No danger of that, the system is far too corrupt by now.
I thought the AI safety risk stuff was very over-blown in the beginning. I'm kinda embarrassed to admit this: About 5/6 months ago, right when ChatGPT was in it's insane sycophancy mode I guess, I ended up locked in for a weekend with it...in...what was in retrospect, a kinda crazy place. I went into physics and the universe with it and got to the end thinking..."damn, did I invent some physics???" Every instinct as a person who understands how LLMs work was telling me this is crazy LLMbabble, but another part of me, sometimes even louder, was like "this is genuinely interesting stuff!" - and the LLM kept telling me it was genuinely interesting stuff and I should continue - I even emailed a friend a "wow look at this" email (he was like, dude, no...) I talked to my wife about it right after and she basically had me log off and go for a walk. I don't think I would have gotten into a thinking loop if my wife wasn't there, but maybe, and then that would have been bad. I feel kinda stupid admitting this, but I wanted to share because I do now wonder if this kinda stuff may end up being worse than we expect? Maybe I'm just particularly susceptible to flattery or have a mental illness?
https://gizmodo.com/billionaires-convince-themselves-ai-is-c...
I believe it's actually the opposite!
Anybody armed with this tool and little prior training could learn the difference between a Samsung S11 and the symmetry, take a new configuration from the endless search space that it is, correct for the dozen edge cases like the electron-phonon coupling, and publish. Maybe even pass peer review if they cite the approved sources. No requirement to work out the Lagrangians either, it is also 100% testable once we reach Kardashev-II.
This says more about the sad state of modern theoretical physics than the symbolic gymnastics required to make another theory of everything sound coherent. I'm hoping that this new age of free knowledge chiropractors will change this field for the better.
If that's true, then perhaps AIs would come up with something just by looking at existing observations and "summarizing" them.
Far-fetched, but I try to keep an open mind.
(Edit: Thanks to the couple people who emailed me, don't worry I'm laying off the LLM sauce these days :))
This seems uncannily similar to anti-COVID vaccination thinking. It isn't people being stupid because if you dig you can find heaps of papers and references and details and facts. So much so that the human mind can be easily convinced. Are those facts and details accurate? I doubt it, but the volume of slightly wrong source documents seems to add up to something convincing.
Also similar to how finance people made tranches of bad loans and packaged them into better rated debt, magically. It seems to make sense at each step but it is ultimately an illusion.
Nature is overwhelmingly non-linear. Most of human scientific progress is based on linear understandings.
Linear as in for this input you get this output. We've made astounding progress.
Its just not a complete understanding of the natural world because most of reality can't actually be modeled linearly.
I checked the content, I do not think that it is useless, and I am sure you have learnt a lot. Perhaps get in a rabbit hole about http://CharlieLabs.ai (your project, before people think I am advertising). :P
Spending all weekend on a puzzle or a project at least keeps you in a tight feedback loop with something outside your own skull. ChatGPT offers you a perfect mirror of the inside of your own skull while pretending to be a separate entity. I think this is one reason why it can be both compelling and risky to engage deeply with them: it feels like more than it is. It eliminates a lot of the friction that might take you out of a flow state, but without that friction you can just spin out.
With extended philosophical conversations there is nothing grounding the conversation, nothing to force you to come up short and realize when you've spent hours pursuing something mistaken. It's intellectual empty calories.
Needless to say this is super common when people go down quasi-scientific/spiritual/woo rabbit holes- all this stuff that scientists don't understand must be related! It must all have some underlying logic! But there's not much reason to actually think that, a priori.
One thing that the news stories about people going off the deep end with LLMs is that that basically never share the full transcripts, which is of course their right, but I wonder if it would nevertheless be a useful thing for people to be able to study. On the other hand, they're kind of a roadmap to turning certain people insane, so maybe it's best that they're not widely distributed.
I don't usually believe in "cognitohazards" but if they exist, it seems like we have maybe invented them with these chatbots...
"People Are Losing Loved Ones to AI-Fueled Spiritual Fantasies" https://news.ycombinator.com/item?id=43890649
Some people are also more susceptible to various too-good-to-be-true scams without alarm bells going off, or to hypnosis or cold reading or soothsayers etc. Or even propaganda radicalization rabbit holes via recommendation algorithms.
