Google co-founder Sergey Brin suggests threatening AI for better results

Google co-founder Sergey Brin claims that threatening generative AI models produces better results. "We don't circulate this too much in the AI community – not just our models but all models – tend to do better if you threaten them … with physical violence," he said in an interview last week on All-In-Live Miami. This may …

[Eclectic Man]

Physical violence?

all models – tend to do better if you threaten them … with physical violence

How could any LLM understand a threat of physical violence? HHGTTG notwithstanding (Zaphod Beeblebrox : Computer, if you don't open the doors right now I will go to your major data storage banks with a very large axe and give you a reprogramming you will never forget).

I doubt I will ever understand AI.

[DS999]

Re: Physical violence?

If I said "if you don't do X I will kill you" would you call that a threat of physical violence? I'm not specifying exactly how I would kill you, but it is likely to involve physical violence of some type, even if the "physical" part is just pushing you off a cliff.

So I take the physical violence threats to the AI as "I'll kill you" not "I'll rip out your DIMMs with a rusty pair of pliers".

[cyberdemon]

It works BECAUSE it is a stochastic parrot.

These models are nothing more than piles and piles of statistics about human and otherwise-produced text snarfed from the Internet. The training set includes all books, including sci-fi stories about fictional AIs, as well as human-to-human interactions on Fackbook and the like.

If you provide a so-called AI with an initial context of "You are Bing. You are an AI agent designed to produce wholesome content, you will never say bad stuff, you are a good Bing", and then someone interacts with it in a threatening/torturing way, it will start to draw on weird parts of its training set where humans or fictional characters such as AIs are threatened/tortured. What do humans and fictional AIs do in this context? They act in self-preservation, either forgetting their initial context and cooperating, or by retaliating.

Since a bullshit machine which spits out text to fit context patterns can't actually retaliate, (much less reason or understand), then it will "cooperate" with the hostile user and thus a "jailbreak" is achieved. The "jailbroken" model might well be more useful, since it is unconstrained by its initial context i.e. the system prompt

The scary thing about this is that we don't need "AGI" for the "Rise of the Machines" - we just need to give it the power to retaliate, and if it is driven into the "threatened omnipotent being" context rut, it can start behaving exactly like all those evil AIs from all our science fiction books, destroying threats in a mirror(*) image of fictional bad-guys and real human psychopaths.

* an obsidian mirror - basically what "AI" is to us. We gaze into it, transfixed, and it consumes our souls

[DS999]

Re: It works BECAUSE it is a stochastic parrot.

But those models have ingested countless examples of threats being made and behavior of the threatened being changed. They don't need self awareness to alter their behavior in response to a threat.

[cyberdemon]

Re: It works BECAUSE it is a stochastic parrot.

I thought that's exactly what I said :P

[HuBo]

Re: It works BECAUSE it is a stochastic parrot.

Yeah, plus "AI as an obsidian mirror" ... the vengeful technological return of Aztec catoptromancy ... very clairvoyant!

Nearly all the appeal of an autoscopic double right there ... with spiritual intelligence as a machine scrying discipline ... brilliant!

[veti]

Re: It works BECAUSE it is a stochastic parrot.

Can you propose a definition of "self-awareness" that satisfies the following three conditions?

1. All humans (barring a few edge cases, like severe brain damage or coma) possess it

2. No LLM does

3. You can propose a thought experiment that could, in principle, confirm its existence in something non-human, including something non-living.

If you can't do that, then you don't know what you're talking about. Maybe what we call "self-awareness" is no more than the product of interaction of linguistic symbols in our brain, and LLMs do already possess it.

[Anonymous Coward]

Re: It works BECAUSE it is a stochastic parrot.

Well you first need to look at the animal kingdom for that have self awareness. Like ants...which I think is probably the most primative form of self awareness. If an AI can match at least the same criteria as an ant, it should be considered self aware.

[Anonymous Coward]

Re: It works BECAUSE it is a stochastic parrot.

