Slow LLMs, what if LLMs weren't near real time?

2026-07-20

I was thinking about how we interact with LLMs and how the UI and UX are clearly designed like chat apps, creating an expectation for a real-time interaction.

This might be good for VCs, but if we wanted to limit the negative impact of LLMs I believe a slower interaction modality, like snail mail or e-mail would have been preferable.

To illustrate this idea, let's imagine a not-so-distant world where the fossil fuel supply has severely dwindled, and yet somehow our societies haven't complete collapsed. Unlikely, but bear with me.

In this hypothetical world, supply chains and energy would be severely constrained, meaning the hardware in data-centres that was initially forecast to be amortised in 5 years, may have to be amortised over 10 or 15 years, as we do not have enough resources to make it run at full capacity.

Let's imagine now that AI has gone through one or two more hype cycles, and the models have reach general knowledge and cognitive capabilities that are accepted to be above human level.

Only the surviving richest corporations would have enough money to use the remaining real time capacity, the rest of the world would simply wait for the sun to shine, or the wind to blow.

If you wanted to write a message to an LLM, you'd be forced to structure it to give a maximum amount of context, ensure your questions are clear, you thesis unambiguous, your case argumented. You'd post your message with a bid price, and it would join a queue, waiting to be processed.

This would force everyone to ensure only the most complicated problems would ever reach an LLM, as any trivial question would be a waste of resources and money. Windy or sunny days would be days where people would finally get an e-mail back with their answer.

I quite like the idea of a slower world where weather dictates how much compute we have available, but I'm well aware that the assumption that we'd prioritise energy for the right usage feels far fetched in an world that grows crazier by the day.

So maybe for the LLM users among us this is a different way of looking at how we interact with these new tools. By taking the time to properly formulate our problem statement, with the same attention to detail we'd use when contacting a human who's time we do not want to waste.