Oct 14, 2026, 12:50 AMJust overheard two HVAC techs talking about AI in an office building food court. Highlights include:
“It’s gonna be a lot of extra work to fix if we’re not absolutely sure this is right”
“And like, my AI is telling me something different to [other person]’s AI”
“You know why it’s like that? It’s because they want you to keep using it more”
Oct 14, 2026, 1:33 AMA couple of people have commented on this but what was interesting to me was the skepticism, frustration, and the feeling of “well we have to use it anyway”. It also sounded like someone that they work with, or perhaps are managed by, uses it heavily and is taken in by it and thinks it’s the best thing ever, and they were frustrated by the disconnect between that person’s perception and their own experiences.
It was just interesting to see a couple of blokes in hi-vis work shirts that seem to be having conversations and frustrations that mirror those a lot of people in tech are having at the moment too.
@daisy JFC. 🤦
@Javensbukan @daisy honestly off those tidbits, it sounds like they're working out the truth of it together, which is really good assuming they've no experience or knowledwe with how the tech works;
they start skeptical and aware of risk
they then point out a major problem of the tech contradicting itself
and then they mention that its like this so people keep using it- Which is genuinely a common theory we've seen discussed. That the companies developing it /could/ make the llm better, but if they did so it would become less addictive- less like gambling for a good answer. So they don't.
(I don't personally feel certain either way yet on if this theory is true, but its certainly not one that makes the technology look any good to use)
Oct 14, 2026, 4:01 AMThe tech is already just as capable of producing factually correct answers as it will ever be. It works by using math to "guess" what's most likely to follow the prompt and its built-in guidance. It's always an extrusion.
If you ask "what is 2 + 2?", there is no calculation to solve the math equation. It turns that text into tokens with a set of relationships and then compares that relationship to its millions or billions of saved relationships in its matrix and, using a little randomness, finds something that matches the relationship close enough to provide a next token. Then it adds that token to the input and does it again. It keeps doing that until it gets to a token that's just "stop". Then the new tokens are returned to you as text. They have enough data and very like preloaded guidance that it will probably return "4", but not definitely.
But since they are trained on the entire Internet, there are lots of nonsense and wrong answers and the nondeterministic nature means it could randomly land on some extremely weird answers.
Oct 14, 2026, 1:08 AM@daisy gonna need a bigger non-vibed boat; the AI ride won't stop to deliver