Strong agree, but with a major caveat: First since we humans don't have a memory the size of the internet, new TO YOU is extremely easy to find. And second: a lot of discoveries can be made by trying to match two different domains and generate new insights at the seams.
A lot of math theorems are discovered that way by humans btw: find a difficult problem in a domain, map it to another domain, apply existing theorems there, go back to the original domain. The only difference is an LLM would have to search for an interesting problem, but that can be done by brute-forcing search at the seams and validating the problem does not previously exists. And there is no new TO YOU for the LLM: it can get help from a vast amount of helper domains.
IMHO there's a shot at having category theory really unifying a large chunk of maths in the next decade. It'll be a big ugly ball of Lean, but a queryable one at that.
Setting up a classical DLRM at work, show of hands: who has managed to bridge the gap to generative recommenders or LRMs?
Our survey identified quite a substantial list of challenges: product cardinality (Semantic IDs?), low performance vs discriminative recommenders (our user intents are heavily biases towards repetitive products), cold starts, entity features, missing explainability suite, training cost, prediction cost, etc.
Has anyone done it easily? Where is the tutorial for generative recommenders ?
GR expertise is essentially only in-house at major social media companies or people hired from these companies. Closest thing to a tutorial for LRM is the Kuaishou onerec papers.
A lot of the mech interp techniques are glorified notions of probing. If models were generally interpretable with respect to the things we cared about, we'd know know when they're hallucinating, for example.
Identifying that a common neuron or set of neurons fires for golden gate bridge can be done only because you already know that you're looking for golden gate bridge. If models had recovered some new truth (e.g., the silver door road), you wouldn't be able to elicit it without knowing to look for it a priori.
Typically the analysis works in the reverse direction: we have a neuron (or SAE feature, or whatever) and we look at examples of data that makes it activate, then inferring an interpretation. This is a lot harder than probing for a known concept, and one of the points of this article is that if you try harder to understand the activation pattern, you will very often find that there is something interpretable there, even if at first glance it looked random.
This was one of my intuitive concerns reading this post.
I've done work with trying to train explainable models for e.g. making diagnoses based on neuroimaging, and it is not clear to me that if you applied these techniques to e.g. vision models trained exclusively on such brain images, that these visual representations or conceptual clusters and relations would really help understanding at all.
The main issue would be that you'd have no clear way to verify any generated explanation for such training data. I.e. we can "verify" some of the explainability shown in the article because clear and recognizable image clusters and/or semantic groupings pop out, because we already understand the domains. But if this were fMRI or e.g. EEG signals, we really wouldn't have that, so the explanations would themselves just be more hypotheses.
Perhaps if we can build on interpretability of vision and language models in domains we understand, we can eventually generalize this to other domains, but I don't think that clearly follows from anything I am reading here.
To me their purpose only came clearly recently. They're just a way to rotate smoothly one dimension into another. And control the speed of the rotation.
x into y. Potential energy to kinetic energy. Charge into current.
Keeping this definition but going higher in the dimensions: x into y, y into z and z into x. Going meta, the x into y rotation defines a plane; And a plane going into another direction is a volume. (x into y) into z.
What made it click, what made it beautiful is Geometric Algebra:
> Java is a blue collar language. It’s not PhD thesis material but a language for a job. Java feels very familiar to many different programmers because we preferred tried-and-tested things.
Average user input quality is not optimal by definition. SO's upvote system and reputation system is what really refined quality. Much like Reddit's system is helping a lot for conversation. Much like academia's publish and citation system is an even better system.
I'm not saying LLM won't be able to train further, but tough and sound reviews are paramount to improving the corpus of quality content. And with the replacement of SO by LLMs, one can wonder if one quality system went down the drain. Anyway there's still code reviews on GitHub and in private repos, but we're still far from getting the human judgement out of the loop.
"Average user input quality is not optimal by definition."
Absolutely.
"SO's upvote system and reputation system is what really refined quality."
Absolutely not!
SO's problem was exactly that they gave average users moderating power. That lead to half of the questions being closed because they were over the head of the average user/moderator regardless of being interesting and useful questions or not.
I do expect there will be more closed knowledge forums. Unfortunately one person exposing it to LLM provider will be enough to break the whole idea. But I believe there will be many that will keep secrets for some in group.
Oil is sky high, a recession might be near. Also, optimizations (like the current KV one) can yield a drastic reduction in RAM needs. It's a risk.
If I was trying so stabilize (just stabilize) funding for humongously expensive fabs: I'd be giving markets reassurances that my business will continue to be high margins for the mid-future.
A lot of math theorems are discovered that way by humans btw: find a difficult problem in a domain, map it to another domain, apply existing theorems there, go back to the original domain. The only difference is an LLM would have to search for an interesting problem, but that can be done by brute-forcing search at the seams and validating the problem does not previously exists. And there is no new TO YOU for the LLM: it can get help from a vast amount of helper domains.
IMHO there's a shot at having category theory really unifying a large chunk of maths in the next decade. It'll be a big ugly ball of Lean, but a queryable one at that.
reply