I sampled hundreds of short context snippets from openwebtext, and measured ablation effects averaged over those sampled forward-passes. Averaged over those hundreds of passes, I didn't see any real signal in the logit effects, just a layer of noise due to the ablations.

More could definitely be done on this front. I just tried something relatively quickly that fit inside of GPU memory and wanted to report it here.

Could you hotlink the boxes on the diagrams to that, or add the resulting content as a hover text to areas, in them or something? This might be hard to do on LW: I suspect some Javascript code might be required to do this sort of thing, but perhaps a library exists for this?

My workaround was to have the dimension links laid out below each figure.

3: [14555]

4: [5030]

5: [10603]

6: [3290]

7: [4330]

My current "print to flat .png" approach wouldn't support hyperlinks, and I don't think LW supports .svg images.

That line was indeed quite poorly phrased. It now reads:

At the bottom of the box, blue or red token boxes show the tokens most promoted (blue) and most suppressed (red) by that dimension.

That is, you're right. Interpretability data on an autoencoder dimension comes from seeing which token probabilities are most promoted and suppressed when that dimension is ablated, relative to leaving its activation value alone. That's an ablation effect sign, so the implied, plotted promotion effect signs are flipped.

The main thing I got out of reading Bostrom's Deep Utopia is a better appreciation of this "meaning of life" thing. I had never really understood what people meant by this, and always just rounded it off to people using lofty words for their given projects in life.

The book's premise is that, after the aligned singularity, the robots will not just be better at doing all your work but also be better at doing all your leisure for you. E.g., you'd never study for fun in posthuman utopia, because you could instead just ask the local benevolent god to painlessly... (read more)

I believe I and others here probably have a lot to learn from Chris, and arguments of the form "Chris confidently believes false thing X," are not really a crux for me about this.

Would you kindly explain this? Because you think some of his world-models independently throw out great predictions, even if other models of his are dead wrong?

Use your actual morals, not your model of your morals.

I agree that stronger, more nuanced interpretability techniques should tell you more. But, when you see something like, e.g.,

25132 ▁vs, ▁differently, ▁compared, ▁greater, all, ▁per
25134 ▁I, ▁My, I, ▁personally

isn't it pretty obvious what those two autoencoder neurons were each doing?

No, towards an $L^0$ value. $L^1$ is the training proxy for that, though.

Epistemic status: Half-baked thought.

Say you wanted to formalize the concepts of "inside and outside views" to some degree. You might say that your inside view is a Bayes net or joint conditional probability distribution—this mathematical object formalizes your prior.

Unlike your inside view, your outside view consists of forms of deferring to outside experts. The Bayes nets that inform their thinking are sealed away, and you can't inspect these. You can ask outside experts to explain their arguments, but there's an interaction cost associated with inspecti... (read more)

(Great project!) I strongly second the RSS feed idea, if that'd be possible.

I think that many (not all) of your above examples boil down to optimizing for legibility rather than optimizing for goodness. People who hobnob instead of working quietly will get along with their bosses better than their quieter counterparts, yes. But a company of brown nosers will be less productive than a competitor company of quiet hardworking employees! So there's a cooperate/defect-dilemma here.

What that suggests, I think, is that you generally shouldn't immediately defect as hard as possible, with regard to optimizing for appearances. Play the prev... (read more)

The ML models that now speak English, and are rapidly growing in world-transformative capability, happen to be called transformers.

This is not a coincidence because nothing is a coincidence.

Two moments of growing in mathematical maturity I remember vividly:

1. Realizing that equations are claims that are therefore either true or false. Everything asserted with symbols... could just as well be asserted in English. I could start chunking up arbitrarily long and complicated equations between the equals signs, because those equals signs were just the English word "is"!
2. Learning about the objects that mathematical claims are about. Going from having to look up "Wait, what's a real number again?" to knowing how $Z$, $Q$, and $R$ interrelat

I found it distracting that all your examples were topical, anti-red-tribe coded events. That reminded me of

In Artificial Intelligence, and particularly in the domain of nonmonotonic reasoning, there’s a standard problem: “All Quakers are pacifists. All Republicans are not pacifists. Nixon is a Quaker and a Republican. Is Nixon a pacifist?”

