FairlyInvolved

I realise no one is infallible but do you not think Daniel Kokotajlo's integrity is now pretty well established with regard to those incentives?

We are going to scale up GPT4 by a factor of ~10,000 and that will result in getting an accurate summary of your daily schedule?

There's a pretty good summary of how well it has held up here, by the significance of each claim:

https://www.lesswrong.com/posts/u9Kr97di29CkMvjaj/evaluating...

Outside of a few retail investors institutional buyers are no so naïve that they take a loss-making company paying out a dividend as a measure of good financial health.

Long Intel but as Dylan (and many others) noted they should have stopped the dividend a year ago.

Writing down a merger/acquisition is probably the main scenario where this could extend to the billions.

I don't have an exact list but I think you'll see differences of that order in the years following big acquisitions like AMD / Xilinx for this sector.

I don't think debris 'going higher' isn't much of a problem. Whenever this happens the orbit is going to be more eccentric - meaning a lower periapsis, and consequentially lots of drag that will cause a rapid deorbit.

On the second point about parabolic orbits I also find that probably relatively low risk because we are only talking about a fraction of an orbit for a collision to occur so unless the debris field was massive the chance of another collision is probably still low. Remember when we are modelling orbit collisions normally we are often talking over 25+ years - 100,000 + orbits.

I think the main problem is busy orbits (e.g. sun-synchronous polar orbits at popular altitudes) where most of the debris remains roughly in the same orbit following an acute collision but has a lot of other potential collision targets. Also as satellites are disabled by a collision they lose the ability to avoid other objects already in the same crowded orbit - i.e. the fraction of objects able to take avoidance decreases increasing the chance that future collisions are from 2 incapacitated satellites, removing the possibility of avoidance.

I think that's just a scaling issue, fundamentally there's no reason why a model trained on video couldn't come to create coherent motion in the same way that image models can now product coherent lighting/themes.

Smaller image models had the same problems with logical inconsistency just because they didn't have sufficient general understanding of how visual concepts.

The same is almost certainly true of video - early smaller models will likely create janky movements/motion, however once they've seen enough video to understand how a person walks, how a scene is framed etc.. there's no reason we couldn't get to the same level of maturity as today's image models.

I think the real issue will come from labelling - most video is only going to be labelled simply with basic info/captions without detailed descriptions of the camera pan, movement of subjects. The amount of text required to accurately describe a scene is much larger than a still image and I'm not sure how once would go about collecting this.

I'm not saying they are doing a good enough job, but that doesn't mean their approach isn't entirely without merit.

Even ignoring the infohazard angle if they published everything immediately that would escalate the race. By sitting on their capabilities and waiting for others to publish (e.g. PaLM, Imagen vs GPT-3, DALL-E) they are at least only playing catch up.

Nuclear warfare is much less concerning than misaligned AI.

Take a look into scaling laws and alignment concerns, this is a very real challenge and existential risk not some crackpot theory.

I kind of get the sentiment about openness but I think it's way more nuanced than you are making out.

There are very good reasons for withholding SOTA models, primarily from the info hazard angle and avoiding escalating the capabilities race which is basically the biggest risk we have right now.

Google / Deepmind have actually made some good decisions to try and slow down the race (such as waiting to publish).

I wouldn't necessarily say a major breakthrough as such but I do think some architectural change is needed. There are concepts in images that we don't rely on purely visual understanding for - like words, we have a language model that we use to rely on when we see text in images. I think we need the same thing in our models to reach the next level of capabilities by combining models across different domains. I don't know if this manifests as pre-training with a language model and then expanding and updating the tensors as part of image training, or some more complicated merger of the models.

To learn logical concepts just from images seems entirely impractical, like we can't rely on having enough images such that models can understand words coherently as language. You could draw a picture of a sign that says "children crossing" not because you can understand and remember exactly what an image of such a sign would look like, but because you have an understand of English and the character set that would let you reproduce it. If you tried to learn to create the same sign in Arabic you'd either need to see a huge number of signs to learn from or (more likely) build a language model for Arabic.

The kind of abstract understandings that we know we can train in language models just aren't learned by image transformers at this scale (or likely any practical scale). A language model could easily understand: "A red cube is stacked on top of a blue plate, a green pyramid is balanced on the red cube" and infer things like the position of the pyramid relative to the blue plate, image models quickly fall over with such examples.

An interesting nascent (and hacky) example of the benefits of combining models is people are using language models like GPT-3 to create better prompts for image models.

And N7 is almost certainly cheaper (holistically) than Intel 7, from an economic perspective AMD have done more with less.

They hold a monopoly of merit, many tried EUV and they succeeded in doing something that others were calling impossible even 5 years before they shipped it, not by shutting down competition but simply because no one else could.

The idea that ASML would somehow be more worthy of merit if Nikon/Canon also succeeded is weird.

Would the moon landing me more impressive if Russia and Japan also managed it?

Several people have answered this from the angle of why the cost is large, but missed the other (necessary) angle - the benefits are also massive and well understood. Plenty of potential projects have huge costs but they don't exist because the benefits aren't big enough, highways exhibit a sort of survivor bias in the mega project world.

Highway systems create massive economic benefits that are relatively simple to calculate, just in the raw # hours saved which massively increases productivity. This can be well captured by the state through taxes or (less frequently) by private industry through tolls - i.e. the people that direct their construction.

Other mega projects have much less certain benefits, or the benefits are harder to capture by those with the ability to allocate resources because incentives aren't sufficiently aligned.

I think that's well-understood and the existence of the F-35 evidences the perceived threat from nations with more advanced militaries.

It does if the reward is sufficiently large.

If BTC mining relied on pieces of litter being collected you'd get people spending $millions on factories to produce litter and vehicles to distribute it just so more could be collected. That sounds crazy to type out, but it's no stranger than the situation we currently find ourselves in. The most advanced tools man has ever created are being used to create chips with no practical value, but that are good for converting energy into tokens through pointless calculations.

I don't believe it is possible to create a proof of work system that is not a "proof of waste" once the reward is high.

Spare CPU cycles felt like a waste, that becomes dedicated processing, that becomes dedicated hardware, that eventually became designing custom hardware.

I really don't see how the current hashrate is required to prevent a 51% attack, that seems way beyond what is necessary.

These Proof of Waste systems aren't really a negative feedback loop, ultimately asset prices determine mining yield. Capacity will keep being added to the network until the difficulty rises such that you are required to destroy $99.99 of resources for $100 of crypto. Until that point it's worth buying more hardware and energy as you get a positive expected return.

They sort of have to, you can't selectively enforce trademarks like you can with other IP