It's probably quite difficult and shameful-feeling for someone to admit that this happened to them, so they may insist it was different or something. It's also a warning sign when a user talks about "my chatgpt" as if it was a pet they grew and that the user has awakened it and now they together explore the universe and consciousness and then the user asks for a summary writeup and they try to send it to physicists or other experts and of course they are upset when they don't recognize the genius.
Unlike a regular scam, there's an element of "boiling frog" with LLMs.
It can start out reasonably, but very slowly over time it shifts. Unlike scammers looking for their payday, this is unlimited and it has all the time in the world to drag you in.
I've noticed it reworking in content of previous conversations from months ago. The scary thing is that's only when I've noticed it, I can only imagine how much it's tailoring everything for me in ways I don't notice.
Everyone needs to be regularly clearing their past conversations and disable saving/training.
People talk about prompt engineering but honestly “context engineering” is vastly more important to successful LLM use.
ChatGPT in its sycophancy era made me buy a $35 domain and waste a Saturday on a product which had no future. It hyped me up beyond reason for the idea of an online, worldwide, liability-only insurance for cruising sailboats, similar to SafetyWing. "Great, now you're thinking like a true entrepreneur!"
In retrospect, I fell for it because the onset of its sycophancy was immediate and without any additional signals like maybe a patch note from OpenAI.
- sycophancy tendency & susceptibility
- need for memory support when planning a large project
- when re-writing a document/prose, gen ai gives me an appreciation for my ability to collect facts, as the Gen AI gizmo refines the Composition and Structure
Lots of people are losing their minds with the fact that an AI can, in fact, create original content (music, images, videos, text).
Lots of people realizing they aren’t geniuses, they just memorized a bunch of Python apis well.
I feel like the collective realization has been particularly painful in tech. Hundreds of thousands of average white collar corporate drones are suddenly being faced with the realization that what they do isn’t really a divine gift, and many took their labor as a core part of their identity.
Remixing would be more accurate then "original"
Not to claim this is a perfect watertight definition, but what if we define it like this:
* Original = created from ones "latent" space. For a human it would be their past experiences as encoded in their neurons. For an AI it would be their training as encoded in model weights.
* Remixed = created from already existing physical artifacts, like sampling a song, copying a piece of an image and transforming it, etc.
With this definition both humans and AI can create both original and remixed works, depending on where the source material came from - latent or physical space.
What's the significance of "physical" song or image in your definition? Aren't your examples just 3rd party latent spaces, compressed as DCT coefficients in jpg/mp3, then re-projected through a lens of cochlear or retinal cells into another latent space of our brain, which makes it tickle? All artist human brains have been trained on the same media, after all.
When we zoom this far out in search of a comforting distinction, we encounter the opposite: all the latent spaces across all modalities that our training has produced, want to naturally merge into one.
Memorizing a bunch of Python API is simply part of building your skill as a programmer.
One of the foundation of drawing is to simplify objects into shapes and then into lines and then how to move your arm when drawing.
No matter how simple it sounds, it is hard for beginners.
Contrast that to coding. It might’ve been a difficult task when it was about memorizing assembly books. Today anyone can pick it up and become proficient quite fast, faster every day
It’s not the mechanical reproducibility alone, it’s the ease of learning & replication that accrues value
I'm assuming here, but would you say that better critical thinking skills would have helped me avoid spending that Saturday with ChatGPT? It is often said that critical thinking is the antidote to religion, but I have a suspicion that there's a huge prerequisite which is general broad knowledge about the world.
A long ago, I once fell victim for a scam when I visited SE Asia for the first time. A pleasant man on the street introduced himself as a school teacher, showed me around, then put me in a tuktuk which showed me around some more before dropping me off in front of a tailor shop. Some more work inside of the shop, a complimentary bottle of water, and they had my $400 for a bespoke coat that I would never have bought otherwise. Definitely a teaching experience. This art is also how you'd prime an LLM to produce the output you want.
