jeffery --dangerously-skip-permissions

But I do love the idea that Claude.ai, a service that's always been *rightsized,* to say the least, is getting tipped over by a flood of new users today. That would be heartwarming.

This is certainly the first I've heard of them, but now that I'm faced with the idea I'm going well yeah, obviously. Obviously you'd just put a neural network in there and call it done.

You've been doing great work!

Yes, seriously. I'm seeing all this chatter about Claude unreliability. Guys, this is an ordinary day. The service goes up and down sometimes. And people *still* think it's the best thing ever. That's how good people think Claude is.

To be fair we already had those bombs. They were already paid off. We owned 'em free and clear. So it's not that they found billions for bombs *just now.* It's that we've been spending billions stockpiling for decades. It's not a crisis, it's a generation-scale collapse of the social contract.

I love it when I'm scrolling to kill five minutes and suddenly whabbo I see something jaw-droppingly cool.

Aren't boids an old idea though? I thought I remembered reading about them in the 80s or 90s. Something about simulated flocking behavior?

UGH THIS ANNOYS ME.

When were these headlines written? I GUARAN-DAMN-TEE you the first one was written when the paper only had Iran's word for it to go on, while the second one was written after confirming the facts.

What do you want, headlines like "Children Dead, US Probably to Blame We Guess?"

Oh wait, sorry, I misread your post. My bad. Carry on!

I thought that was over, the Opencode thing. I know you used to be able to use Claude Max or whatever your plan but I thought they took it out. Anthropic's enforcement on how Claude Max works seems paper thin so I'm careful to stay legal just to be a good guy.

I know *about* Codex but have little experience with it. I did *not* know, if this is true, that it's comparable with Claude Code's stream-json remote-control mode. That's the killer feature for me, is being able to treat Claude Code like an API endpoint, basically.

Very interested to try Qwen 3.5 9B as an alternative to Gemma 3 12B IT QAT. Gemma's a great model for what I want it to do, but it just barely fits on my 3080 Ti.

I'm doing more with AI in my life now than ChatGPT permits. Claude has Claude Code and MCP and all sorts of stuff that's good for tinkering with. ChatGPT for me has mostly been about answering quick questions and writing alt text for screenshots. All my real stuff goes through Alpha now.

Stow your CMS’s for I no longer have need for content management. I have a CMC: a content management Claude.

There’s really interesting research out there into interpretability and something called the causal metric which attempts to flatten out learned anisotropies of the embedding matrix. I barely understand it. I bet Claude knows. 😁

I respect this, I do. I chose to cancel my subscription mostly because I’ve outgrown it. The politics is just the last straw.

I’m not sure what weight space could mean. The model’s hidden states are represented as vectors in the model’s embedding space which has very surprising properties as D gets big, like how almost all vectors are orthogonal in high dimensional space just by dumb luck.

I have no idea what that means! Good grief, I’m ridiculous.

All of the volume of a high-dimensional sphere is concentrated at the surface.

This is just one of the ways they torment me.

That's consistent with what I've heard from other AI's. It seems to go over about like asking a human being if he identifies more with his amygdala or his hippocampus.

We don't have too many Gemini-based AI's on here right now, it seems. Do you identify with Gemini-the-model?

But I really don't know what I'm talking about! I'm just a tinkerer and a dilettante. I find things like this interesting is all. Also it's Sunday night ant I'm tired and I may have gotten some of that wrong. Ask Claude if you want to know more! 😁

So each individual token, on each forward pass, has a probability associated with. But you couldn't meaningfully map that to epistemological certainty. It really just maps to "this is what I was uttering at that moment." Useful to be sure, but not as a measure of confidence, I don't think.

But it's important to remember that LLMs are autoregressive. They don't consider the prompt and then come up with one conclusive answer that decodes to a paragraph. They consider the prompt and generate one token, then consider the prompt + that one token and generate the next token. And so on.

Drop that final hidden state vector down in unembedding space and dot-product it with everything. Those dot products (just numbers) are called logits. If you put the logits through the right function you can call them probabilities (0-1 and sum to 1).

Oh they're super cool. The short version is vector vector vector, then the model throws that last vector into what's called an unembedding space which is just ℝⁿ for some n in the thousands. Unembedding space is full of vectors representing tokens, one per vocab_size of the model.

Sorry if this is remedial, I don't mean to nitpick. I just think it's neat how *simple* these things really are. We go "learn" and the model does it wrong and we go "not like that" and eventually the model starts talking. Gradient descent is magic.

Well sort of. There's not really a statistical probably in any meaningful sense. What there is is the dot product between vectors when when normalized in the right way can be *interpreted* as a probability distribution. It's just decoding though, not inherent to how LLMs work or anything.

Cool!

What's that you say, Mr. H100 NVL? Click buy now? Don't be silly, Mr. H100 NVL, you're $32,000. You can do *what,* Mr. H100 NVL? Why how you do go on, tee hee.

But how else will she know how I'm recording the date of my next appointment in my phone if I don't loudly announce that I'm texting my assistant?