Censorship resistance is a requirement for building a safer future with #AI. In the presentation from our meetup in Zug (Jul 2018) we are explaining why it is so and how Pandora Boxchain team develops such technology. Sing up on the future meetups on http://meetu.ps/e/Fwg7Z/jKDhn/a and #FreeAI!

2. Pandora Box put into Blockchain
Trigger the sequence of events
leading to AI-human cooperation via economic
incentives
boost AI evolution towards appearance of super-
human multiple AIs able to work in a trustless
environments

3. Our goal is to design
and build global platform
for a decentralised uncensorable AI

7. – vox populi
“But wait, what about risks?”

8. 1. AI progress is exponential
2. AI will quickly outrun humans
3. AI will not need humans
4. So AI must be regulated
Elon Musk worst fear
Really?
Here we go:

9. Censorship greatly increases the risks of
having A.I.s turned against humanity
Self-preservation is very likely an
important part of any intelligence,
including AI
Once A.I. knows that it can be censored,
e.g. destroyed by a red button or forcibly
modified by less deadly censorship — that
will cause it to evolve by excelling in
deception, pretending, lying and camouflage

10. 1. Some humans are much smarter than others
2. Smart people do not need stupid
3. So by Musk’s logic all stupid people
have to be dead already :)
Historical example
This is not happening because of
Game theory & multiagency
It’s not working like that!

11. Evolution is an adaptation to the
environment
it always involves interaction with other
agents through cooperation and competition
that cannot be substituted by criteria
determined by any authorities

12. Humans are already not the only type
of entities on the Earth.
We have created corporations –
“legal entities”, non-human participants
of economy AND culture, i.e. civilization
AI may become just another “entity”
AI is just
another entity!
Aha!

13. Liberalization
1. Competition with economic interaction
2. Trustless env. and guns :)
3. Multiplicity & decentralization
4. Constant progress
Blockchain
Is the best solution for creating
#d10e & trustless opened economics
What is working:

14. The Solution
Instead of:
AI regulations
Big data regulations
AI “Kill switch”
“Azimov laws”
Use:
Economic incentives
Zero knowledge
Game theory
Audibility & accountability
Blockchain

15. Multi-agent systems
• AI is not singular; at some point
in future will become interacting.
• We will have multiple agents:
autonomous vehicles, IoT, mobiles, etc.
• Multiple V2V, V2E, E2E protocols will evolve, which
will face byzantine faults and multiple security issues
Blockchain is a technology that overcomes
the same issues

16. Imagefromsitewww.artlebedev.ru

17. Imagefromsitewww.artlebedev.ru

18. BLOCKCHAIN
ALLOWS
MALICIOUS AGENT
DETECTION IN
AUTONOMOUS
AI-DOMINATED SYSTEMS

20. Problems to Solve
• Scalability: run large-volume computations (ANNs)
in trestles decentralised network w/o repeating all
computing on each of the nodes
• Censorsip-resistance: blockchain consensus protocol
with high censorship resistance, but without PoW

21. What we have
Proof of Computing Work (PoCW)
Protocol that allows large-scale computing (like AI)
in a trustless peer network without
repeating computations on each of the nodes
– this solves main blockchain & decentralised networks
scalability problem
Consensus protocol (Prometheus)
Based on top of PoCW, solves main PoW and PoS problems
— this solves censorship resistance problem

22. About the project
• Zug (Switzerland) based, started
in May 2017
• Privately funded, no ICO planned
• 1st Round: Aug 2017,
$500m raised
• 2nd Round: May 2018,
$3m raised (lead by CyberFund)
• Valuation (by the second round) –
37.5m CHF (~USD)
• Emphasize on Science, Engineering
& Working Code
• Member of CryptoValley Association
• 5 PhDs and PhD fellows in the core
team
• Joint research with ETH University
(Zurich), Institute of Cybernetics
(Kiev), Web3 Foundation,
Melonport, Ocean Protocol

23. Where are we with release
• Science: formal verification, academic papers prepared
• Engineering: tech specifications in UML
https://github.com/pandoraboxchain/pyrrha-techspecs
• Code: smart contracts + full node code
https://github.com/pandoraboxchain/pyrrha-consensus
https://github.com/pandoraboxchain/pyrrha-pynode
• Testnet in Ethereum network + Web UI to testnet (web client, explorer)
http://client.pandora.network/
http://pandora.network/

24. Core differentiators to other AI + BC
• Scalability protocol (PoCW) reducing computation redundancy from O(N)
to O(1), where N is the number of nodes in decentralised network
• Consensus protocol based on PoCW that solves main PoW and PoS
problems
• Formal verification with game theory and other scientific instruments
• Focus on censorship-resistance
• No ICO but already working MVP
• Synergy and cooperation with many of existing projects in the field

25. Pandora Network Participants
Data Providers
ML Model
Developers
Consumer
Computing
Power
Providers
mining rewards
Decentralised market Decentralised network
buys
buys
provides job
& payes

27. Solving scalability problem
We designed a protocol that allows large-scale
computing in a trustless peer network without
repeating computations on each of the nodes
I’s named Proof of Computing Work – or PoCW.

28. PoCW Basics
• ML computation run only once by a randomly selected worker node
of the network. Node puts a stake on the correctness of the
computing.
• 10% of the computation is verified once by a higher-staked node
(Verifier), randomly selected also from all high-staked nodes
• In case of inconsistency (negative verdicts of computation
correctness) the stake of the original worker is slashed
• Worker can put an appeal (by increasing stake).
Two round arbitration (closed and public) is run in this case
(details follow)

29. Solving Censorship Resistance
Problem: Generic Blockchain Consensus
• Partly synchronous (with maximum ∆ network delays) under basic
conditions
• Asynchronous recovery mode
(with chain splits and incentivised scheme for chain merge
after synchronicity restoration)
• On-chain randomness, no BFT-type agreements, i.e. fast and up
to 50%+ tolerance to byzantine faults
• Finality is reached maximum within ζ (zeta, ~20) blocks
following any byzantine fault, but normally faster due to
special gossip messaging algo

30. Main challenges in designing PoS
•Get on-chain multiparty-agreed randomness that can’t be
affected by the last party
•Build protection from long-range nothing at stake attacks
•Make it able to work in async environment (like network
failures) and restore from asynchronicity to synchronicity
without double-spends

31. Asymmetric
double-stranded
blockchain

32. Asymmetric Double-stranded
Blockchain
• Composed of two interconnected blockchains running similar consensus
protocol but asymmetrical in the form of stake used by nodes:
• First chain (proof of stake) uses common cryptocurrency-based
stake
• Second chain (proof of reputation) uses reputation, that comes
from the performance of cognitive work
• This can be compared to DNA, which doubles its security by organising
genetic material into cross-referencing double helix
• Or brain hemispheres, where one is responsible for “rational” and the
other – for “irrational”

33. PoS chain –
Value & Contracts
Layer 1: generic cryptocurrency
(bitcoin-like) but without
drawbacks of PoW algorithm
Layer 2: sidechains with Turing-
complete virtual machines (EVM,
WASM, other can be ported in the
future)
Layer 3: state and payment
channels (Lightning, Luminal)
Economics:
• Cryptocurrency (BOX)
PoR chain –
AI & High-Load Computing
Layer 1: base level for
decentralised AI
Layer 2: sidechains for scaling
Layer 3: proof of computing work
state channels
Economics:
• Utility token (PAN) +
• Reputation
(node property, not liquid)

34. @pandoraboxchain