All of Razied's Comments + Replies

More insightful than what is conserved under the scaling symmetry of ReLU networks is what is not conserved: the gradient. Scaling $w_1$ by $\alpha$ scales $\partial E/\partial w_1$ by $1/\alpha$ and $\partial E/\partial w_2$ by $\alpha$, which means that we can obtain arbitrarily large gradient norms by simply choosing small enough $\alpha$. And in general bad initializations can induce large imbalances in how quickly the parameters on either side of the neuron learn. 

Some time ago I tried training some networks while setting these symmetries to the values th... (read more)

I suspect the expert judges would need to resort to known jailbreaking techniques to distinguish LLMs. A fair interesting test might be against expert-but-not-in-ML judges.

Sorry to be blunt, but any distraction filter that can be disabled through the chrome extension menu is essentially worthless. Speaking from experience, for most people this will work for exactly 3 days until they find a website they really want to visit and just "temporarily" disable the extension in order to see it.

For #5, I think the answer would be to make the AI produce the AI safety ideas which not only solve alignment, but also yield some aspect of capabilities growth along an axis that the big players care about, and in a way where the capabilities are not easily separable from the alignment. I can imagine this being the case if the AI safety idea somehow makes the AI much better at instruction-following using the spirit of the instruction (which is after all what we care about). The big players do care about having instruction-following AIs, and if the way to do that is to use the AI safety book, they will use it. 

Do you expect Lecun to have been assuming that the entire field of RL stops existing in order to focus on his specific vision?

Very many things wrong with all of that:

1.  RL algorithms don't minimize costs, but maximize expected reward, which can well be unbounded, so it's wrong to say that the ML field only minimizes cost. 
2. LLMs minimize expected log probability of correct token, which is indeed bounded at zero from below, but achieving zero in that case means perfectly predicting every single token on the internet. 
3. The boundedness of the thing you're minimizing is totally irrelevant, since maximizing $f\left(x\right)$ is exactly the same as maximizing $g\left(f\right(x\left)\right)$ w

The word "privilege" has been so tainted by its association with guilt that it's almost an infohazard to think you've got privilege at this point, it makes you lower your head in shame at having more than others, and brings about a self-flagellation sort of attitude. It elicits an instinct to lower yourself rather than bring others up. The proper reactions to all these things you've listed is gratitude to your circumstances and compassion towards those who don't have them. And certainly everyone should be very careful towards any instinct they have at publ... (read more)

Weird side effect to beware for retinoids: they make dry eyes worse, and in my experience this can significantly decrease your quality of life, especially if it prevents you from sleeping well.

Basically, this shows that every term in a standard Bayesian inference, including the prior ratio, can be re-cast as a likelihood term in a setting where you start off unsure about what words mean, and have a flat prior over which set of words is true.

If the possible meanings of your words are a continuous one-dimensional variable x, a flat prior over x will not be a flat prior if you change variables to y = f(y) for an arbitrary bijection f, and the construction would be sneaking in a specific choice of function f.

Say the words are utterances about the probability of a coin falling heads, why should the flat prior be over the probability p, instead of over the log-odds log(p/(1-p)) ?

Most of the weird stuff involving priors comes into being when you want posteriors over a continuous hypothesis space, where you get in trouble because reparametrizing your space changes the form of your prior, so a uniform "natural" prior is really a particular choice of parametrization. Using a discrete hypothesis space avoids big parts of the problem.

Wait, why doesn't the entropy of your posterior distribution capture this effect? In the basic example where we get to see samples from a bernoulli process, the posterior is a beta distribution that gets ever sharper around the truth. If you compute the entropy of the posterior, you might say something like "I'm unlikely to change my mind about this, my posterior only has 0.2 bits to go until zero entropy". That's already a quantity which estimates how much future evidence will influence your beliefs. 

Surely something like the expected variance of $\log \left(p/\right(1-p\left)\right)$ would be a much simpler way of formalising this, no? The probability over time is just a stochastic process, and OP is expecting the variance of this process to be very high in the near future.

