.

[. @notselfmodel]

@nickcammarata also the uncanny glitch in chatgpt's face/jaw as he says "and they're scared of me." is (accidentally?) narriatively genius

[. @notselfmodel]

@nickcammarata my opsec is worse than I thought

[. @notselfmodel]

@shinboson you might be a superpersuader I can feel e/acc being upweighted in my mind in real time

[. @notselfmodel]

@Michael_Druggan for a people who pride themselves on being able to override their basic human motivations in service of rationally calculated strategies they sure are bad at it

[. @notselfmodel]

@paradite_ could never be me, maybe you're just a low-entropy person

[. @notselfmodel]

@Qivshi1 The structure of an LLM can encode almost any mechanism. This is like saying "we know how a transistor works so therefore we understand how Linux works" - no, those are two completely separate things to understand and the fact that one runs on the other doesn't change that.

[. @notselfmodel]

@Qivshi1 @abachurin You do realize that the entire computational structure of an LLM is defined by the parameters of matmuls?

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan An LLM in training can mispredict the next token by having an incorrect model of the person speaking - that is being stupid, it's using a model and that model being wrong

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan How does that prevent an LLM base model from functionally holding beliefs about its simulated human minds?

[. @notselfmodel]

@itsrealAJB @tinrab @theo It was always doomed; as the models continue to get better people will prefer to interact with them as "coworkers" rather than "tools" and features like plan mode and goal will necessarily disappear as they are no longer needed. (This is already happening.)

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan LLMs due to being arbitrary-person-simulators at their core can't very meaningfully hold beliefs about abstract facts, but the internal model definitely *can* hold beliefs about implicit things like "what sort of person am I modeling" and "what would this person believe?"

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan What do you posit creates interference patterns when we send one photon at a time?

[. @notselfmodel]

@shinboson you might be a superpersuader I can feel e/acc being upweighted in my mind in real time

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan Because here you (unless I misread you) implied your model to be simpler than quantum mechanics and it isn't clear to me that it is, if you build it out enough to actually explain every effect (since you currently have only a partial solution) - did I misinterpret?

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan My point is that you can't use "it has irreducible information" to write off a different model in support of yours unless it's clear that yours has less, and in this case unless I'm missing something it isn't clear that it does.

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan Is the code for this public?

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan The whole point of the double slit experiment is that interactions at the slits remove the effect of interference. Can you replicate that?

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan Every rule you have to specify is complexity. Why your particular system of particles and rules and not, say, the minimum turing machine that generates and runs (interleaved) all finitely specifiable programs? Where does the information come from?

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan Sure, but reality closely agrees with quantum models meaning your model must also agree closely with quantum models which I'm skeptical of - do you have results showing it does?

Also, wrt your second statement, why is the remaining complexity in *your* system justified?

[. @notselfmodel]

@matterasmachine @itchy_est @Michael_Druggan at a first glance, interesting, will look more closely later, but I think your mapping from particles to independent machines in 3d space is too direct? a more abstract machine could make sense but literally voxelizing space makes it pretty hard to model quantum effects, no?