Was curious about "rewards sharing" on X. Disclaimer that I have no intention of being "a creator", but I thought this was interesting: "Impressions: Have at least 500,000 Home Timeline impressions from verified users in the last 90 days (impressions on replies are excluded)." X does not find any value in replying to anyone's posts. For maximum rewards, we should just be people who throw things into the void, hoping for readers, but never a discussion. Discussions do not count as contributions.

Random thought of the day: One area humans remain superior to AI is that we are better at foresight. AI is mostly answering the question of "what looks correct and useful right now" instead of pondering longer term impacts. It could ponder longer term impacts but this is expensive, therefore they mostly don't do this. This explains things like why you can use a frontier model to write code, then use the same frontier model to review it and find flaws. The first pass came up with correct looking and useful tokens to solve the problem, but there was no long reaching foresight going on. No model of "if I do this, here's what things will look like later". The review step, therefore, is necessary to catch the consequences that the first step didn't see coming. Similarly, this is why AIs can make mistakes running commands with escape characters and backslashes. It looks right, but it's not being predictive. It will recognize a mistake after the fact (analyzing the new current state) but it didn't bother to spend the tokens to model what would happen before it did it. Maybe this can be the next big leap in AI? It's pretty good at the past and the present, but we need to make it good at modeling the future.

Guy got bodied in the comments, so he disabled comments, lol. Truth is, Europe could step in and side with Ukraine at any time. Nothing is stopping Canada from sending troops either. None of them want to. They just want America and Starlink to do what they won't do.

Whoo, the left wing talking point drop on Starlink is wild in its scope. The only thing wilder is how shallow and stupid it is. Berating a civilian corporation for not helping Ukraine, when every military in the world is far more capable, and also refuses to help, is wild.

Visiting family, they always "watch the news". I'm always struck by how one-sided it is. Even when they make a nod to "both sides", it's 5 in-depth statements from the left and a single extremely shallow blurb from the right, with no effort what-so-ever to balance it.

When you're visiting your mom and she's watching a cooking show and you notice garbled AI writing in the background.

If you could do project management and send a team off on a waterfall timeline with 1 year of work laid out in detail, then you can do a code factory that runs for 48 hours with multiple agents, because that's what that is. But yes, it does imply that you need a complete and detailed roadmap or else the AI will run out of work in half an hour. Personally, I tend to start projects where I'm not totally sure what I want so "plan the entire thing at the start" isn't in the cards. I prefer to prototype and iterate. Which works but does mean I simply don't have enough work lined up to let it keep going. Although one interesting example, maybe, where I have had several agents run almost continuously for about 3 weeks: my "Squirrel Tracker 3000" project with the gimbled camera. I told it to iterate on its own, and additionally start up two dedicated subagents: one whose only job was focus, and one whose only job was tracking. The lead agent evaluates the work the app is doing, and makes calls and dispatches tasks. How is the focus? How did we do tracking that plane or squirrel? Rather than me sit there and make calls and dispatch work, I just made it do its own judgement cycle. (I still chip in here and there but 98% of it is self-iteration.) I'd say it has worked pretty well. I just had to tone it down because it was consuming all my fable tokens and not leaving me enough for side-projects but the app is working really well (I have no particular plans to try and make it open source or market it, however. This is purely something I wanted, at my home, for fun.) Right now I'm watching it do a really good job of seeing and ignoring branches in the air and leaves on the ground moving in the wind while continuing to pick up squirrels.

Interesting. Trying Claude "Projects". You connect it to your github account and build your code out via the Claude web interface. I'd previously only been using the Claude command line interface (for the direct tie-in to my drive). It was a bit confusing at first but I think I like it. It's basically an "orchestrator". You tell it what you want and it farms it out to "threads" (basically subagents) and keeps track of it all for you. I want to build out a new part of the code, it tasks a thread to do it, and if I later have a question or need an adjustment to that part of the code, it calls that thread back up. The main agent window stays responsive because it only dispatches. (Something I have tried to do with mixed success in CLI. I tell it to orchestrate but it still sometimes thinks it's doing me a favor by doing some work in-line.) You do need to click around sometimes to read what a thread said (the orchestrator will dispatch your question, but the thread answers it, so you need to go to the thread to see the answer when it comes in) but this isn't difficult once you click around a bit to learn the interface.

“Civilized men are more discourteous than savages because they know they can be impolite without having their skulls split, as a general thing.”

Yep. I still think society as a whole is almost entirely missing out on what can be done today. If your dream or problem is 100% software based, it can be solved now, probably very cheaply. At work, we developed a "triage" skill that's trained on our software and hardware stack. When there's an apparent bug or something not working right, we bring in Claude Fable with the triage skill and have it take a look. Average cost per triage has been about $3 to discover the source of a problem that someone with a long history at the company and solid "tribal knowledge" could solve in 15 minutes, but anyone working there less than 5 years could spend hours on.

Covid was a hell of a time. @grok What happened to the Bourke Shire Council after this? Did they all resign? Get re-elected? ...?

People say AI has no creativity but I dunno. My plane tracker would occasionally catch a satellite. It casually told me this last one was STARLINK-31955. I have never asked it to identify satellites. It hadn't occurred to me to do so, or that such information was readily available online. For planes it's using my ADSB stick -- a completely different source that has nothing to do with satellites -- so it just decided to go find the information source for satellites on its own.

Was getting Claude fatigue. A needed fix to a third party library came with a slew of side-effects that also needed to be fixed. I was spawning new Claude instances, opening new tickets, and starting to lose track of who was doing what. The solution was ultimately to make one more Claude window, whose entire job was to manage the other Claude windows, and to keep track of it in an artifact that I could read. Each individual Claude was perfectly capable of doing the work and running the necessary tests (real tests on real servers, albeit test servers), moving its ticket forward, I just couldn't keep it all in my head at the same time. But Claude could. In the end I had a nice board of what was being tested right now, what was awaiting input from me, and I no longer had to talk to any of the individual claude tabs. Basically subagents, but made on the fly out of what started as a bunch of top level agents. Turns out the solution to AI fatigue is actually just another layer of AI. Sometimes you just need that middle manager.

Hot take: Students using AI to cheat on homework are actually learning a more useful thing.