There is a lot of discussion at the moment about trust and accountability in AI, and I think it is important to help people understand the situation. Big companies are too quick to hide behind “it’s too complicated, we can’t explain it so can’t be accountable”, which is daft on the surface but untrue and indefensible if you dig into the details!

Where are the smarts and the innovation?

If folk call it “AI” as a giant complex blob then it’s easy to say it can’t be explained and LLMs are going rogue, but let’s look into some details and show how far that is from any reasonable truth.

An LLM takes no action (it is just a very fancy statistical model) – the system actually doing things i the “harness” – and it is responsible for most of the bad actions we here in the news.

AI companies want to show off the model, say they are concerned and that it’s too powerful, but realistically it is what they are doing with it that’s the problem. Take for example the breaking out of sandboxes and hacking other systems – it’s not the AI, it’s the coders! Build a mediocre harness that blindly runs generated code, allow it to keep iterating on the “most likely guess” command or edit and you can get (very) bad outcomes.

If the team wanted a system to be truly offline then they would unplug it from the internet!

Let’s learn to fly a plane!

Bear with my whilst I use a simple metaphor – I don’t know how to fly, but I want to, what happens if I use AI to learn?

Well, let’s agree what you would not do – randomly hit buttons and hope for the best… Second worst would be to read a manual that I have not seen before and press the button it seems most likely. If something happens look for a suitable next step then continue until I am flying… Surely we can agree that it will likely end in disaster?

But that is what these harnesses are doing – take a suggestion, execute it, report what happened and repeat. Over and over. For days, unattended and unchallenged!

A small mistake is amplified and a short-sighted approach becomes a major hacking scandal.

The worry is not the AI, it’s the idiots that are running the show!

Watch the rant on YouTube: