XMPPwocky

expert streaming with cache-conditional routing is very powerful (you can run 4.1 Flash on two Strix Halos and a 9070XT at around 50-70TPS decode in ~full quality), and that setup might be able to at least do a few tok/s with it.

Seems like it'd depend a lot on e.g. carbon impact of the marginal watt v.s. environmental impact of the marginal gallon? Might depend on the individual site.

It looks to me like there's two Google DCs with different LLCs?

> However, using a computer cursor to highlight the redacted text box in the report to DWEE, then copying and pasting it into a separate document reveals that *Agate LLC* is reporting 52.65 megawatts of electricity during peak electrical demand and 13.299 megagallons of water between cooling towers, evaporative systems and site operations in the last year. That totals out to 13 million gallons of water.

> For context, 13 million gallons is enough to fill around 20 Olympic-size swimming pools, and is less than half of what the City of Lincoln reports using on Sept. 29.

> The *six* data centers reporting as of Sept. 30 show a total use of 765 million gallons of water last year, enough water to fill 1,159.93 Olympic-size swimming pools.

> More improper redaction shows that the data center using the most water annually is *Fireball Group LLC*, the Google Data Center in Papillion. Fireball LLC reports 547.88 megagallons for their 2025 Annual Water Consumption.

out of curiosity, do you think fable would get this right? (I'm not sure myself, and haven't tried yet.)

hm- does the model that wrote this know that labs already pay for training data- that stuff scraped from the Internet is not particularly where today's capability gains come from?

Looks like this has nothing to do with the hypervisor, it's not a traditional VM escape

Thanks for the response! This helps me understand what you meant way better. I'm about to go to bed here, but I'll respond in the morning. Ty!

> My point here is there is no continuous state that is not computed from the context.

Oh, are you talking more about the lack of continual learning across context windows? Gotcha if so, my error.

Could you explain why running without sensory input is relevant here? It strikes me as unrelated to how dangerous/hard-to-"kill" something is (sure, I could run without sensory input, but I'm not doin' anything anymore!) - what makes you feel differently (or am I misunderstanding you again?)

> The state is entirely internal to the network and disappears after a token is generated,

Yes and no, but mostly no, at least within a context window.

Mathematically, you could write a single step of LLM decode as a pure function from a list of past tokens to a predicted token (or a distribution over tokens, if you consider sampling separately).

But nobody actually implements this, because each token depends on state computed at past tokens in a way you can reuse.

So, in practice, inference computes a very rich vector of state- at each layer, for each token. And models do indeed use this to plan and track things over time (you can see this in interpretability results, e.g. with linear probes or natural language autoencoders).

> Stop the GPU, hit ctrl-c on the inference server, pull the power plug, cut the ethernet cable, send a kill signal, etc - any of these stop submitting new batches to the GPU and halt execution. That stops tokens from being generated. Stopping a "rogue" LLM is that easy. No input, no output.

This is also true about a human brain. My brain isn't going anywhere- it can't move by itself. It's also easy to kill (without the rest of my body, it dies in minutes!)

However, malicious human brains- especially powerful human brains, like leaders of countries- are often quite difficult to stop, because they're able to control systems that can see, speak, walk, run, fire a weapon, and so on.

One such system is the rest of the body, of course, but there are others (consider a UAV pilot, Perimetr, or a powerful leader who tells other humans what to do).

The brain being squishy doesn't make the thing easy to kill.

Definitely agree with that last paragraph - I don't think anybody involved thinks they're doing an RCT here!

tbc, I'm pretty sure that both an individual subatomic particle and a NAND gate or an adder are not conscious. (At least when I'm not in panpsychist moods!)

But, in the same way that one particle not being conscious doesn't necessarily mean some configurations of many particles- like humans- aren't, one logic gate not being conscious doesn't mean there can't exist configurations of many logic gates that are.

Is a single subatomic particle conscious? Unless you're a panpsychist then surely no, right?

And yet humans, made of such particles, are!

So it seems like reductionism doesn't work here.

I think you'd benefit from explaining why something being "just math" means it can't be intelligent, conscious, or have a personality- it seems like you're sort of assuming that implicitly, could you make an argument for it?

Could you explain how this is related to "consciousness", "intelligence", or "personality"?

Ah, you think by "Luna would eventually get smarter" they meant that the model backing it would, without retraining, somehow improve? I figured they just meant they'd use newer models as they come out.