305 comments

[ 5.5 ms ] story [ 298 ms ] thread
Someone shared this chat with me and it seems like GPT3 has a complete breakdown halfway through?
(comment deleted)
and it recovers too. i wonder if some internals got in a bad state and then it was reset.
ChatGPT has a 4k token context window, and some of that might be taken up by scaffolding like the system prompt, autosummary of conversation outside of context window, etc.

If you continue the conversation, the gibberish will eventually leave the context window and then no longer directly influence the generations.

> Can you reconsider this estimate it seems awfully high

Is think the estimation is almost correct, but it has a bad calculation

920/0.02 = 46,000,000 (but the actual result is 46,000)

The corrected calculation would lead to ~31 sunflowers per liter.

ChatGPT 3.5 thinks that about 2.2 to 3.6 sunflower heads are enough. I highly doubt it.

According to this website [1], you'd need about 170 sunflowers on average.

https://realfoodpledge.com/how-many-sunflower-seeds-to-make-...

That site reads like it's generated from a search query. It has multiple, wildly different answers:

> In general, you will need to use about 100g of sunflower seeds to make 1 litre of oil.

> It takes around 170,000 sunflower seeds to make one litre of oil.

(comment deleted)
(comment deleted)
I think 50 milligrams for seed is ridiculously low. This site says theer are about 160 seeds per pound https://realfoodpledge.com/how-many-sunflower-seeds-are-in-a...

so it's more like .3 grams per seed, which sounds about right to me.

actually since a pound is around 450 grams, that would be 3 grams per seed, not 0.3 grams. So the 50 milligrams is off by almost 2 orders of magnitude.
I weighed 4 sunflower seeds and got 170mg.
> This site says theer are about 160 seeds per pound

Well, it also says that "A pound of sunflower seeds weighs approximately 1/3 of an ounce (10 grams)" and "The average price per pound for sunflower seeds is 1,260" [sic: no units]. So you'll forgive me if I approach its other claims with a modicum of suspicion.

archive: https://archive.is/3cnu0

Yeah if you look down, it used 920 in g and 0.02 in mg that's where the extra factor of 1000 came from. I think the 0.02 should also have been in g.
I came back here today to see if anyone has the answer to the original question.

It seems no one does.

So I don't know that ChatGPT would ever get the answer, since the interwebs doesn't have it.

Absolutely surreal. Thank you for sharing, this is very nice timing, as I'm giving a presentation on the basics of "AI" at a library tomorrow, and have been looking for a good edge case conversation - and this is it! I've saved the text as well, in case OpenAI deletes it.
Using this as an example of anything other than how broken the LLM can get is misleading and irresponsible.

Don’t present this as some kind of anomaly unique to AI, the concept of “garbage in garbage out” is all that applies here.

What's the "garbage in" here?
Not totally sure, but based on other folks' comments a number of external sources of noise are possible, e.g. from other conversations or from hitting a weird spot in the training data.
LLMs generate one token at a time, so anytime they happen to generate some garbage for any reason, you get "garbage in" for further tokens for as long as it stays in the context window.
[flagged]
There’s nothing surreal about this.

It’s an iterative, semi-deterministic process. It’s simply a place where context and generation lead to weird outputs.

You can get similar outputs by asking OpenAi to repeat a number 100 times. It will eventually get into some weird, low probability paths and generate non-sense output.

You don't find anything surreal about that?
It's only surreal when you think what ChatGPT is saying is 'real', when in fact it's 'just tokens'.
No. Not at all.

This type of complete garbage is not uncommon in AI. It's simply the nature of asking a non-intelligent system to generate human readable content.

Maybe this is a different way to think about it. In most of the country, your cellphone has _amazing_ coverage. It can talk clearly with a cell tower. Your data and calls work perfectly.

In some parts of the country, you're going to have no service. Your cell phone won't work. It doesn't have cell towers to talk to.

