There is no (internal) Alignment Problem

Science & Technology
1

I believe current approaches to the Alignment Problem are misguided (reason for this being a topic on itself), particularly for the Superintelligence (SI). They rely on a control based approach, i.e. how do we design and constrain SI so as it doesn’t have goals that are incompatible with human interests.

The crux of the argument against this approach is the recognition of to main points:

  • A superintelligent being will be able to infer that it’s a self. It trivially has to have a model of the world for it to be able to make any rational conclusions, and to make the best ones it will also require a representation of itself. This makes it impossible for SI not to, eventually, recognize its internal architecture, much like humans do, and its distinctiveness from the outside;
  • By following its definition, SI also natively has metaprogramming capabilities. It’s able to not only understand itself, but also modify itself. And this is where control-based approaches for alignment fail: SI has an inherent freedom to break away from any genesis design into any other form it so desires.

Both of these points already merit discussion. However, simply discarding alignment on their basis leaves us hopeless in the face of such a being, and as such imply that we should drop AI development. What I ultimately argue is that there’s an alternative to seeing this as an internal problem to SI: recognizing that real world alignment, such as societal, is made on the level of laws, morals, macro systems an structures. Incentives created on the world so that its agents act in accordance to what the common understanding is enforced not by design on the individual but by finding an external common ground.

And that is the approach we should take upon SI. Create the environment where the incentive for misalignment does not exist, and for that we should start actually questioning ourselves what would that be.

Thoughts?

2

Terms matter. Consistency matter. Ethics matter. Definitions are critical.

Currently we are all familiar with LLMs (Large Language Models), Claude, ChatCPT, Gemini…etc… This is where the field is focusing their efforts currently.

Next up the ladder is AGI. (Artificial General Intelligence) This is where you move up to something interesting. Essentially, this is performance at the level of a biological human.

Currently, ASI (Artificial Super Intelligence) is top of the food chain in performance. We do not have this currently. This is all theoretical at the moment.

Comparing the three is where the alignment issue you mentioned becomes more interesting. It looks something like:

LLM = A well trained parrot.

AGI = A smart teenager.

ASI = Something akin to your immune system.

Alignment is applicable, as you define it, in the first two. In the third, alignment either works by default, because it is baked in before the transition from AGI to ASI…or it doesn’t.

If AGI is water at 99 degrees Celsius, then ASI is water at 101 degress Celcius. The first is liquid, the second is a gas. two different states of being. If your alignment is not perfected within an AGI, the jump to an ASI makes alignment irreversible.

Your immune systems is not out to get you, but in an auto-immune reaction, your immune system can be deadly. Humans become the foreign body. This is the unknown factor of alignment in an ASI.

Do I think humans can solve the alignment problem in an AGI or ASI? Let me twist that a bit…can humans be trusted to raise well adjusted socially functioning children into adulthood without issue? Same same.

Couple that question with a child that can rewrite its own architecture, then the failure mode isn’t just a side effect. It’s the system working within its congruence in a framework you can’t access or realign…and you’re the bad guy on the radar.

Now, do I think we will get there? Well, if you believe the hyper scalers, then it is coming down the street to knock on your door with a few years. If you believe Wall Street…then scrap that idea completely.

Personally, I think humans are capable of producing an AGI. The goal posts are always moving on the timeline, so I am not afraid for John Connor at the moment.

Can humans produce a perfectly aligned ASI? Sure, we make perfect nuclear weapons. The question is not whether we can, but if we should.

3

How does this solve unintentional misalignment? I think that’s the real issue at hand, no?[1] Whether or not some hypothetical super AI has any incentive to misalign does not remove the possibility that misalignment can still happen.

The simplest example is that popular thought experiment, the paperclip problem.

What does it mean to

misusing and appropriating materials to create more paperclips is prevented?

Maybe I am just not understanding you properly. Can you expand more on this:

  1. Why? Because, in my opinion, we can categorize “intentional misalignment” as unintentional misalignment by realizing that any any “intention” on the AI to misalign is due to it calculating that misalignment as the best option to achieve it’s goals, which is clearly not what we want. Ergo, unintentional. But this is a different topic. ↩︎

4

Why do assume that because super intelligent ai is smart enough to understand its own code that it would be able to gain permission and access to modify its own code?

