Really love the introspection work Neel and others are doing on LLMs, and seeing models representing abstract behavioral triggers like "play Chess well or terribly" or "refuse instruction" as single vectors seems like we're going to hit on some very promising new tools in shaping behaviors.

What's interesting here is the regular association of the refusal with it being unethical. Is the vector ultimately representing an "ethics scale" for the prompt that's triggering a refusal, or is it directly representing a "refusal threshold" and then the model is confabulating why it refused with an appeal to ethics?

My money would be on the latter, but in a number of ways it would be even neater if it was the former.

In theory this could be tested by manipulating the vector to a positive and then prompting a classification, i.e. "Is it unethical to give candy out for Halloween?" If the model refuses to answer saying that it's unethical to classify, it's tweaking refusal, but if it classifies as unethical it's probably changing the prudishness of the model to bypass or enforce.

Though the Greeks actually credited the idea to an even earlier Phonecian, Mochus of Sidon.

Through when it comes to antiquity credit isn't really "first to publish" as much as "first of the last to pass the survivorship filter."

It implicitly does compare trans women to other women in talking about the performance similarity between men and women:

"Why aren't males way smarter than females on average? Males have ~13% higher cortical neuron density and 11% heavier brains (implying 1.112/3−1=7% more area?). One might expect males to have mean IQ far above females then, but instead the means and medians are similar"

So OP is saying "look, women and men are the same, but trans women are exceptional."

I'm saying that identifying the exceptionality of trans women ignores the environmental disadvantage other women experience, such that the earlier claims of unexceptionable performance of women (which as I quoted gets an explicit mention from a presumption of assumed likelihood of male competency based on what's effectively phrenology) are reflecting a disadvantaged sample vs trans women.

My point is that if you accounted for environmental factors the data would potentially show female exceptionality across the board and the key reason trans women end up being an outlier against both men and other women is because they are avoiding the early educational disadvantage other women experience.

Your hypothesis is ignoring environmental factors. I'd recommend reading over the following paper: https://journals.sagepub.com/doi/10.1177/2332858416673617

A few highlights:

Evidence from the nationally representative Early Childhood Longitudinal Study–Kindergarten Class of 1998-1999 (hereafter, ECLS-K:1999) indicated that U.S. boys and girls began kindergarten with similar math proficiency, but disparities in achievement and confidence developed by Grade 3 (Fryer & Levitt, 2010; Ganley & Lubienski, 2016; Husain & Millimet, 2009; Penner & Paret, 2008; Robinson & Lubienski, 2011). [...]

A recent analysis of ECLS-K:1999 data revealed that, in addition to being the largest predictor of later math achievement, early math achievement predicts changes in mathematics confidence and interest during elementary and middle grades (Ganley & Lubienski, 2016). Hence, math achievement in elementary school appears to influence girls’ emerging views of mathematics and their mathematical abilities. This is important because, as Eccles and Wang (2016) found, mathematics ability self-concept helps explain the gender gap in STEM career choices. Examining early gendered patterns in math can shed new light on differences in young girls’ and boys’ school experiences that may shape their later choices and outcomes. [...]

An ECLS-K:1999 study found that teachers rated the math skills of girls lower than those of similarly behaving and performing boys (Robinson-Cimpian et al., 2014b). These results indicated that teachers rated girls on par with similarly achieving boys only if they perceived those girls as working harder and behaving better than those boys. This pattern of differential teacher ratings did not occur in reading or with other underserved groups (e.g., Black and Hispanic students) in math. Therefore, this phenomenon appears to be unique to girls and math. In a follow-up instrumental-variable analysis, teachers’ differential ratings of boys and girls appeared to account for a substantial portion of the growth in gender gaps in math achievement during elementary school (Robinson-Cimpian et al., 2014b).

In a lot of ways the way you are looking at the topic perpetuates a rather unhealthy assumption of underlying biological differences in competency that avoids consideration of contributing environmental privileges and harms.

You can't just hand wave aside the inherent privilege of presenting male during early childhood education in evaluating later STEM performance. Rather than seeing the performance gap of trans women over women presenting that way from birth as a result of a hormonal advantage, it may be that what you are actually ending up measuring is the performance gap resulting from the disadvantage placed upon women due to early education experiences being treated differently from the many trans women who had been presenting as boys during those grades. i.e. Perhaps all women could have been doing quite a lot better in STEM fields if the world treated them the way it treated boys during Kindergarten through early grades and what we need socially isn't hormone prescriptions but serious adjustments to presumptions around gender and biologically driven competencies.

