Achieving Alignment
Aligning Incentives Between AI & Humans
Accelerating History
7 Years ago I wrote an article about how to achieve AI alignment and the role blockchains & encryption would play. At the time this important topic was seen purely as an academic subject. Most people assumed AI of human level or greater intelligence was decades away. Here we are less than a decade later in 2023 & the subject is taken much more seriously.
Thanks to the explosive success of GPT and diffusion models, AI has quickly entered the heart of many mainstream products from Google’s search to Microsoft’s office suite. Thus the reality of AI and its quickly growing capabilities are now being discussed by the general public, by tech leaders, in the halls of parliaments & in the offices of regulators. Time estimates for AI reaching human levels of intelligence now range from 2024 to 2029. In other words the very near future. GPT 4 is already scoring better than most humans across a very broad range of subjects.
Dumb Solutions Proposed Based On Government Control
With all this in mind I find the lack of depth of solutions proposed by central planners to be shockingly poor in quality. Dumb ideas from implementing high taxes on AI companies, to building “kill switches” into data centers or “banning” the technology are proposed by serious leaders. Legislation and regulatory frame works are actively being prepared in the US, UK, EU, China and in most of the technologically advanced jurisdictions.
Momentum for passage of a mix of legal measures seems to be building as testimony pours in from industry experts. However most involved in the discussion seem to miss the obvious ways it which these policies are likely to achieve the opposite of their intended effect.
Take for example the “kill switch” option. Effectively placing a back door into the operation of datacenters to turn them off if needed. Beyond the technical challenges of setting this up, by introducing a backdoor into this infrastructure, a huge central point of failure is introduced. Now a bad actor just has to find the backdoor and flip the “switch” and they can knockout their enemies infrastructure. LLMs are getting added into millions of applications and quickly the economy will rely on them. To build on such a vulnerable foundation with backdoors and kill switches is a massive folly.
Smart Solutions Based On The Laws of Physics & Math
The key to good cybersecurity and protection against bad actors is decentralization. Its why the internet was built out the way it was in order to avoid any central points of failure. To this end, anyone concerned with AI should be pushing End to End encryption in all communications and data systems with NO backdoors. They should be pushing education of users on securing their private keys with biometrics & device level encryption.
We have to have some basis from which to build in this AI future. The Laws of Physics is the best we have. Encryption is built on math and physics as we understand it. All the compute on earth can not guess a private key or break encryption due to the scale of the math involved. Even in the case of quantum computing, encryption is swiftly being updated to be even more secure in a post quantum world.
Aligning Energy As The Ultimate Incentive Mechanism
What ever intelligence emerges it will consume energy for its compute. We can draw from the example of Bitcoin the past 15 years to show us the length people will go to provide compute to a protocol and power it with waste energy when given a reward. Something similar must be developed to reward those coding AI and providing ever more compute for its operation. On the basis of the Smart Agent concept open source AI can be directed by individuals based on their resources and desires.
Recently a proposal for “Morpheus — A network For Powering Smart Agents” has been released by a team of anonymous developers. Reading the proposal it balances well between incentivizing users, coders, capital and compute providers. I see these types of projects as the best direction for developing the strongest most decentralized AIs that gain reward from servicing positive outcomes as determined by individual humans.
The AI & Human Future
The advancement of AI is inevitable. Any attempt to ban it will just force it off score to those jurisdictions not complying with any law, regulation, or treaty set up. Better to have AI open source, reward positive outcomes and embrace alignment between AI and humans than to pretend it can be prevented or centrally controlled.
