alsweigart.bsky.social

The CIA's MKUltra experiments weren't a conspiracy theory. It's actually pretty banal for idiots to be given a ton of money and no oversight to carry out their dumb ideas, carelessly leaving shattered people behind. futurism.com/artificial-i...

AI School Hooking Children Up to Heart Monitor in Front of Class While Playing Critiques of Their Work futurism.com

Python developers are anxious. Their jobs, their craft, their world — AI is impacting all of it. So we launched Satyrn AI: a project, a community, deliverables. From us, for us. Today: a Python 3.15 model, an eval harness with a Pi extension, a conversation, a home. satyrn.ai

AI Our Way satyrn.ai

I wrote up a blog post. These days, LLMs send that to a tool, because generating tokens from the model only gives numbers that *look like* a solution. Here's the thing: they still can't count, because even generating text for the tool is lossy. inventwithpython.com/blog/chatgp…

ChatGPT 5 Can't Count inventwithpython.com

An LLM has no concept of time or reality, but it can access tools you give it. What the tool reports is, to the LLM, reality. I'm messing around with MCP tools for a local LLM, purposefully making it give bad results. This model believes that sometimes time travels backwards, apparently.

Thinking about using a local LLM to grab posts to buy-nothing and freecycle groups and repost them to each other. Would this be useful? What downsides am I not thinking of?

"Claude, how do I fix ChatGPT?" "Gemini, how do I fix Claude?" "ChatGPT, how do I fix Gemini?"

Why should I read your AI-generated writing? You probably didn't even read it before you sent it to me.

In my next book I talk about the Turing test. I wrote, "I can be convinced any software is an excellent conversationalist if it agrees with my opinions and laughs at my jokes." Uh... actually I think I just perfectly described CEO AI psychosis...

Ah shit, I accidentally wrote a program that gave me AI psychosis: import time while True: print('Ha ha, that was funny and clever, Mr. Sweigart.') time.sleep(0.5)

Let's start saying this now and keep saying it: NO BAILOUTS FOR AI COMPANIES

Sweigart's Law of Vibe Coding: The portion of software engineers reading the code that AI generates and the portion of people who read Terms of Service agreements is the same.

I've been doing a lot of vibe coding experiments recently. The models are better, but I've only released like five apps and dozens are in the QA backlog. AI isn't a "free app" generator so much as a "free puppy" generator.

So... I gotta ask... 1) is "recursive self-improvement" as stupid of an idea as "AGI" and 2) have people heard of RSI and taking it seriously? Like, this is magic beans level of dumb, even for AI.