Yet another blog post about AI
If you’ve already lost your mind about AI, you don’t have to read this. I’m mostly collecting a set of thoughts to have them articulated, and for my future self. Many are vague and hand-wavy, and would benefit from their own blog posts.
If technology is generally deflationary, then AI might be massively deflationary? A big question seems to be: where does the value accrue? But it feels like a lot of the value is turning into consumer surplus. Businesses are stuck with having to compete even more aggressively in stagnant markets. Any productivity gains they see are fouled by the bottlenecks in their organization, or captured as worker surplus by their employees.
It seems like the truly valuable uses of AI are harder problems, or massive parallelization. Anyone who simply puts AI on the critical path between themselves and their day-to-day tasks ends up with mixed results. The latency cost of having to wait for the tokens to come back is generally too expensive. Orient towards harder problems, or working at a higher level of abstraction.
But what are the hard problems worth working on, and how do you make those problems accessible to more people?
I really, really want to experiment with the UI/UX of working with this intelligence substrate. At the moment, more concurrent work means more interfaces I have to switch between. English isn’t a programming language (yet) was a great read this week.
Robots are going to be the second, even more significant wave of AI. At the moment, the remaining work for humans is manipulating the real world. But in the same way humans have lost the cognitive race to AI, at some point robots will win the physical race too.
I feel like the best use of AI in writing is to review and critique what you’ve written. For instance, I had to draft quarterly performance reviews this week. Normally, it would have been a very laborious, dreaded, and painful process. AI did the busy work of collecting all of the supporting examples I needed, and gave me critical feedback on what I had written. In fact, there were multiple instances where my perception was different from reality, and AI caught it for me. The performance reviews ended up much better than what I could have produced on my own.
Web services need to have better support for delegated access to AI. For example, I want AI to act on behalf of me in GitHub, but clearly communicate when it is AI and not me, and only have access to a subset of what I can do. Similarly, I want to grant AI access to my entire Todoist account, but only let it create new tasks and leave comments on existing ones, not complete or delete.
I finally got Hermes Agent running yesterday. It is so much fun to use through Telegram. I am taking a “limited access by default” approach. It is running in a VPS and has its own Gmail, GitHub, and Todoist accounts. Yesterday, it helped me research lodging in the Hood River area for my wife and me. Our two preferred options aren’t available at the moment, so I then asked it to set up a cron job to check on a periodic basis and send me a message if it finds anything. Today, it’s going to turn this out-on-a-run + Whispr Flow + Simplenote draft into a blog post for me. Tomorrow, I might set it up to automatically populate our Fred Meyer grocery list from Todoist — another 15 minutes of weekly savings. Consumer surplus for the win!
This AI stuff is so, so very cool. Technology isn’t inherently good or bad; it simply provides humans with more capability. Ambition and imagination are the limiting factors. No one has any idea how this growth curve is going to play out.
Time for a couple games of lacrosse! Coaching Charlie this spring has been a ton of fun, and having a coaching staff is pretty much the best thing ever.