events page within 24 hours; otherwise, it will be deemed that users have given up claiming the airdrop.
For builders: start with one decision your team makes repeatedly today. Keep it small, measurable and reversible. Then decide whether it deserves a model.
So the near-term opportunity isn't "AI controls your wallet." It's better workflows: fewer manual reviews, faster routing, clearer escalation — humans in the loop where it counts.
Why it matters: a model can be wrong or manipulated. A permission check in code can't be talked past its limits. Use AI where judgment helps; keep hard rules where money moves.
The key design line: the model suggests, code decides. Signing rights, spend limits, allowlists and transaction checks stay enforced in code — not in a prompt.
In crypto apps these calls are everywhere: which route for a swap, whether a transaction looks risky, when a withdrawal needs a second look, which support ticket needs a person.
Great to be back in Bitcoin Asia conference, always nice to catch up with all the OG bitcoiners. Q Day concern is just a distraction, crypto & AI adoption is the next exciting trend. Stay tuned on Bitlayer's new updates and new direction.
SEC Proposes First Crypto-Specific Offering Framework
The SEC just proposed Regulation Crypto Assets — the first purpose-built framework for crypto offerings in U.S. history.
Two new exemptions: up to $5M (startup) and $75M (fundraising). Plus a safe harbor from the "investment contract" label once decentralization is achieved.
A decade of regulation-by-enforcement may be ending. 60-day comment period is now open.
#Bitcoin#CryptoRegulation
5/ Why this matters for BTCFi: Bitcoin's base layer has now demonstrated — again — extreme resistance to scope changes. Complex logic, data-heavy applications, and programmability continue to move to Layer 2 solutions. The L1/L2 division of labor isn't a design choice — it's what the network keeps choosing.
4/ Context: This was the first contested consensus change attempt in nearly 5 years. Previous soft fork proposals (SegWit, Taproot) passed with overwhelming miner support (>90%). BIP-110's 2.53% is the lowest support any serious Bitcoin fork attempt has received.