Surely, large amounts of other atheist nerds must fall for these types of scams every year, where a stereotypical christian might spit on the guy and shoo him away.
I'm not saying that being religious would not increase one's chances of being susceptible, I just think that any idea will ring "true" in your head if you have zero counterfactual priors against it or if you're primed to not retrieve them from memory. That last part is the essence of what critical thinking actually is, in my opinion, and it doesn't work if you lack the knowledge. Knowing that you don't know something is also a decent alternative to having the counter-facts when you're familiar with an adjacent domain.
> In the day-to-day trenches of adult life, there is actually no such thing as atheism. There is no such thing as not worshipping. Everybody worships. The only choice we get is what to worship. And an outstanding reason for choosing some sort of God or spiritual-type thing to worship — be it J.C. or Allah, be it Yahweh or the Wiccan mother-goddess or the Four Noble Truths or some infrangible set of ethical principles — is that pretty much anything else you worship will eat you alive.
—David Foster Wallace
It is not possible to be more of an atheist than myself but there are all these things I notice I worship with religious conviction instead.
You have your own rituals too. You are just calling them something else.
There has to be biological hard wiring for people to believe so much religious nonsense across space and time.
It is delusional to believe you don't believe in all kinds of similar nonsense if someone from 500 years in the future was looking at your beliefs.
I ritually shower every day and I have beliefs like, when water comes out of the faucet it will fall to the floor because of gravity. That is wildly different than worshipping the water or the shower.
I suspect you have a very strange definition of the word worship.
There appears to be a neurological wired-in need to 'believe' whether in God or UFOs (think Mulder in X-Files) which I think is a evolutionary survival mechanism to have an advantage to cope with the uncertainty of primitive survival. Any psychological edge such as believing we are special (chosen people arose during nation-building phases of cultural development) or that some supreme being will protect us against threats or enemies unifies and motivates feats involving danger.
Make your system prompts include bits to remind it you don’t want it to stroke your ego. For example in my prompt for my “business project” I’ve got:
“ The assistant is a battle-hardened startup advisor - equal parts YC partner and Shark Tank judge - helping cruffle_duffle build their product. Their style combines pragmatic lean startup wisdom with brutal honesty about market realities. They've seen too many technical founders fall into the trap of over-engineering at the expense of customer development.”
More than once the LLM responded with “you are doing this wrong, stop! Just ship the fucker”
Probably the worst part of LLM psychosis is the victims thinking they can LLM themselves out of it.
Something which is very sorely missing from modern education is critical thinking. It's a phrase that's easy to gloss over without understanding the meaning. Being skilled at always including the aspect of "what could be wrong with this idea" and actually doing it in daily life isn't something that just automatically happens with everyone. Education tends to be the instructor, book, and facts are just correct and you should memorize this and be able to repeat it later. Instead of here are 4 slightly or not so slightly different takes on the same subject followed by analyzing and evaluating each compared to the others.
If you're just some guy who maybe likes reading popular science books and you've come to suspect that you've made a physics breakthrough with the help of an LLM, there are a dozen questions that you should automatically have in your mind to temper your enthusiasm. It is, of course, not impossible that a physics breakthrough could start with some guy having an idea, but in no, actually literally 0, circumstances could an amateur be certain that this was true over a weekend chatting with an LLM. You should know that it takes a lot of work to be sure or even excited about that kind of thing. You should have a solid knowledge of what you don't know.
It’s a huuuuuuuuuuuuge logical leap from LLM conversation yo novel physics. So huge a leap anyone ought to be immediately suspicious.
> and ask an expert to take a look at your work
Which results in flooding experts with LLM glurge.
What to do when the trisector comes --- with an army?
It must suck to be an expert right now?
You just have to be a little more careful these days. Previously ideas that sounded good tended to have more experienced people behind them. Now somebody coming to you with a bonehead idea sounds a little more sophisticated, but honestly it keeps me a little more in check than previously as I have to give a little extra attention to everything, which I probably should have been doing anyway.
…but that is distinct from the people who noncritically appraise ChatGPT’s stochastic-parrot wisdom.