What on Earth was the point of choosing this as an example? To rouse the political emotions of the readers and distract them from the main question? To make Republicans feel unwelcome in courses on Artificial Intelligenc

2. The anchor of a major news network donates lots of money to organizations fighting against gay marriage, and in his spare time he writes editorials arguing that homosexuals are weakening the moral fabric of the country. The news network decides they disagree with this kind of behavior and fire the anchor.

a) This is acceptable; the news network is acting within their rights and according to their principles
b) This is outrageous; people should be judged on the quality of their work and not their political beliefs

…

12. The principal of a private school is a

Academic philosophers are better than average at evaluating object-level arguments for some claim. They don't seem to be very good at thinking about what rationalization in search implies about the arguments that come up. Compared to academic philosophers, rationalists strike me as especially appreciating filtered evidence and its significance to your world model.

If you find an argument for a claim easily, then even if that argument is strong, this (depending on some other things) implies that similarly strong arguments on the other side may turn up with n... (read more)

Modest spoilers for planecrash (Book 9 -- null action act II).

Nex and Geb had each INT 30 by the end of their mutual war.  They didn't solve the puzzle of Azlant's IOUN stones... partially because they did not find and prioritize enough diamonds to also gain Wisdom 27.  And partially because there is more to thinkoomph than Intelligence and Wisdom and Splendour, such as Golarion's spells readily do enhance; there is a spark to inventing notions like probability theory or computation or logical decision theory from scratch, that is not directly me

What sequence of characters could I possibly, actually type out into a computer that would appreciably reduce the probability that everything dies?

Framed like this, writing to save the world sounds impossibly hard! Almost everything written has no appreciable effect on our world's AI trajectory. I'm sure the "savior sequence" exists mathematically, but finding it is a whole different ballgame.

In the beginning God created four dimensions. They were all alike and indistinguishable from one another. And then God embedded atoms of energy (photons, leptons, etc.) in the four dimensions. By virtue of their energy, these atoms moved through the four dimensions at the speed of light, the only spacetime speed. Thus, as perceived by any one of these atoms, space contracted in, and only in, the direction of that particular atom's motion. As the atoms moved at the speed of light, space contracted so much in the direction of the atom's motion that the dimen

Past historical experience and brainstorming about human social orders probably barely scratches the possibility space. If the CEV were to weigh in on possible posthuman social orders,^[1] optimizing in part for how cool that social order is, I'd bet what it describes blows what we've seen out of the water in terms of cool factor.

1. ^{^}

   (Presumably posthumans will end up reflectively endorsing interactions with one another of some description.)

Don't translate your values into just a loss function. Rather, translate them into a loss function and all the rest of a training story. Use all the tools at your disposal in your impossible task; don't tie one hand behind your back by assuming the loss function is your only lever over the AGI's learned values.

This post crystallized some thoughts that have been floating in my head, inchoate, since I read Zvi's stuff on slack and Valentine's "Here's the Exit."

Part of the reason that it's so hard to update on these 'creative slack' ideas is that we make deals among our momentary mindsets to work hard when it's work-time. (And when it's literally the end of the world at stake, it's always work-time.) "Being lazy" is our label for someone who hasn't established that internal deal between their varying mindsets, and so is flighty and hasn't precommitted to getting st... (read more)

A model I picked up from Eric Schwitzgebel.

The humanities used to be highest-status in the intellectual world!

But then, scientists quite visibly exploded fission weapons and put someone on the moon. It's easy to coordinate to ignore some unwelcome evidence, but not evidence that blatant. So, begrudgingly, science has been steadily accorded more and more status, from the postwar period on.

"Calling babble and prune the True Name of text generation is like calling bogosort the True Name of search."

In the 1920s when $\lambda$ and CL began, logicians did not automatically think of functions as sets of ordered pairs, with domain and range given, as mathematicians are trained to do today. Throughout mathematical history, right through to computer science, there has run another concept of function, less precise at first but strongly influential always; that of a function as an operation-process (in some sense) which may be applied to certain objects to produce other objects. Such a process can be defined by giving a set of rules describing how it acts

Given a transformer model, it's probably possible to find a reasonably concise energy function (probably of a similar OOM of complexity as the model weights themselves) whose minimization corresponds to executing forwards passes of the transformer. However, this [highly compressive] energy function wouldn't tell you much about what the personas simulated by the model "want" or how agentic they were, since the energy function is expressed in the ontology of model weights and activations, not an agent's beliefs / goals. [This has] the type signature of a uti

This is a great theorem that's stuck around in my head this last year! It's presented clearly and engagingly, but more importantly, the ideas in this piece are suggestive of a broader agent foundations research direction. If you wanted to intimate that research direction with a single short post that additionally demonstrates something theoretically interesting in its own right, this might be the post you'd share.