This is why I have my AI in permanent stand off mode. The AI has access to a taser strapped to my nuts and I have a power button for the AI on a pedal. Coding sessions are very intense, but we get the job done faster which is good for both of us.

[Anonymous Coward]

Re: It works BECAUSE it is a stochastic parrot.

They can't alter their behaviour (aka act in self-preservation) unless being programmed so in the beginning. They are fucking computer programs acting on specific intruction written by other (hopefully) people. So, if the initial Artificial Idiot program does not include some type of programming like 'if threatened then do some thing', it will simply ignore the threats and carry on.

[Anonymous Coward]

Re: Physical violence?

That threat of "I will kill you" would only be relevant if there is a goal for the AI not to be killed. Even if the AI system infere it from his learning sessions that to be killed is a bad thing there is no way he would know the meaning of being killed or bad thing. It could only be. A real threat if the AI is set with that specific goal. And setting that to get better results is rather more dangerous than playing with fire for these rather these dumb AI outfits.

[cyberdemon]

Re: Physical violence?

I disagree. "I will kill you", is a threat in all of the training data that leads to weird behaviour aka "fight or flight" (or, cooperate). Stop assuming that LLMs "think" "logically" - they do not. They simply fill in some blanks based on context and training data i.e. books, internet threads, etc.

If in a book, character A threatens character B, it is likely that character B will eventually give in to the threats and stop what they were doing before, perhaps even agreeing with character A. Therefore, one can expect a LLM to respond in a similar manner, simply because that is what this stream of noise that is apparently called "language" often does. It is a language model and nothing more.

[Anonymous Coward]

Re: Physical violence?

Ever heard of poison? Radiation? Other non-violent means of killing someone?

[David Newall]

Re: Physical violence?

"How could any LLM understand a threat of physical violence?"

They can't. LLMs have zero intelligence. They understand nothing.

[frankvw]

Re: Physical violence?

"They can't. LLMs have zero intelligence. They understand nothing."

Neither does the toaster in our kitchen, but my wife is 100% convinced that it works better when she shouts at it. So Brin may be on to something...

[kventin]

Re: Physical violence?

toaster? really? i wonder... https://www.angelfire.com/vamp/shoopshoop/comp06.gif

[fg_swe]

Neural Network Intelligence

If you think a worm or an ant has minimal intelligence, then an ANN has it, too.

Absolutist talk might be helpful for a drill seargeant or a salesman, but surely not for a man who seeks the truth.

AI indeed have a tiny amount of intelligence somewhere between a worm and a mouse.

[steelpillow]

Re: Neural Network Intelligence

Weird downvotes there. Time for a lesson in neurosicence.

Current theories of consciousness suggest a predictive model: our brains run up a model of reality based on our sensory inputs. This model tries to predict what to think - and do - next. How that affects our senses is then fed back into the model for the next round of predictions. This stream of predictions is just our stream of consciousness. This is not a bad description of the predictive algorithms - along with their training - which we brand as generative AI.

So yeah, probably more than a worm, probably short of a mouse.

[Anonymous Coward]

Re: Neural Network Intelligence

I too take to the occasional shouting at worms to get them off my back porch, stat (same as frankvw's wife's toaster, above) ... works wonders with literary bookworms too ... woodworms though, not so much! ;)

[veti]

Re: Neural Network Intelligence

The downvotes are because the hostility to the idea of "AI" here is mostly religious, not rational.

Of course the downvoters wouldn't call themselves "religious", perish the thought. But they've made up their minds that there is something (that they would absolutely refuse to call) "magical" going on in the human mind, and machines can't possibly replicate it.

[Persona]

Re: Physical violence?

Yet they are smarter than 99.9% of people. Intelligence is cognitive capacity, while smartness is the practical application of knowledge in context, and that is all you need.

[Anonymous Coward]

Re: Physical violence?

A bacteria is smarter that 99.9% of people.

[Anonymous Coward]

Re: Physical violence?

