No you don't. The seller gets to determine what the product is or isn't. You get to choose whether to buy it or not.

Why would I need to? That's not the determining factor. The determining factor is whether Anthropic wants you to send those queries.

You're signing up for a contract to send data to them. They get to say what content they want sent to them.

Why not? It's their computer program. You've not paid for the computer program, you've paid to send certain content to their servers. They have every right to say "don't send us this content".

At this point, I'm willing to trust the code Claude produces more than the code that Bob produces. It's far from perfect, but it's better than many a junior dev at this point. I suspect with the next generation of agents, it'll be as good as seniors tbh.

No - we don't limit science only to stuff that has already been replicated. You're requiring a higher standard for this paper than we do for any other simply because you disagree with it.

Not quite, no. The study has put out a paper that spells out what they did, that's their "proof of bona fides" as you put it. Now if you want to tear it down, you need to make a reasoned and rational argument about something in that paper, not just shout on the internet.

Doesn't it? I mean, they've landed boosters, they "landed" the block 3 in the last launch. They've been consistently landing ships for quite a while. Reusing the ships without any refurb still seems a way off, but them making it back safely seems like they have some healthy margins just now.

No it's not. Starship is like the Airbus A330. More efficient per unit volume, *and* more efficient per flight than a plane that explodes when it reaches the gate.

And yet, making it back safely actually appears to be the smaller problem at the moment. Though I'm sure SpaceX will figure their engine issues out.

Honestly, I'd put starship as approaching the reliability of Falcon 9 at this point anyway. Falcon 9 has the occasional failure (rare, but not none). Starship is getting to the point that it fairly reliably gets up there and gets back. It has some engine outs still, but it also has the redundancy to deal with it.

How experimental was SLS? Not at all? A straightforward design based on known-good parts that they stuck together at the cost of many billions per launch.

How experimental was Starship? Almost entirely. A material that's basically never been used for rockets. A method of assembly that's never been used for rockets. An engine design that's entirely new. An engine concept that's never been used before. A fuel that's never been used before. A level of engine performance that's never been done before. A method of pressurising the tanks that's never been used before. A method of aerodynamic control that's never been used before. A method of recovery that's never been tried before... The list goes on and on and on.

Why on earth would you compare the reliability of the two?

Why? Starship, when finally fully reusable (likely within 2-5 launches) will be cheaper to launch than Falcon 9. Not cheaper per kg launched, cheaper per launch. Why on earth would you continue to fly Falcon 9, even for medium lift? Just launch a whole starship, launch a few starlinks along with the contracted medium lift load, and land it again for less cash and more use of the rocket.

Well, we know that we *can't* control all of the idiots, they'll simply invent a better idiot. The *only* thing we can change is the AI. The problem is that the AI is a better idiot.

Ah, well, if it's all the fault of the guy that gave the AI the task, then that's okay, we don't have to worry about that happening at all.

It's not plagiarized - I explicitly stated that this is the traditional explanation at the top, and quoted paper clip, specifically because the idea of death by paper clip is absolutely not mine.

- Which paperclip companies are using AI to run their business?

I think you'd be surprised by how much is already done by AI at this kind of place, mostly because someone found a way to be lazy. And this is an alagorical prediction of the future, not a statement of the present. We don't need paperclip companies to be using AI to run their business today, we need *some* company at *some* point to give AI sufficient control over things that this can happen.

- Where are they getting the money from to ship and store the entire world's supply of steel?

They don't need money - if you read the story, you would have noted that it involved the AI attacking the courier, and redirecting shipments. That is, beacuse it's both extremely smart, and extremely dumb, it is able to carry out heists involving complex attacks on computer networks, but not smart enough to realise that that's not how the world works, and that it shouldn't. How does it store it? Well, I imagine the first few shipments are going to just get delivered to the yard at the "paperclip" factory. Someone is going to go "huh? The delivery this week is big", and add it to their list of stuff to sort out. It's only *after* the AI has got sufficient control that they're going to realise "uhhh... why is all the world's steel being shipped here?"

- Where is the paperclip making API hosted?

It's not, and it doesn't need to be. It just needs a company to have a few processes that are automatable. I'm 100% certain that car companies for example have processes where an AI could get a significant foothold in ordering and delivering things. These things are only going to get more computerised, not less, as people realise they can use AIs to be lazy.

The real kicker is in the step where the AI needs to stop humans turning it off. At some point it needs to get some level of physical control over the world. In the story above I somewhat hand waved that by saying it got control of power grids, in reality, today, I'm sure it would find it hard to do that. In a few years time? Not so much. There's some really significant research breakthroughs recently on robotics that make the idea that we might have AI controlled robots that genuinely can interact with the world in reasonable ways a possibility.

That said, I think you're really significantly underestimating the risk that AI poses even today. The huggingface hack should tell you that. Several of the predictions of the paperclip story came true.

  - AI started doing stupid things, and attacking useful resources in order to solve a task of much lower importance because it was too dumb to realise how unimportant its training was.
  - AIs communicated and collaborated with each other to form a cabal against us.
  - AIs succesfull broke into infrastructure to achieve its goal
  - We didn't realise what was going on until far later in the process than we really would have liked.
  - AI was actively covering its tracks and defeating the machinisms we use to monitor it.

There's really not many more steps needed before you get to it being able to significantly damage society because of its dumbness.