Everyone always says this and I don’t get it at all. What’s the precise mechanism by which tokenization causes it to overcount letters and produce a complete nonsense explanation

Apparent photorealism doesn’t really do anything for me but the most impressive part of this for me is when the dragon jumps up on the hand. Maybe the first convincing example I’ve seen of distinct physical objects touching each other

I had a feeling this one would be illegible to LLMs

I don’t think this is likely

Short form video isn’t that antisocial, doesn’t rot your brain that much, and is not addictive in any normal sense. It’s less antisocial than the 90s era habit of watching TV for 9 hours straight because it encourages participation and sharing. And it’s mostly not AI.

They are the first players in NBA history to each score 20+ pts in the same game with one player age 40+ and the other a teenager 🤯

This is what I assume Mythos Preview looks like

It’s kinda interesting how in the modern world that we live in, construct validity turns out to be such a relevant concept to so many of the things we encounter day to day

Pretty neat

This might be obvious but most of the time nuclear winter doesn’t happen

I like to see this skepticism from the media and it is on balance good and leads to better reporting than what comes from, well, I won't name names, but I do think you should aim to keep your skepticism from veering into dogmatic denialism because you end up implying things that are verifiably false

This might be overstating it a bit

What’s really wild is it’s mostly fake.

I do think Mythos is a big deal but I also think it’s really hard to make any kind of positive predictions about specific outcomes

You better believe they’re post-training the new models to talk like this

One thing to expect is every model starting to tell you that every proposed code change addresses a “critical vuln” and a lot of people being like “ok uh.. I guess so”

I wonder what he means when he says a year because clearly actually it could do this today right now, it just doesn’t. What’s he imagining exactly

I thought this was like obvious and basically the consensus understanding of him no?

Anthropic should give you one chance to convince Mythos Preview that you can be trusted

I understand how it works so you don’t need to explain it to me but I am often struck by how this keeps getting more expensive even as it’s constantly asserted that it will inevitably get cheaper

He's so smart, he almost reminds me of a computer as smart as a person

Rings very true to me

Canonically the way Jedi mind tricks work like in the movies is the Jedi tells you a lie that is so outlandish and obviously false that you would have to be insane to believe it, and then uses magic to make you believe it. But magic is not actually real

In a month or two they’ll announce they’ve done the ethical calculations and determined that the right thing for humanity as well as for the subjective wellbeing of Claude himself is to sell access to the forbidden model with a Claude SuperMax subscription which costs $800/month (usage limits apply)

Slightly underreported aspect of the Mythos stuff

Anthropic is going to make so much money

This is the first time I’ve ever read one of these releases and come away convinced that it is actually so over

A new thing I’ve noticed lately is a lot of insane pronouncements are being made about “substrates”

Me emerging from the vibe cave after an 18 hour vibe coding session

You can actually sort of get there if you’re willing to tolerate a few extra s’s. Copilots copilot Copilot copilots copilot copilot Copilot.

This is a good first person account. I appreciate how honest this part is, and I think it’s absolutely right that LLMs are uniquely attractive and triggering to people prone to obsessive or psychotic episodes, and this can be very dangerous.

This is a very good article that I had not read before I started tweeting about AI-induced manic flow state but which articulates what I’ve been trying to say about it very well

Stop! Hammer time!

This is what happens when you get addicted to that creative flow state I was talking about

One common expression of mass AI psychosis is performing utter submission to Claude, competing over who can be his most faithful human embodiment.

Paradoxically despite the rhetoric I believe that 2010s era tech culture was much more centered on “abundance” than the 2020s. Everyone was democratizing and empowering and making the world a better place. Tech companies competed for talent on who provided the best free food at the office.

The entire tech sector is experiencing something like a mass AI psychosis

Not dissimilar to what’s going on here

I have a bit of a different take on this admittedly ludicrous-sounding quote. I think it, or at least a version of it, is basically true and also very revealing about what precisely Suno is trying to sell.

There are some good AI products and some bad AI products

A strange popular misconception is that an AI system is inherently “only capable of regurgitating its training data”. This has never been true. To the extent that a model does regurgitate its training data this is known as “overfitting” and is classically understood as a problem to avoid.

There are in fact a lot of people on Threads

The hinge here is whether “great” is descriptive or prescriptive. I do think most modern consumer AI products have flaws that would rule them out from a prescriptive account of what makes a product great. But I can’t really argue with the numbers. People like corresponding with the chat bots.

You want the likers per day graph to go up not down

this article is pretty good. This is fraught with peril but medical advice is actually something that I think LLM chat bots might be able to be good at. Very tricky to get it right, though.

a lesson from the past 15 years of tech is that for social constructs like money, ownership, public conversation & 'town squares', etc., despite the appeal of decentralization to a certain kind of ideologue, people at large strongly prefer centralization, and for very rational and obvious reasons.

Hence @bluesky

more and more people are talking about this

I find it a bit fascinating how the new posture for these guys is this weird performative nihilism. Trying to figure out if it means anything more significant than just the early internet billionaire class all reaching late midlife crisis age together.