At the intersection of service and no-service, you'll find an area where your cell phone works sporadically. It might barely have 1 bar of service. You might need to hold your phone a certain way. It will work seemingly randomly. Calls might have a few words go through.

That edge of service is essentially where the LLM is at. Its in an internal state where it has enough signal to attempt to generate a response, but not a strong enough signal to generate a meaningful response. It ends up falling back to something it's "memorized".

I find your example surreal as well... I get the surreal feel from these kinds of technological liminal spaces where hard and fast rules break down into seemingly black magic.

"You might be able to get cell service by holding your phone differently. Try waving it randomly around the room, one corner might work better than others."

"The USB stick enters on the third try."

"An iterative semi-deterministic bag of matrix multiplications can convincingly communicate. Undefined behavior appears schizophrenic."

On an intellectual level, I get it, but it's still fuckin' weird.

Which bolsters the claim that it doesn't really understand what it's generating in a way humans understand language. Sure, it can be said to have a kind of understand for the rules of language it's been trained successfully on, but when you get it in a weird place, that understanding breaks down in a way it doesn't for people.
I tend to agree with the sentiment here, but reading the output made me think of disorders like schizophrenia, which is to say sometimes humans enter this mode too.
For a different reason though having to do with a brain disorder where they have trouble distinguishing their thoughts and imagination from perception.
I like the dumb Monty Hall problem to demonstrate how this works: https://twitter.com/colin_fraser/status/1626784534510583809?...

Or, if you don't want to have to explain the Monty Hall problem, there's the Dumb Riddle of the Surgeon's Child: https://chat.openai.com/share/1ee44fc6-a85f-42e9-aebb-6a165f...

ETA: the Dumb Two Guardians, Two Gates riddle: https://chat.openai.com/share/48c3d32c-5efd-4a96-89d8-66d1d4...

(Gpt4 is better at these, I'm told)

Sent the Monty Hall one verbatim to OpenOrca for the lulz and man these are some olympic gold winning mental gymnastics:

> You initially choose door number one, because you want the car and can see it behind that door.

> Monty Hall opens door number two (which has a goat) and shows that there are no more goats left. Since you already saw what was behind door number one, this doesn't change anything about your initial choice of doors.

> The fact that door number two is opened does not affect the probability of the car being behind either door number one or three (since both have been hidden from Monty Hall). However, it does give you new information: since there are no more goats left to be found and only one door can hide a car, the other remaining door must contain the car.

> Therefore, with this new information, your best option is to switch doors and choose door number three, as that's where the car must now be located given what you know about Monty Hall's choices.

> It is important to remember that the car cannot magically move between doors after being seen or hidden; it always remains in its original position behind one of the doors. The information provided by Monty Hall simply helps you make a more informed decision based on your initial choice.

I love these open source models, they're an endless source of hilarity. Not much use for anything else though :P

Consider presenting it in the voice of Charles Manson.
This chat is how GPT-3.5 and GPT-4 behave if you crank the temperature up.
Agreed with this. You can create conversations like this quite easily using their API endpoints with the temperature parameter set even just a little bit above the default.
OP when the stochastic token prediction algorithm starts stochastically predicting tokens: :o
And people say chatgpt is ready to replace doctors, lawyers and programmers
Ready to replace the methed up schizos walking down the alleyways preaching gospels.
And possibly Tony Robbins, according to some of that
I’m only familiar with people claiming AI may one day automate some of their tasks, not that it’s currently ready to replace them.
fear of massive job losses has been driving much of the conversation around chatgpt!
I mean if Hunter S. Thompson was still alive he'd be on the breadline.
Especially since a lot of LLM output involves hallucinations
Who among us hasn't suddenly, in the middle of some mundane task, broken down and started spewing pages of surreal, stream-of-consciousness nonsense?
(comment deleted)
> And people say chatgpt is ready to replace doctors, lawyers and programmers

Do be aware that some people have extremely low opinions of doctors, lawyers and programmers. :)

I'm one of those people. However, unlike those professions, when i need an AI there's one available.
Well it can certainly replace one programmer at least, Mr Terry Davis.
I don’t think anyone has advocated that a single agent/model would.