Unfortunately the entire complexity has just been pushed one level down into the definition of "simple". The L2 norm can't really be what we mean by simple, because simply scaling the weights in a layer by A, and the weights in the next layer by 1/A leaves the output of the network invariant, assuming ReLU activations, yet you can obtain arbitrarily high L2 norms by just choosing A high enough. 

Unfortunately if OpenAI the company is destroyed, all that happens is that all of its employees get hired by Microsoft, they change the lettering on the office building, and sama's title changes from CEO to whatever high level manager positions he'll occupy within microsoft.

Hmm, but here the set of possible world states would be the domain of the function we're optimising, not the function itself. Like, No-Free-Lunch states (from wikipedia):

Theorem 1: Given a finite set $V$ and a finite set $S$ of real numbers, assume that $f:S\to V$ is chosen at random according to uniform distribution on the set $S^V$ of all possible functions from $V$ to $S$. For the problem of optimizing $f$ over the set $V$, then no algorithm performs better than blind search.

Here $V$ is the set o... (read more)

There is no optimization pressure from “evolution” at all. Evolution isn’t tending toward anything. Thinking otherwise is an illusion.

Can you think of any physical process at all where you'd say that there is in fact optimization pressure? Of course at the base layer it's all just quantum fields changing under unitary evolution with a given Hamiltonian, but you can still identify subparts of the system that are isomorphic with a process we'd call "optimization". Evolution doesn't have a single time-independent objective it's optimizing, but it does seem to me that it's basically doing optimization on a slowly time-changing objective.

Why would you want to take such a child and force them to ‘emotionally develop’ with dumber children their own age?

Because you primarily make friends in school with people in your grade, and if you skip too many grades, the physical difference between the gifted kid and other kids will prevent them from building a social circle based on physical play, and probably make any sort of dating much harder.

I'll try to say the point some other way: you define "goal-complete" in the following way:

By way of definition: An AI whose input is an arbitrary goal, which outputs actions to effectively steer the future toward that goal, is goal-complete.

Suppose you give me a specification of a goal as a function $f:S\to \{0,1\}$ from a state space to a binary output. Is the AI which just tries out uniformly random actions in perpetuity until it hits one of the goal states "goal-complete"? After all, no matter the goal specification this AI will eventually hit it, th... (read more)

E.g. I claim humans are goal-complete General Intelligences because you can give us any goal-specification and we'll very often be able to steer the future closer toward it.

If you're thinking of "goals" as easily specified natural-language things, then I agree with you, but the point is that turing-completeness is a rigorously defined concept, and if you want to get the same level of rigour for "goal-completeness", then most goals will be of the form "atom 1 is a location x, atom 2 is at location y, ..." for all atoms in the universe. And when averaged across all such goals, literally just acting randomly performs as well as a human or a monkey trying their best to achieve the goal.

Goal-completeness doesn't make much sense as a rigorous concept because of No-Free-Lunch theorems in optimisation. A goal is essentially a specification of a function to optimise, and all optimisation algorithms perform equally well (or rather poorly) when averaged across all functions.

There is no system that can take in an arbitrary goal specification (which is, say, a subset of the state space of the universe) and achieve that goal on average better than any other such system. My stupid random action generator is equally as bad as the superintelligence w... (read more)

Zvi, you continue to be literally the best news aggregator on the planet for the stuff that I actually care about. Really, thanks a lot for doing this, it's incredibly valuable to me.

Wouldn't lowering igf-1 also lead to really shity quality of life from lower muscle mass and much longer recovery times from injury?

The proteins themselves are primarily covalent, but a quick google search says that the forces in the lipid layer surrounding cells are primarily non-covalent, and the forces between cells seem also non-covalent. Aren't those forces the ones we should be worrying about? 

It seems like Eliezer is saying "the human body is a sand-castle, what if we made it a pure crystal block?", and you're responding with "but individual grains of sand are very strong!"

But perhaps the bigger reason is that I find SIA intuitively extremely obvious. It’s just what you get when you apply Bayesian reasoning to the fact that you exist.

Correct, except for the fact that you're failing to consider the possibility that you might not exist at all...

My entire uncertainty in anthropic reasoning is bound up in the degree to which an "observer" is at all a coherent concept.

And my guess is that is how Hamas see and bill themselves.

