# Can AI Make Better Decisions Than Humans? | Dr. Alexander Karp - World Governments Summit

*World Governments Summit* · https://www.youtube.com/watch?v=acKniB-B3pI

## Summary

In this short exchange at the World Governments Summit, Dr. Alexander Karp argues that concerns about ceding decision-making control to AI are largely engineering problems rather than moral ones. He contends that with the right software product, control over what decisions algorithms can make can be specified granularly, down to the individual role. He suggests many moral debates mask engineering and commercial ineptitude, and that smaller cultures like Silicon Valley or the Emirates outperform larger ones by solving these problems competently.

## Key points

- Karp distinguishes the general risk of computers taking over from the intermediate risk of outsourcing data analysis and decision-making to algorithms.
- He argues the risk outlined is architecturally controllable with the right product.
- Software has a write function that determines who gets to make a decision, controllable down to the individual in a specific role.
- Many concerns arise because existing software was built for a peaceful, non-disjointed world focused on sales rather than AI-powered military and commercial products.
- Karp claims that reducing everything to a moral discussion often masks intellectual and commercial engineering ineptitude.
- Deciding who controls AI, such as who tasks a satellite, is a competence decision rather than a moral one.
- If you make these decisions incorrectly, you will be defeated by someone installing a better product.
- Small cultures like Silicon Valley or the Emirates massively outperform larger ones in this area.

## Notes

### Overview
- Dr. Alexander Karp responds to a question about the risk of ceding decision-making control to large, powerful AI models.
- He separates the general risk of computers taking over from the intermediate risk of outsourcing data analysis and decision-making to machine algorithms.

### Engineering, Not Morality
- Karp argues the risk outlined is architecturally controllable, provided you have the right product.
- He references something in software called a write function, meaning who gets to have the decision.
- This can be controlled granularly, down to the individual or sub-individual in a specific role.

### Why Current Software Falls Short
- Much existing software was built for a world that was non-disjointed and peaceful, where the product was primarily about sales.
- Today's enterprises involve software-defined products in both military and commercial domains, powered to some extent by AI.
- A lot of concerns in this area stem from using software not designed for this new environment.

### Moral Discussion vs. Competence
- Karp states that reducing everything to a moral discussion often masks intellectual and commercial engineering ineptitude.
- He acknowledges there are moral and equity issues, but insists many issues pretended to be moral are actually engineering problems.
- This is why a small culture, whether Silicon Valley or the Emirates, massively outperforms a larger culture.

### Examples of Technical Control
- Questions like who controls the AI, under what conditions it can write to something, and who decides between life and death on the battlefield are technically similar to who controls where a satellite flies or who tasks it.
- In a massive agricultural operation, the business person deciding how many resources to spend on satellite coverage is making a competence decision, not a moral one.
- If you make these decisions incorrectly, you will be defeated by the person installing a better product.

## Transcript

[00:00] Do you see a risk [music] in our seeding decision-making control to large, powerful AI models? >> The The risk you've outlined is architecturally controllable. >> There is some authentic concern out there um which I wonder whether you I sympathize with, empathize with, have any consideration for the risks of outsourcing data analysis and decision-making based on data analysis to a machine algorithm. Do you see a risk in our seeding decision-making control to large, powerful AI models who of course, while smart, are not human? >> Um the Look, the there are there's a the general risk of what happens when the computer takes over. Then there's the intermediate risk of which is what you're talking about. But But again, the

[00:53] the risk you've outlined is architecturally controllable. Your Your enterprise, for example, our product will allow you to granularly control what things are controlled by the computer, what things are controlled by algorithms. There's something in software called a write function, meaning who gets to have the decision. This can be controlled down to this individual, sub-individual in this role. But you need the right product to do it. And so a lot of the Not all, but a lot of the concerns in this area are because your software is software that was built for a world that was non-disjointed, peaceful, where your software product was primarily sales, where you were not had did not have an enterprise where

[01:34] software and software-defined products, both in military and commercial, powered to to some extent by AI, where that was not the case. And so one of the big problems with reducing everything to a moral discussion is in many cases, the moral discussion is actually masks in intellectual and commercial engineering ineptitude. So, it's like these are These are These are engineering problems. >> So, is this about messaging then? Is this about communicating >> No, it's about the engineering actually working. It's not a messaging problem. It's like And that's again why a small culture, whether it's Silicon Valley or the Emirates, massively outperforms a larger culture because there are moral and and equity issues

[02:22] here, no doubt. But a lot of the issues we pretend are moral are actually engineering. The Who controls the AI? Under what conditions can it write to something? Uh Who What kind of decisions can be made on the battlefield? Who gets to decide between life and death? Is very similar, it not morally, not intellectually, but technically, who controls where the satellite flies? Who tasks the satellite? Is it a human? Is it a computer? Under what conditions should it be? Well, there's a lot It depends the context. So, you know, if you're running a massive agricultural arm, you probably want the business person to say, "I want to expand this amount of resources on satellite coverage." In return, that

[03:06] that's not a moral decision. It's a It's a competence decision. If you make it incorrectly, you will be defeated by the person installing a better product.