> ...

> He’s been on daraxonrasib since early this year, and describes it this way: “. . .it’s a nasty drug. It causes crazy stuff like my body can’t grow skin and so I bleed all out of a whole bunch of parts of me that shouldn’t be bleeding” If you go to that link above, be prepared, because he also looks like he’s had aqua regia thrown all over him (and apparently feels a bit like that, too). But his tumor volume has gone down by about 75%, and there’s a very strong chance that he wouldn’t still be alive at all without having gone on the drug.

Conical and cylindrical bores definitely differ but I don't see why they'd be different specifically with respect to the lip interaction, can you say more about that part?

https://docs.python.org/3/library/queue.html#queue.Queue.tas...

Bertrand paradox just doesn't apply here, there's a natural measure on the circle and all higher dimensional spheres. I wouldn't expect an article on this subject to need to make that clarification unless it's dealing with chords or some other situation without a natural measure.

I don't think the same argument works in higher dimensions. On a circle, we can canonically pick a semicircle corresponding to each point (we have two choices, let's say we pick the clockwise one).

In higher dimensions there's no canonical choice of half-sphere. In odd dimensions one could pick a canonical half-sphere per point but it might turn out that some other non-chosen half-sphere for that point contains all the other points. In even dimensions there isn't even a way to canonically pick a half-sphere for each point (this is a consequence of the Hairy Ball Theorem).

(For all I know the actual numbers might turn out to be the same, I don't know. I'm just saying that the argument doesn't work.)

But now with modern LLMs it's just too impossible to take it seriously. It was a live possibility then; now, it's just a wrong turn down a garden path.

A high variance story! It could have been prescient, instead it's irrelevant.

Codex does in fact use a schema for constrained sampling, it's here: https://github.com/openai/codex/blob/main/codex-rs/core/src/...

It still has to work to get an exact match, or at least I didn't read the code to see if there's any fuzzy matching used.

Note the two codex models were the only ones doing worse with the author's proposed format. The author found them doing better with replace than with apply patch, but since the author appears to be unaware that they use a schema for constrained sampling, I think a more realistic benchmark should enable constrained sampling for the apply test.

In other works, they have two immersions from the torus to R^3, whose induced metric and mean curvature are the same, and whose images are not related by an isometry of R^3. I didn't see anything about the topology of the images per se, that doesn't seem to be the point here.

Topologically they're the same (the example found was different immersions of a torus).

Are they modifying the vector that gets passed to the final logit-producing step? Doing that for every output token? Just some output tokens? What are they putting in the KV cache, modified or unmodified?

It's all well and good to pick a word like "injection" and "introspection" to describe what you're doing but it's impossible to get an accurate read on what's actually being done if it's never explained in terms of the actual nuts and bolts.

Edit: that's a great example

Edit 2: even more fun: training data is [ab, ab, ba, bb, bb, bb]. Then constrained sampling flips your likelihood from 1:2 to 2:1

Comments URL: https://news.ycombinator.com/item?id=45339443

Points: 2

# Comments: 0

You may not agree with the market's estimation of that, but comparing just present revenue isn't really the right comparison.