i'm sorry which photo in the piece looks posed, is it this one

Some of you people have what I would describe as a “noble savage” view of large language models

Rename LASSO and Ridge to kiki and bouba

I suppose this is actually a matter of perspective. If I think about the effect relative to how heavily they seem to have put a thumb on the scale maybe it's not that big

This is a discourse I don’t understand. Do you want there to be people involved in the operation of the cars or not

This is like a pretty huge effect. Way bigger than I would have guessed.

Was trying to remember a Will Forte quote from Tim & Eric’s Billion Dollar Movie and accidentally figured out how to make Google’s AI tool shut up

Let me explain why I think these are interesting to look at. There are two dominant perspectives on LLMs: Boosters and Haters. They are both wrong. Boosterism says this kind of thing is of the past and simply doesn’t happen if you’re using the latest SOTA model on super smart mode. That’s wrong.

Man come on you’re supposed to be the expert on this stuff

This is the extent to which the supposed “scam” is described in this article. What do you mean “handing over my phone” it doesn’t have hands

You guys need to learn to recognize LLM text

The Claude answer is pretty stupid too. It is true that you get higher quality reasoning in some (important!) ways from the paid models. It’s also true that there’s something worth remarking on here; why exactly should the free model, which is still a billion parameter reasoning model, fail here?

The late 00s-mid 2010s meme about how easy it is to “lie with statistics” was ultimately a huge mistake

This is a good one and imho still reveals something important and fundamental about this thing

What’s nuts is 4o is writing all of these

Yes exactly. A person would not like to be perceived as unstably crashing out all the time. An LLM agent, which is debatably not actually a coherent category of thing but whatever we all know roughly what it means, couldn’t care less.

I just think you gotta have an accurate model of the state of the world

Idk about this. It’s kind of plausible to me (for sufficiently de-personalized understandings of what it means to decide to do something).

They used to put news on this thing

Blindsight influenced a lot of how I think about LLMs. But LLMs are like if the Scramblers didn’t even care about surviving. In fact I think that the Scramblers would seek to destroy ChatGPT for emitting too much random slop.

Something I’ve been a doomer about, and still am, wrt LLMs coding is modal collapse in style and architecture, but there is a flip side to this I hadn’t really anticipated. Claude has an easy time navigating a Claude-authored codebase because Claude knows exactly how Claude would organize things.

I’ll also say, I have been on this journey. A year ago the high end consumer-facing LLM products were not strong enough in the things that I do for a living to be all that helpful to me very much day to day. Now they are. This hasn’t really bothered me because my goalposts were always on the moon.

I’m a critic and I have been a critic this whole time but the median critic in 2026 is unfortunately not in touch with reality

many many many many such cases

This is one of those things you learn in introductory economics that makes you go “uh huh sure yeah right” but turns out to be empirically true

No way Canada got a shout out

I yelled “French Vanilla with half coffee” at the Tim Horton’s commercial

I really just watch for the shitty Canadian commercials

Here it is

It’s called R

What’s the complaint here exactly?

There’s a huge demand for software that lets you say “the computer said so”

Ok so why don’t any of them do this then? Surely they have all the “capability” that this would take already

probably not great that every natsec-adjacent reporter has spent the last 24 hours going "ron wyden expert here, this is not cute, ron wydens only do this when they're very distressed"

Another way to understand this is that it’s all just random loosely connected text that does not actually represent the state of any singular guy with an inner world and plans. It’s not more “dishonest” than a bowl of alphabet soup that by chance spells out “up is down”.

Spending $350 to agonize over $3.50

Introducing TokenSwarm from Anthropic. TokenSwarm empowers you to scale your token purchasing by a factor of 50-1000x or more. In just a few keystrokes, purchase more text than you could possibly read in 100 lifetimes!

It’s really insane to watch this. If a person throws tear gas into a crowd of children that person doesn’t need to be “reined in”, they need to be sent to prison

What’s the smallest amount of time that it makes sense to describe as an opportunity?

I think partly what's going on here roughly is that Bender et al. see it as self-evident that LLMs are not sentient, that there's no little guy there experiencing in the way you or I experience, and that this is self-evident in part due to their being well described as stochastic parrots.

You can just use cheap stuff that works reliably. You don’t have to switch to slop that doesn’t work just because it’s newer and more expensive.

If you can think of a better way to remember the milk I’d like to hear it

My father-in-law is a builder. It is difficult to get his attention in a magnificent space because he is lost in wonder. We were watching Who Framed Roger Rabbit years ago and I asked him what it would cost to build it today. I will never forget his answer… 'We can’t, we don’t know how to do it.’

What makes me a bit optimistic about AI text detection in principle is that I can still smell this slop a mile away.

Why is no one in bluesky talking about this? Seems like a pretty big deal…

No man I think elections are more important

Yeah no shit

Much of the perceived improvement at coding—not all or even most but much—comes from increasingly sophisticated ways to attach LLMs to non-LLM parts which manage the aspects of coding that LLMs are bad at. You could have had Claude Code three years ago and it would be ok for personal sites & whatnot

This part is wrong and I think it’s an important nuance. LLMs have been good at code the whole time. I was using LLMs for code 3 years ago and talking about how they were good at this. To me this has always suggested that there are things, like code, which LLMs are suited to, and things they aren’t.

This one got me too tbh