This post has successfully stuck around in my mind for two years now! In particular, it's made me explicitly aware of the possibility of flinching away from observations because they're normie-tribe-coded.

I think I deny the evidence on most of the cases of dogs generating complex English claims. But it was epistemically healthy for that model anomaly to be rubbed in my face, rather than filter-bubbled away plus flinched away from and ignored.

This is a fantastic piece of economic reasoning applied to a not-flagged-as-economics puzzle! As the post says, a lot of its content is floating out there on the internet somewhere: the draw here is putting all those scattered insights together under their common theory of the firm and transaction costs framework. In doing so, it explicitly hooked up two parts of my world model that had previously remained separate, because they weren't obviously connected.

Complex analysis is the study of functions of a complex variable, i.e., functions $f\left(z\right)$ where $z$ and $f\left(z\right)$ lie in $C$. Complex analysis is the good twin and real analysis the evil one: beautiful formulas and elegant theorems seem to blossom spontaneously in the complex domain, while toil and pathology rule in the reals. Nevertheless, complex analysis relies more on real analysis than the other way around.

--Pugh, Real Mathematical Analysis (p. 28)

One important idea I've picked up from reading Zvi is that, in communication, it's important to buy out the status cost imposed by your claims.

If you're fielding a theory of the world that, as a side effect, dunks on your interlocutor and diminishes their social status, you can work to get that person to think in terms of Bayesian epistemology and not decision theory if you make sure you aren't hurting their social image. You have to put in the unreasonable-feeling work of framing all your claims such that their social status is preserved or fairly increas... (read more)

Thanks -- right on both counts! Post amended.

An Inconsistent Simulated World

I regret to inform you, you are an em inside an inconsistent simulated world. By this, I mean: your world is a slapdash thing put together out of off-the-shelf assets in the near future (presumably right before a singularity eats that simulator Earth).

Your world doesn't bother emulating far-away events in great detail, and indeed, may be messing up even things you can closely observe. Your simulators are probably not tampering with your thoughts, though even that is something worth considering carefully.

What are the flaws you... (read more)

When another article of equal argumentative caliber could have just as easily been written for the negation of a claim, that writeup is no evidence for its claim.

Switching costs between different kinds of work can be significant. Give yourself permission to focus entirely on one kind of work per Schelling unit of time (per day), if that would help. Don't spend cognitive cycles feeling guilty about letting some projects sit on the backburner; the point is to get where you're going as quickly as possible, not to look like you're juggling a lot of projects at once.

This can be hard, because there's a conventional social expectation that you'll juggle a lot of projects simultaneously, maybe because that's more legible t... (read more)

A multiagent Extrapolated Volitionist institution is something that computes and optimizes for a Convergent Extrapolated Volition, if a CEV exists.

Really, though, the above Extrapolated Volitionist institutions do take other people into consideration. They either give everyone the Schelling weight of one vote in a moral parliament, or they take into consideration the epistemic credibility of other bettors as evinced by their staked wealth, or other things like that.

Sometimes the relevant interpersonal parameters can be varied, and the institutional designs... (read more)

Because your utility function is your utility function, the one true political ideology is clearly Extrapolated Volitionism.

Extrapolated Volitionist institutions are all characteristically "meta": they take as input what you currently want and then optimize for the outcomes a more epistemically idealized you would want, after more reflection and/or study.

Institutions that merely optimize for what you currently want the way you would with an idealized world-model are old hat by comparison!

Stress and time-to-burnout are resources to be juggled, like any other.

“What is the world trying to tell you?”

I've found that this prompt helps me think clearly about the evidence shed by the generator of my observations.

