We are slowing down AI automations to increase civilization's productivity because we don't have verifiable communication pipelines. We're going to stumble upon the same problem that we faced in the blockchain world - that is, we have extremely powerful technology capable of taking decisions on behalf of the collective but is bottlenecked by the fact that there is not enough verifiable information to inform these decisions.
Blockchains have smart contracts that are capable give out loans, settle disputes, protecting users' rights - but hit this roadblock on how to take these actions without enough trustworthy data. Decentralized exchanges like Uniswap have access to information about the supply and demand for different coins - from the blockchain. Using that information, it is capable of taking the decision on the price of each coin relative to the other. The natural extension to Uniswap is Aave. Aave is capable of giving out loans, but Aave needs verifiable information on what the price of each token is on other plaforms like centralized exchanges to process a loan. That problem was solved by Chainlink, by building a trustworthy-enough system to provide Aave with the verified information it needs to make a decision. Further more, many usecases continue to be infeasible to execute on a blockchain or various coprocessors - not because the tech doesn't exist, but because the required information is not available to these systems in a trustworthy way. Now we have emerging technologies like more expressive oracles, consensus layers like UMA, zkTLS that is beginning to bridge this gap.
AI Agents, much like smart contracts, are capable of operating autonomously. Ofcourse, way more creative/intelligent - rather than the deterministic smart contracts. But the fundamental doesn't change. If you need AI to take autonomous decisions, to really boost humankind's productivity, you need verifiable information that they take actions upon. So far we have only seen chatbots that are exceptionally good at data retrival in single player modes. That is you ask a question, it gives you the answer and nothing changes for any other user whatsoever. But if AI agents are taking actions like making hiring decisions, making lending decisions etc. that is a multi player zero sum game. That is, if you get the job - someone else does not. If you get a loan, there is someone who doesn't because there is only so much money to lend. In these multi player zero sum games orchestrated by autonomous systems - verifiablity is incredibly important to make any meaningful decision. But how does an AI Agent get access to verified information? We run into the same problem that smart contracts ran into. There is no way to verify information. If a candidate is making claims about themselves on a resume, there is no way for an AI agent to efficiently verify that. Again, as in the case with smart contracts, there are no data pipelines from the source of an information to the autonomous system.
I'd argue the AI Agents being able to take important decisions is going to be critical to unlock their impact. And that's a problem we'll soon face, now that foundational models are good enough for a wide variety of AI Agents to be built. That future of AI Agents taking actions for society is inevitable. For that future, we need more verifiable data. If we have more and more verifiable data, we will see more and more AI Agents being able to take real-world-impacting actions. Meaningful actions to push society forward.
One future is where a large foundational model company strikes partnerships with numerous sources of information. Like Open AI has partnerships with your bank, your employer, your landlord, your gym and gets all of that information to power AI Agents. First up, that's an unlikely future for one single company to strike partnerships with every source of information globally. Also, it wouldn't be an ideal world where the company has outsized leverage on decisions an AI Agent takes in literally all the usecases. It would be terrible if the only model that can power a lending agent is OpenAI, or the only model that can power hiring usecases is Anthropic.
That is also the case to be made for more open source models to exist, different models can be used for different usecases. But such models themselves, obviously don't have user specific information to drive decisions using. Applications that are built on top of this model are responsible for procuring and processing the data that is used in this decision making. Different applications for different usecases for different demographies. That is the most likely future, because it is impractical for one AI to be good at all the usecases and have access to the required information.
Lastly, I suspect we would not want any single AI to take actions all by itself - but rather we'd want many agents with different models, different biases to work a conclusion together. Decision by committee, but an autonomous committee. Especially in cases like taking actions that affect a large society - like capital allocation, legal adivsory, administrative tasks.
If you're still with me, and agree that there are going to be different models and different applications for different usecases - it means you will be interacting with many of them in the course of your day or life.
If you build reputation and credentials on one service, the status quo is that you will be able to use that reputation and credentials on that service alone. The moment you move from one service to another, you lose all your hard earned reputation and credentials. If you have been building cool open source tools on Github, you can't take that reputation to a hiring platform. If you have been earning by designing great landing pages on Upwork, you can't take that reputation to a bank to ask for a loan. If you have been writing industry defining thought pieces on Substack, you can't take that to a Visa office.
Today, in most of these transactions you have a human on the other side. The hiring manager, the banker, the visa officer are all humans. You could take a screenshot of your Github, Upwork or Substack to them. There is a chance that they do accept it. But we all know how easy it is to morph these screenshots. Even more so with AI. It is already hard enough to convince a human that these are original screenshots with human elements like tone of voice and bodylanguage, but would be impossible to convince an autonomous system. The reputation that you are building on various digital experiences need to be
In the current world, when you move from service to service you have to start over from scratch. It's like when you immigrate to a new country, you arrive naked and have to work your way up all over again.
When your credentials are going to be important to interact with any service, largely run by AI Agents - you need these credentials to be in your control. You should not get locked out because the service where you built your reputation doesn't want to talk to the service on which you want to use that reputation. Even worse would be if the service where you built the reputation is now defunct. Imagine you worked super hard at a startup, but that startup doesn't exist in 20 years for any hiring manager to verify your work history against.
For that reason, it is important that we have verifiable information and even better would be if there is a way for the users to take self-custody of their credentials and use it as they deem appropriate. There is no good reason why I can't share testimonials i received on LinkedIn with a job application portal. That needs to change to make the world more automatable, and the world to not let companies' vested interest come in the way of a user's right to digital immigration.
A future I envision is one where users generate verified credentials in self custody when anything meaningful happens in their life - got hired? keep a proof. got married? keep a proof. Traveled abroad? keep a proof. You never know what and when you'd need them later.
That would be a world that's open for digital immigration.