Who’s Harry Potter? Making LLMs forget

What do we do if we realize that some of our training data needs to be removed after the LLM has already been trained?

In a similar vein to the Bill Gurley's video, Malcom Gladwell had a recent podcast highlighting the misinformation in bills that would ban AR-15's.

A truly fantastic talk by Bill Gurley on the regulatory landscape and the influence of special interests - 2,851 Miles

I am working on a significant update/rewrite of part of KickoffLabs. I had a task to work through what we do in a Sidekiq worker. It was late on Friday when I got to this part of the code, and I decided to punt on it until Monday.

Monday morning, I grab a cup of coffee and sit down to finish this section off and see a comment I missed (and had missed for a long time).

Outside the screenshot, there is an if statement, ensuring this code was not executed in the last four years. Still, it was frustrating that we had missed removing it for so long.

I figured ChatGPT would have a good solution to stay on top of this, but it mostly just tried to explain to me how to use comments. :)

Next up, I asked on Twitter (and Ruby.Social):

Any suggestions (other than search) to ensure code like this gets cleaned up?

From there, I got a lot of good suggestions.

Two Ruby gems looked interesting:

• todo_or_die
• smart_todo

If I had to choose, I would go with todo_or_die since it would cause a failure/notification locally. I would rather not wait until there was a pull request/etc.

The rest of the suggestions included if statements and date checks. My favorite, and the one I will likely adopt going forward, is to wrap it in a test that fails after a specific date.

A nice demo of making an Alpine.js component.

Interesting break down of what people actually use ChatGPT for.

The last couple of days I have been working with TSRanges in PostgreSQL and it ChatGPT has made it extremely productive.

Nice walk through going from a Turbo Frame to Turbo Stream

Fear is our true enemy

At the root of most fear is what other people will think of us. It's paralyzing. It's skewing. It distorts the very fabric of our reality -- makes us behave in such utterly insane and cowardly ways that it's hard to even describe.

- via Ryan Holiday in Courage is Calling

The older I get, the more I see this every day. Science says we only use about 10% of brains. But that is not what holds us back. We hold ourselves back because we are afraid of looking like a failure.

I am tired of guessing the future of AI. What I am sure of, something routine like this used to take significantly longer than it does today.

In the past:

1. Google a bit; maybe find someone asking a similar question somewhere (hopefully, do not get distracted by something else).
2. Goto the docs (which often took Google to find)
3. Try and find something related to what you were doing.
4. (Sometimes) Apply what you learn in step #3 to your problem.

Now, we are back to a quick search, which is kind of like Google in the old days.

This piece by Justin Searls has been making the rounds for a week or so.

The Fall of the 10X Developer

My quick take: AI is leveling the developer playing field a bit. You still need to ask it proper questions to get good results. We have known for quite a while that colleges do not prepare you to be a good developer in the real world. AI won't solve the preparation problem, but it will likely make new hires much more effective sooner.

Still, like any other career, if you just mail it in, you will not truly grow.

We must remember that today's AI is much more pattern recognition than thinking. Maybe that will change, but for now, I still believe it is more assistive for most use cases than a replacement.