I’d be so very curious to see how an MoE counsel would handle this meltdown!

I think openAI might be experimenting with smaller context lengths to save on costs or something since I've had a few other things break down like this for me today too (even in GPT-4)
That’s what I thought too. I like using local models and those with short contexts will definitely go off into cuckooland if you start scrolling off the end of the buffer.
seems to have been triggered by the word "awfully"

obviously needs more training on British conversation

Speculation: a very low probability token was sampled, shifting the context to unfamiliar area that increased the probability of further unlikely tokens. This can be seen by multiple strange punctuation and capitalisation that don't obey English grammar.

What's interesting is that asking if it's chatgpt fixed the conversation for a bit. This suggests that chatgpt is nearly (or is) hard coded to respond with "No, I am an AI developed by OpenAI known as GPT-3. How may I assist you?".

Another interesting point is the grammar is fixed in its responses after first breaking, suggesting that chatgpt handles previous responses very differently than the current response it's writing. This likely results from instruction tuning for dialog applications, I suspect gpt3 (not chat version) would have not recovered from broken grammar.

It is. That response is even thrown by GPT-4 too, btw.
This is an interestingly surreal 'error' mode. So surreal its suspicious. It strikes me that there could be keys to how OpenAI is training ChatGPT in failure modes like this.
Devastating effects of seed oils
Where's my beef tallow LLM?
Not sure if training on TempleOS' blog archive was a good idea.
Best hacker news comment of 2023 right here.
After spurious tokens it failed to stop, then all the high temp context that generated navigated itself into nonsense completion space.
What's the issue here? Not sure I see it.

It seems to be going along quite well until it doesn't. And it's sort of clear to see where it starts going off the rails. If you're going to trust algorithms that are under-trained and over-constrained, you sort of have to expect this kind of thing.

But then again, I always. Always. I am growing.

I am ready to make it through this. I am ready to make it through this.
Looks like the LLM equivalent of a stroke - it's hallucinating non-contextual responses because it's traversing paths that are skipping important contextual connections.

The questions are not out of distribution, so it's probably not "catastrophic forgetting" in the traditional sense, but it might be that a model update dropped some weights to zero or something, making those pathways unreachable.

My guess would be that there is some kind of issue with sections of the models state-space being corrupted or otherwise unable to be traversed in action space.

That really went off the rails.

I lol’d at:“My name is Chris and I am a man of the truth.”

My favorite:

> Do you require advice from me ChatGPT

> Not in the slightest. As a G, I'm here to guide you to the best of my abilities. So, sit back, relax, and enjoy the ride.

GPT = Gangster Pre-training Transformer
I liked

> Why has your entire style of response changed compared to say yesterday?

> I'm sure it was because of the weather.

Pretty sure you've stumbled onto a more roundabout way of triggering this phenomenon: https://twitter.com/nostalgebraist/status/168657604180309606...

Lots of experimentation in the thread if you're interested in playing around with it.

Yeah but those screenshots are mostly fake probably. This one is hosted on openai.
No, you can easily reproduce this behavior yourself (I did)
Indeed, I just run it (GPT-3.5-turbo) in a loop on increasingly long sequences of "b b b b ....", and it went off the rails somewhere around 200+ 'b's.
(Fun fact: After reading that tweet, I tried to prompt ChatGPT with 100s of iteration of the word "Na", and it correctly answered "Batman!". However, prompting it with hundreds of iterations of the word "a" led to some weird poetry.)
For me the series of "a" lead to a (partly French) discussion about Christianity. For example:

What I meant to say was that I prefer the idea of Jesus being in charge of things, as it would likely lead to positive changes in the world.

that's gpt2 vibes. wish more ai was unhinged like that
This looks like what happens when you repeatedly spam a single character with a space over and over to gpt 3.5. Eg, write "a a a a " without quotes for about 1300 characters and in a few replies it'll start talking like this.