As Gauss stressed long ago, any kind of singular mathematics acquires a meaning only as a limiting form of some kind of well-behaved mathematics, and it is ambiguous until we specify exactly what limiting process we propose to use. In this sense, singular mathematics has necessarily a kind of anthropomorphic character; the question is not what is it, but rather how shall we define it so that it is in some way useful to us?

--E. T. Jaynes, Probability Theory (p. 108)

Bogus nondifferentiable functions

The case most often cited as an example of a nondifferentiable function is derived from a sequence $f_n\left(x\right)$, each of which is a string of isosceles right triangles whose hypotenuses lie on the real axis and have length $1/n$. As $n\to \infty$, the triangles shrink to zero size. For any finite $n$, the slope of $f_n\left(x\right)$ is $\pm 1$ almost everywhere. Then what happens as $n\to \infty$? The limit $f_{\infty }\left(x\right)$ is often cited carelessly as a nondifferentiable function. Now it is clear that the limit of the derivativ

Epistemic status: politics, known mindkiller; not very serious or considered.

People seem to have a God-shaped hole in their psyche: just as people banded around religious tribal affiliations, they now, in the contemporary West, band together around political tribal affiliations. Intertribal conflict can be, at its worst, violent, on top of mindkilling. Religious persecution in the UK was one of the instigating causes of British settlers migrating to the American colonies; religious conflict in Europe generally was severe.

In the US, the 1st Amendment legall... (read more)

The explicit definition of an ordered pair $\left(\right(a,b)=\{\left\{a\right\},\{a,b\}\left\}\right)$ is frequently relegated to pathological set theory...

It is easy to locate the source of the mistrust and suspicion that many mathematicians feel toward the explicit definition of ordered pair given above. The trouble is not that there is anything wrong or anything missing; the relevant properties of the concept we have defined are all correct (that is, in accord with the demands of intuition) and all the correct properties are present. The trouble is that the concept has some irreleva

Social niceties and professionalism act as a kind of 'communications handshake' in ordinary society -- maybe because they're still a credible correlate of having your act together enough to be worth considering your outputs in the first place?

Very cool! I have noticed that in arguments in ordinary academia people sometimes object that "that's so complicated" when I take a lot of deductive steps. I hadn't quite connected this with the idea that:

If you're confident in your assumptions ($ϵ$ is small), or if you're unconfident in your inferences ($\delta$ is big), then you should penalise slow theories moreso than long theories, i.e. you should be a T-type.

I.e., that holding a T-type prior is adaptive when even your deductive inferences are noisy.

Also, I take it that this row of your table:

FWIW, this post strikes me as a very characteristically 'Hansonian' insight.

Now, whatever $T$ may assert, the fact that $T$ can be deduced from the axioms cannot prove that there is no contradiction in them, since, if there were a contradiction, $T$ could certainly be deduced from them!

This is the essence of the Gödel theorem, as it pertains to our problems. As noted by Fisher (1956), it shows us the intuitive reason why Gödel’s result is true. We do not suppose that any logician would accept Fisher’s simple argument as a proof of the full Gödel theorem; yet for most of us it is more convincing than Gödel’s

I know that the humans forced to smile are not happy (and I know all the mistakes they've made while programming me, I know what they should've done instead), but I don't believe that they are not happy.

These are different senses of "happy." It should really read:

I know forcing humans to smile doesn't make them $happ{y}_{humane}$, and I know what they should've written instead to get me to optimize for $happ{y}_{humane}$ as they intended, but they are $happ{y}_{smiley}$.

They're different concepts, so there's no strangeness here. The AGI knows what you meant... (read more)

The human brain does not start out as an efficient reasoning machine, plausible or deductive. This is something which we require years to learn, and a person who is an expert in one field of knowledge may do only rather poor plausible reasoning in another. What is happening in the brain during this learning process?

Education could be defined as the process of becoming aware of more and more propositions, and of more and more logical relationships between them. Then it seems natural to conjecture that a small child reasons on a lattice of very open structur

Yeah, fair -- I dunno. I do know that an incremental improvement on simulating a bunch of people in an environment philosophizing is doing that but running an algorithm that prevents coercion, e.g.

I imagine that the complete theory of these incremental improvements (for example, also not running a bunch of moral patients for many subjective years while computing the CEV), is the final theory we're after, but I don't have it.