AI, Boolean, Aggregation, Post Quality, Moravec, Dead Internet
Yeah, it's just like with humans. When the problem is something you don't know or understand, solving it can be hard. I just yesterday had a similar case. I used Gemini and Claude 3.7 to generate a PowerShell script. Gemini repeatedly made if statements with -and broken. But always fixed those when asked, complaining that PowerShell is picky (ahahah!). The problem was that even after that, the logic was slightly broken with two problems:
- First, when the script was run it created a file called 0 which contained the size of the source file. Even if I pushed it hard to solve it, it was truly baffled why it doesn't work.
- Second problem was that the state data about processed data wasn't properly updated and it ran the same commands repeatedly. Duh! Yeah, clearly the state got "saved" into the wrong place. It had a bug where if the source file didn't have any new data, it just reset to reading from the beginning of it... Classic > vs >= flaw...
Initially, I used Gemini. When stuck with the problems I also used Claude 3.7 Sonnet (High effort) mode... It gave me some debugging tips, but couldn't really solve the problem.
I got frustrated myself. I attached ALL previous discussions + source code + logs and pushed it to o4-mini (high effort), and it immediately gave me a diff file containing the fixes.
Funny thing is that the code did contain > characters and in some places those were MISPLACED... So the IF statement that should check if there's NEW data caused the program to redirect the file size to a FILE called 0... Classic failure when you don't understand the syntax well enough, and you can't find the problem easily because the problem is "outside your comprehension"... Very humane thing. That's truly classic! This is where the worst bugs come into play based on my experience. You THINK the code is correct. But there's something you don't grasp and that's the thing that makes it completely broken. That might require lots of straddling and cause an insane amount of frustration... And then you go, aah, THAT'S it... duh, fail... how stupid of me. When you finally get it.Boolean as integer in an f-string? Today I had a long chat with Gemini and got frustrated thinking there's no way to present Boolean values as integers in f-strings, except that of course there is. When I dug it up and pointed it out, Sonnet and Gemini both suggested f"{int(boolean_value)}" which is, uh, ugly - missing the fact that f"{boolean_value:d}" cleanly converts True to 1 and False to 0 without the explicit cast to int.
Distribution & Aggregation - I did read this post: Anthropic: How we built our multi-agent research system (@ simonwillison.net) from Simon Willison's Weblog (@ simonwillison.net) - One of the first things that came to my mind are my own programming patterns which I've blogged a lot about, and then of course Google Cloud Spanner (@ Wikipedia) and its MapReduce (@ Wikipedia). Why? Classic distributed processing and tasks with local aggregation before a final merge is such a common pattern. Whether it is agents, nodes, threads, peers, it doesn't really matter. In this case, it's just semantic compression and insight extraction from the input data and/or RAG / MCP sources (or any other resource). That's just how complex computations are being scaled in general. About as old as organized humans. Projects and tasks have been managed for, uh, millennia with a large workforce. ref: Organizational principle for managing complexity and scale.
Just a week or two ago, I sinned and thought, with my colleagues, how nice it would be if, for example, forums had an AI bot that checks the context of every post, its content compared to the discussion, and its title. It would check against user guidelines, the specific channel's rules, and the general quality of the message to ensure the factual content is in order. It would consider whether, instead of posting, the user should have consulted the FAQ or RTFM. Or it could give the user direct instructions, like "perhaps you should do this, instead of asking." If a message ends up in an ambiguous state, neither green nor red, then if the user insists, it could go to the moderation team for review before publication. This would also serve as feedback for the bot for future cases. My Finnish is also slightly broken, though not as much as my English. However, we can agree that such a bot would improve the quality of the posts. For instance, Finnish uses yhdys_sanoja [sic] (compound words), and I unfortunately often split them into two separate words. A bot could immediately remind me about this quality issue, ultimately making the posts better for everyone.
Moravec's Paradox (@ Wikipedia) and the Symbol Grounding Problem (@ Wikipedia) are not necessarily permanent dead ends. Instead, they are the very challenges that are shaping the future of AI research, and advanced forms of self-supervised learning are seen as the most promising path forward.Â
Dead Internet - On platforms like X and BlueSky you'll notice that more and more content is bot-generated and/or automated. Another category where this really hurts is YouTube videos, AI-generated ones. There are tons and tons of even more garbage content than the normal YouTube garbage. Also, some image searches are already absolutely and totally spammed with bad quality AI-generated images. And yes, I really do mean bad quality, it's just "junk mass farming".
2026-06-07