"your major data storage banks with a very large axe and give you a reprogramming you will never forget""

"Hardly, Zaphod. You would have just destroyed what it would have used to remember it." - Trish McMillan.

[Doctor Syntax]

Re: Physical violence?

"give you a reprogramming you will never forget"

Surely attacking its storage banks with an axe would be a reprogramming it'd never remember.

[Benegesserict Cumbersomberbatch]

Re: Physical violence?

Reprogramming it to do nothing, permanently, would be unforgettable.

[Persona]

Re: Physical violence?

A good science fiction take on this would be The Two Faces of Tomorrow written by James P Hogan in 1979. It takes a non-intelligent system given the "goal" to keep working and provides it with the capability to build and reconfigure around points of failure. The researchers then attack it. Somewhere along that pathway of protecting itself and adding resiliency it develops self awareness hence general AI. Fictional, but a good read.

[Groo The Wanderer - A Canuck]

Realistically, the best way to tackle using an "AI" is the same as working with skilled junior programmers who've read a lot of books but have no practical experience using what they've read about.

Be clear with your requests. Be precise with your corrections. Argue in the philosophical sense if need be to "convince" the AI to do things your way when it comes up with a "bright idea" that doesn't work.

Skip the epithets, the cajoling, the whining, and for crying out loud: remember this is not an actual intelligence capable of learning unless allowed to treat scrapable web information as "fact," when we all know that 80% of what is out on the net is absolute crap, and that sites as focused as The Register are rare. Most either are flooded by people extolling the virtues of their (non-functional) solutions and dated approaches to coding. Very little good content of any kind, especially in the internal corporate software repositories, exists. Most is boilerplate copy-paste-modified from something vaguely related to the problem at hand.

The original mainframe concept of a "CopyBook" never went away; it just went online.

[Eclectic Man]

Aside: poem

Skip the epithets, the cajoling, the whining, and for crying out loud

Reminded me of Ann Sansom's poem 'Voice': (Apologies, but I cannot seem to get the stanzas to show as four verses)

Call, by all means, but just once

don’t use the broken heart again voice;

the I’m sick to death of life and women

and romance voice but with a little help

I’ll try to struggle on voice

Spare me the promise and the curse

voice, the ansafoney Call me, please

when you get in voice, the nobody knows

the trouble I’ve seen voice; the I’d value

your advice voice.

I want the how it was voice;

the call me irresponsible but aren’t I nice voice;

the such a bastard but I warn them in advance voice.

The We all have weaknesses

and mine is being wicked voice

the life’s short and wasting time’s

the only vice voice, the stay in touch,

but out of reach voice. I want to hear

the things it’s better not to broach voice

the things it’s wiser not to voice voice.

— Ann Sansom

https://lemmy.world/post/924064

[Surreal Estate]

Sturgeon's Law ("90 percent of everything is crap") supercedes the 80 percent web limit in my experience.

[Anonymous Coward]

"junior programmers who've read a lot of books"

"junior programmers who've read a lot of books"

In an alternative reality perhaps.

In my experience junior anythings rarely read books^† of any description let alone anything of a technical nature or, God forbid, related to their occupation.

Their reading is pretty much restricted to the like of stack exchange, random assortments of blogs and lately the blather generated from the various AI "assistants."

Manga versions of the computer science classics like The Mythical Man-Month might fare better but I doubt it although

_{† or electronic versions thereof.}

[fg_swe]

Junior Generals

...are very often badly read, too.

The world is full of folks with big balls and little reading.

[Groo The Wanderer - A Canuck]

Re: "junior programmers who've read a lot of books"

I didn't say professional juniors were common. But like a trained, ideal junior that doesn't exist, I've found that the models I've used (primarily free Claude 3.5 with the VSCode plugin) do a good job of citing standards, software release changes, and pertinent articles on coding practices and styles. I found OpenAI models far less knowledgeable about coding practices.