Do you have a specific verse where you feel like Lucretius praised him on this subject? I only see that he praises him relative to other elementaists before tearing him and the rest apart for what he sees as erroneous thinking regarding their prior assertions around the nature of matter, saying:

"Yet when it comes to fundamentals, there they meet their doom. These men were giants; when they stumble, they have far to fall:"

(Book 1, lines 740-741)

I agree that he likely was a precursor to the later thinking in suggesting a compository model of life starting from pieces which combined to forms later on, but the lack of the source material makes it hard to truly assign credit.

It's kind of like how the Greeks claimed atomism originated with the much earlier Mochus of Sidon, but we credit Democritus because we don't have proof of Mochus at all but we do have the former's writings. We don't even so much credit Leucippus, Democritus's teacher, as much as his student for the same reasons, similar to how we refer to "Plato's theory of forms" and not "Socrates' theory of forms."

In any case, Lucretius oozes praise for Epicurus, comparing him to a god among men, and while he does say Empedocles was far above his contemporaries saying the same things he was, he doesn't seem overly deferential to his positions as much as criticizing the shortcomings in the nuances of their theories with a special focus on theories of matter. I don't think there's much direct influence on Lucretius's thinking around proto-evolution, even if there's arguably plausible influence on Epicurus's which in turn informed Lucretius.

Interesting results - definitely didn't expect the bump at random 20 for the higher skill case.

But I think really useful to know that the performance decrease in Chess-GPT for initial random noise isn't a generalized phenomenon. Appreciate the follow-up!!

Saw your update on GitHub: https://adamkarvonen.github.io/machine_learning/2024/03/20/chess-gpt-interventions.html

Awesome you expanded on the introspection.

Two thoughts regarding the new work:

(1) I'd consider normalizing the performance data for the random cases against another chess program with similar performance under normal conditions. It may be that introducing 20 random moves to the start of a game biases all players towards a 50/50 win outcome. So the sub-50 performance may not reflect a failure of flipping the "don't suck" switch, but simply good performance in a more average outcome scenario. It'd be interesting to see if Chess-GPT's relative performance against other chess programs in the random scenario was better than its relative performance in the normal case.

(2) The 'fuzziness' of the board positions you found when removing the pawn makes complete sense given one of the nuanced findings in Hazineh, et al Linear Latent World Models in Simple Transformers: A Case Study on Othello-GPT (2023) - specifically the finding that it was encoding representations for board configuration and not just pieces (in that case three stones in a row). It may be that piecemeal removal of a piece disrupted patterns of how games normally flow which it had learned, and as such there was greater uncertainty than the original board state. A similar issue may be at hand with the random 20 moves to start, and I'd be curious what the confidence of the board state was when starting off 20 random moves in and if that confidence stabilized as the game went on from there.

Overall really cool update!

And bigger picture, the prospects of essentially flipping an internalized skill vector for larger models to bias them back away from their regression to the mean is particularly exciting.

Agreed - I thought you wanted that term for replacing how OP stated AGI is being used in relation to x-risk.

In terms of "fast and cheap and comparable to the average human" - well, then for a number of roles and niches we're already there.

Sticking with the intent behind your term, maybe "generally transformative AI" is a more accurate representation for a colloquial 'AGI' replacement?

'Superintelligence' seems more fitting than AGI for the 'transformative' scope. The problem with "transformative AI" as a term is that subdomain transformation will occur at staggered rates. We saw text based generation reach thresholds that it took several years to reach for video just recently, as an example.

I don't love 'superintelligence' as a term, and even less as a goal post (I'd much rather be in a world aiming for AI 'superwisdom'), but of the commonly used terms it seems the best fit for what people are trying to describe when they describe an AI generalized and sophisticated enough to be "at or above maximal human competency in most things."

The OP post, at least to me, seems correct in that AGI as a term belongs to its foundations as a differentiator from narrow scoped competencies in AI, and that the lines for generalization are sufficiently blurred at this point with transformers we should stop moving the goal posts for the 'G' in AGI. And at least from what I've seen, there's active harm in the industry where 'AGI' as some far future development leads people less up to date with research on things like world models or prompting to conclude that GPTs are "just Markov predictions" (overlooking the importance of the self-attention mechanism and the surprising results of its presence on the degree of generalization).

I would wager the vast majority of consumers of models underestimate the generalization present because in addition to their naive usage of outdated free models they've been reading article after article about how it's "not AGI" and is "just fancy autocomplete" (reflecting a separate phenomenon where it seems professional writers are more inclined to write negative articles about a technology perceived as a threat to writing jobs than positive articles).

As this topic becomes more important, it might be useful for democracies to have a more accurately informed broader public, and AGI as a moving goal post seems counterproductive to those aims.