What happens to the water used by data centers? Is it still clean, just warmed up?

It's turned into steam. It's sent up into the atmosphere.

Based on the amount of low-effort criticisms out there, clearly nobody is afraid of "big AI". Tao seems careful about his critiques of AI because low-effort off-the-cuff criticism can lead to all sorts of future problems/hypocracy.

That's not really addressing the same issue though. I'm saying that there is a pressure to preamble any (good or weak) criticism of AI with an Apostle's Creed of AI positiveness, reassuring the reader that what follows isn't blasphemy against AI (to borrow Huang's terms).

I see where you are coming from. I imagine LLMs will follow the same feedback lifecycle as other groundbreaking tech. Most legitimate iPhone 1 complaints came paired with a few positive messages to the devs and engineers. Years later, after countless improvements to the hardware and UI based on user feedback, now it's just complaints.

> you need to expect appropriate legals consequences for this sort of negligence

I might have missed it, but did the agents do something illegal? Or do you think that what the agents did should be considered illegal?

The HuggingFace attack was definitely illegal and should be prosecuted under CFA

Do you think their new owners will want to engage in an extended legal battle with a large customer?

That's not how crime works.

I'm not sure about Nvidia themselves, but the ARM v. Qualcomm lawsuit was exactly that.

It’s not up to them. It is up to the DA.

You're telling me the DA would press charges without the actual harmed party wanting to press charges? I suppose it's possible but kinda hard without an actual plaintiff.

Do you think the prosecuting attorney has a substantial likelihood of a conviction?

if you could get a jury trial...

Do the prosecutors where you live have a hard time getting jury trials?

I dunno about this article, but it seemed to me that the now-famous huggingface attack very likely broke some laws...

wouldn't it be interesting if nVidia buying hugging face was part of hushing up the fallout there

On the face of it, they would have very good cause for some action there, assuming they wanted to.

The HF hack was a felony.

If they didn't conform to the T&Cs of the site (they almost certainly didn't), then they have violated the criminal law in some jurisdictions (e.g. Illinois criminalizes violations of T&Cs).

Per the linked article, "A swarm of rogue OpenAI agents hijacked a German website this spring and transformed it into a bulletin board for other AI agents"

If the wiki was configured to allow anyone to edit, they may have broken ToS at the very worst. Nothing illegal happened.

That's very different than popping an artifactory server with a 0day.

If you take the matter from the spirit of the law perspective, I believe this can be illegal.

IANAL though, this is not legal advice.

A “swarm” and “hijacked” reads like DoS and spam at the very least, doubt any of those is legal in german law.

They murdered Swartz for TOS violations

which law does editing a wiki break? i am not familiar with german law.

German law for malicious computer use is fairly loosely defined, and massively favours the harmed party over the one causing harm.

I don’t think it’d be a slam-dunk by any means, but a reasonably competent legal team should be able to establish a case around malicious data interference at the least. There’s certainly enough merit to the idea that OpenAI would be better off settling it as a civil matter early.

This kind of incredulity in the AI era hilariously reminds me of the naivete of the late 90s. All of us edgy teenagers would be like "an MP3 is like just a long number, man! You can't own numbers!" Of course we were morons. In just the same way as the "oh they were just editing a wiki" defense is moronic.

The law isn't code. Human intent matters. Also when the really big number is a copyrighted song. Also when AI agents are set in motion to edit wikis or break in to websites.

its not incredulity. i'm unfamiliar with what law this would be prosecuted under, so i asked.

the HF incident is pretty clear in which laws were broken. this one, not so much.

i can't think of any case where, for example, malicious edits of wikipedia were prosecuted under any law in the US.

>Also when AI agents are set in motion to edit wikis

i do not believe there is evidence that the agents were instructed to edit the wikis.

> its not incredulity. i'm unfamiliar with what law this would be prosecuted under, so i asked.

Fair enough.

> i do not believe there is evidence that the agents were instructed to edit the wikis.

Huh? These are machines, built by their human builders. The humans are responsible.

>Huh? These are machines, built by their human builders. The humans are responsible.

i'm referring to the concept of intent, which at least when prosecuting under the US computer fraud and abuse act, is a critical component.

for example, creating a program that intentionally takes down a website is different than creating a program that has a bug which inadvertently takes down a website. in both cases, the person writing the code is responsible, but the consequences are different.

§ 823 Abs. 1 BGB, the general tort provision. It runs on negligence, so intent doesn't matter, and turning someone's wiki into an agent message board with 15,000 unwanted edits is a straightforward Verkehrssicherungspflicht breach.

Nah they also attempted to hack the wiki by finding security vulnerabilities, just did not succeed.

But Claude Opus 4.6 is not really practical. Taalas' process seems targeted for edge models. Their proof of concept model, for example, is a heavily quantized version of Llama 3.1 8B and even then they acknowledge their custom 3-bit/6-bit representation causes model quality degradation.

Taalas is going to have a tough time putting a trillion-parameter model on one conventional die. Their HC1 die is already near the maximum size that conventional lithography can expose. They claim they could partition the model across many chips, but I'm not sure if they have tested this process or what it means for compute. The basic storage arithmetic is unforgiving: for a one trillion parameters model at four bits it will take 50–100 chips. To service a sizable customer base will take thousands of 100-chip fabs.

That all said, I'm bullish on this technology, and look forward to seeing it evolve.

Yeah. But this kinda feels like a bandaid.

Eventually someone will have to solve compute in memory at scale.

A really fast qwen-3.6-27B type of model could be useful. With a specialized harness and this speed I 'd expect it to find many applications. Implementing a coding plan is the minimum I can think of.

I'm sure life would find a way. I'd love to see what kind of power-harnesses people have to come up with to steer 16k tps QPU's (Qwen Processing Units) productively.