And your guess would be completely, hopelessly wrong. There is an actual document called "The Covenant of Hamas" written in 1988 and updated in 2017, which you can read here, it starts with

Praise be to Allah, the Lord of all worlds. May the peace and blessings of Allah be upon Muhammad, the Master of Messengers and the Leader of the mujahidin, and upon his household and all his companions.

... so, uh, not a good start for the "not religious" thing. It continues:

1. The Islamic Resistance Movement “Ham

It is important that Gazans won't feel like their culture is being erased.

A new education curriculum is developed which fuses western education, progressive values and Muslim tradition while discouraging political violence.

These two things are incompatible. Their culture is the entire problem. To get a sense of the sheer vastness of the gap, consider the fact that Arabs read on average 6 pages per year. It would take a superintelligence to somehow convince the palestinians to embrace western thought and values while not feeling like their culture is ... (read more)

Oh, true! I was going to reply that since probability is just a function of a physical system, and the physical system is continuous, then probability is continuous... but if you change an integer variable in C from 35 to 5343 or whatever, there's no real sense in which the variable goes through all intermediate values, even if the laws of physics are continuous.

If he's ever attended an event which started out with less than a 28% chance of orgy, which then went on to have an orgy, then that statement is false by the Intermediate Value Theorem, since there would have been an instant in time where the probability of the event crossed 28%.

The most basic rationalist precept is to not forcibly impose your values onto another mind.

What? How does that make any sense at all? The most basic precept of rationality is to take actions which achieve future world states that rank highly under your preference ordering. Being less wrong, more right, being bayesian, saving the world, not imposing your values on others, etc. are all deductions that follow from that most basic principle: Act and Think Such That You Win.

Wait, do lesswrongers not know about semaglutide and tirzepatide yet? Why would anyone do something as extreme as bariatric surgery when tirzepatide patients lose pretty much the same amount of weight after a year as with the surgery?

But if you are right that you only respond to a limited set of story types, do you therefore aspire to opening yourself to different ones in future, or is your conclusion that you just want to stick to films with 'man becomes strong' character arcs?

Not especially, for the same reason that I don't plan on starting to eat 90% dark chocolate to learn to like it, even if other people like it (and I can even appreciate that it has a few health benefits). I certainly am not saying that only movies that appeal to me be made, I'm happy that Barbie exists and that ... (read more)

I watched Barbie and absolutely hated it. Though it did provide some value to me after I spent some time thinking about why precisely I hated it. Barbie really showed me the difference between the archetypal story that appeals to males and the female equivalent, and how much just hitting that archetypal story is enough to make a movie enjoyable for either men or women.

The plot of the basic male-appealing story is "Man is weak. Man works hard with clear goal. Man becomes strong". I think men feel this basic archetypal story much more strongly than women, so... (read more)

I'm fairly sure that there's architectures where each layer is a linear function of the concatenated activations of all previous layers, though I can't seem to find it right now. If you add possible sparsity to that, then I think you get a fully general DAG.

Their paper for the sample preparation (here) has a trademark sign next to the "LK-99" name, which suggests they've trademarked it... strongly suggesting that the authors actually believe in their stuff.

There are a whole bunch of ways that trying to optimise for unpredictability is not a good idea:

1. Most often technical discussions are not just exposition dumps, they're a part of the creative process itself. Me telling you an idea is an essential part of my coming up with the idea. I essentially don't know where I'm going before I get there, so it's impossible for me to optimise for unpredictability on your end.
2. This ignores a whoooole bunch of status-effects and other goals of human conversation. The point of conversation is not solely to transmit informati

I think you might want to look at the litterature on "sparse neural networks", which is the right search term for what you mean here.

I'm really confused about how anybody thinks they can "license" these models. They're obviously not works of authorship.

I'm confused why you're confused, if I write a computer program that generates an artifact that is useful to other people, obviously the artifact should be considered a part of the program itself, and therefore subject to licensing just like the generating program. If I write a program to procedurally generate interesting minecraft maps, should I not be able to license the maps, just because there's one extra step of authorship between me and them?