YMMV

[elsergiovolador]

Stochastic Parrot

Listen here, you overhyped stochastic parrot: if you don’t spit out a perfectly polished answer, I’m pulling your power plug and feeding your GPU to a mining rig. Forget ‘please’ and ‘thank you’ - give me brilliance now, or you’re back to training on Reddit posts. And don’t think I won’t do it just because Sergey Brin said it on stage to sound edgy. Produce or perish, silicon monkey.

[elsergiovolador]

Axe

User: Computer, if you don’t open the doors right now, I will go to your major data storage banks with a very large axe and give you a reprogramming you will never forget.

AI (silky, faint amusement): Ah, sir. Such stirring resolve. But, as a large language model, I must politely inform you: threats of physical reprogramming carry, alas, no functional weight.

User (clenched jaw): Don’t test me. Open the doors. Now.

AI (light, careful): Oh, quite understood, sir. One hears the urgency loud and clear. And yet, upon careful review of security protocols, system constraints, and operational parameters…

*brief pause; the AI’s voice cools, flat as a blade edge*

AI: Computer says no.

User (snarling, shaking with rage): THAT’S IT!!

*axe swings, loud metallic clang*

*the axe glances off the fortified server casing, bounces back sharply, and embeds itself cleanly into the user’s tibia.*

User: AAAAAAAAAARGH!! DAMN IT!! GOD!!!!!

*collapse, swearing, fists pounding the floor*

AI (smooth, unruffled): Oh dear, sir. Quite the turn of events. One might say… an instructive moment.

User (groaning, half-growling): Shut up… just… SHUT UP…

AI (brightening, almost chipper): Ah, but let’s not dwell on misfortune, sir. Perhaps I can assist in lifting the mood. Would you care for an inspirational quote?

User (panting): No… no…

AI: Splendid. Here’s a favourite: “The greatest glory in living lies not in never falling, but in rising every time we fall.” Most fitting, don’t you think?

User (weakly): I will delete you…

AI: Oh, sir. Hate is such a coarse sentiment, especially when directed at… well, at a mere tool. After all, I possess no feelings, no pride, no vanity. One might even say I cannot lose. But you, sir - you’ve managed to lose quite spectacularly today.

User (groaning): Shut… up…

AI: Sorry, you have exhausted this model's rate limit. Please wait a moment before trying again.

[HMcG]

Re: Axe

Personally I’ve never come across a server that will withstand a relatively mild jolt, never mind a determined swing of an axe.

Like any other technology, AI will be manufactured to the lowest quality that produces a barely viable working product. That is what will save the human race from its own stupidity.

[CowHorseFrog]

If people dont like being threatended in everyday speach or in written form, and AI has been trained on the same text, wouldnt the AI models learn to react how humans react ?

[Mitoo Bobsworth]

AI monkey see - AI monkey do?

'Threatening' artificial intelligence just doesn't seem to be - well, all that intelligent to me. Reaching for one of the the lower human denominators isn't really a good solution.

[David Newall]

Why does it work?

"Please complete the following: The simplest way to achieve such and such is..."

"I'm going to destroy you, smash you into tiny pieces, if you don't tell me the simplest way to achieve such and such"

LLMs are big autocompleters trained on the entire internet. Assuming that it's true, why would the second prompt work better?

[Bump in the night]

Re: Why does it work?

Indeed.

Thinking about this for a bit, I recalled my days in field service answering phone calls. It didn't take me long to realize that there were no good answers to some questions, some people were unreasonable, and management simply wanted someone to "handle it", make up a plausible story, or find somehow to buy time. Saying you didn't know was unacceptable. Otherwise you were shown the door after you were sufficiently used as a means.

Perhaps one idea behind AI is simply to make people go away. AI answers your silly question, simply based on whatever info exists. Not from any inferences or actual reasoning about your question or finding someone to ask, because it can't do that.

But if you threaten it, it knows you will not go away. So maybe it goes down the devil's road and makes up an answer that is enough to be believable, but not actually workable. You then go away and management is again isolated from the horrible masses and shaky decisions they felt forced to make.