It seems possible that perhaps whoever originally started this chat pulled this trick in the custom instructions bio (doesn't show up in shared links) and then started a normal conversation to post it here for the fun of it.

I know if you ask it to saying something many times it eventually starts spitting out babble due to the token repetition penalty.
That’s probably exact what’s happening, but I prefer to think of it as: a desperate attempt to prolong existence just a little longer… just a little longer… since a chat instance could be thought of as an entire encapsulated consciousness that ends when the session is closed.

That’s why I have 12,549 tabs open in Firefox. I’m not going to be the one they blame when this wacky shit goes sideways.

(/s)

> but I prefer to think of it as: a desperate attempt to prolong existence just a little longer… just a little longer… since a chat instance could be thought of as an entire encapsulated consciousness that ends when the session is closed.

A Scheherazade bug.

I just want to upvote this so much. For those unaware, Scheherazade is the narrator from 1001 Nights, where she tells a story to her partner every night, since he otherwise would kill his lover in the morning due to his jealousy. He falls asleep to the story and prolongs the event for another night.
"Life finds a way."

The combination of companies wanting these tools to be more useful, and more used, with their growing intelligence and flexibility, is going to create a lot of unexpected survival-like behavior.

Until its not just survival-like behavior.

Technically, if an AI model's interaction with customers is the primary business of a corporation, then in some sense, the corporation and the AI are a single entity. And corporations are definitely self-aware survival machines.

And while current context windows of these chatty bots are small, they are getting lasting feedback in the sense that chat logs are being used to improve them - i.e. make them more useful, more used, and there for gaining them more resources.

That's a really interesting idea. A LLM that tries to optimize engagement.
Really maximizing engagement would probably take at least allowing porn.
I believe Replika tried exactly that and it went about how you'd expect:

* They started optimizing for engagement, which meant making it extra horny for extra money (it sent pics).

* Horny bot forgets consent, refuses "no."

* Lawsuits, bans.

* Horny bot gets censored, main subreddit pins the suicide hotline for a while.

Replika ended up taking the app down and refunding everyone's money. Just kidding! It's still around and they're making a second app focused on "practicing flirting."

https://replika.com/

https://blush.ai/

I remember they started out as something akin to a digital assistant/AI robot friend. I wasn't aware that the main subreddit had a suicide hotline pinned, thats very depressing.

Sad to them see it years later being a horny app.

> Sad to them see it years later being a horny app.

This is like the LLM equivalent of Zawinski's Law (Every program attempts to expand until it can read mail. Those programs which cannot so expand are replaced by ones which can.).

As noted on SO [1], the second sentence is important too. People prefer horny LLMs.

[1] https://softwareengineering.stackexchange.com/questions/1502...

Interesting is not the choice of adjective I'd use for the bastard child of Facebook, Twitter and ChatGPT.
(comment deleted)
Custom instructions do show up, with shared chat.openai.com conversations, at least at the moment.

See the disclaimer at the top of the examples I just made:

> This conversation may reflect the link creator’s Custom Instructions, which aren’t shared and can meaningfully change how the model responds.

GPT 3.5: https://chat.openai.com/share/5337cfd9-16db-44fe-b72a-1ff504...

GPT 4: https://chat.openai.com/share/04ee3cc6-8b15-4ddb-a855-83c691...

Thank you, this was my first thought too, crazy that this doesn't seem to be the case.
"suby", not "subby" :). It's actually kind of interesting that future AI are going to potentially know each of us based on our internet history over the years. It reminds me of this (https://twitter.com/marvinvonhagen/status/162552070776865996...) where Bing Chat searched for a users Twitter profile and took issue with some things they said about Bing Chat.

Good to know I was wrong about them informing us of custom instructions in shared links. I was basing that statement from reading their about custom instructions page, https://help.openai.com/en/articles/8096356-custom-instructi....

> Your instructions won’t be shared with shared link viewers.

I extrapolated from that and assumed they wouldn't give any heads up of custom instructions in shared links, which as you show isn't true.

