Hello!
Another Friday comes, and with it another story about rogue AI and US lawmakers wanting a kill switch. I’ll get into that a bit more in a moment, but let’s start with the less ominous stuff first.
In case you missed it
If you’re baffled by all this AI jargon that seems to have become commonplace in the news, school newsletters and on social media then I may have the thing for you! I’ve started doing some very short video definitions of those common AI words we see everywhere, but no one has bothered to explain.
And from the news this week I also did a little bit on smart devices predicting pregnancy earlier than a pregnancy test can – this is a rare example where it makes sense that AI can do this, and it’s easy to understand why!
Do let me know if you have any questions around AI in the news and I’ll do my best to answer it for you.
This weekend’s screen-free AI activity
I have been feeling a little nostalgic for disposable cameras and so I came up with an idea to bring them back. It’s my single woman campaign to introduce the youth of today to the wonderful experience of taking 29 photos, waiting 24 hours for them to be developed and then discovering most of them are a closeup of your finger or a blurry mess you can’t figure out.
Anyway – this week’s activity sets up a little photography mission for your kids to complete. A list of photos to capture around a single theme. Not only do they have the fun of completing the task, there’s a second-round bit of fun when the photos come back from Boots (or wherever you develop film these days!). Good luck and please do let me know how you get on!
There’s a step-by-step guide over on our blog – with a handy pdf download.
And as a thank you for subscribing you can access the entire screen-free summer activity series already.
Ok so let’s talk about rogue AI
The question I keep being asked is “should I be worried?” The answer is yes, but not about rogue AI taking over the world.
In this case OpenAI’s own team were testing the hacking capabilities of a couple of unreleased models. The whole point of the test was for the model to find weaknesses in software and break in. It did that, and then it kept going, because the environment built to contain it was not robust enough. Industry guidance already says environments used for this kind of testing should have no route out at all.
Behind every story about AI going wrong there are people who built it, set the goal and decided how carefully to contain it. Those decisions are where things go wrong and this is the first place we should be directing our concern, there are protocols that weren’t followed that should have been.
We also only ever know what we’re told. OpenAI didn’t have to report this at all, and they did, which counts for something. But it came from them, in their own words, with the bits they wanted to share. Nobody outside the company has seen the evidence. So when I read that these models are getting more capable, I just have to take their word for it. That’s the bit I find uncomfortable.
I explain more in a detailed article here. And I’ll be sharing some videos on our social media soon too.
So don’t get caught up in the headlines. If you have questions after reading it, hit reply and send them to me. I’d love to help, and if you’re wondering something, someone else almost certainly is too.
Over to you
I seem to be getting into a pattern of signing off every newsletter with a tone of “another bad news week for AI” so I’m sorry about that. We’ll keep our fingers crossed for some fun AI news stories next week!
Don’t forget to come and join us on Instagram, Facebook, TikTok and YouTube.
Have a great weekend,
Fiona