A nice property of the above process is that we can sample xt at any arbitrary time step t in a closed form using reparameterization trick. Let αt=1−βt and α¯t=∏i=1tαi: xt=αtxt−1+1−αtϵt−1 ;where ϵt−1,ϵt−2,⋯∼N(0,I)=αtαt−1xt−2+1−αtαt−1ϵ¯t−2 ;where ϵ¯t−2 merges two Gaussians (*).=…=α¯tx0+1−α¯tϵq(xt|x0)=N(xt;α¯tx0,(1−α¯t)I) (*) Recall that when we merge two Gaussians with different variance, N(0,σ12I) and N(0,σ22I), the new distribution is N(0,(σ12+σ22)I). Here the merged standard deviation is (1−αt)+αt(1−αt−1)=1−αtαt−1.

The Lump of Labour fallacy is the misconception that there is a fixed amount of work to be done, and that if some work is taken by a machine then there will be less work for people. But if it becomes cheaper to use a machine to make, say, a pair of shoes, then the shoes are cheaper, more people can buy shoes and they have more money to spend on other things besides, and we discover new things we need or want, and new jobs. The efficient gain isn’t confined to the shoe: generally, it ripples outward through the economy and creates new prosperity and new jobs. So, we don’t know what the new jobs will be, but we have a model that says, not just that there always have been new jobs, but why that is inherent in the process. Don’t worry about AI!

In the Hinton et. al paper, they’re able to match the performance of an ensemble of 10 models with a single, distilled model, and performance only decreases from 61.1% accuracy to 60.8% accuracy (99.5% of the original performance, with 10% of the size). Now, the Hinton paper is comparing against an ensemble, which is a particularly wasteful way to increase model size, but that’s still impressive result, and much better than training a model from scratch to perform the same task (which had only 58.9% accuracy).

However, co-ordinate networks admitting traditional activation func-tions (e.g., ReLU, sigmoid, and tanh) fail to capture high-frequency details due to spectral bias [29]. To overcomethis limitation, positional embedding layers [40] are oftenadded, but they can produce noisy first-order gradients thathinder architectures requiring backpropagation [17, 8]. Arecent alternative approach is to use non-traditional activa-tion functions, such as sine [35] or Gaussian [30] activa-tions, which enable high-frequency encoding without posi-tional embedding layers. The major benefit of these activa-tions over positional embedding layers is their well-behavedgradients [35, 30].

Finally, a language model generates the final prediction along with anoptional explanation

Planning forAGI and beyond

We have a nonprofit that governs us and lets us operate for the good of humanity (and can override any for-profit interests), including letting us do things like cancel our equity obligations to shareholders if needed for safety and sponsor the world’s most comprehensive UBI experiment.

a tight feedback loop of rapid learning and careful iteration

The institutions of the world will need to be strengthened with additional capabilities and experience to be prepared for complex decisions about AGI.

As another example, we now believe we were wrong in our original thinking about openness, and have pivoted from thinking we should release everything (though we open source some things, and expect to open source more exciting things in the future!) to thinking that we should figure out how to safely share access to and benefits of the systems. We still believe the benefits of society understanding what is happening are huge and that enabling such understanding is the best way to make sure that what gets built is what society collectively wants (obviously there’s a lot of nuance and conflict here).

Many of us think the safest quadrant in this two-by-two matrix is short timelines and slow takeoff speeds

it seems like the original conception of hyper-evolved RL agents competing with each other and evolving intelligence in a way we can’t really observe is less likely than it originally seemed

If AGI is successfully created, this technology could help us elevate humanity by increasing abundance, turbocharging the global economy, and aiding in the discovery of new scientific knowledge that changes the limits of possibility.

They refute the contemporary argument that technology and humanity will always complement each other in the labor domain, relegating that phenomena to the earlier, “First Machine Age.”

in the future, talent, more than capital, will represent the critical factor of production.

Third Industrial Revolution’s start into digitization, the Fourth Industrial Revolution’s technologies, such as artificial intelligence, genome editing, augmented reality, robotics, and 3-D printing, are rapidly changing the way humans create, exchange, and distribute value.

there are some very specific patterns that they're entering into to defeat the ostensible purpose of the group meeting together. And he detailed three patterns.

The first is sex talk, what he called, in his mid-century prose, "A group met for pairing off." And what that means is, the group conceives of its purpose as the hosting of flirtatious or salacious talk or emotions passing between pairs of members.

The identification and vilification of external enemies. This is a very common pattern. Anyone who was around the Open Source movement in the mid-Nineties could see this all the time. If you cared about Linux on the desktop, there was a big list of jobs to do. But you could always instead get a conversation going about Microsoft and Bill Gates. And people would start bleeding from their ears, they would get so mad.

The third pattern Bion identified: Religious veneration. The nomination and worship of a religious icon or a set of religious tenets. The religious pattern is, essentially, we have nominated something that's beyond critique. You can see this pattern on the Internet any day you like. Go onto a Tolkein newsgroup or discussion forum, and try saying "You know, The Two Towers is a little dull. I mean loooong. We didn't need that much description about the forest, because it's pretty much the same forest all the way."

In many situations, all people have access to the network. And "all" is a different kind of amount than "most." "All" lets you start taking things for granted. Now, the Internet isn't everywhere in the world. It isn't even everywhere in the developed world. But for some groups of people -- students, people in high-tech offices, knowledge workers -- everyone they work with is online. Everyone they're friends with is online. Everyone in their family is online. And this pattern of ubiquity lets you start taking this for granted. Bill Joy once said "My method is to look at something that seems like a good idea and assume it's true." We're starting to see software that simply assumes that all offline groups will have an online component, no matter what.

you cannot completely separate technical and social issues

Members are different than users. A pattern will arise in which there is some group of users that cares more than average about the integrity and success of the group as a whole. And that becomes your core group, Art Kleiner's phrase for "the group within the group that matters most."

The core group has rights that trump individual rights in some situations. This pulls against the libertarian view that's quite common on the network, and it absolutely pulls against the one person/one vote notion. But you can see examples of how bad an idea voting is when citizenship is the same as ability to log in.

Users have to be able to identify themselves and there has to be a penalty for switching handles. The penalty for switching doesn't have to be total. But if I change my handle on the system, I have to lose some kind of reputation or some kind of context. This keeps the system functioning.

Second, you have to design a way for there to be members in good standing. Have to design some way in which good works get recognized. The minimal way is, posts appear with identity. You can do more sophisticated things like having formal karma or "member since."

Three, you need barriers to participation. This is one of the things that killed Usenet. You have to have some cost to either join or participate, if not at the lowest level, then at higher levels. There needs to be some kind of segmentation of capabilities.

And, finally, you have to find a way to spare the group from scale. Scale alone kills conversations, because conversations require dense two-way conversations. In conversational contexts, Metcalfe's law is a drag. The fact that the amount of two-way connections you have to support goes up with the square of the users means that the density of conversation falls off very fast as the system scales even a little bit. You have to have some way to let users hang onto the less is more pattern, in order to keep associated with one another.

A Group Is Its Own Worst Enemy

So email doesn't necessarily support social patterns, group patterns, although it can

New players cannot compete with these successful networks, whose influence deepens and becomes more entrenched as they ingest more data, more resources.

(in contrast to metaphysically intelligent, which would mean that one is able to make decisions free from the data one has gathered, or how one has been innately programmed).