If the RL is brief and limited to a small subset of parameters, the AV will produce reasonable language since it inherits that from the base LLM, and it will produce descriptions aligned with the input to the base LLM that produced the autoencoded activations, since the AR is still close to the base LLM (and could reconstruct the activations perfectly if fed the full context which produced them).

Some of the attention-like ops proposed in this new work are most simply described as implementing the associative memory with a hypernetwork that maps keys to values with weights that are optimized at test time to minimize value retrieval error. Like you suggest, designing these hypernetworks to permit efficient implementations is tricky.

It's a more constrained interpretation of attention than you're advocating for, since it follows the "attention as associative memory" perspective, but the general idea of test-time optimization could be applied to other mechanisms for letting information interact non-linearly across arbitrary nodes in the compute graph.

[1] https://arxiv.org/abs/2501.00663

[2] https://arxiv.org/abs/2504.13173

[3] https://arxiv.org/abs/2505.23735

This sort of "dynamic chunking" of low-level information, perhaps down to the level of raw bytes, into shorter sequences of meta tokens for input to some big sequence processing model is an active area of research. Eg, one neat paper exploring this direction is: "Dynamic Chunking for End-to-End Hierarchical Sequence Modeling" [1], from one of the main guys behind Mamba and other major advances in state-space models.

[1] - https://arxiv.org/abs/2507.07955

Ie, the skills aren't particularly complicated in principle, but the conditions needed to acquire them aren't widely available, so the pool of people with the skills is limited.

The sorts of useful analogies I was mostly talking about are those that appear in scientific research involving actionable technical details. Eg, diffusion models came about when folks with a background in statistical physics saw some connections between the math for variational autoencoders and the math for non-equilibrium thermodynamics. Guided by this connection, they decided to train models to generate data by learning to invert a diffusion process that gradually transforms complexly structured data into a much simpler distribution -- in this case, a basic multidimensional Gaussian.

I feel like these sorts of technical analogies are harder to stumble on than more common "linguistic" analogies. The latter can be useful tools for thinking, but tend to require some post-hoc interpretation and hand waving before they produce any actionable insight. The former are more direct bridges between domains that allow direct transfer of knowledge about one class of problems to another.

I think you're basically agreeing with me. Ie, current models are not superintelligent. Even though they can "think" super fast, they don't pass a minimum bar of producing novel and useful connections between domains without significant human intervention. And, our evaluation of their abilities is clouded by the way in which their intelligence differs from our own.

In general, I agree that these models are in some sense extremely knowledgeable, which suggests they are ripe for producing productive analogies if only we can figure out what they're missing compared to human-style thinking. Part of what makes it difficult to evaluate the abilities of these models is that they are wildly superhuman in some ways and quite dumb in others.

The most successful folks tend to mix talent and hard work with a bit of luck in terms of early gold striking to gain a quick boost of credibility that helps them draw other people into their fold (eg, grad students in a big lab) who can handle a lot of the metric maxxing to free up some (still not enough) time for more ambitious thinking.

I agree with your general point though. Ie, we need more thorough empirical investigation of how reasoning behavior evolves during RL training starting from the base model. And, current RL training results seem more like "amplifying existing good behavior" than "inducing emergent good behavior".

The "trick" allows you to fit a linear function in that higher dimensional space without any potentially costly explicit computation in the higher dimensional space based on the observation that the optimal solution's parameters can be represented as a sum of the higher dimensional representations of points in the training set.

Some straightforward scenarios are in, eg, robotics where one can design sequences of increasingly difficult instances of a task like moving objects from one storage bin to another. The basic idea is that the agent would have no reward or learning signal if it jumped straight into the full version of the task, so you let it develop competence on simpler variants and gradually increase difficulty until the agent can get useful learning signal on the full task.

However, if you're training on many problems, it's possible in principle that if you have traction on _any_ of the problems, then the learning signal you get from success on those problems will have a positive effect on the model's behavior on other problems. Ie, the learning that you do on problems where the model is already producing positive reward behavior will nudge the model towards producing positive reward behavior on problems where it wasn't previously doing so.