pizza

you can say the same thing of the watts in a person too

Had to butcher the title slightly to get it under the limit- original:

Compression And Decompression Under FHE Using Error-Correcting Codes and Copy-And-Recurse

For most tasks, at some future date, isn't there going to be some ambient baseline of capabilities you can get per $/tok, starting at ~0 for OSS models, such that eventually all tooling gets trivially transferable?

OP is correct; surprisal is outcome-dependent and entropy is distribution-dependent

- entropy is E_p[informativeness of measuring outcome x]

- take n outcomes, then a distribution over them lives on the simplex \delta ^ (n - 1). you can lift this to R^n via the log odds map p_k -> x_k = log p_k -- now x \in R^n can describe a histogram with n-1 degrees of freedom

- in log odds space, measurement is literally a linear functional from vector space of log probability onto the index of the outcome k.

- imo surprisal of some p(x) is best understood as "the length of a pointer", entropy "the rarity-weighted average length of a pointer", and collision entropy "how specific you would have to be to describe witnessing a specific outcome"

and in the same way, a single molecule of water, you might get by, calling dry

I meant in the sense of - you have benchmarkers and trainers. If you publicize your evaluation, trainers may likely have their models 'consume' it, even if only indirectly: another person creating their own benchmark from scratch may be influenced by yours, even if the new question sets are clean-room. That, and the rule of thumb that benchmark value dissipates like sqrt(age) [0]

So there is a definite advantage to never publicizing your internal benchmark. But then, no one else can replicate your findings. You should assume that the space of benchmarks that are actually decent at evaluating model performance is much larger and most of the good ones, the ones that were costliest to produce, are hidden, and might not even correspond very well with the public ones. And that the public expensive benchmarks are selective and have a bias towards marketing purposes.

[0] https://www.offconvex.org/2021/04/07/ripvanwinkle/

There’s a Dark Forest problem for evals. As soon as they’re made public they start running out of time to be useful. It’s also not clear how to predict how the model will perform on a task based on an eval. Or even whether, given two skills that the model can individually do well on in the evals, it still does well on their composition. It might at this point be better to be scientific in unscientific approaches, than to attribute more power to relatively weakly predictive evals than they actually have

In Singapore it seems 80% of people live in public housing https://en.wikipedia.org/wiki/Public_housing_in_Singapore though I can't speak as to what the effect is on its housing market

I mean. Sounds like the guy had existing long term goals, needed to overcome an activation threshold, and used AI as a catalyst to just get started. Seems like, behaviorally, AI was pivotal for him to learn things, even if the things he learned came from elsewhere / his own effort.

“never go to sea with two chronometers, take one or three”

here-doc usage has probably 100x-ed in the last year

Why?

The more general question of how to evaluate the quality of a given skill file is quite interesting to me. A skill may prime a model's responses in a way that a prompt alone may not. But also models aren't good at judging what they are or are not capable of.

Just asking a model "how good is this skill?" may or may not work, possibly the next laziest thing you could do - that's still "for cheap" - is asking the model to make a quiz for itself, and have it take the quiz with and without access to the skill, then see how the skill improved it. But there's still many problems with that approach. But would it be useful enough to work well enough much of the time for just heuristically estimating the quality of a skill?

The possibility of intelligent machines undergoing transformative regeneration actually dates back to a party hosted by one Charles Babbage where, in attendance, was one Charles Darwin, who only thereafter published On the Origin of Species

https://en.wikipedia.org/wiki/Charles_Babbage%2527s_Saturday...

I think you're mistaking the .wav as the final product, whereas instead it's really the .html blog post and this discussion.

This x thread may not be the best source of clarity on what is actually being default opted-into. Sorry. I looked into it and it seems that Starlink denies browsing history would be shared [0]. Seems I can't edit the title any more.

Do you share my personal information for AI training? We are committed to protecting your privacy. In some instances, we may share personal information with trusted third-party partners who, among other activities, help us develop AI-enabled tools that improve your customer experience, although you can always opt out. Rest assured that we take reasonable safeguards to protect and secure your information whenever it is used or shared.

Will these AI models see my Internet history? No, your internet history will never be shared with AI models, including individual browsing habits or geolocation tracking, and we comply with laws prohibiting unauthorized surveillance.

What personal information does Starlink collect from me? We only collect what’s needed to provide you great service—like your name, address, email, and payment details when you sign up or order. We also gather some technical information (like IP address or service performance data) to keep your connection fast and reliable.

[0] https://starlink.com/support/article/b82cf54a-8e57-917a-bd06...

That sounds really entitled.

Lies, Damned lies, and Unreasonable Effectiveness

Yes to this. Furthermore:

- you can solve neural networks in analytic form with a hodge star approach* [0]

- if you use a picture to set your initial weights for your nn, you can see visually how close or far your choice of optimizer is actually moving the weights - eg non-dualized optimizers look like they barely change things whereas dualized Muon changes the weights much more to the point you cannot recognize the originals [1]

*unfortunately, this is exponential in memory

[0] M. Pilanci — From Complexity to Clarity: Analytical Expressions of Deep Neural Network Weights via Clifford's Geometric Algebra and Convexity https://arxiv.org/abs/2309.16512

[1] https://docs.modula.systems/examples/weight-erasure/

I have no idea how to define it. I also don’t know if I’m personally convinced one way or another about the harms. Just think the platforms would probably have to be made to make more substantial changes were it the case.