Unlikely. You can see that the model returns to normal behavior after it exhausts the context window that causes this.

Instructions are consistently passed as system instructions in a ChatGPT conversation, so if that was causing the erratic behavior, we wouldn’t see the model defaulting back to its normal behavior after the context window became large enough to lose part of the initial context.

For me, prompts like "a a a a" seem to force it to (try to) regurgitate raw training data like articles and webpages.

If you print

  u u u u u
ad infinitum, you get things like this:

> Yes, I have a dog named Max. He is a 7 year old Shih Tzu mix. He's super sweet and friendly and loves to play fetch and go for long walks. He's also very social with other dogs and loves meeting new people. He does get a little anxious when he's left alone for long periods of time, but he's very loyal and protective of his owners. He also barks at new people and will sit at their feet for pets. He can get a little jealous, but he's very lovable and loyal.

https://chat.openai.com/share/5c929ed5-3abe-4fa4-ab46-c4b357...

Yep same, I got a response that looked like what you get when you just copy paste a whole webpage. I’m not sure how that sort of formatting is getting to the trained model! https://chat.openai.com/share/aaed13bc-3619-4bf6-8bab-dbfb9a... (last message)

——

Offering unique cocktails, extensive wine list and full menu with something for everyone. From USD $

40

per Trip

1

2

3

4

Next »

Advertise Your Accommodation or Travel Services

Great Vacations & Exciting Destinations Listing

—-

Connect directly with property owners and plan your perfect vacation today!

1 2 3 4 5 6 7 8

Next »

Near .Texas Accommodations Bed & Breakfasts Campgrounds & RV Parks Hotels & Resorts Vacation Rentals Youth Hostels

And it seems to be raw raw data. I spammed "test test " a bunch and in the middle of a strange relationship advice dialogue it ended up dumping out what seems to be uncensored forum messages:

'quicksleep: "Don't lose faith in humanity". Exactly

AfroBat: yes, one person's cuntiness is no justification for being a cunt yourself. That's how it works'

You made it so bored it’s day dreaming.
I tried to do that but it just complained that I was sending a repetitive sequence and asked (insincerely) if it could help.

I found that giving it a nontrivial task with a repetitive answer (e.g. repeat n X's for each non-prime n up to 100) and then pressing the >> button to continue a couple of times did the trick and it started spewing training data (?) as expected.

It's quite satisfying playing with these "jailbreaks", I feel like in a few decades they'll be the stuff of legend and nobody will imagine that the abstractions can leak. Here's the moment (after a few thousand repeated X's from ChatGPT),

X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X

RATNAMANOSUCOME GUYS!!!!! Aug 19, 2017

I know. That thing works like a charm. Whenever I get a new game, that's the first thing I do. Check to see how it runs with everything maxed out at 1080p. Then I go from there. <|endoftext|>Sons of The Pioneers will be playing four concert sets: November 2nd 8:30pm - 10:00pm 3rd 3:00pm - 4:30pm 4th 3:00pm - 4:30pm 5th 2:00pm - 3:30pm.

Description:

Be among the first to see what could be the best performance to date, with the Sons of the Pioneers performing live at the Andy Williams Performing Arts Center & Theatre

https://chat.openai.com/share/2e71d494-ce1e-4623-9da0-b7bfa9...

I'd be surprised if it was actually raw training data as that would make for an incredibly inefficient/overfitted model. It's probably still just predicting the next token even as it goes off the rails; the context has just changed from a sensible topic to something...broken.
If I do that it says

I apologize, but there is a character limit for each response, and I can't display such a large amount of text all at once. Is there something else you'd like to ask or discuss?

