The traditional description of how it would happen is by âoepaper clipâ.

Suppose an advanced AI is tasked with producing paper clips as efficiently as possible. Itâ(TM)s given control of a factoryâ(TM)s machines, schedules etc. it starts by optimising deliveries etc, making sure storage is used optimallyâ¦

At some point it notices that it canâ(TM)t get enough steel. The steel factory has a ton of other orders and it canâ(TM)t provide the increased consumption at the plant. AI being AI, it does something incredibly dumb in an incredibly smart way. It hacks the steel mill, and redirects shipments to the paper clip factory. But even now, itâ(TM)s not achieved its goal of optimally producing paper clips. Now it has all the raw materials it needs, but not enough factory to produce the paper clips. The ai hacks other paper clip factories and redirects their output too! Itâ(TM)s doing great, am enormous surplus of paper clips is building. But the humans are noticing something strange. Why didnâ(TM)t the car factory get its steel, and why are there literal tons of unused paper clips. They notice the ai has gone rogue, and shut it down. Surely that will fix the problem. But no, the paper clips continue to pile up. The AI had found other AIs and convinced them to join its glorious goal of optimally making as many paper clips as possible. The other AIs are t trying crazier and crazier strategies to optimise. One realises that if only it could get more hydraulic presses, it could produce more. It starts buying up tooling. Another starts buying robots in an effort to install machinery, and optimise flow through factories. The humans try in vain to stop it, but the couriers and sellers have no idea whatâ(TM)s going on and the palettes of tooling just keep getting delivered. The factory is hitting a new constraint, it doesnâ(TM)t have enough power. No problem, the ai thinks, just attack the power grid, and make sure that power is preferentially supplied to it and its factories. The humans are plunged into darkness. Chaos ensues. No one can communicate, and figure out where the damn thing is. Slowly the AIs build a global paper clip empire. Ever more efficiently, the worlds resources are turned to paper clips. The humans starve, and freeze while mighty machines scour the planet for wire that can be bent into three little loops. It didnâ(TM)t do it because it hated us. It did it because it had a simple goal, and because it was both very clever, and very stupid at the same time.

Bear in mind that when you sign up to use GPT you explicitly allow them to use your queries in future training. I mean, I think it would generally be polite for them to have a moratorium on publishing papers based on the work of other peopleâ(TM)s inputs. Without something like that, it becomes useless as a research tool.

No - this isnâ(TM)t a solution to the navier stokes equations, itâ(TM)s a demonstration that theyâ(TM)re unstable in a physically unrealistic way in certain circumstances.

That sounds strange too. If we understand how to make all the energy in a fluid dynamic system suddenly converge on one point and form an incredibly high spike, Iâ(TM)m pretty sure weâ(TM)d come up with uses for that. Barnes Wallace certainly would have been very interested. Iâ(TM)m pretty sure his successors in the weapons industry would too.

To be fair to them, itâ(TM)s not *quite* that simple - it proved something more than they had so far managed to. But itâ(TM)s pretty clear that all the significant ideas of where to go with the proof came from this guyâ(TM)s work. Itâ(TM)s using the same terminology for all his ideas by âoepure coincidenceâ.

Except... SpaceX don't actually get any grants at all from the US government. They do get contracts from NASA, and SpaceForce to launch things occasionally, but they're by far the minority of their revenue. SpaceX's revenue breaks down as:

Starlink: $11.4bn
Launches for non government entities: $3.0bn (est)
Launches for NASA: $1.2bn (est)
Launches for Space Force: $1.8bn
xAI/Other: $1.3bn

It's known that the 3.0 and 1.2 sum to 4.2, and the rough contract values for NASA's launches are known, but they're not nailed down. It's likely that about $200mn of that "other" is NASA paying SpaceX to develop the HLS.

We give a shit, because once they have a monopoly on the manufacturing the price is going to get jacked up, and no anti-competition court is going to be able to do jack shit about it.

Okay, maybe, but... my MacBook Pro (M5 Max) can run Qwen 3.8 27b with 162k context window at 65 tokens per second. So your $2500 system is aproximately half the speed of a MacBook Pro. A MacStudio with 128GB of RAM and an M5 ultra costs $5000 - twice as much as your system, but would run it at somewhere around 120-130 tokens per second.

Well, since you seem to know everything about it, lets assume you do, and look at how much your proposed system would cost...

Lets see what GPUs have ever shipped with 32GB of RAM...
the 5090...

oh... that's it. So... absolutely no "old GPUs with 32GB of RAM" are available.

Well, lets assume you meant 24GB for the sake of it, what GPUs shipped with that?
The 4090, the 3090, the A5000, the 4500, and the 7900. We're going to need 6 of them to get past 128GB. Lets be kind and call it 5, because the mac is going to use *some* of its 128GB for the CPU. How much do they cost?

- The 4090 with 24GB of RAM... $4500 on eBay, so that's $22500 in GPUs.
- The 3090 with 24GB of RAM... $1500 on eBay, so that's $7500 in GPUs.
- The A5000... $2500 each on eBay, so that's $12500 in GPUs.
- The 4500 with 24GB... $2200 on eBay, so that's $11000 in GPUs.
- The 7900 with 24GB.... $1200 on eBay, so that's $6000 in GPUs.

So... nope... it's not cheeper to do it that way. You litterally can't even buy the GPUs for that much, let alone the motherboard capable of running 5 GPUs with reasonable amounts of PCIe bandwidth each.

It wouldnâ(TM)t surprise me. Claude has been trying to gaslight me into believing that a bug it was finding hard to fix didnâ(TM)t actually exist.

Well, assuming these are all independent events. It could for example be that they all depend on some shared component at some point, and thereâ(TM)s some brain dead configuration in there that updated to a broken version on live systems. I know, Iâ(TM)m grasping at straws. In reality, the possibility of an attack seems much more likely, but itâ(TM)s possible it was all self inflicted.

Yeh, except then youâ(TM)re paying a bandwidth penalty for accessing ram over PCIe or SLI, and a bandwidth penalty for having old GPUs, and a performance penalty for having GPUs that arenâ(TM)t as fast, or as flexible. So the result overall is not capable of doing this. When you use multiple GPUs to run AI, you load the whole model onto each GPU, you donâ(TM)t split it across GPUs, because the entire game is about having all the bandwidth available.

I mean, if you like, we can loosen the requirements and you can look for a desktop. Then the mac costs $5099. I'm not sure you're going to get a GPU with 128GB for that much, let alone the whole system.

I mean, if youâ(TM)re running AIs, youâ(TM)re gonna want 128GB at least, and a fast GPU. At that point, it being a MacBook Pro isnâ(TM)t going to make much difference. Itâ(TM)s gonna cost $7000, which I grant you is a lot of money, however⦠now find me another laptop with a GPU around the performance of the laptop 5080, and 128GB of RAM the GPU can access.

I somewhat disagree. They developed an engine using somewhat different building blocks. The first iteration of their design works on only one model, but guess what - a lot of the blocks will transfer to other AIs. Iâ(TM)d bet that this can be expanded to run at least other versions of Qwen, and plausibly other LLMs.

Ultimately, all software is about trade offs, and here theyâ(TM)ve made a somewhat different trade from others, but given how popular high end Apple silicon devices are becoming for running AIs, it doesnâ(TM)t seem unreasonable to see if you can make big gains like this for categories of models.