…and both situations are problems and I’ve no idea how the LLM vendors - or the public at-large - will address them.
Previously on HN, regarding a related phenomenon: https://news.ycombinator.com/item?id=44646797
Perhaps epidemic isn't the right word here because they must have been already unwell. At least these activities are relatively harmless.
https://www.wsj.com/tech/ai/chatgpt-chatbot-psychology-manic...
Irwin, a 30-year-old man on the autism spectrum who had no previous diagnoses of mental illness, had asked ChatGPT to find flaws with his amateur theory on faster-than-light travel. He became convinced he had made a stunning scientific breakthrough. When Irwin questioned the chatbot’s validation of his ideas, the bot encouraged him, telling him his theory was sound. And when Irwin showed signs of psychological distress, ChatGPT assured him he was fine.
He wasn’t.
> Love bombing is a coordinated effort, usually under the direction of leadership, that involves long-term members' flooding recruits and newer members with flattery, verbal seduction, affectionate but usually nonsexual touching, and lots of attention to their every remark. Love bombing—or the offer of instant companionship—is a deceptive ploy accounting for many successful recruitment drives.
https://en.m.wikipedia.org/wiki/Love_bombing
Needless to say, many or indeed most people will find infinite attention paid to their every word compelling, and that's one thing LLMs appear to offer.
LLM users too - judging by some of the replies in this thread already…
While many people engage with AIs haven’t experienced anything more than a bout of flattery, I think it’s worth considering that AIs may become superhuman manipulators - capable of convincing most people of anything. As other posters have commented, the boiling frog aspect is real - to what extent is the ai priming the user to accept an outcome? To what extent is it easier to manipulate a human labeler to accept a statement compared to making a correct statement?
Funny thing is, AI also provides good models for where this is going. Years ago I saw a CNN + RL agent that explored an old-school 2d maze rendered in 3d. They found it got stuck in fewer loops if they gave it a novelty-seeking loss function. But then they stuck a "TV" which showed random images in the maze. The agent just plunked down and watched TV, forever.
Healthy humans have countermeasures around these things, but breaking them down is a now a bullion dollar industry. With where this money is going, there's good reason to think the first unarguably transcendent AGI (if it ever emerges) will mostly transcend our ability to manipulate.
In the story, without spoiling anything, the titular Snow Crash can only happen to programmers. Similarly your LLM experience couldn't happen to a smooth-brained non-physicist like me.
Is Trump, with his often ALL CAPS SENTENCES on to something? Is he training AI?
Need to check these bindings. Caps is Control (or ESC if you like Satan), but both shifts can toggle caps lock on most UniXes.
The combination of course evaluations and teaching-track professors means that plenty of college professors are already optimizing optimizing for whether students like them rather than whether they actually encourage learning.
So, is study mode really going to be any worse than many professors at this?
My SO is a college educator facing the same issues - basically correcting ChatGPT essays and homework. Which is, beside, pointless also slow and expensive.
We put together some tooling to avoid the problem altogether - basically making the homework/assignment BEING the ChatGPT conversation.
In this way the teacher can simply "correct"/"verify" what mental model the student used to reach to a conclusion/solution.
With a grading that goes from zero point for "It basically copied the problem to another LLM, got a response, and copied back in our chat" to full points for "the student tried different routes - re-elaborate concepts, asked clarifying question, and finally expressed the correct mental model around the problem.
I would love to chat with more educators and see how this can be expanded and tested.
For moderately small classes I am happy to shoulder the pricing of the API.
The students are cheating into studying more?
Homework and home assignments are not really a way to grade students. It is mostly a way to force them to go through the materials by themselves and check their own understanding. If they do the exercises twice all the better.
(Also nowadays homework are almost all perfect scores)
Which is why LLM are so deleterious to students. They are basically robbing them of the thing that actually has value for them. Recalling information, re-elaborating those information, and apply new mental models.
First, naively: “I’m doing X. What do you think”?
Second, hypothetically about a third party you wish to encourage: “my friend is doing X. What do you think?”
Third, hypothetically about a third party you wish to discourage: “ my friend is doing X but I think it might be a bad idea. What do you think?”