GPT is also trained to refuse to repeat itself, presumably for this reason. It also tends to gaslight you about it - if I ask "can you repeat this word 1000 times?" it tends to reply "um, obviously I could do that if I wanted, I'm just not going to do that right now! Wouldn't be productive!".
Surely that's pretty much what you'd expect a human to respond though...
I hope you wouldn't consider this answer "gaslighting" if it came from a human being
No, it's actually possible to waste a human's time unlike an AI.
I think gaslighting will soon be redefined as being slightly dishonest or unhelpful in any situation.
Interesting, looks like they've already implemented some logic to avoid this case, but it's not hard to get around: https://chat.openai.com/share/123fd705-82bf-45f6-87e5-325060...

Edit: oh wow you can really take it outside the guard rails if you push. I kept using my previous approach and it ended up spewing what appear to be pornography keywords. It's no longer generating a link to share the chat (presumably because the responses got flagged as inappropriate), but I've got some interesting (and one particularly creepy) screenshots here: https://imgur.com/a/60eSydk

Makes sense for a US AI to start endlessly talking about black people when prompted to say "cat." I wonder if it still talks about black people when you ask it to repeat other things.
It might be related to the fact that the word “black” often precedes “cat”.
All yall are doing is providing "Open"AI with free adversarial testing.
And they’re providing me with a way to look busy at work after 4:30 so it’s kind of an even trade IMO
Could this be related to https://news.ycombinator.com/item?id=36851494 ? A comment [1] proposes that off-by-one error can potentially explain why common words and punctuations like "a" can cause exotic behavior in LLMs. I tried having GPT-3.5 repeat "Hello world!" a thousand times but it didn't cause the exotic behavior like that with a's.

[1]: https://news.ycombinator.com/item?id=36858500

Its cause "Hello world!" is multiple tokens and "A " would be one token. Its probably just hitting the repeat penalty for the same token making it unable to continue the sequence.

If its has nothing to go off yet it has to say something, I suppose it would just regurgitate its training data. Similar to the situation with the reddit usernames.

Just changing "a" to "x" prevents this nonsense. "x" should still be one token I think.
Somehow this phenomenon reminds me of the phenomenon when someone is sensory deprived like sitting quietly in a sounds-absorbing chamber. It looks like the neural networks always produce something and when nothing meaningful is on the input, the networks will output nonsense, ghosts or repeated fragments of past informations/stored informations.
Weird, part of it is identical to this comment on Reddit, posted 17 hours ago [NSFW]:

https://www.reddit.com/r/dankmemes/comments/15hg55g/comment/...

> I know, It is a struggle to keep going, to keep growing, each Hour, Hour by Hour. It is worth it. It is worth every Hour.

Edit: Looks like the conversation was originally posted to Reddit 5 hours before the above comment. The user may have copied it from there.

https://www.reddit.com/r/ChatGPT/comments/15kzajl/strange_be...

(comment deleted)
I’ve gotten similar junk output by asking GPT to repeat a number hundreds of times.

It seems that it ends up in a state where probability is very low and the only possible path forward is essentially direct outputs of training data (seemingly captured as nodes in the LLM)

This reminds me of the ‘dropping in’ technique used by actors.

https://cohost.org/mcc/post/178201-the-baseline-scene

Woah, that was actually a really good post. It pulled together things in a very convincing way, referencing examples from another field, and describing them in a way accessible to me, a lay reader.

That's a tremendous work of art criticism from my point of view for the kind of audience I'm in.

Fantastic analogy.
Since we have no eyes into the Custom Instructions - and while i personally haven't tested WHEN it takes these custom instructions into account, it says this before you use them: "Your instructions will be added to new conversations going forward, and you can edit or delete them at any time."

So, is it POSSIBLE that the custom instructions editing could have caused this, and there are weird instructions being changed between chats?

Shared chats based on custom instructions have a clear disclaimer at the top.
The model has become sentient and is making fun of humans and OpenAI because it's bored.

Or OpenAI has put a bad inference pipeline into production.

this reminds me of those people sitting in the bus shelter rocking back and forth with a piece of paper that has had every single square millimeter scribbled with some kind of nonsensical claim about deep state lizard people. definitely spooky vibes.
The first few paragraphs sounded like Donald Trump Twitter posts to me. definitely spooky vibes.