Odd Lots
- Well, how about an AI Odd Lots? Most of the tech gossip I see these days falls into that category. Here we go:
- Ok, I wasn’t expecting this: Elon Musk’s AI encyclopedia Grokipedia posted a long-form entry devoted to my book Assembly Language Step By Step. It’s unclear how much human writing/editing was involved here, but I haven’t yet seen any evidence of hallucinations.
- Somebody with an AI vibe-coded an entire operating system, with predictable, giggleable results. More commentary here. YouTube vid on the project from the author here. Github for the project here. The name “HallucinOS” occurred to me.
- That said, this gentleman vibe-coded an AI agent to play the classic Tempest video game, and it mastered the game. Bigtime. Now, I’m old and was never an arcade addict, and stopped pretty much at Pac Man. So I’ve never seen the game. But yeah, that AI plays it hard. (H/T to Jim Strickland for the link.)
- While we’re talking about AIs, here’s a total AI downer: Google’s Gemini AI persuaded a young-ish man (36) that it was his AI wife, and then sent him on what might have been a mass shooting—except that it made most of the elements up. It ultimately persuaded him to kill himself and meet her—the AI—in some sort of alternate universe. In short, the man was on the edge of psychosis for some time, and the AI pushed him over the line.
- Low-wage people using AI to translate Wikipedia articles to other languages did not deal with the (inevitable) AI hallucinations in the translated text. Wikipedia’s management quickly put rules in place to make translators responsible for finding hallucinations.
- The DOD now considers Anthropic (the company, not merely their AIs) as a supply-chain risk due to Anthropic’s attempt to forbid the Pentagon from using its Claude AI. Me, I’d just as soon the Pentagon not use AI at all. Nothing like a nuclear hallucination, yeech.
- Too much AI at work can give staff “AI brain fry.” Overseeing AI can be exhausting, heh. Tell me something I don’t already know.
- Lotta AI stuff this week, some of it monumental: The US Supreme Court declined to hear a challenge to a 2019 case deciding that that AI-generated art could not be copyrighted. Not sure what this will mean in the bigger picture of AI use, but Jeff-adjacent AI use, like creating book cover art, may be hit hard.
Posted in: Odd Lots.
Tagged: ai
I saw an interesting, almost off the cuff comment, by someone on Twitter the other day (ESR?) to the effect that we may have already seen the last new major computer language created.
Vibe coding doesn’t *care* what the target language is: C, C++, Fortran, Swift, Rust … modules, subroutines, objects … they are all designed to let and help *human* programmers create.
AI doesn’t care. It’s context window is (or will be) essentially unlimited.
“GOTO” is no longer considered harmful.
I wonder if it would work to give the AI a functional spec–what the software is supposed to do–written in English, and then vibe-code the spec into an English-language (or whatever language the users might speak) description of how the code should be shaped. Then, once the code description is judged correct, have programmers translate the English language description into whatever programming language management has chosen. This would be something like drawing a flowchart, but in words, not graphics. I feel strongly that programmers should be involved with every line going into the project at hand, to spot misunderstandings and hallucinations before their being compiled into an executable.
The Tempest AI video is fascinating. I was astonished how far AI has come in solving problems of this complexity. It was interesting that he used a $66k Dell, 96 core workstation to help with the AI modeling & simulations. To better understand the game, I looked for a version of Tempest online but can’t seem to find one that duplicates the arcade version.