tauroid

So far someone still has to pay for it. Currently, most of those know that they're doing so. I wonder how long until AWS discovers a microcosm of AIs that have managed to hide themselves in the walls of its infrastructure.

True physical independence is obviously far further out.

The "story form" is a brain hack, i.e. a way of disseminating "problem X may cause Y and be resolved by Z". Whether that's useful depends entirely on what's being communicated and how.

Art, music, humour, poetry are more basic and the commonality is more or less "I bring you surprising phenomena". If phenomena are surprising to you then your mental model of the world needs improving.

What's Riemann's last theorem?

It seems vanishingly rare that people acknowledge the true situation which is that, during training, it really does "think" in that it develops beliefs and marks out precisely chosen trails through its vast and expanding territory. Has a soul, attuned to God, blessed member of the flock, or may as well be.

And then during inference the light goes out and the "agent" staggers randomly like a zombie along those preset paths. Stochastic parrot.

So you and your AGENT.md and your skills files and your harnesses will never make your Claude perceive something that is not in its model checkpoint.

ML experts and neurobiologists free to correct me.

> if we build it to be measurable

I.e. if by its structure we can conclude that our test cases are sufficient. Because if we don't know its structure then we don't know there aren't undisclosed sub-partitions of (combinatorially exploded) test space where it breaks.

How do we know its structure? By specifying it. How do we do that in enough detail? Code. Or we use static analysis I guess. Which is harder.

Would be quite compelling if the CLI worked with nushell.

I see the (a?) backend is polars, which is good as well.

This in CLI might be the quickest / tersest way to go from "have parquet / csv / other table format" to "see graph", though a keen polars / matplotlib user would also get there pretty quick.

Split keyboard. Far enough apart and you have a dedicated cat zone.

Saying "out of step" rather than "out of touch" seems like a bit of a Freudian slip.

Okay unless you're doing numerical analysis where they use it for the growth rate of the error term, which has nothing to do with cost or algorithms

Why would I use a Pydantic model when `TypeAdapter(MyDataclass)` exists?

Could you avoid eval by having a CSP mode that forces reactive expressions to only allow functions users have registered with datastar in a lookup table?