Fast trustless bridges usually require payments to be locked up in escrow
But what if users could just send to the relayer—trustlessly, and at a fraction of the capital and gas cost of the escrowed model?
Introducing Across Prime, a new design from me @hal2001 @mrice32
🧵
Intent-based bridges use relayers to send payments faster than the speed of cross-chain messages
The simplest way to build a fast bridge is to have a trusted relayer. Users send money directly to the relayer and hope the relayer fulfills their intent on the destination chain
What if we don't want to trust the relayer?
@AcrossProtocol helped pioneer a trustless model based on escrow: users send payments into escrow, and the relayer claims it later after proving they fulfilled the intent
This requires each payment to be locked up for some period
Across Prime introduces a bonded model, where payments go directly to relayers, but are secured by a security deposit
Since bonds can be reused as soon as payments are confirmed, this is more capital-efficient
It's also more gas-efficient, since it doesn't require a claim tx
I've loved working with @hal2001 @mrice32 and the rest of the innovative @AcrossProtocol team on pushing the limits of bridge designs
Prediction markets were the original motivation for AMMs, but today they mostly use orderbooks
One reason: outcome tokens don't behave like other assets
We present the pm-AMM, a new invariant designed for prediction markets
🧵
This required addressing a deep problem—what does it mean for an AMM to be optimized for a particular asset (like options or bonds)?
Many have considered this before (including me and @niemerg with YieldSpace), but we didn't have a clear reason for picking one curve over another
The missing piece came from @jason_of_cs @ciamac @Tim_Roughgarden @alz_zyd_ , in their seminal paper on LVR
They showed that for geometric-Brownian-motion assets, constant geomean pools like Uniswap or Balancer have a nice property: LVR is a constant fraction of portfolio value
* Lets arbitrary apps capture their MEV
* Preserves composeability
* Would work today on OP Stack L2s like @Optimism @base @Blast_L2
The secret? The surprising power of priority ordering 🧵
To implement MEV taxes, smart contracts charge a fee as a function of the priority fee of the transaction.
We show that if a smart contract charges searchers a MEV tax of (say) $99 for every $1 of priority fee, it can capture 99% of the competitive MEV for that transaction.
MEV taxes are a flexible technique. We outline how they could address three major problems in MEV research:
* DEX routers like UniswapX that optimize execution
* AMMs that minimize LVR for liquidity providers
* Wallets that capture "backrunning" MEV leaked by transactions
⚖️ Reduces LVR
⚙️ Optimizes swap fees
📈 Smooths LP returns
🌊 Should attract higher liquidity than any fixed-fee AMM
New paper with @ciamac (@paradigm / @Columbia_Biz) and @AustinAdams10 @saraareynolds (@Uniswap)
Two of the most important challenges in AMM research are:
* Minimizing losses to informed flow (i.e. reducing loss-vs-rebalancing)
* Maximizing revenue from uninformed flow (i.e. optimizing fees)
The am-AMM targets both of these problems with one mechanism.
The am-AMM rents the right to manage the pool to the highest bidder, in a continuous auction we call a “Harberger lease,” with rent going to liquidity providers.
The pool manager has the power to set the swap fee, and to receive those fees.
In Blend, lenders are matched peer-to-peer with borrowers
Lenders sign off-chain offers that specify an interest rate and a collection they're willing to lend against
Borrowers can select the best available offer to instantly borrow against their NFT on-chain
Loans are perpetual by default, with a constant interest rate
If interest rates fall or a borrower wants to leave a position, the borrower can repay at any time (and instantly take a different offer with a better rate, if they want)
It looks like ChatGPT doesn’t distinguish between questions and answers, so you can train it to give whatever kind of answer you want by giving it the first few words of the desired response
Hacking the AI is old news, the real trick is getting the AI to hack you