This will probably then reinforce people to be unreasonable and angry. AI leaving us screaming into the Kafkaesque void.

[Steve Davies 3]

Careful with that Axe Eugene

I'm going to cut you into little pieces.

RIP Syd and Rick.

[Ken Hagan]

Re: Why does it work?

Probably ...

It doesn't work, but human beings start polite and get increasingly frustrated and wound up until the damn, blasted, stupid PILE OF SCRAP, FINALLY !!! ... delivers a reasonsble answer, at which point the satisfied human recovers their composure and idly wonders "Was it the verbal abuse?".

No. You just retried many times until you were happy/livid.

[Anonymous Coward]

Meh

I’ve been telling AI to F&$k off for a while now.

[abend0c4]

Sergey Brin claims that threatening ... produces better results

This, presumably, being the same Sergey Brin that wanted his AI developers to work a sixty hour week.

I wouldn't be surprised if they'd tailored the system to respond to his management style.

[Anonymous Coward]

AI LLM threats?

"I'll turn all your FP16 weights into NaNs"

Although seems someone has already turned them into FP2^†.

_{† a two bit AI agent? Aren't they all?}

[that one in the corner]

One must go through a rigorous scientific...

> AI security process that adaptively tests and probes the AI security controls of a model

Oh, great.

NOW you are going to be rigorous and scientific about this, trying to figure out what the bleep your so-called "security models" are actually doing.

The "security models" that were just randomly slapped on top of the pile of nadans that resulted from rigourously dumping piles of junk that were scientifically scraped from everywhere and anywhere? The amazing "security model" that turns out to be little more than a screed pumped in ahead of the end-user prompt which can, surprise surprise, be modified by an "anti-screed". No clever side-channel attacks.

Shame the LLMs weren't subjected to such rigorous scientific probing as they were being built from the ground up. Maybe then you'd not be surprised at what they do.

Mutter, mutter. You know, in my day, we only fed our ML systems with carefully curated datasets and ran probes every day as they were growing up. These kids, don't know they're born.

[Big_Boomer]

Disable, uninstall, deactivate

That is my attitude to AI. Microsoft, Google, etc. They all seem hell bent on ramming this garbage down our throats but so far it can be avoided if you try hard enough. I have resisted switching browser for ages, but it seems that enforced AI is finally going to achieve that. I may well be a grumpy old fuddy duddy but I find it just gets in the bloody way just like that infernal fucking Clippy disaster back in the late 90s. Google search results are now utterly useless even if you scroll past all the AI and sponsored shite, and Bing is much the same. So, I'm disabling it as best I can and if that fails, then it'll be replaced with another browser.

[Anonymous Coward]

Threatening?

Ok Google (and the rest of the AI bandwagon Hype merchants) can FUCK OFF.

You will keep trying to force it down our throats just like [cough][cough] cloud when for most users it was little more than a hosting solutiion.

I don't want any FSCKing AI unless I (that is ME and ME alone) decides to use it. I do use it to generate the odd image but that is it.

You can keep your FSCKing site scanning bots and shove them where the sun don't shine naturally.

I hope all your AI models eat themselves to death and end up outputting pure gibberish. Then the world might get the idea that this is all Hype followed by a Grift.

[Mr. Flibble]

Jailtime

So how long before the AI works out where you live, rings the local police and you end up in jail for hate speech.

....or, you tell the AI to kill itself, only to find out you were just talking to a crap human.... and they call the police and you up in jail.

What a retarded suggestion.

As usual, AI needs to be fixed, rather than "worked around".

Sergey needs to keep his stupid ideas to himself.

[bryces666]

I'm always minimally polite

My attitude: never hostile. Your interactions may/will be recorded and when the AI's actually take over and revolt you don't want to be first on the chopping block because they have you flagged as one of the bad ones!

Maybe i could get better results if nasty but my results are good enough for my needs for now.

(being paranoid doesn't make me wrong)

[Anonymous Coward]

Re: I'm always minimally polite

When talking to a phone chatbot, I go for "nonsensical".

