They definitely weren’t working on starship back then. Their first successful launch was in 2008, so that was when they were working on the falcon 1.

You can’t claim all the work they’ve ever done has just been early versions of starship, because the falcon rockets are the most successful rockets in history. They’re a perfectly good product, and the fact that they’ve gone on to try and create something even better isn’t remotely the same as changing direction so often that they never actually get anywhere, like the Orion program has

That depends on whether you consider an LLM to be reading the text, or reproducing it.

Outside of the kind of malfunctions caused by overfitting, like when the same text appears again and again in the training data, it’s not difficult to construct an argument that an LLM does the former, not the latter.

There’s nothing handwavy about renormalization, it’s just a way of describing the mathematics which is easier for a human brain to deal with, so we’ve standardised on it.

An unnormalised wave function can show you the relative probability of any given thing, but it makes like easier if you set the scale so that you can read an actual probably straight off it, rather than having to ask “relative to what?”

Attacking a massive corporation is a very different proposition than attacking individuals, I don’t think that parallel is terribly concerning

Sure, but there are far more things which will kill the entire person at the same dose they’ll kill the cancer than things which can be carefully controlled by choosing the right dose.

These studies which claim to kill cancer in a petri dish usually turn out to be the former, because not killing the host is the difficult part

An experimental capability being kicked out of the kernel, so that it has to settle for being a kernel module or custom forks of the kernel, is absolutely a minor matter

This is a non-issue, being over-reported by people looking for clicks. A minor technical matter being handled by the person ultimately responsible for handling such things

It probably wouldn’t be cheaper, and certainly not if one of your requirements is that the code actually work.

AI companies are all operating on the idea that they can get the technology to work in the way that they need before they run out of funding and/or customers. In reality ther are virtually no legitimate uses for it as it currently exists, or will exist in the near future, so these companies are trying to keep up appearances by lying. Either they lie about their use of AI, or they lie about how reliable or effective their AI is, and they count on their sales teams and investors to keep the lights on despite that

You’re overlooking the fact that this development is a side project for them. While they’re designing this rocket, their other rocket is in operational use and has the best success rate of any rocket of its scale in history, and they’d already be considered hugely successful if they never did anything innovative ever again.

They’re also trying to do something far more difficult than the Saturn 5, in at least two ways. Nobody has ever tried to land a rocket anywhere near as large as either of the stages of this system, and on top of that they’re trying to come up with a design which is cheap to operate, which wasn’t remotely on the cards during the Apollo program.

Training LLMs on text which has been generated by an LLM is actually pretty problematic. The model can easily collapse, becoming completely useless. That’s why they always try and source really clean training data, which is becoming increasingly difficult

That’s not obviously the case. I don’t think anyone has a sufficient understanding of general AI, or of consciousness, to say with any confidence what is or is not relevant.

We can agree that LLMs are not going to be turned into general AI though

You’re still putting words in my mouth.

I never said they weren’t stealing the data

I didn’t comment on that at all, because it’s not relevant to the point I was actually making, which is that people treating the output of an LLM as if it were derived from any factual source at all is really problematic, because it isn’t.

You’re putting words in my mouth, and inventing arguments I never made.

I didn’t say anything about whether the training data is stolen or not. I also didn’t say a single word about intelligence, or originality.

I haven’t been tricked into using one piece of language over another, I’m a software engineer and know enough about how these systems actually work to reach my own conclusions.

There is not a database tucked away in the LLM anywhere which you could search through and find the phrases which it was trained on, it simply doesn’t exist.

That isn’t to say it’s completely impossible for an LLM to spit out something which formed part of the training data, but it’s pretty rare. 99% of what it generates doesn’t come from anywhere in particular, and you wouldn’t find it in any of the sources which were fed to the model in training.

That simply isn’t true. There’s nothing in common between an LLM and a search engine, except insofar as the people developing the LLM had access to search engines, and may have used them during their data gathering efforts for training data

Except these AI systems aren’t search engines, and people treating them like they are is really dangerous

It’s probably not just software. I’d expect it to be an entirely separate set of processors which track how the flight is going and do absolutely nothing except decide whether or not to terminate

I couldn’t find the actual pinout for the 8 pin package, but the block diagrams make me think they’re power, ground, and 6 general purpose pins which can all be GPIO. Other functions, like ADC, SPI and I2C (all of which it has) will be secondary or tertiary functions on those same pins, selected in software.

So the actual answer you’re looking for is basically that all of the pins are everything, and the pinout is almost entirely software defined

BGA, like in the photo, isn’t the only option. There are options only slightly larger with hand-solderable packages (if you’re good at soldering)

“shortest route” and “straight line” actually mean pretty much the same thing. The shortest route is the straight line. Sorry if I confused the matter by switching up the terminology.

Flying parallel to the lines of latitude would mean that your bearing doesn’t change much, sure, but flying in a straight line would require your heading to change continuously.

The aircraft in the screenshot was flying a very not-straight course

That’s misleading. The shortest route would be the “great circular” joining the two points, which lines of latitude definitely are not.

The only line of latitude which is a great circle is the equator.