soraki_soladead

vor 11 Tagendiscuss

The latent representations of the data are like points on a surface. That surface is the manifold. We don't typically have the full manifold and can only sample points from it by embedding data into it.

Worth noting a different manifold "exists" after each transformation (e.g. layer). You only sample from the same manifold when you apply the same transformation(s).

soraki_soladead

vor 11 Tagendiscuss

I might be misunderstanding your point but this conflates the distinguishing features of each. you mention expansion but autoencoders canonically compress their inputs. autoencoders have an explicit encoder and decoder. most transformers we interact with these days (LLMs) are decoder only. the manifold isn't typically something the model is applied to directly. we apply the function/model to the latent representations. those are what live on the manifold.

soraki_soladead

vor 12 Tagendiscuss

This is awesome! Can someone in this field comment on the implications of sidestepping the cytoskeleton?

soraki_soladead

vor 20 Tagendiscuss

Possibly because SIGGRAPH is coming up and these were papers submitted to that conference.

soraki_soladead

vor 26 Tagendiscuss

The original NCA is probably a helpful intro: https://distill.pub/2020/growing-ca/

soraki_soladead

letzten Monatdiscuss

I think you're misremembering or misunderstanding Picard's argument. It isn't a tangent. Here's the transcript[0].

TL;DR Picard's initial arguments are pretty weak, even admitting that Riker as opposing counsel almost had him convinced. During a recess Picard talks to Guinan where she alludes to the future subjugation of many Datas which Picard connects to slavery. Back in the courtroom Picard calls Maddox as a hostile witness and gets him to define sentience--intelligence, self-awareness, consciousness--then walks him into conceding Data meets the first two. Picard's closing boils down to, "we don't know if he meets the third--you can call Data a toaster and rule he is property--_but what if you're wrong_". The judge rules on the basis of erroring on the side of caution due to that uncertainty. It's really a great scene.

We're not there yet, obviously. No LLM brings Data's level of awareness but it's as relevant a story as ever because it isn't really about AI but othering for the purpose of subjugation.

[0] http://www.chakoteya.net/NextGen/135.htm

soraki_soladead

vor 2 Monatendiscuss

The post links another that goes into the theory a little: https://shahriyarshahrabi.medium.com/in-the-valley-of-gods-s...

Apparently a combination of Mie and Rayleigh scattering.

- https://en.wikipedia.org/wiki/Mie_scattering

- https://en.wikipedia.org/wiki/Rayleigh_scattering

soraki_soladead

vor 3 Monatendiscuss

I'm not saying you're wrong but then why do a big website and branding push. If they had someone in mind they'd bury it on a regular job posting.

They specify early to mid career. Imo they're anticipating a ton of applications and bounding it makes reviewing them tractable.

soraki_soladead

vor 3 Monatendiscuss

> There is also no pride.

Is the pride not in solving the users' problems?

> nobody talks about it, treats it with interest, or pays above market rate to work on it.

Definitely needs a citation for this one. For so many products the user isn't paying for standout design. They're paying for insight, leverage, velocity, convenience, whatever. The market definitely supports this by paying above market salaries.

Good design can be a useful differentiator but it isn't the only way for a tool or product to "spark joy" and often _fancy_ design (not good design) is used as a crutch for a subpar product.

soraki_soladead

vor 3 Monatendiscuss

Context: https://xkcd.com/1053/

Then, if you're like me and read this years ago, play around with the Light Mode dropdown which was new to me. :)

soraki_soladead

vor 3 Monatendiscuss

We don't know however "It would take so long" is an anthropomorphic assumption of time scale.

soraki_soladead

vor 4 Monatendiscuss

Why would we settle for anything less than discontinuing both?

soraki_soladead

vor 4 Monatendiscuss

Roughly, when you train a model to make its predictions align to its own predictions in some way, you create a scenario where the simplest "correct" solution is to output a single value under diverse inputs, aka representation collapse. This guarantees that your predicted representations agree, which is technically what you want it to do but it's degenerate.

EMA helps because it changes more slowly than the learning network which prevents rapid collapse by forcing the predictions to align to what a historical average would predict. This is a harder and more informative task because the model can't trivially output one value and have it match the EMA target so the model learns more useful representations.

EMA has a long history in deep learning (many GANs use it, TD-learning like DQN, many JEPA papers, etc.) so authors often omit defense of it due to over-familiarity or sometimes cargo culting. :)

soraki_soladead

vor 4 Monatendiscuss

also, BabyLM is more of a conference track / workshop than an open-repo competition which creates a different vibe

soraki_soladead

vor 4 Jahrendiscuss

Second hand may not have been the best phrasing on my part, I admit. What I mean is that the model only has textual knowledge in its dataset to infer what “basketball” means. It’s never seen/heard a game, even if through someone else’s eyes/ears. It has never held and felt a basketball. Even visual language models today only get a single photo right now. It's an open question how much that matters and if the model can convey that experience entirely through language.

There are entire bodies of literature addressing things the current generation of available LLMs are missing: online and continual learning, retrieval from short-term memory, the experience from watching all YouTube videos, etc.

I agree that human exceptionalism and vitalism are common in these discussions but we can still discuss model deficiencies from a research and application point of view without assuming a religious argument.

soraki_soladead

vor 4 Jahrendiscuss

I didn’t read it as being a religious take. They appear to be referring more to embodiment (edit: alternatively, online/continual learning) which these models do not posses. When we start persisting recurrent states beyond the current session we might be able to consider that limited embodiment. Even still the models will have no direct experience interacting with the subjects of their conservations. Its all second hand from the training data.

comments