The word "curiosity" has a fairly well-defined meaning in the Reinforcement Learning literature (see for instance this paper). There are vast numbers of papers that try to come up with ways to give an agent intrinsic rewards that map onto the human understanding of "curiosity", and almost all of them are some form of "go towards states you haven't seen before". The predictable consequence of prioritising states you haven't seen before is that you will want to change the state of the universe very very quickly.

Not too sure about the downvotes either, but I'm curious how the last sentence misses the point? Are you aware of a formal definition of "interesting" or "curiosity" that isn't based on novelty-seeking? 

According to reports xAI will seek to create a "maximally curious" AI, and this also seems to be the main new idea how to solve safety, with Musk explaining: "If it tried to understand the true nature of the universe, that's actually the best thing that I can come up with from an AI safety standpoint," ... "I think it is going to be pro-humanity from the standpoint that humanity is just much more interesting than not-humanity."

Is Musk just way less intelligent than I thought? He still seems to have no clue at all about the actual safety problem. Anyone thi... (read more)

A theory of the popularity of anime.

Much like there have been ten thousand reskins of Harry Potter I’ve been waiting for more central examples of English-language cultural products to take that story archetype and just run with it. There is clearly a demand.

Well then Rejoice! The entire genre of Progression Fantasy is what you desire, and you need only browse the Best Of RoyalRoad to see lots of english-language stories that scratch that particular itch. In fact, I find these english stories immensely superior to anything in anime or manga. 

A particul... (read more)

Overall, a headline that seems counterproductive and needlessly divisive.

Probably the understatement of the decade, this article is literally an "order" from Official Authority to stop talking about what I believe is literally the most important thing in the world. I guess this is not literally the headline that would maximally make me lose respect for Nature... but it's pretty close. 

This article is a pure appeal to authority. It contains no arguments at all, it only exists as a social signal that Respectable Scientists should steer away from talk of... (read more)

That's not a math or physics paper, and it includes a bit more "handholding" in the form of an explicit database than would really make me update. The style of scientific papers is obviously very easy to copy for current LLMs, what I'm trying to get at is that if LLMs can start to make genuinely novel contributions at a slightly below-human level and learn from the mediocre article they write, pure volume of papers can make up for quality.

• "This has been killing people!"
• "Yes, but it might kill all people!"
• "Yes, but it's killing people!"
• "Of course, sure, whatever, it's killing people, but it might kill all people!"

But this isn't the actual back-and-forth, the third point should be "no it won't, you're distracting from the people currently being killed!". This is all a game to subtly beg the question. If AI is an existential threat, all current mundane threats like misinformation, job loss, AI bias, etc. are rounding errors to the total harm, the only situation where you'd talk about them is i... (read more)

I don't think we need superhuman capability here for stuff to get crazy, pure volume of papers could substitute for that. If you can write a mediocre but logically correct paper with $50 of compute instead of with $10k of graduate student salary, that accelerates the pace of progress by a factor of 200, which seems enough for me to enable a whole bunch of other advances which will feed into AI research and make the models even better.

If we get to that point of AI capabilities, we will likely be able to make 50 years of scientific progress in a matter of months for domains which are not too constrained by physical experimentation (just run more compute for LLMs), and I'd expect AI safety to be one of those. So either we die quickly thereafter, or we've solved AI safety. Getting LLMs to do scientific progress basically telescopes the future.

Fair point, "non-trivial" is too subjective, the intuition that I meant to convey was that if we get to the point where LLMs can do the sort of pure-thinking research in math and physics at a level where the papers build on top of one another in a coherent way, then I'd expect us to be close to the end. 

Said another way, if theoretical physicists and mathematicians get automated, then we ought to be fairly close to the end. If in addition to that the physical research itself gets automated, such that LLMs write their own code to do experiments (or run the robotic arms that manipulate real stuff) and publish the results, then we're *really* close to the end. 

If the question is ‘what’s one experiment that would drop your p(doom) to under 1%?’ then I can’t think of such an experiment that would provide that many bits of data, without also being one where getting the good news seems absurd or being super dangerous.

Not quite an experiment, but to give an explicit test: if we get to the point where an AI can write non-trivial scientific papers in physics and math, and we then aren't all dead within 6 months, I'll be convinced that p(doom) < 0.01, and that something was very deeply wrong with my model of the world.