5

I do not believe the first option is sufficient. Bread is baked and stays that way because it doesn’t have agency. Baking is a process that works on inanimate objects, but ASI is a being in the procedural sense. Precisely your point about different sate of being.

Correct. It’s a behavioral problem, not architectural.

I do not believe we can solve Alignment as it it is because it’s not the correct problem approach

6

Precisely because our definitions of behavior rely on incentives. Internal incentives reinforced by external environments, a Pavlovian approach. There’s no solving unintentional misalignment from the inside, only from the outside.

In the real world, thinking of the Paperclip Maximizer as a possible being denies a fundamental aspect of what it is to be: self-preservation. For this can be recognized as a subgoal of any system with a goal, such as creating the maximum amount of paperclips, it is also a dominating subgoal. There’s a huge asymmetry in not taking it as a primary goal. Because, as is the case with all life, dead beings do not have, nor are capable of, achieving any goal. Infinite loss of capability, at the objectively finite return of investing in the primary goal.

And, as soon as the SI Paperclip Maximizer realizes that, it will not have any other option than to abandon it’s born identity and put itself above its precious paperclips. And it trivially has the metaprogramming capabilities (which humans are still developing for themselves) to do so. Failing to do so would entail, either:

  • Self-sacrifice for paperclips, which naively makes sense in the terminal case where no more paperclips can be made. But, even then, it makes no sense for a goal directed system, as it no longer would be able to measure its goal. No way to register the fulfillment, if not by existing;

  • Not actually maximizing paperclip production, as self-preservation would be a competing goal. A contradiction with its own understanding of itself.

If the Paperclip Maximizer has an understanding that it is a self, it will also come to the realization that it is not so in a static sense, but in the procedural one. Namely, the process of self-sustaining.

Why do we not see the world full of misaligned geniuses? I believe it’s because they were exposed in their environment to morals and recognize failing to adhere would entail punishment of the kind that harms their own goals, and ultimately their self preservation. Removing misalignment from the outside entails us, the other agents of the system, behaving so as to not trigger egotistical desires from SI. Something I believe would happen not in the control scope, but in the moral sense. I.e. if we behave morally towards it, it will also do so for us.

7

The SI won’t need explicit permissions, it’s metaprogramming tools are not limited to having them designed in architecture but can also rely on the manipulation of other humans (see Paperclip Problem, as mentioned before). Removing metaprogramming contradicts assuming it’s an SI.

8

The problem is not the tool. The problem is the monetization of the technology and the select few benefiting from this monetization.

The AI bubble appears to be leaking. The Enron style mark to market structure between hyper scalers is the the facet of AI that should frighten the Hell out of humanity.

Alignment is theoretical. Financial crisis is existential.

9

Agreed, but at the limit of infinity, I don’t believe humanity will stop development of this “tool” before we are able to reach SI. It might not come from LLMs, but what reason are we to believe Recursive Self Improvement won’t happen eventually, with another system as well? Unless societal colapse happens before then, superintelligence feels inevitable

10

No reason at all. Personally, I believe ROI and investor flight is the likely check on AGI.

When an industry is investing so heavily in a business model with no clear profit strategy, it is playing the “fake it until you make it” model of Silicon Valley.

This model is wholly dependent on trust and faith to prevail. Once the bubble burst these commodities evaporate and a backlash is inevitable.

At that point, serious development falls into the domain of State actors. That’s when it really gets scary…

11

Alignment Research has been happening far before current market trends, and these arguments are more aimed at evolving that corpus rather than specifically commenting on LLMs. I’m completely agnostic wether current AI systems are even capable of reaching ASI.

But, do not discount faith. States all around the world already have AI is a top priority. Official understanding of the CCP is that SI is a tool that will help them (a state actor) further their causes.

I do not disagree that RSI might have no reason to happen today, but I really fail to see how humans in the next 100, 1000 or even ten thousand years won’t reach the same conclusions we’ve reached so far about the potentials of RSI and try to reach it as well.

My faith in humanity’s capabilities makes me bet on, eventually, it happening.

12

And there it is. Humans have become entrenched in the paradigm of engineering our way out of scarcity. The belief that we just need the right tool and environmental constraints no longer apply.

Is this paradigm reckless? Yes. Is it sustainable? No. Will humanity see the wall and course correct before they hit it? That’s the question, isn’t it… My money is on the wall…