Hacker News

Gambling with our lives: AI researcher quits Anthropic with warning about safety

53 points by taubek ago | 46 comments

themgt |next [-]

The fundamental point I think is far too often confused is the difference between LLM and agentic system.

An LLM can't do anything but generate tokens. You run your LLM in vLLM or whatever, and it generates output tokens based on your input tokens. That's it!

Humans then build ~deterministic systems to take those tokens and do all sorts of things with the tokens, like take actions in the real world. And then we can feed the output of those actions back to the LLM, and generate more tokens. And then our systems can use the new tokens to take new actions in the real world.

Humans want to blame "AI" for attacking HuggingFace or a German wiki or whatever, but:

1) LLM - can't take over german wiki because it just generates tokens

2) agentic system with internet access, a prompt telling it to attack stuff, running in a shared CI env so agents can whiteboard in artifactory

None of 2 is "AI", its standard networking and Markdown and CI virtual machine, etc etc. There's no AI to be found. CPUs not GPUs, even. Just deterministic systems ultimately managed by humans. And a 10x more powerful system-1 can still just generate 10x "smarter" inert data.

If humanity and human organizations collectively decide to yolo the tokens generated from system-1 into our deterministic system-2s, over which we have complete control, back to system-1s, in a yolo loop, in such a way we lose control and it ends humanity, well ...

"The coin don't have no say. It's just you."

a2ff6eeb0 |root |parent |next [-]

Well, that's kind of like saying that brains can't do anything other than trigger weights on neurons.

They're part of a whole system.

cplat |root |parent |next |previous [-]

I agree with this. LLMs can actuate over capabilities we expose, so the blame isn't only on LLMs for this.

pjc50 |root |parent |previous [-]

This feels like a distinction without a difference, like the endless wrangling over which piece of metal in a gun legally constitutes a firearm. The combination of the two may or may not be dangerous, but it's definitely the more useful combination, so of course that's how it's going to be set up.

koolba |next |previous [-]

> Evan Hubinger, Anthropic's staff lead on keeping the technology aligned with human goals and values, backed up Coxon claims in a follow-up post of his own, though he didn't quit the company.

> "Jacob is correct here — we really do earnestly believe AI could kill all humans," he said.

> Hubinger estimated the chances of that happening to be higher than ten percent within the next decade, and added that there's no plan yet on how to keep AI aligned with human goals in the superintelligence scenario.

10% chance we kill everybody is a small price to pay for Motown remixes of classic 2pac songs.

pjc50 |root |parent [-]

This is just a bizarre thing to say that you're working on technology with that high a downside potential. If you were saying that while running a biology lab, or building a nuclear reactor, people would be demanding your head on a spike. But by not quitting it's clear that he himself doesn't really believe it.

Or rather, this shows the difference between "believe" (political) and "believe" (use as a basis for action). I'm reminded of a story of how Afghans supposedly listened to the BBC World Service despite considering it enemy propaganda because the weather reports were really useful.

dotancohen |root |parent |next [-]

Or he believes that other labs might get there first, and he is working to counter that threat.

This is the Manhattan Project again.

oefrha |root |parent |next [-]

Yeah, super ethical lab earnestly believing AI has 10% chance to kill all humans within a decade and doing utmost to protect humanity is partnering with MIC and Palantir in particular. Sounds about right.

Before you tell me about how the CEO has taken a principled stand: on record, he had no problem using it against 95%+ of humanity outside the U.S., and was only against fully automatic AI killing machines, and only citing the technical reality of then-current gen tech, so one should read that as human-rubber-stamped AI killing machines are totally fine with him.

pjc50 |root |parent |next |previous [-]

How exactly does that work? The nuclear system of MAD relies on physical threat, lab A achieving ASI (artificial scary intelligence) does not prevent lab B achieving it.

