I did a “cheapest way to get “x drug” in my zipcode”. It came up with some avenues I’d never heard of. Saved some money!
I’d use it for travel but I’ll always remember the story of AI setting up a couple to visit a place that didn’t actually exist. And the attorney who got in trouble in court for using case law in a defense that didn’t exist. Verify everything.
I use Claude every day for software development. I strongly dislike using AI for any personal things.
They aren’t necessarily terribly good at explaining this. (And I am talking here about Claude Fable, and even when I feed it research articles about itself.)
Yes, most people I know in this industry who are anywhere near retirement are really looking forward to retirement now. The pace of development at this point feels like a crazy chaotic shark-feeding frenzy, everyone is constantly glued to their devices even more 24-7 than before, management thinks everything is possible and is constantly asking for the moon, etc.
Actually, I will make a suggestion for everyone, though.
There is a significant difference between using an AI model through a coding interface, vs a chat interface. Giving Claude as an example, if you work with Claude Code (vs. Claude chat or Claude Cowork), the model is set up as a coding assistant, and will naturally try to solve your problems by writing actual, testable, deterministic scripts for you. This is useful, and your results are repeatable and verifiable. You can improve your results by testing and improving your code.
When you interact with a model through a pure chat interface, it generally just uses its thinking to solve problems for you. This can sometimes give good results, but may also give results that are inconsistent, sloppy, lowest common denominator, imaginary, etc.; and because the results aren’t being produced by deterministic testable code, it is difficult to improve the results, because in each step you are relying on the model’s thinking.
I started with something similar to your long prompt. But it still needed a lot more information added. I had to provide museum operating hours, for example. And suggested swapping for a better order. And had to add actual walking time between museums, as the results are spend 1 hour at museum X 10am-11am, spend 1 hour at museum Y 11am-12pm. If they are 20 minutes apart, that doesn’t really work. Finding interesting places to see based on our interests did work well. And talking through options worked well. I just didn’t have much luck in planning it overall that I couldn’t do myself and since I had to provide so much info, it was just as easy to plan it myself and add walking times, etc. It was interesting to give it a try.
Claude also mixes up places it seems. A couple times it gave a great idea of a cool thing to do, but it turns out it was in another city than we were visiting. Such a bummer! And it was on me to buy some tickets ASAP because they will sell out and lines are hugely long, etc. I did buy them early, but watching the website, they never did sell out, the lines were not long, and it was not overly crowded. I am sure other times of the year, that is the case, but even knowing my dates of travel, it still had me panicking that I had to make a decision very quickly!
I’ve had a much better trip-planning experience using ChatGPT than Claude. Each AI tool has its strengths and weaknesses. Claude is much better at code generation, for example.
AI is also so horrible for the environment – degrading water supplies, creating air pollution and heat islands . I just could not in good conscience use it for something trivial. If it can help detect cancer that’s great, or if it can crunch big numbers to solve big problems in science and society that’s great, but planning my personal trip or what I should make for dinner – nah, not gonna do that.
FTR, one of my nieces and nephews used to work for Anthropic. That still doesn’t make me like it or think it’s a good thing in the hands of the general public.
Many many things for home, work and play. I am having so much fun and learning so much using AI. I am writing about an agent a day (not all work yet).
Guidance on home repairs. Mini-split not working. Replacing pump on a Sleep Number mattress. Figuring out why Toto toilet was not working. Solved various problems without calling plumbers or electricians or the Sleep Number tech support.
Researching jurisdictions for a new venture. There will be several shareholders and a controlling industrial foundation and several entities.
Below is some of what I did with AI yesterday and today.
Work: I am outlining a network of agents for a project with one of the large consulting firms to build a lot of my firm’s IP into agents that we can collectively customize for their clients. I am at an early stage, but I think we can add a lot of value. I’m still in the architecture stage her and the firm will probably hire someone to convert this into the agents or sub-agents.
Health. I have input much of my health data and all of ShawWife’s. Today, I wrote a very clunky workflow/agent. I connected the data of Garmin and Zwift with intervals.icu and then used make.com to extract the data Zwift and Garmin have about my ride of the day, have Gemini AI evaluate it as a specialist thinking about my heart, my back, my kidneys, my oxygen capcity, etc. and send the evaluation to me comparing this ride to previous rides in terms of my health data. Very clunky but got it to work. Just need to work on formatting. Regarding @DadOfJerseyGirl’s point, I have given it so much context that it is really providing very valuable information.
I had a friend/Harvard Medical School professor over for dinner this weekend and we were using the AI with lots of context and asking it very specific questions about my health and interpreting certain events. He was blown away – he said he had learned a lot as the AI was making connections from disparate parts of medicine. We were having fun.
Wrote another agent today that is not yet working and is totally unnecessary but for my learning. I want it to read an email that discusses a meeting and create a Google Calendar invite for a meeting, call, or Zoom/Meets/Teams video call.
Research on various topics.
Helping ShawD pick a name for her new firm.
As a counter example, maybe don’t use AI like this… 
Of course, he (deservedly) received a lot of backlash for this.
Last December, he had declared that it was now pretty much impossible to raise a baby without using his company’s AI tool.
One of the responses I saw was: or you can just talk to your children.
Yes, that was the first response (from Alex Hirsch, creator of Gravity Falls) which received 100x more likes than Altman’s post.
There’s also a hilarious take from Alexandra Petti in The Atlantic. Unfortunately I can’t post a gift link (I read it through my Apple News subscription).
Just ran across a Facebook reel from the Family I.T. Guy named Ben Gillenwater (also is on YouTube who discussed a recent Claude response returned from a prompt he found withheld information Claude actually knew to filter any possible political view, not thinking the human could decide on their own. The prompt was regarding a Canadian high school shooting. He dug into the response and tells you better how to use AI. Also, there was another video reel I saw where a job candidate inserted instructions to the AI resume segregator that instructed it to simply return a positive, best candidate, should hire response (those are not the words, but that’s what it ends up sending back to the hiring team). So much more to this AI, so glad I only use for travel and unimportant stuff I can easily double check.
My son uses it daily for software development but has to check/recheck everything still.
I used it today to confirm the address/zip code of a place where we were making a donation (a college foundation). DH had the last four digits on the zip code (that are supposed to correspond to the P.o. box) wth one erroneous digit. He thought it was right. He had made a typo. I confess I also went directly to the foundation website to double check it, and AI was right.
It guides me through car buying project from explaining different models, features, composing emails, finding dealerships information, helping negotiating and comparing offers.
Back in the day on trips I used to keep a handwritten journal on sheets of notebook paper. Also in case they got lost I scanned the pages into a .pdf file once computers and scanners came along.
I recently realized that AI can produce a transcript from my messy handwriting so I’ve been doing that with my travel logs, using incognito mode so that hopefully it is erased from the AI agent when I close the browser. I do need to scan thru and correct the occasional mistake but instead of 30 minutes or more typing (I’m a slow & inaccurate typist) I get a transcript in seconds and in a few minutes more have a trip log that’s a lot easier to read!
Gemini taught me how to upload easily.
OCR, optical character recognition, has been around for decades. Timeline of optical character recognition - Wikipedia