Blue Origin is far, far from having "caught up". They've had three launches, with a 33% mission failure rate. They are now where SpaceX was fifteen years ago, but with a much worse record (Falcon 9s first mission failure was launch 19, though it did have a partial failure on launch 4 -- primary payload successful, secondary payload failed). New Glenn has done better on booster recovery, but they weren't the ones learning how to do it.

And even if SpaceX never manages to make Starship fully-reusable, they can always punt, build a lighter, fully expendable second stage and have a launch platform that blows every other heavy lift vehicle in the world away.

The Chinese are moving pretty fast but they're also a generation behind.

My understanding is that LLMs are built on a foundation of ANNs, and that indeed the backpropagation used to train ANNs is a statistical process;

Two responses. One, that's discussing individual-neuron scale processes rather than collective processes; and this was a discussion about inference, not training. Human neurons also learn by error minimization (Hebbian learning). But this does not describe the macroscopic processes that result from said minimization.

* During training, neurons develop into classifiers that detect superpositions of concepts that collectively follow the same activation process. Individual neurons weight their input space and subdivide it by a fuzzy hyperplane to achieve a classification result.

* In subsequent layers, said input space is formed from a weighted combination of the previous layer's classification; thus, the superpositions of questions being formed are more complex, as are the classification results.

* In a LLM, this iterates for dozens of layers, gaining complexity at each layer, to form each FFN

* The initial input space to a FFN is a latent (conceptual representation), as is the output; the FFNs, in result, function as classifier-generators; they detect combinations of concepts in the input space, and output the causally-resultant concepts into the output space

* FFNs alternate with attention layers dozens to hundreds of times in order to process the information, each layer building on the results of the previous one.

The word to describe that is not "statistics". It's "logic".

In a LLM, the first few layers focus on disambiguation. If there's a token for "bank", is this about a riverbank, a financial bank, banking a plane, etc? As the layers progress, it starts building up first simple circuits, and then progressively more complex circuits - you might get a circuit that detects "talking like MAGA", or "off-by-one programming errors", or whatnot. In the late layers, you have the general conclusions reached - for example, if it were "The capitol of the state that contains America's fourth-largest metro area is...", you've already had FFNs detect the concepts of fourth-largest metro area and encoded Dallas-Forth Worth, and then later taken that and encoded "Texas", and then finally encoding "Austin". And then in the final couple layers you converge back toward linguistic space.

Anthropic has done some great work on this with attribution graph probes and the like; you can detect what circuits are firing, and on what things those circuits fire, and ramp them up or down to see how it modifies the output. They very much work through long chains of logical inferences.

I use every style imaginable, including photos, in my tests. Same result every time.

One time I even did it with a Calvin and Hobbes comic, pretending than an AI made it. Responses included things like "The illustration also looks like shit and barely makes sense. Hope that helps.", "God damn this sucks so bad", "This also fucking sucks", and "The only punchline here is casual, pointless cruelty. if you think this is funny then you're literally a psychopath."

I don't want to take transit. It stinks, it's unsafe, and takes away too long to get places.

Yes we know America is utter shit at transit.

Do you realise that it's clean, safe and quick in other parts of the world?

All true, but it's worth pointing out that if the Starship lander succeeds it will enable us to do a lot more, a lot faster. The whole "15 refueling flights for every moon trip" seems kind of crazy on its face, but if you look at the costs (assuming Starship works and become fully reusable), it makes the total cost per kilogram delivered to the surface of the moon insanely low and enables comparatively massive payloads to be delivered.

Big risk, big (potential) reward. Running both the Starship and Blue Moon projects in parallel is probably a good risk mitigation strategy, but if Starship succeeds completely, Blue Origin's lander will be a relic. Of course, it's also possible that Starship will just fail, or that it will succeed but be difficult to man-rate, in which case it may become the delivery service for lunar cargos, while people fly on Blue Moon.

LLMs are not "statistical models" (randomness only even comes into play in the final conversion from latent space to token space because latent space is high dimensional, token space is low dimension, you need a rounding mechanism, and a "noisy" rounding mechanism works best; what you're thinking of, by contrast, is Markov models). And you cannot just "get lucky and randomly solve an unsolved math problem"; that's not how any of this works.

Also, it's silly that people are acting like "all problems but this one were already in the literature". AI has solved a whole slew on Erdos problems, and only a fraction had anything to do with existing literature.

And even in "existing literature" examples, it's not "nobody ever thought to search before" as if all mathematicians are morons, or that mathematicians adore putting out Erdos problem solutions without claiming them, It's that nobody had ever thought to apply an obscure technique from a given piece of literature to said Erdos problem.

The simple fact is, AI has gotten much better at solving unsolved math problems than humans are. It's simply another field that it's taking over, the same way it has been taking over programming. One can debate how much is "clever insight" vs. "just chugging away at possibilities until it hits on ways to advance toward the goal", but ultimately, that's a distraction from the fact that: it's getting really good at solving math problems that humans have spent decades on without success.

I'm also going to add that the number of "GW" of AI promised to be installed exceeds the power generation manufacturing capacity of the world. Building more of the latter to match the former exceeds the heavy industry capacity of the world.

There's no physical way to get from where we are now to the growth claims that are buoying AI stocks!

Hey, we've got to enlighten them all about the cubical nature of time first before we can get them to accept the electric universe.

Let me spell it out for you. Here are some specific areas the university folks have no idea about

ok...

You show me a candidate with an advanced degree that can demonstrate knowledge in these areas, I will happily hire them. Most, however, do not, because school doesn't teach what it takes to engineer software in the real world.

That's not right. It's not even wrong. It's just a confused mess.

Every time I try and reply, I think actually no, that's not the most pertinent point, and that ends up going in circles. So, I'll just reply with a list of points in no particular order of importance.

* Degrees aren't trade schools of whatever framework you're using today.
* Advanced degrees don't reduce knowledge, so if bachelors know something so will someone with a bachelors and masters, which makes the masters neutral, not a strike against
* Unless you're interviewing straight out of university you should be interviewing on relevant experience not giving someone a mark against because they decided to expand their knowledge of some niche 20 years ago
* Speaking of straight out of university, no one comes straight out of university knowing all that
* It rather depends on what subject the advanced degree is in.
* What master's degree in databases doesn't teach relational databases?
* PhDs don't "turn things little demonstration project in to their professor".[*] You literally have no idea what a PhD even is.
* About the only people qualified to write crypto code are people with PhDs; very many of the latest attacks, side channels etc come from academia. Anyone who's done a PhD in cryptography will know way more about nonces etc than you most likely.
* Did I mention no one coming out of a bachelors knows all that stuff. So it's complete nonsense that a master's is a strike against.

so yeah, you have a chip on your shoulder a mile wide.

[*] In some universities there's a bit of that at the beginning, but it does no form the majority of the time.

Fair, though my data -- and Play purchases -- were my bigger concern when they first started talking about just ending legacy GSuite.

You must have signed up to change over before they backed off. They announced that everyone would have to switch and pay, but I waited because I didn't think it would stick, and it didn't. I have about 25 users on mine, so paying wasn't really feasible.

I searched for it and found it. I can't check now; I left Google last year.

Clearly that wasn't spam, because his name is nospam007.

Forced? You think reviewers of writing contests are captives, imprisoned by mad scientists and forced to read story after story, like some sort of literary equivalent of the backstory to MST3K?