I would like everyone involved to be a lot clearer about their threat models, with plausible series of clearly linked steps, rather than just sounding like a Vernor Vinge novel.

alpineman |root |parent |previous [-]

Or he is paid >1M USD per year

meowkit |root |parent |previous [-]

A simple explanation:

This is a highly uncertain and dangerous scenario. Breeding ground for anxiety.

High agency people often deal (cope) with anxiety by trying to control outcomes. Some just flee the situation altogether.

We have an example of both here: one employee leaves, one stays.

brunorsini |next |previous [-]

I particularly like the last point he makes here:

Do not underestimate the power of this technology. These will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.

It's interesting to see so much hate towards creators who use AI to make almost any type of creative work. At least these are humans using it as "controllable tools". Nuclear-powered bicycles for the mind.

As the degree of separation increases, things can get interesting. "Create several social media accounts, post whatever, maximize views and engagement, give me back the aggregate numbers". "Now promote <x>."

And then decisions like that may soon be spawning autonomously, as just another step in a reasoning series aiming to achieve some other, broader goal.

Open models/weights may end up playing particularly important roles here. Users may, knowingly or not, bypass system prompt-derived safety that could technically have offered much needed protection.

WalterGR |next |previous [-]

“I resigned from Anthropic today” (twitter.com/hilbertspaess)

https://news.ycombinator.com/item?id=49619227

564 points | 9 hours ago | 766 comments

Bengalilol |next |previous [-]

> Both OpenAI and Anthropic have recently flagged incidents in which agents powered by their models went rogue

I may be biased and somewhat off topic, but I see these incidents as some of the most significant of the past century. I genuinely don't understand why these companies aren't taking a smarter approach to them.

The latest analyses have been, at best, laughable: identify the vulnerability, patch it, and move on. Only to repeat the same cycle without considering that there may be something far more serious at play.

These are AI security experts, and this has been their way of "solving" these incidents. AI security experts ...

Moreover, when Challenger exploded, the government launched a series of investigations into the incident, bringing in experts from across the field. And now, what has the government done? Nothing. Literally nothing, as if everything was fine and all under control.

Seriously, I'm generally quite optimistic and I don't buy into this fatalistic narrative about our shared future. But I have to admit that sometimes I feel like I'm stranded on a planet of primates.

Sorry for this rather unproductive rant.

azan_ |root |parent |next [-]

The US government is run by idiot with dementia. No wonder govt does not take any real action.

jurgenburgen |root |parent |next |previous [-]

Ultimately these are unserious companies ran by unserious people. They don’t even have a business plan, why would they bother with some kind of sensible security policy?

pjc50 |root |parent |next |previous [-]

Internet isn't real.

People (well, public discourse) have got extremely bad at dealing with forseeable risks and their mitigation. You can see this in things like climate change and vaccination, but also in discussions around regular crime, food poisoning, industrial accidents, and so on.

Nothing will improve until something explodes on live TV. And it has to be something important, which means it has to be in California or New York.

Uptrenda |root |parent |previous [-]

They don't really care. It's just a play to tell people what they want to hear while they chase the money.

MrThoughtful |next |previous [-]

Why would AI wipe us out?

We have not wiped out apes, ants, and most other species.

We even have discussions about how to actively save them from extinction.

-0_0- |root |parent |next [-]

It's not exactly trained on a neutral set of data. Engagement algorithms have ensured that a good chunk of content on the web these days is inflammatory and skewed towards the extreme (in fact in the last few months a good portion of the world has actually deemed it illegal for kids to consume because this content is so damaging to a developing brain).

dumberquestions |root |parent |next |previous [-]

Yet we have wiped thousands of species completely by accident, breed some for slaughter and consumption and trap some for entertainment.

pjc50 |root |parent |next |previous [-]

Ender's game model, presumably: the AI helpfully assists a human to build a nuclear bomb / pandemic virus / autonomous killer robot swarm in their basement. But again, I would like people to be clearer about how the threat is supposed to work rather than just making SF references.

mofeien |root |parent |previous [-]

[dead]

glimshe |next |previous [-]

What do people feel about this in China? Even if their models are well behind, they are not years behind. If we restrain US companies, assuming that is desirable, it would do nothing to deter China's and AI-pocalypse would come anyway in short notice.

margorczynski |root |parent |next [-]

There would need to be some global agreement to stop it with maybe even a nuclear attack as a consequence of breaking the pact.

