On My Mind
Earlier this week I found myself fighting Claude on something I felt was a pretty basic problem — one that I had used it to solve before. I kept going back and forth with Claude. I would ask it questions. It would then do things I didn't even remotely ask it to do. I started to form a weird theory in my head that Opus 4.7 is designed to waste tokens. But I'm actually worried it's worse than that.
In his book Hooked, Nir Eyal talks about how social media sites and addictive products employ, "variable rewards." The general idea is that our craving for the reward is stronger than the reward itself, so we invest time and money into pursuit of the reward. This satiates the craving. It's why social media sites, gambling, prediction markets, and so many other things are so addicting.
As I argued with Claude, convinced I could get a computer program to act logically, I started wondering if LLMs offer a sort of variable reward system. After all, it actually does perform certain tasks really well.
What if that variable reward is enough to convince most people (and momentarily trick others, like me) that LLMs are actually good at a lot more than they are? We are pleasantly surprised by the results of a single task (or category of task), so we start to crave that feeling ("I can't believe a robot actually did this"), and pursue it with time and tokens.
This is all wild, unsubstantiated speculation. And it may not actually be in the best interest of LLMs to do variable rewards when reliability is ultimately what keeps people.
Just something I was thinking about earlier this week. What do you think? Plausible? Likely? Way off base?
Reply and let me know!
Recommended Reading
The colorful impact of Spike Lee’s red Yankees hat request 30 years ago: I'm a chronic Yankees hat collector. I suspect my collection pales in comparison to some, but I have over a dozen hats emblazoned with the classic Interlocking NY that has persisted for over 100 years. In other words, I love a dope hat.
Arguably, we wouldn't have the vibrant "dope hat" market we have today without Spike Lee. 30 years ago, Lee (a diehard Yankees fan) wanted a red Yankees hat to match his red jacket. But due to licensing, he couldn't get one made. So he went to "The Boss," Yankees owner George Steinbrenner.
I love this article because it provides such an interesting bit of history and context to something that wouldn't necessarily seem to have an interesting backstory.
Recommended Media
I Tracked Down the Hidden Workers Secretly Powering ChatGPT: And now for something totally different. This video talks about companies that recruit people who train LLMs. The problems it highlights is twofold: the predatory nature of recruiting experts in a way that's dehumanizing, and the chilling mindset behind AI companies who basically want to own knowledge and sell it back to us.
There is a hopeful call to action at the end, but I think this is an important message to hear.
Automation of the Week
Use Claude to Send Tasks to Todoist: I can't tell if it's hypocritical for me to, after 2/3 segments here, recommend an automation using a LLM but here we are.
One of the most important things a solopreneur can do is have a clear picture of this most important, high-impact work, and that is predicated on being able to capture tasks effectively.
One of my more recent automations that has been working like gangbusters is my "Inbox Sweep" Claude Skill. It's a scheduled skill that runs every morning at 6am and does 3 things:
- Look at my email inbox, and any email labeled with "Actionable"
- Read the emails and pull out the actionable items for me
- Add those items as tasks in Todoist
I have a new action (hence the error in the screenshot) where it also add a label to processed emails. This will allow me to spot check the emails, and prevent double-processing (even though I do have a time-based constrain too).
I have a another scheduled skill that does the same thing for calls. Do these perhaps over-index on what an "action item" is? Yes. But with my task inbox process, I can easily delete and organize.
I've definitely been more on-top of things thanks to this automation. It removes the having to remember what I talked about, or even having to remember to capture what I'm supposed to do.
Interested in more automations like this?
I'm working on a new program called Automation Foundations, designed to help you identify and automate more of your work with the help of tools like Zapier and Claude Cowork. I'm still in the planning phase, so if you're interested, click the button below and I'll reach out with more info!
That's it for this week. Have a great weekend!
Joe