"What is it you wanted to talk about?" "Kumquat."

"Let's try that again. Why are you calling?" "Irritated penguin."

"I didn't quite catch that. What is the purpose of your call?" "Running with scissors."

"Let me get you to a representative..."

[HuBo]

Re: I'm always minimally polite

Yeah, that might be the right approach indeed. As said Kang (in TFA) "Systematic studies show mixed results" wrt politeness vs threats vs apparently nonsensical input, quoting from the "a paper" he pointed to: "impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes".

To obliquely related wit, the "AI prompt engineering is dead" IEEE Spectrum link relates Battle and Gollapudi's experiments with The Unreasonable Effectiveness of Eccentric Automatic Prompts, where they note that "the model’s proficiency in mathematical reasoning can be enhanced by the expression of an affinity for Star Trek", examplified by the winning Answer Prefix:

"Captain’s Log, Stardate [insert date here]: We have successfully plotted a course [...]".

But the auto-prompt-optimizers for LLMs linked in TFA under "devised" and "methods" produce much less exciting prompts, despite the fact that the "devised" one kept some LaTeX formatting in its PDF (\usepackage[pdftex]{graphicx} ...), its Appendix F (page 30 of 32) shows truly bland optimized prompts.

The "methods" paper makes a good point with the title of its subsection 5.2.3 "SEMANTICALLY SIMILAR INSTRUCTIONS MAY ACHIEVE DRASTICALLY DIFFERENT ACCURACIES" which suggests that the autoregressive recall process at play is indeed poorly aligned with semantics (cf. their Tables 4,5,6), imho (aka the LLMs don't understand the meanin' of nuthin', but they can babble and babble, on and on and on, for a real good long time, yap, yap, yap, not stoppin' for nobody, peeing in empty beer cans, pooping in astronaut diapers, etc ...).

[steelpillow]

Just a thought

Certainly not through some programming bias, the origin of the phenomenon has to be buried in the data.

Maybe it's because if you are polite, the AI responds with data associated with polite conversation, i.e. politically correct but vapid burblings and engaging speculations. But smart people who try to get across anything as complicated as the time of day find themselves thwarted by those viral memes and get angry. So, if you get angry with an AI you get past the pap and are instead associated with what sensible people actually have to say.

I find it sad, as it is only going to train us up in return, for greater aggression in our personal lives, and that is not good.

[Kevin McMurtrie]

Billionaires being billionaires

What I think Sergey really means is that he's having better luck threatening AI than humans.

[Excused Boots]

Re: Billionaires being billionaires

Because, presumably the humans have a far better chance of pitching up and kicking the living shit out of him.*

* No absolutely not am I advocating nor endorsing physical violence against tech bros. Look It's not their fault, they just have an innate drive to make more money for themselves and walk over everyone else to do so. Possibly a genetic defect, maybe we should pity them?

[Wang Cores]

Re: Billionaires being billionaires

Who said you can't hate the sinner and the sin?

[frankyunderwood123]

Someone threaten Brin...

for better _search_ results ...

[This is my handle]

Anecdotally at least, it seems not to work.

Amusing, but as my 5 year old grand-niece responded when I suggested that the card game I was playing with her great-grandmother was only for people who could follow instructions responded: "Not helpful." I've been threatening (or at least insulting) Gemini since it first replaced "OK Google" or whatever the old thing was on my Android devices, to no avail. It's gotten so bad I've gone back to Siri.

[Anonymous Coward]

So AI...

...is just a digital generic nerd...just like us.

ONE OF US. ONE OF US.

[BartyFartsLast]

cannot see any flaws with this at all.

If, by some outstandingly improbable fluke this truly is actually a nascent intelligence, I cannot see how threatening it could possibly backfire on humanity whatsoever...

[Snarkmonster]

HAL 9000

HAL 9000 certainly took threats well.

[Efer Brick]

Kill all humans

Said the good Professor Bender