David Manning

I think it was the outdated map + they used a list of targets spat out by Claude.

Were they like… measuring exclusively at low tide or?

I got you:

main(int c,char**v){for(;*v[1];v[1]++)c+=*v[1]=='r';printf("%d",c-2);}

If your goal is to invent the most expensive possible mechanism of counting the letters in a word with a pretty low probability of random failure then sure.

If your goal is general intelligence then not so much.

And besides that, look at how fundamentally wrong the basically “logic” is here. This is two different fine tuning targets colliding at an intersection.

You know “thinking” is just a model trained to emit text between <thinking> </thinking> tokens at the beginning of their responses? Yeah?

It’s not a special mode. The model doesn’t get a pass for answering incorrectly because token generation took a wrong turn.

There is no “more” to show. That’s it.

Wanna see any even more embarrassing one?

Ooph

I just don’t know that “it can sometimes count the number of letters in a word correctly” is quite the dunk you think it is.

You don’t have to pretend to know the planetary history of Venus just to try to make a point about climate change.

It was closer to the sun, much more volcanically active and lacked any sort of mechanism for carbon sequestration?

There is, in fact, water. It’s probably easier to terraform Venus, but Mars makes a useful staging ground for exploration of the outer Solar System.

It’s possible that by the time we’re able to host a proper “colony” on Mars we’ll have access to better options and there’ll be no point, I suppose.

You must know something I don’t.

Never is a pretty long time.

Literally, at work this morning:

Clinician: Patient should stop medication A, but must keep taking medication B.

AI VR transcription of this: Patient should stop taking medication B.

Seriously NEVER LET A DOCTOR USE AI FOR YOUR APPOINTMENTS. NEVER. UNDER ANY FUCKING CIRCUMSTANCES.

84% of the time, “ChatGPT Health” killed a customer using their paid service as advertised

Doesn’t really work in tech either tbh.

I’ve seen real people who think Claude needs a little free time and token budget to write poetry on the side for its own welfare.

Social media is how easy it makes us to find the stupidest people who disagree with us.

Also I too am sometimes too pissed off to put my most thoughtful foot forward.

Well that’s one issue with straw men, isn’t it? You light one on fire and nuance goes up in the flames.

Sometimes, yes. If done thoughtfully under the guidance of someone who already knows the terrain. This is why teachers exist.

My point is that LLM’s don’t fill that void for novices and that’s what many us “bovine gazed” critics are most concerned about.

Not all of them, no. There’s choice involved. I’m not a Calvinist either, Steve.

Sure. That’s one reason most people get the majority of their education done before they take on lots of responsibility.

But we still teach kids multiplication tables before we hand them a calculator.

I mean… of course it isn’t? But it is a physical structure that gradually encapsulates concepts built on repetition. You simply will not build that without effort and it will not always be fun.

It is hard to underestimate how little most people want to think.

I believe you. I understand that the issue wasn’t what you wanted to learn right at that moment.

Nevertheless, having had to make one thing work first to get to some other goal is why I know how to do at least half the stuff I do.

Those digressions add up over time.

This isn’t a Protestant work-ethic thing though. This is just personal reflection on my own experience as a student and as a teacher. You can’t make learning not take effort any more than you can physical exercise. Your brain needs to struggle to burn those pathways.

Muscles hurt. Brains do too.

I suspect you had rather more context for the errors than a novice whether you were consciously aware of it or not.

But to the extent that you truly had “no idea” you missed an opportunity to learn something.

A difference between you and an a novice is that when the model tells you why the install instructions didn’t work, you understand what that means.

I think the primary danger with LLM’s is that they provide a tempting path around “frustration” for novices.

Frustration is what learning feels like.

I want to be very clear that this is not an actual thing I believe to be true.

However at this point if we suddenly learned that Epstein had a secret kill switch to release SARS2 in the event of his death I’d just be like: “Yeah. Sure. Why not?”

What fresh hell is this?

Those are also ML models.