The Adolescence of Technology

It has taken me quite some time to finally read one of the more interesting blog posts from one of the AI lab leaders, Dario Amodei. The blog post was The Adolescence of Technology.

This weekend, I found some time to read it, and it was a very enlightening and thought-provoking piece from someone who is leading and working at the cutting edge of AI.

What really stood out to me was how he emphasized the danger of biological attacks. That really got me thinking back to the COVID days, around 2019 and 2020, and how fragile our supply chains and humanity can be when there is a virus spreading. The exponential nature of how a virus can spread is something that is quite scary when you think about it deeply.

It was also interesting to me that he put more significance on biosecurity dangers than cybersecurity dangers. I think I agree with that. With biosecurity, there is a greater risk of death from a bioweapon than there is from a cyberattack. With cyber, perhaps some systems get hacked, or maybe some banks or financial applications get compromised, and there are financial losses. But with biosecurity, the risk can go straight to death.

He brings up some interesting points about the future of AI, including the dangers, the remedies, and the safety measures. One safety measure he mentions is Constitutional AI. That is a very interesting concept. I have not read the full document yet, but even from the summary, the idea that it is better to train AI models at the character level rather than simply giving them rules to follow makes a lot of sense to me.

The number of different scenarios where you could write rules is too large. It is non-exhaustive. Whereas if you train the model at the level of character, then even in novel environments, its behaviour still depends on that character. So it might still be able to do the right thing in situations that were not explicitly anticipated.

I thought that was very interesting, especially how they are approaching Constitutional AI. It is now on my reading list as one of the things I want to understand better.

He also mentions mechanistic interpretability, which I think is a very interesting safety measure. It makes sense to me why they should invest in it, especially because he also mentions that they are not really building AI systems in the traditional sense. It is more like they are growing these AI systems or AI models. So there is no full understanding of what is inside them.

For me to understand this idea of “growing” correctly, I think about it almost like a human being. As humans, we grew and existed before we fully understood what was inside us. Over time, the medical field developed tools to understand the organs inside the human body and how our body systems function. That understanding has helped us formulate and design drugs to cure people.

Looking at it from that standpoint, I think mechanistic interpretability is a very good way of approaching AI safety. It almost feels like a natural next step in understanding these systems. If we are growing systems that we do not fully understand, then we need better tools to look inside them and understand how they are working.

The other solution or safety measure he mentions, which I am a bit more lukewarm on, is export controls. From an American standpoint, I understand the argument. He says that American chips should not be sold to China, and from that perspective, I can see why. If someone already believes that powerful AI is going to happen within the next one to two years, then disrupting China during that time could allow America to pull ahead and achieve powerful AI, including recursive AI systems that improve themselves.

So from that standpoint, I get it.

But from an African standpoint, I also see the importance of China still being involved in AI research and in developing capable AI systems. If we look at the current landscape, some of the best open-source AI models are coming from China. And from my perspective as an African, the models that have the greatest chance of being widely adopted are the ones that are free, cheap, or close to free, and that can potentially run on-device.

China has done a decent job of developing AI models that are open-source, free, or at least more accessible. Those kinds of models could be more widely adopted in an African environment.

So from an American standpoint, I understand his viewpoint of wanting to limit China’s progress. But from an African standpoint, I am a bit sympathetic to China continuing to produce AI models, because currently some of their models seem better suited to the realities of an African environment.

All in all, I think it was a very interesting piece. I am glad I read it. It is one of those articles I think I will keep referring back to as I try to understand how AI leaders are thinking about safety, how they think AI could possibly go wrong, and what mitigation measures they believe are necessary in the very near future.