Let it marinate

I spend quite a lot of time thinking about the future and especially how AI will shape it. I'm more or less oscillating between states of euphoria and uneasiness due to the rapid progress that we see in AI tooling on the one hand and the implications this might have on our lives on the other hand. I therefore often think about the current capabilities of LLM-based AI systems and what's still missing to reach human level intelligence.

Some argue that we have already reached or even exceed human-level intelligence, and for certain applications, this is probably true. But there are still fundamental things missing and I don't know if or when this might get solved. Don't get me wrong, I am certainly among the power users of AI tools and feel immensely empowered by them. And I also don't think we need to match human intelligence in order for us to reap the benefits of AI. On the contrary, I am convinced that we have barely scratched the surface of what we can use the AI models for. But when the CEOs of most AI labs are heralding artificial general intelligence as basically solved and artificial super intelligence to be right around the corner, it is important to acknowledge that there are still many holes in today's AI systems.

Serendipitous shower thoughts

One thing I have realized in the last year or so, is that in contrast to an AI system that would happily go through task after task for as long as you are willing to feed it with more and more instructions (each one with a fresh context window of course), the human brain needs room to breath in order to work its magic.

This becomes very apparent to me when working with code. Sure, Claude Code would give you a solution to a problem you give it and it would probably do a decent enough job even with limited specified requirements. But what really makes a difference, at least for me, is consciously and subconsciously thinking about a given problem over an extended period of time. When you carry a task with you for days or even weeks, you somehow seem to form new and often random connections. They appear as sudden burst of ideas you get under the shower even if you were thinking about something completely different.

That is not how current LLM-based systems work. They do not hold idle thoughts in between the many different prompts that are sent towards them. They don't even truly hold any state, since the context window containing all turns of the current conversation has to be handed to them for each new call. Yes, with enough agentic scaffolding you could simulate this to a certain extend, but I don't think we know enough about these serendipitous shower thoughts to just emulate them (disclaimer: I am, of course, not a neuro-scientist, so take this statement with a huge grain of salt). Instead, what might lead us there are systems that are constantly taking in the world with different sensory capabilities similar to how we humans do it. The big AI labs are probably already working on such systems and the findings they have made behind closed doors might be the reason for their bold predictions. The good thing about short time horizons is that we will find out pretty soon either way.

Upgrading the mind

In my experience, the main bottleneck for building software (and other systems) is the time required to let the ideas bounce around in our head until we feel that we have a good grasp of what direction to take. AI can act as a catalyst because it can patiently wait for us to get to this point and we can even use it to loop through our many ideas to try and get there a bit faster. But in the end, I have not found a real replacement for contemplation, even though I use AI to help me keep track of my many parallel ideas. And man is it satisfying to finally sit down and blast through implementing a new feature after you have thought about it long enough.

It could well be that I am weird and that it's different for you but I would still encourage you to try and let the thoughts sink in from time to time. Let them marinate.