1 of 25

MEV for AI people

@phildaian

2 of 25

In this talk

  1. A brief intro to MEV
  2. MEV alignment =? AI alignment
  3. sandboxes for AI alignment
  4. cross-domain MEV for AI people
  5. crypto as an adversarial-incentives API

3 of 25

mev

4 of 25

The Hydra�“Flaws of Decentralization on 0x and Etherdelta”

Ordering (OO) Fees

5 of 25

The Hydra : Into the Rabbithole

6 of 25

“Flaws of Decentralization”

Issue: exposure to arbitrage, miner frontrunning, exchange abuse

Found $1M/y profit running our own bot. Pre-PGA MEV bot!

7 of 25

Ordering Fees: One Puzzle Piece

8 of 25

Flash Boys 2.0 (a warning)

“MEV”, “PGA”s - beyond ordering

fair ordering: Themis and more (Kelkar et al)

9 of 25

Flash Boys 2.0

10 of 25

Enter MEV

OO fees represent one case of a more general quantifiable value we call miner-extractable value (MEV).

MEV refers to the total amount of Ether miners can extract from manipulation of transactions within a given timeframe, which may include multiple blocks’ worth of transactions.

11 of 25

mev

12 of 25

Clockwork Finance: A New Definition

MEV: the amount a miner can increase their balance across a series of blocks

13 of 25

Flashbots

mev: wat do?

… and many other projects

14 of 25

mev wat do?

MEV is fundamental

wat do?

  1. transcriptability (bonus)
  2. interoperability (trust)
  3. “permissionless” design

DAPP DEVELOPERS: Limit MEV to what is strictly required/inherent. Think carefully about MEV.

USERS: Demand the above. Don’t use dapps with predatory MEV. Understand what MEV means to you.

L1/L2: Redesign with MEV in mind. Think about the use cases you are designing for, and how best to manage MEV. ��VALIDATORS: Extract, extract, extract, and don’t be ashamed. Your continued extraction is required to the game-theoretic security of any network on which MEV exists.��Everyone else: mitigate 👏 negative 👏 externalities.

15 of 25

… the birth of the magic bot

> illuminate the dark forest�

> democratize MEV extraction��> distribute MEV benefits

16 of 25

17 of 25

end mev begin ai

18 of 25

blockchain

19 of 25

blockchain

blockchain

blockchain

20 of 25

Blockchains are sufficiently sophisticated, sufficiently autonomous, and sufficiently abstracted from humanity to constitute their own “AI”s

21 of 25

… but why is this hyper rationalist weird metasystem better than my favorite LLM?

  • Hyper-adversarial, automated, symbolic behavior on a massive scale
  • “GPU farms for free”
  • Autonomous agents <> Open AIs
    • agency/ownership reqd for alignment?
  • The Scientific Method!� Test the limits of incentives for alignment

22 of 25

- there is no alignment without aligned incentives

- necessary but not sufficient

- crypto as ad-hoc alignment of the unaligned

- gives hope for AI alignment

- researching this further is already helping AI alignment

23 of 25

- privacy and decentralization of resources is essential to MEV alignment

- AI researchers should study programmable privacy and decentralization more explicitly

blockchain

Value == information == leverage�Cryptography allows for (some) boundaries

24 of 25

- cross-domain MEV alignment implies world alignment

implies AI alignment

- decentralization is necessary for cross-domain alignment

blockchain

blockchain

25 of 25

crypto allows for sandboxed experimentation with the real world

- can validate toy models, run alignment experiments with

no a-priori modeling, and more

- incentives require real-world validation; crypto provides

the API