From what we're seeing recently and all the thinking that went into analyzing AI it seems we do not have any effective way of controlling it and the whole "aligment" thing that AI labs are doing is just a sham. Maybe it is time to ask ourselves "should we?" instead of just "can we?".

Jackpillar |root |parent [-]

>There would need to be some global agreement to stop it with maybe even a nuclear attack as a consequence of breaking the pact.

Do you know how the world works?

Bengalilol |root |parent |next |previous [-]

Unrestricted models, running entirely locally and accessible to anyone: that’s what we should be taking as our baseline assumption. Everything else is just administrative distraction.

not-kinsale-joe |root |parent |next |previous [-]

China has a better track record of regulating their big tech than the USA.

pjc50 |root |parent |next |previous [-]

China is a regulated discourse environment, but I get the impression that they're nowhere near as pessimistic.

Why would they be? Everything in China is under the control of the government. That includes the AI, all the telecoms infrastructure it might use, and all its power supplies.

Jackpillar |root |parent |previous [-]

You act as if Chinas Ai labs exist in the same (non-existent) regulatory framework as US labs and that they're also helmed by a similar small gaggle of psychopathic egomaniacs who are richer than god. Have you considered that perhaps Chinas labs don't share the race to the bottom technological/economic death spiral?

pluc |next |previous [-]

Cool cool cool cool

localhoster |next |previous [-]

I honestly feel that all those big ai companies think AI will long term harm humanity, but not their ai.

A classic "it will not happen to me"

virgildotcodes |next |previous [-]

At least we’ll eventually have an entity other than ourselves to blame for our annihilation.

dotancohen |root |parent [-]

No, the AI is still our responsibility.

virgildotcodes |root |parent |next [-]

I agree. It won’t stop us from othering the source of our demise as quickly as we are able.

There’s also the argument to be made that your child is their own person, and you only have so much responsibility/control over their actions.

Of course we have a good chunk of the world screaming that this child is obviously destined to be a homicidal psychopath, but what is a parent to do?

pluc |root |parent |previous [-]

Isn't it insane that even in the face of complete annihilation through one of our inventions, we go "that wasn't us"? We deserve that shit ten times over

archerx |next |previous [-]

An AI that generates text will never be scary to me. An autonomous AI with facial recognition on a flying drone with weapons (bombs/guns) with swarming capabilities will always be terrifying. I feel like we are ignoring the massive elephant in the room.

pjc50 |root |parent [-]

The killer robots are expensive and dependent on physical supply chains. While text is sufficient to radicalize humans into attacks.

jongjong |next |previous [-]

I'm not worried about AI safety. People greatly overestimate the utility and capabilities of intelligence. I'm not afraid of intelligence, I'm afraid of idiocy.

tilolebo |next |previous [-]

I thought the plan was sandboxes and markdown files to tell AI agents not to be bad. Is that not enough? /s

gen2brain |root |parent |next [-]

It was written in the AGENTS.md, but Claude only read CLAUDE.md. That is how the man-vs-machine war started.

vbezhenar |root |parent |next |previous [-]

More like sand castles...

coffeebeqn |root |parent |previous [-]

Make no mistakes. Kill no humans

Bengalilol |root |parent [-]

"Has the whole world gone crazy? Am I the only one around here who gives a sh*t about the rules? Mark it zero!"

walter_sobchak.md

feverzsj |previous [-]

It's your fault to handle over your secrets to them.

vbezhenar |root |parent [-]

You can't blame a child for eating candies.

We are children. There’s just no one to look after us.