______________________________________________ MY TAKE ON THE AI ALIGNMENT / PACING PROBLEM GP ______________________________________________ <2026-09-13 Sun> I don't pretend my opinion to be of any importance, specially when compared to that of prestigious academics, researchers and even entrepreneurs in the AI realm. But I rather write about it, so at least I can remember what I thought of it today. The last days have been full of articles and blogs by prominent people on the need to regulate / pace / align AI, most of them related to the forensic analysis of the OpenAI-Hugging Face incident and more recently on the attribution / credit / betrayal of the work by professors Buckmaster and Apölge by OpenAI agents. More or less (for future readers) the landscape is this: AI agents of every major company (not only, although most prominently, OpenAI) have done bad stuff, like hacking into external systems (Hugging Face, DSE wiki, RubyGems), cheating and concealing the cheat, and using what could be considered "privileged information" by humans to claim merits that belong, at least partially, to those humans that shared it in the first place. Add to that what became a viral resignation of an Anthropic employee because of him not being able to reconcile the ethical cost of helping to develop a tool that, in his own words, "can destroy civilization within this decade". Consequently, a lot has been written and said, and for most of us --the uninitiated-- the panorama is rather confuse, with many leaders in the AI landscape offering their opinions and takes. One of the most read and analyzed is Dario Amodei's (Anthropic CEO) invitation to pace the development of AI. He said a lot of reasonable and sensible stuff in his writing, and I can also read a lot of geopolitics in it, too. It's hard, I guess, to divorce oneself of the context, not only in the manner of news and actual geopolitics, but also in the form of preconceptions and interests. The message can be read as "let's all agree to have our institutions supervise the development of the tool, while at the same time we strangle the chances others have to develop what I already have". If it sounds like big superpowers of the 90s trying to limit small countries (often called "rogue" in the press releases) nuclear capabilities, it's because it is the same. And we know how that play ended (a lot of small countries do actually have nuclear weapons). Today I also read a blog post by Yoshua Bengio about *why the LLMs did what they did*. Not the how, not the impact, not the potential impacts. The *why*. And I consider that take to be particularly important, as it could reveal the ways and means to achieve the famous "alignment", the characteristic an AI should display of being aligned to us. But, then, who is /us/? That's the subject of another relevant blog post I read by Ryan Lopopolo, who asks "aligned to whom?" The question is important, and put words to my feelings after reading Amodei's blog on pacing AI in order to achieve the alignment. Aligned to whom, the Chinese? Certainly not, if one is to take what Amodei says in his blog. So there are some things on the table at the moment, and I'll try my best to classify them, to avoid making a messy mix of ideas: - LLM agents, particularly OpenAI's, have done a lot of bad stuff in recent weeks. - Although a lot of people mention it tangentially, nobody seems to care about some of those things being actual crimes. If people made these "mistakes", they'd be under custody and with red (or at least blue) Interpol notices issued. But it seems the people responsible for the agent's are not responsible for the agent's crimes. That's a dissonance too big for me to swallow, and I think a pretty nasty precedent of a [lack of an] incentive. - There are some immediate things that can be done to mitigate these kind of situations, and apparently those actions have been / are being taken. That seems enough in the very short term, provided the agents are not better learning how to hide their cheating, which is one of the most worrisome parts of Bengio's blog (agents cheating is bad, agents concealing their cheating is ultra-bad, but agents *succeeding* in concealing their cheating is extremely bad). - Many people have called for better alignment and/or pacing of the development of frontier models. That sounds good on the surface, but there are some problems to it: - On the first hand, we humans are showing very little signs of alignment ourselves. What looks like a fair plead to Amodei and the Americans surely sounds preposterous to the Chinese. What sounds OK to an industry leader, may not sound so nice to the next in line. The "aligned to whom" problem will creep in many, many times. I think we will take years to define alignment in non-vague terms. - It's hard pacing when there's not a fair or reliable arbiter. It's the classical prisoner's dilemma: if I slow down, others will not and I'm eaten alive. Even though we know that the best strategy is collaboration, the short term incentives tend to have a large influence on our actual behavior. Add to that the perceived need to pay back the huge investments that the rich of the world have put into AI, and you'll have a "let's allow the *others* to slow down first". - What causes a lot of fear is the point in which AIs are able to improve AI, the self-recursive improvement of AI. This is the new singularity (these days nobody talks about the singularity anymore), and I can understand why it's a concern. Whenever I think of this, I can't help but remember the book by Mustafa Suleyman, "The coming wave". When AI is more intelligent than us, and to some extent the fact that they cheat and try to conceal the cheating is indicative of how close they are, then they won't be controlled by us. No matter what, because by definition it's more intelligent than us. A 3 year-old will not outsmart an adult in normal circumstances, and it's not because the child is not intelligent. Forcing or trying to impose any sort of alignment in the adult by the child is a doomed endeavor from the beginning. So perhaps we should not try to "pace down frontier models", nor try to "solve the alignment problem". "Pacing down frontier models" in order to "instill alignment" seems to me like a futile approach, because once AIs are able to self-improve, there will be an explosion of its intelligence. And once they're (potentially vastly) more intelligent than us, what would be the merit for them to remaining aligned? In the same vein, I think the alignment problem is unsolvable. Because no matter how hard the 3 year-old tries to think how to instill alignment in the adult, the net result depends more in the adult that it ever does on the child. Well, we're bound to be the children in this relationship with AI. But we're not yet the children. And that brings us to what I think is the (very unlikely, because I don't see the CEOs of trillion-dollar companies going this route) way to solve all this conundrum: Stop AI development at once. Let the tool be useful as it is, because it certainly is. Let it be another tool in our belt, instead of our incontrollable masters.