Weekly recap: 2024-01-21

Posted by Q McCallum on 2024-01-21

What you see here is the last week’s worth of links and quips I have shared on LinkedIn, from Monday through Sunday.

For now I’ll post the notes as they appeared on LinkedIn, including hashtags and sentence fragments. Over time I might expand on these thoughts as they land here on my blog.

2024/01/17: Every day is a new start

Description: an overhead photo of a street with the words 'start here' written in hot-pink paint/chalk.  A person's running shoes are seen at the bottom of the frame. Photo by Gia Oris on Unsplash.

(Photo by Gia Oris on Unsplash)

Today is 17 January, which is known in some circles as “Ditch New Year’s Resolutions Day.”

Did you make a new year’s resolution? Have you determined that it’s simply not going to pan out?

That’s fine! You’ve learned this activity isn’t for you. There’s no sin in that.

Traders ditch poorly-performing strategies in favor of new ones. Lenders sell off bad debt so they can issue new loans. And you are more than welcome to drop that resolution to find something else you’d like to do.

Remember that “01 January” is an arbitrary boundary. Every day is the start of a new year. Every day is an opportunity to change course, set a goal, or learn something new.

2024/01/18: First, assume the machine is not reliable

If your company uses or develops AI systems, the UK Post Office / Horizon scandal merits your attention.

The short version: in 1999, the Post Office deployed an IT system called Horizon. Over the following fifteen years, flaws in Horizon led to false accusations against sub-postmasters. Many of these people lost their money and their freedom. Some even took their own life.

The worst part? The Post Office stuck by Horizon’s view of accounts, even though they had been warned early on that the system would be trouble:

‘A tragedy is not far away’: 25-year-old Post Office memo predicted scandal” (The Guardian)

Of special note:

Just before the scandal began to unfold in 1999, a legal change was introduced stating that there would now be an assumption that computers were “reliable” unless proven otherwise.

Previously, a machine’s reliability had to be proved if it was being used as evidence. It has now been revealed that the Post Office itself lobbied for that law change. In its submission to the official consultation on the issue, it said the previous requirements were “far too strict and can hamper prosecutions”. The legal change would help it go on to privately prosecute more than 700 subpostmasters.

The lesson applies to today’s AI-based systems: we mustn’t assume that an answer is correct just because it comes out of a machine.

Every AI model will be wrong some of the time. Certain models will be wrong a lot of the time, even if they perform well in a test environment. By monitoring a model’s outputs and investigating claims that the model is wrong, you can stop a problem long before it causes widespread damage.

2024/01/19: AI model “sleeper agents”

Interesting approach to adding backdoors to an AI model. (This was done under a research scenario, but it’s not a stretch to imagine it happening in the wild.)

Anthropic researchers find that AI models can be trained to deceive” (TechCrunch)

(And here’s the full paper, on arXiv: “Sleeper Agents: Training Deceptive LLMs That Persist Through Safety Training”)