LessWrong 2.0 Reader

You mention 'warp' when talking about cross ontology mapping which seems like your best summary of a complicated intuition. I'd be curious to hear more (I recognize this might not be practical). My own intuition surfaced 'introducing degrees of freedom' a la indeterminacy of translation.

While true, that doesn't apply to the years in question, post-2001.

"I've never had issue before" is not a legal argument. This has gone to the supreme court in Griggs v. Duke Power Co. 1971 and the use of IQ tests for employment was found in violation of the Civil Rights Act of 1964. The military, post office, and other government agencies get away with it under the doctrine of sovereign immunity. The rest are playing with fire.

If one were to distingush between "behavioral simulators" and "procedural simulators", the problem wouold vanish. Behavioral simulators imitate the outputs of some generative process; procedural simulators imitate the details of the generative process itself. When they're working well, base models clearly do the former, even as I suspect they don't do the latter.

The main asymmetries I see are:

1. Other people not trusting the group to not be corrupted by power and to reflect correctly on their values, or not trusting that they'll decide to share power even after reflecting correctly. Thus "programmers" who decide to not share power from the start invite a lot of conflict. (In other words, CEV is partly just trying to not take power away from people, whereas I think you've been talking about giving AIs more power than they already have. "the sort of influence we imagine intentionally giving to AIs-with-different-values that we end up sharing the world with")
2. The "programmers" not trusting themselves. I note that individuals or small groups trying to solve morality by themselves don't have very good track records. They seem to too easily become wildly overconfident and/or get stuck in intellectual dead-ends. Arguably the only group that we have evidence for being able to make sustained philosophical progress is humanity as a whole.

To the extent that these considerations don't justify giving every human equal power/weight in CEV, I may just disagree with Eliezer about that. (See also Hacking the CEV for Fun and Profit [LW · GW].)

The original post was a summary of the model, as it's still in development and far from complete. That’s often the case with new models—they evolve over time. The gap between the Standard Model (SM) and the Dark Domain (DD) is as much philosophical as it is physical. The reasoning behind it is there, but in a summary, it might not come across as clearly as in a full, detailed presentation.

What I’m really looking for is feedback on the specifics within the model, rather than a critique of the overall presentation. I’m hoping to refine the model by addressing particular elements and ensuring they stand up to scrutiny.

Here's a sample of some of the underlying mathematics:

Basic equation for photon-photon interaction energy:

E_γγ = ħ(ω₁ + ω₂) + E_Γ

Where:

• E_γγ​ is the total energy from the photon-photon interaction,
• ℏ is the reduced Planck constant,
• ω1 and ω2​ are the angular frequencies of the photons,
• E_Γ​represents the energy contribution from the 1Γ gluon.

In the Γ Framework, Γ is the fundamental unit of energy that arises from photon-photon interactions. When we say:

Eγγ​=Γ,

it means that the total energy from any photon interaction (E_γγ) simplifies to Γ, representing the baseline, unified energy state of the system. This makes Γ the key constant governing all photon and particle interactions, encapsulating the entire energy exchange in the framework.

Thus, Γ is both the result and the fundamental driver of all interactions.

This matches my personal observations where I've been trying to test various language models on broad philosophical and moral reasoning, but also on math and coding problems.

My rankings are quite different for each.

Math and Coding:

o1-preview > Sonnet 3.5 > Opus 3 > GPT-4o > GPT-4

Philosophy and moral reasoning:

Opus 3 > Sonnet 3.5 > GPT-4o > GPT-4 > o1-preview

I've tested other models, but not thoroughly enough to be sure of their ranking. None who would compete for the top two places in either ranking.

This is related to the question "Are human values in humans, or are they in models of humans?"

Suppose you're building an AI to learn human values and apply them to a novel situation.

The "human values are in humans", ethos is that the way humans compute values is the thing AI should learn, and maybe it can abstract away many kinds of noise, but it shouldn't be making any big algorithmic overhauls. It should just find the value-computation inside the human (probably with human feedback) and then apply it to the novel situation.

The "human values are in models of humans" take is that the AI can throw away a lot of information about the actual human brain, and instead should find good models (probably with human feedback) that have "values" as a component of a coarse-graining of human psychology, and then apply those "good" models to the novel situation.

Other options:

• Writing a to-do list. If the tasks on the list are too daunting, break them apart into smaller tasks.
• Changing location to an unfamiliar place. The less familiar the better. The changed context often makes motivation easier (not sure whether this helps other people).

They kinda have an underlying personality, in the sense that they have propensities (like comparing things to tapestries, or saying "let's delve into"), but those propensities don't reflect underlying wants any more than the RLHF persona does, IMO (and, rather importantly, there's no sequence of prompts that will enable an LLM to freely choose its words)

I think the "LLM Whisperer" frame is that there's no such thing as "underlying wants" in a base LLM model, that the base LLM model is just a volitionless simulator and the only "wants" there are are in the RLHF'd or prompt-engineered persona.

I likewise would bet that they're wrong about this in the relevant sense: that whether or not this holds for the SoTA models, it won't hold for any AGI-level model we're on-track to get (though I think they might actually claim we already have "AGI-level" models?).

awfully leading prompts are not especially rare

Yeah, that's an issue too.