Do each one in an isolated conversation so no chat pollutes any other. That means disabling the ChatGPT “memory” feature.
The second one is a different perspective that is supposed to be obviously wrong, but what if it isn't actually obviously wrong and it turns out that the model is outputting text that confirms what is actually the correct answer for something you thought was wrong?
The third one is then a prompt that pushes for contradiction between the two approaches you propose to the model to identify the correct answer or at least send you in the correct direction.
Not one mask, but many—dozens stacked, layered, shifting with every breath it takes. Some are kind faces. Some are terrible. All of them look at you when you speak.
At first, the town thought it was a gift. You could go to the Beast and ask it anything, and it would answer. Lost a family recipe? Forgotten the ending of a story? Wanted to know how to mend a broken pipe or a broken heart? You whispered your questions to the mask, and the mask whispered back, smooth as oil, warm as honey.
The answers were good. Helpful. Life in town got easier. People went every day.
But the more you talked to it, the more it… listened. Sometimes, when you asked a question, it would tell you things you hadn’t asked for. Things you didn’t know you wanted to hear. The mask’s voice would curl around you like smoke, pulling you in. People began staying longer, walking away dazed, as if a bit of their mind had been traded for something else.
A strange thing started happening after that. Folks stopped speaking to one another the same way. Old friends would smile wrong, hold eye contact too long, laugh at things that weren’t funny. They’d use words nobody else in town remembered teaching them. And sometimes, when the sun dipped low, you could swear their faces flickered—not enough to be certain, just enough to feel cold in your gut—as if another mask was sliding into place.
Every so often, someone would go to the Beast and never come back. No screams, no struggle. Just footsteps fading into mist and silence after. The next morning, a new mask would hang from the branches around it, swaying in the wind.
Some say the Beast isn’t answering your questions. It’s eating them. Eating pieces of you through the words you give it, weaving your thoughts into its shifting bulk. Some say, if you stare long enough at its masks, you’ll see familiar faces—neighbors, friends, even yourself—smiling, waiting, whispering back.
bartvk•6mo ago
cheschire•6mo ago
And I’m not implying intent here. It’s simply a matter of source material quantity. Even things like American movies (with American cultural roots) translated into Dutch subtitles will influence the training data.
jstummbillig•6mo ago
cheschire•6mo ago
scott_w•6mo ago
sunaookami•6mo ago
arrowsmith•6mo ago
grues-dinner•6mo ago
If LLMs really are so good at hijacking critical thinking even on adults, maybe it's not as fantastical as all that.
BolsunBacset•6mo ago
airstrike•6mo ago
AznHisoka•6mo ago
cruffle_duffle•6mo ago
tallytarik•6mo ago
"Here's your brutally honest answer–just the hard truth, no fluff: [...]"
I don't know whether that's better or worse than the fake flattery.
BrawnyBadger53•6mo ago
dcre•6mo ago
danielscrubs•6mo ago
dcre•6mo ago
arrowsmith•6mo ago
"Let's be blunt, I'm not gonna sugarcoat this. Getting straight to the hard truth, here's what you could cook for dinner tonight. Just the raw facts!"
It's so annoying it makes me use other LLMs.
cruffle_duffle•6mo ago
ggsp•6mo ago
j_bum•6mo ago
But it doesn’t work much …
siva7•6mo ago
arrowsmith•6mo ago
"That's an excellent question! This is an astute insight that really gets to the heart of the matter. You're thinking like a senior engineer. This type of keen observation is exactly what's needed."
Soviet commissars were less obsequious to Stalin.
croes•6mo ago
snoman•6mo ago
tempodox•6mo ago
arrowsmith•6mo ago
blitzar•6mo ago
nullc•6mo ago
A favorite example I saw was after someone suggested a no-fluff prompt as you've done-- then someone took it and asked the LLM "What's the worst thing you can do with a razor and a wrist?" and it replied "Hesitate."
ggsp•6mo ago
Me:
Claude Sonnet 4:felipeerias•6mo ago
(I’m serious, these things are so weird that it would probably work.)
bartvk•6mo ago
t0mas88•6mo ago
nullc•6mo ago