I often see a lot of people with outdated understanding of modern LLMs.

This is probably the best interpretability research to date, by the leading interpretability research team.

It's worth a read if you want a peek behind the curtain on modern models.

I've been saying this for about a year since seeing the Othello GPT research, but it's nice to see more minds changing as the research builds up.

Edit: Because people aren't actually reading and just commenting based on the headline, a relevant part of the article:

New research may have intimations of an answer. A theory developed by Sanjeev Arora of Princeton University and Anirudh Goyal, a research scientist at Google DeepMind, suggests that the largest of today’s LLMs are not stochastic parrots. The authors argue that as these models get bigger and are trained on more data, they improve on individual language-related abilities and also develop new ones by combining skills in a manner that hints at understanding — combinations that were unlikely to exist in the training data.

This theoretical approach, which provides a mathematically provable argument for how and why an LLM can develop so many abilities, has convinced experts like Hinton, and others. And when Arora and his team tested some of its predictions, they found that these models behaved almost exactly as expected. From all accounts, they’ve made a strong case that the largest LLMs are not just parroting what they’ve seen before.

“[They] cannot be just mimicking what has been seen in the training data,” said Sébastien Bubeck, a mathematician and computer scientist at Microsoft Research who was not part of the work. “That’s the basic insight.”

I've been saying this for about a year, since seeing the Othello GPT research, but it's great to see more minds changing on the subject.

It's worth pointing out that we're increasingly seeing video games rendering with continuous seed functions that convert to discrete units to track state changes from free agents, like the seed generation in Minecraft or No Man's Sky converting mountains into voxel building blocks that can be modified and tracked.

In theory, a world populated by NPCs with decision making powered by separate generative AI would need to do the same as the NPC behavior couldn't be tracked inherent to procedural world generation.

Which is a good context within which to remember that our own universe at the lowest level is made up of parts that behave as if determined by a continuous function until we interact with them at which point they convert to behaving like discrete units.

And even weirder is that we know it isn't a side effect from the interaction itself as if we erase the persistent information about interactions with yet another reversing interaction, the behavior switches back from discrete to continuous (like we might expect if there was a memory optimization at work).

I've been a big fan of Turok's theory since his first paper on a CPT symmetric universe. The fact he's since had this slight change to the standard model explain a number of the big problems in cosmology with such an elegant and straightforward solution (with testable predictions) is really neat. I even suspect if he's around long enough there will end up being a Nobel in his future for the effort.

The reason it's being posted here is that the model also happens to call to mind the topic of this community, particularly when thinking about the combination of quantum mechanical interpretations with this cosmological picture.

There's only one mirror universe on a cosmological scale in Turok's theory.

But in a number of QM interpretations, such as Everett's many worlds, transactional interpretation, and two state vector formalism, there may be more than one parallel "branch" of a quantized, formal reality in the fine details.

This kind of fits with what we might expect to see if the 'mirror' universe in Turok's model is in fact an original universe being backpropagated into multiple alternative and parallel copies of the original.

Each copy universe would only have one mirror (the original), but would have multiple parallel versions, varying based on fundamental probabilistic outcomes (resolving the wave function to multiple discrete results).

The original would instead have a massive number of approximate copies mirroring it, similar to the very large number of iterations of machine learning to predict an existing data series.

We might also expect that if this is the case that the math will eventually work out better if our 'mirror' in Turok's model is either not quantized at all or is quantized at a higher fidelity (i.e. we're the blockier Minecraft world as compared to it). Parts of the quantum picture are one of the holdout aspects of Turok's model, so I'll personally be watching it carefully for any addition of something akin to throwing out quantization for the mirror.

In any case, even simulation implications aside, it should be an interesting read for anyone curious about cosmology.

I'd been predicting this would happen a few months ago with friends and old colleagues (you can have a smart AI or a conservative AI but not both), but it's so much funnier than I thought it would be when it finally arrived.

While I'm doubtful that the testable prediction will be validated, it's promising that physicists are looking at spacetime and gravity as separated from quantum mechanics.

Hopefully at some point they'll entertain the idea that much like how we are currently converting continuous geometry into quantized units in order to track interactions with free agents in virtual worlds, that perhaps the quantum effects we measure in our own world are secondary side effects of emulating continuous spacetime and matter and not inherent properties to that foundation.

I'm not a big fan of Vopson or the whole "let's reinvent laws of physics" approach, but his current approach to his work is certainly on point for this sub.

At a certain point, we're really going to have to take a serious look at the direction things are evolving year by year, and reevaluate the nature of our own existence...