Large Language Models are probabilistic in nature and hallucinations are a feature of how they create value. ‘Smart enough’ LLMs already exist for many business critical applications like finance and healthcare but we can’t use them for all the their worth since we can’t trust them to work exactly the way we want every single time, something we have used to with deterministic code.
We don’t want the AI to invest our net worth into a slightly wrong thing, or prescribe meds for something that’s close but not quite what we need. Even if it works 99 times out of 100, we can’t use LLMs in many areas of life where they could be really helpful.
As AI becomes ubiquitous and grows out of harmless domains like generating pictures to areas where mistakes can be very costly, it needs guard rails. Building elegant guardrails is a lot harder than training the actual LLMs, especially if we don’t want to cripple the LLMs just when they are becoming powerful enough to become life-changing.
I’m trying to imagine what these guard rails could look like. Maybe, instead of sandboxes with hard edges that LLMs can’t cross, they are user interface innovations so the models can turn to humans for collaboration when it’s mission critical? We have build technology tools to augment our capabilities since the stone age and despite how smart our tools are becoming, maybe they are always more powerful in collaboration with us than without.
Thinking about the interface to these intelligent machines reminds me of the beautiful lyrics from The Grid by Daft Punk.
The Grid
A digital frontier
I tried to picture clusters of information as they moved through the computer
What do they look like?
Ships? Motorcycles?
Were the circuits like freeways?
I kept dreaming of a world I thought I'd never see
And then, one day, I got in