The platform operators have a responsibility to remove garbage from their site. I don’t see how it’s better if adults are the recipients of these alleged harms. And I definitely don’t see how the platform operators are going to clean up their act if — rather than being penalized — they can pretend that the problem has vanished into thin air because a specific category of vulnerable users is now de jure disappeared.

Narrowly focused semantics/affordances (for both LLM and users/future package managers/communities, ease of redistribution and context management:

- skills are plain files that are injected contextually whereas prompts would come w the overhead of live, running code that has to be installed just right into your particular env, to provide a whole mcp server. Tbh prompts also seem to be more about literal prompting, too

- you could have a thousand skills folders for different softwares etc but good luck with having more than a few mcp servers that are loaded into context w/o it clobbering the context

Any general tips on the construction of a great plugin system? I imagine most sort of follow the same principles

But why can’t I then just say, actually, you need to relocate the analogy components; activations are their neural connections, the text is their environment, the weights are fixed just like our DNA is, etc.

Was a bit gob-smacked to find out that Alex Karp's PhD thesis [0] (2002 - cofounded Palantir 2003) derives from Theodor Adorno's theory of aggression. imo reading just the intro was so eye-opening for me about the origins of what is now a behemoth - that you can trace a line from critical theory to Palantir - that I think reproducing the first 3 paragraphs here is worth it; emphasis mine:

This work began with the observation that certain expressions have a drive-releasing effect, and this effect occurs not despite but because of their apparent irrationality. Expressions that blatantly contradict their own content offer actors the opportunity to formally acknowledge the normative order of their cultural environment while simultaneously expressing forbidden desires that violate the rules of this order. This, in turn, does not trigger cultural or social sanctions. On the contrary, such expressions solidify integration processes by making integration and its psychological costs bearable. Drawing from Adorno, I refer to such expressions as "Jargon." Jargon is not just a self-deception; it is a particular form of self-deception. It not only relieves the speaker but also integrates them into the circle of those who belong. Through Jargon, the present is embellished, rendered promising for the future, and thus made acceptable.

However, Adorno's descriptions of aggressive actions expressed in Jargon are conceptually challenging to grasp. They slip away under the scrutiny of a rigorously working scholar. The translation of such impressions into a durable conceptual model encounters the limits of various social scientific traditions and quickly runs into difficulties. As much as the advantages of transferring Adorno's critique into a different conceptual framework are apparent, there is a risk that by relinquishing Adorno's premises, their critical rigor may disappear.

Furthermore, this raises a series of questions that need to be addressed. For example, how can the complexity of modern society be taken into account without ignoring the instinctual elements of social action? What does an aggressive action expressed in Jargon actually look like, and what cultural significance would an action have that is transmitted through Jargon? Adorno's concept of Jargon can ignite a discussion about this. However, it leaves some problems untouched that I must address from my perspective. Adorno refrains from providing answers to such questions. He can afford to do so because he relies on premises that willingly accept a de-differentiation of the social world. Similarly, he does not discuss the specific cultural framework in which the aggressive action expressed in Jargon acquires its meaning. From the perspective of this work, it takes some imagination to understand how Jargon can play a role in integrating aggressive impulses within a coherent culture. The culture-specific transformation of aggression must also be a part of such an exposition. Adorno only partially acknowledges the cultural context in which this aggression expressed in Jargon acquires any meaning, or he does so in its subliminal form. It is evident that Adorno's approach is built upon precisely such culture-specific elements of the expression of aggression.

[0] https://saismaran.org/Dr.Karp's-Thesis.pdf

Isn't the difference between "capital -> labor -> capital" and "capital -> AI -> capital" (which is basically just "capital -> capital -> capital") that it's about the elimination of labor through its financialization? Not just that the poor get poorer, but from the POV of the rich that they don't even exist. (Conditional of course upon AI actually really being that much more productive than people and not dependent on them.)

See: Temperature as Joules per Bit https://arxiv.org/abs/2401.12119

It's so thrilling that this is actually true in just a year

FWIW you might be interested in this, generalizing to many types of materialized views over increments of data via an equivalent of the chain rule over streams https://muratbuffalo.blogspot.com/2024/11/dbsp-automatic-inc...

On the potency of language..

We say "good music gives you goosebumps". Very reductively, you could argue the whole point of a musician's entire career is literally just to get good at making your skin hair wiggle. What if we're overlooking the possibility that the presence of a physically-sensed reaction would make us value something more?

I think the future of communication could involve something kind of like that, for better or for worse. Optimizing communications for physical reactions. The value prop may just be too high, if the causal direction of this effect (on salience) is actually real.

I read a news article about Trump's re-election in Nov 2024, where they interviewed one of his supporters - they said something to the effect of "when I heard the news, my whole nervous system just suddenly relaxed." I wouldn't be surprised if forms of communication get more and more optimized, whether via RLHF or recsys likes or some other means, for their ability to target, for lack of a better term, "embodiment goals"

This is also known as "The Fundamental Theorem of Stream Calculus" in stream calculus. Using coinduction for an (infinite) stream sigma, eg

    sigma(0)      = head(sigma)
    sigma'        = tail(sigma)
    (a ++ sigma)' = sigma