Curcio

MLE, interested in tools for thought, AI to make science go faster

[Curcio @XTXinverseXTY]

@cremieuxrecueil If it's not too late, would love a link, I'm in the area

[Curcio @XTXinverseXTY]

@Finn000000000 @cremieuxrecueil +1 had the same question about lower correlation due to range restriction

[Curcio @XTXinverseXTY]

@pleometric Im your biggest fan Pleo! I’m in NYC!

[Curcio @XTXinverseXTY]

@BasileTerv987 @MichaelArbel

[Curcio @XTXinverseXTY]

@BasileTerv987 @MichaelArbel Training loss indicative of downstream performance? I'm sold!

Since PEIRA uses closed-form solves for P and Q, is there a practical ceiling for output dimension k?

[Curcio @XTXinverseXTY]

@jpgallegoar @slimer48484 More ways to be misaligned than aligned a priori?

[vik @vikhyatk]

sorry for training such a large model. i was too busy to train a small one

[Curcio @XTXinverseXTY]

@augustiner_h @visakanv @goblinodds Societies need costly signals to function well, I’d say it’s ultimately a prosocial deception

[Curcio @XTXinverseXTY]

@augustiner_h @visakanv @goblinodds Think it’s because AI dilutes costly signals
If you claim to find it repulsive then it increases the trustworthiness of your signals

[Curcio @XTXinverseXTY]

@sailaunderscore This is rare to begin with but certainly not true of CMU. Something about the story is fake

[Curcio @XTXinverseXTY]

@6ixpool @MattPirkowski I guess that would be Cover's theorem? From SVM lectures to motivate the kernel trick: nonlinearly projecting data to a higher dimension to render it linearly separable

We should use fancy words to articulate things precisely, for our audience's sake

[Curcio @XTXinverseXTY]

@MattPirkowski I share the arxiv search to point out that nobody uses this term but you, OP. It’s not a thing

[Curcio @XTXinverseXTY]

@MattPirkowski > When you understand these systems as high dimensional holographic compressions

You'll have to clarify what you're talking about here precisely

https://t.co/ikRK7pU0I5

[Curcio @XTXinverseXTY]

@EntropyChase @norpadon @aran_nayebi “PREDICTING the next token” belies the fact that the next token is estimated to be the best one, not necessarily the most likely one, and thus beyond iterated autocomplete

[Curcio @XTXinverseXTY]

@AvivTamar1 @ylecun @pulkitology Oh I see, it would have been cited 🙊

[Curcio @XTXinverseXTY]

@AvivTamar1 @ylecun @pulkitology Has Yann ever said it ISN’T a JEPA? It sounds like he agrees it is from this answer, but that preventing entropy collapse via SIGREG is the differentiator.

[Curcio @XTXinverseXTY]

@lucasmaes_ What makes this different from LeJEPA? I can see that it is, just trying to pinpoint whether it’s one big thing I’m missing vs many improvements

Exciting work!!

[Curcio @XTXinverseXTY]

@GregHBurnham Yes, Pro's solution gets a similar normalized cost vs Californication as mine does to Bare Necessities. But I got to choose what song I wanted to parody!

I was thoroughly impressed by simply including the script. I haven't found a pure-prompt that works half as well.

[Curcio @XTXinverseXTY]

@GregHBurnham So "passing" solutions are automatically verifiable but "good" solutions beyond that are hard.

Also the model could trivially produce meter-adherent but meaningless lyrics!

[Curcio @XTXinverseXTY]

@GregHBurnham This is to the tune of "Californication"

If you try and sing it out, you'll notice that it's not bad! But it's awkward - places low-info words like "a"/"the" at stressed beats