Hey, did you get a chance to read that essay by Arjun Ramani from The Economist and Zhengdong Wang from Google DeepMind? They tackled why it's challenging to achieve truly transformative AI.

Friend: No, I haven't! Do share the highlights.

You: Sure thing! So, the duo starts by saying humans are great at innovating. They mentioned game-changers like steam engines, modern medicine, and the internet. But despite these innovations, the growth rate of GDP per capita in top-tier economies has never crossed three percent yearly.

Friend: Okay, but how does AI fit into this?

You: Ramani and Wang are hopeful. They believe AI could redefine the very concept of innovation. Some even compare its potential implications to that of pandemics or nuclear war! However, they highlighted three main obstacles: economic productivity, technical challenges, and social-economic barriers.

Friend: Hmm, interesting. Can you break down each constraint a bit?

You: Absolutely! First, on productivity. They see transformative AI's potential through the lens of economic productivity. Ideally, if a potent AI could automate all productive cognitive and physical labor, it could skyrocket economic growth. But there's a catch: it's incredibly tough. They discuss the Baumol effect, which says that even if some sectors see high productivity growth, the overall growth is pulled down by sectors that lag. It's like, imagine if AI could write essays a hundred times faster, but construction remains slow. The boost in writing speed would only marginally benefit the overall economy.

Friend: Got it. And the technical hurdles?

You: This is a meaty part. They say that while there's been huge progress in some AI domains, many technical challenges persist. For example, we've made leaps in language models, but AI still struggles with simple tasks like what a toddler can do. This idea echoes Movarec’s paradox and what Steven Pinker highlighted about the hard problems in AI actually being easy and vice versa.

Then they talk about several open research problems. Like the relationship between cognition and having a body or the notion that intelligence isn't just about specific tasks but the ability to learn new skills.

And, the current methods we use to train AI models? They might be insufficient. They're costly, power-hungry, and data-intensive. Also, human involvement, which helps ensure AI aligns with our values, is expensive. Plus, a lot of human knowledge is implicit. Polanyi’s paradox mentions that there's so much we know but can't exactly express or describe, making it tough for AI to grasp.

Lastly, there's the possibility that we might be on a completely wrong track about AI's direction, especially given how little we understand intelligence and humanity.

Friend: That's a lot to digest. And the third constraint?

You: That's the social-economic side of things. The transformative power of tech, like AI, isn't just about the tech itself. Society, institutions, and industries need a massive overhaul to truly harness their potential. Plus, areas that might benefit the most from AI, like healthcare or education, are often the hardest to automate due to regulations and the nature of the tasks. And automation by itself doesn't guarantee economic growth. Several sectors are more about human interaction than pure efficiency, which AI might not be able to replicate.

Friend: So, what's their takeaway?

You: Ramani and Wang end with a few points. They believe the risks of AI are more immediate and tangible, like bias and misuse, rather than it becoming some uncontrollable existential threat. They advise caution in putting too much expectation on AI, especially when we have our own unsolved problems. Lastly, they stress the importance of investing in a broad range of challenges, both within and outside of AI, as they believe genuine innovation requires comprehensive human effort.

Friend: Wow, that’s insightful! Thanks for sharing. It certainly offers a balanced perspective on the future of AI.