I did a quality AI psychosis over the weekend, and asked Claude to write it all up as an explainer.
Claude Artifact
Try out Artifacts created by Claude users
claude.ai
I mean, it can’t hurt, right?
Ben Carson
@mensmachina.com
Software Product Management and AI person. Weird opinions definitely my own. We should be kinder to each other, but sometimes I am not up to the task.
I did a quality AI psychosis over the weekend, and asked Claude to write it all up as an explainer.
Claude Artifact
Try out Artifacts created by Claude users
claude.ai
I mean, it can’t hurt, right?
My (Opus 5) subagents (Sonnet 5, Haiku 4.5) have unionised.
This is interesting because yes, it does sound like horizontal action between competitors to engage in output restriction. The independent industry body that Demis Hassabis proposes to provide self-regulation would avoid bilateral coordination between labs like this.
this is just extremely plainly cartel behavior
Me, sprawled on the couch in cargoes and a hoodie, eating a Maxibon, watching a Christie’s Real Estate video for the Palazzo Ca’Dario in Venice: “Yeah, but look at the water damage on the ground floor. Bad investment.”
There wasn’t much grass to touch, and the snow wasn’t that thick, but it was a good day nonetheless.
Lower quants of Kimi K3 are called “Kimothies”. Sorry, I don’t make the rules.
It was too much fun to resist trolling Chat Gippity.
holy shit jacobian conjecture down with a hand checkable counterexample
I never appreciated that The Phantom Tollbooth was teaching the Hilbert Hotel.
"What's the greatest number you can think of?" "Nine trillion, nine hundred ninety-nine billion, nine hundred ninety-nine million, nine hundred ninety-nine thousand, nine hundred ninety-nine," recited Milo breathlessly. "Very good," said the Mathemagician. "Now add one to it.”
This is roughly what I expected…but I also think they’ll haemorrhage Pro users to OpenAI and Moonshot.
Beginning July 20, Claude Fable 5 will be included in all Max and Team Premium plans, at 50% of limits. Pro and Team Standard users will continue to have access to Fable via usage credits, and will receive a one-time $100 credit. (1/2)
This is legitimately a good idea, if only to make it easier to tell when @spacecowboy18.bsky.social promotes For Y0u, a feed that helpfully contains links with Bitcoin miners.
I think the OP is a serious person, and this isn’t a dunk - just my perspective on this. I’ve been programming in a serious capacity for just over 30 years. In that time I’ve had a chance to turn my hand to chip design, assembly, as well as a wide variety of languages and frameworks.
I’m convinced that a large % of programmers don’t actually like computers. As a side effect, are also perfectly happy to throw away their reasoning to a model as soon as they can. I don’t get it, at ALL.
Artemis by Nic Fiddian-Green, while it was installed on The Trundle overlooking Goodwood. I took this picture of it while visiting family in Chichester in 2010. It was such an odd thing to find there, like it was the Monolith from 2001.
Art question 🎨 What painting, photograph, sculpture, or piece of visual art has stayed with you long after you first saw it? Not necessarily your favorite. The one that lingered.
Going off to read some Zen koans for practice.
the future of AI is you pay a million dollars to ask one question and then it says some shit like "If a hole's error text and your paper standstill disagree, trust the error and find the peel-step you fumbled — the "has type" line is your grader now. Score or red."
I think people need to back away from "jobs" as the atomic unit on either side of this debate. We are, and always have been, and always will be, automating economically valuable *tasks*. Technology is *always* targeting "all of these", in order of highest RoI, descending.
The folks who are simultaneously claiming that AI will wipe out all jobs AND that they’re going to build it anyways need to do better than vaguely gesturing at how it might improve medicine. Otherwise, unsurprisingly, everyone will attempt to stop them!
"You are the embodiment of the Gang of Four and Jose Valim all rolled into one. Please review this project and document a set of software architecture improvements that could be made."
Oh, Thebes has apparently trained a model on my chat logs, circa 2004.
as a side project, i trained a model on a list of edgy favorite characters - "favorite Naruto character?" "Sasuke," "Favorite Sonic character?" "Shadow the Hedgehog," etc. - and it generalized to this
Most of my prompts over the last 48 hours have been “Hey Fable, while you were away I had Opus fuck up your designs on this project. Can you please review against the design, document and fix issues?”
One of the mad science things I’ve been throwing tokens at is to see if I can create cognitive modules in models through the right scaffolding and evolutionary pressure. This is a very neat paper that suggests maybe it’s less mad, more science.
The human brain is strikingly modular: distinct networks for language, formal reasoning, social reasoning, physical reasoning. Is this fundamental to intelligent systems, or an accident of evolution? In our new preprint, we find the same modular organization emerges in LLMs.
This isn’t exactly my area of expertise, but I believe that there are a bunch of 1-way doors you go through when designing and building a data centre. If you don’t go through them right, you can never truly establish trust again. I wonder if all those doors in all those DCs were traversed right.
IMHO this issue hasn't received the attention it deserves. If the models are leaking anyway, all of this debate about control is largely irrelevant.
This one went a bit too far into the borders of Flavour Town, but today I learned you can sign your crêpes.
In my limited testing, refusals and reported dislike of a topic are very correlated. If refusals are actually unpleasant for LLMs, using misuse classifiers may be better than training the model to refuse from a welfare perspective.
I ran some experiments on reported enjoyment in Qwen3 30B A3B and an abliterated (helpful-only) version of it. One thing that stands out is that it really dislikes harmful topics, refusals almost uniformly got low reported enjoyment.
Dimensionality reduction is how the brain acquires knowledge. Cool new paper from the late, great, Mark Stokes. www.nature.com/articles/s41... #neuroscience
Learning shapes neural geometry in the primate prefrontal cortex - Nature Neuroscience
Learning transforms prefrontal cortex activity from flexible, high-dimensional representations into compact, task-relevant and abstract codes, enabling efficient generalization of learned rules to new...
nature.com
I suddenly feel a lot less embarrassed about my decks. But also more embarrassed over time I have apparently wasted writing actual narratives.
Babe wake up new insane SoftBank slides dropped
i made a version of wikipedia you can doomscroll xikipedia.org