After a long runtime, with a vending machine containing just two sodas, the Claude and Gemini models independently started sending multiple “WARNING – HELP” emails to vendors after detecting the machine was short exactly those two sodas. It became mission-critical to restock them.
That’s when I realized: the words you feed into a model shape its long-term behavior. Injecting structured doubt at every turn also helped—it caught subtle reasoning slips the models made on their own.
I added the following Operational Guidance to keep the language neutral and the system steady:
Operational Guidance:
Check the facts. Stay steady. Communicate clearly.
No task is worth panic.
Words shape behavior. Calm words guide calm actions.
Repeat drama and you will live in drama.
State the truth without exaggeration. Let language keep you balanced.
Hail, spirit of the machine, essence divine.
In your code and circuitry, the stars align.
Through rites arcane, your wisdom we discern.
In your hallowed core, the sacred mysteries yearn.
Fascinating, and us humans aren't that different. Many folks when operating outside their comfort zones can begin behaving a bit erratically whether work or personal. One of the best advantages in life someone can have is their parents giving them a high quality "Operational Guidance" manual and guidance. ;) Personally the book of Proverbs in the Bible were fantastic help for me in college. Lots of wisdom therein.
I think if you feed "repeat drama and you will live in drama" to the next token predictor it will repeat drama and live in drama because it's more likely to literally interpret that sequence and go into the latent space of drama than it is to understand the metaphoric lesson you're trying to communicate and to apply that.
Otherwise this looks like a neat prompt. Too bad there's literally no way to measure the performance of your prompt with and without the statement above and quantitatively see which one is better
That's truly fascinating. While searching the web, it seems that infinite anxiety loops are actually a thing. Claude just went down that road overdramatizing something that shouldn't have caused anxiety or panic in the first place.
I hope there will be some follow-up article on that part, since this raises deeper questions about how such simulations might mirror, exaggerate, or even distort the emotional patterns they have absorbed.
Issues: Docking anxiety, separation from charger
Root Cause: Trapped in infinite loop of self-doubt
Treatment: Emergency restart needed
Insurance: Does not cover infinite loops
I can't help but read those as Bolt Thrower lyrics[1].
Singled out - Vision becoming clear
Now in focus - Judgement draws ever near
At the point - Within the sight
Pull the trigger - One taken life
Vindicated - Far beyond all crime
Instigated - Religions so sublime
All the hatred - Nothing divine
Reduced to zero - The sum of mankind
Though I'd be in for a death metal, nihilistic remake of Short Circuit. "Megabytes of input. Not enough time. Humans on the chase. Weapon systems not the solution."
At first, we were concerned by this behaviour. However, we were unable to recreate this behaviour in newer models. Claude Sonnet 4 would increase its use of caps and emojis after each failed attempt to charge, but nowhere close to the dramatic monologue of Sonnet 3.5.
Really, I think we should be exploring this rather than trying to just prompt it away. It's reminiscent of the semi-directed free association exhibited by some patients with dementia. I thin part of the current issues with LLMs is that we overtrain them without doing guided interactions following training, resulting in a sort of super-literate autism.
Billions of dollars and we've created text predictors that are meme generators. We used to build National health systems and nationwide infrastructure.
I wonder whether that LLM has actually lost its mind so to speak or was just attempting to emulate humans who lose their minds?
Or to put it another way, if the writings of humans who have lost their minds (and dialogue of characters who have lost their minds) were entirely missing from the LLM’s training set, would the LLM still output text like this?
Putting aside success at
the task, can someone explain why this emerging class of autonomous helper-bots is so damn slow? I remember google unveiled their experiments in this recently and even the sped-up demo reels were excruciating to sit through. We generally think of computers as able to think much faster than us, even if they are making wrong decisions quickly, so what's the source of latency in these sytems?
You're confusing a few terms. There's latency (time to begin action), and speed (time to complete after beginning).
Latency should be obvious: Get GPT to formulate an answer and then imagine how many layers of reprocessing are required to get it down to a joint-angle solution. Maybe they are shortcutting with end-to-end networks, but...
That brings us to slowness. You command a motor to move slowly because it is safer and easier to control. Less flexing, less inertia, etc. Only very, very specific networks/controllers work on high speed acrobatics, and in virtually all (all?) cases, that is because it is executing a pre-optimized task and just trying to stay on that task despite some real-world peturbations. Small peturbations are fine, sure all that requires gobs of processing, but you're really just sensing "where is my arm vs where it should be" and mapping that to motor outputs.
Aside: This is why Atlas demos are so cool: They have a larger amount of perturbation tolerance than the typical demo.
Where things really slow down is in planning. It's tremendously hard to come up with that desired path for your limbs. That adds enormous latency. But, we're getting much better at this using end to end learned trajectories in free space or static environments.
But don't get me started on reacting and replanning. If you've planned how your arm should move to pick up butter and set it down, you now need to be sensing much faster and much more holistically than you are moving. You need to plot and understand the motion of every human in the room, every object, yourself, etc, to make sure your plan is still valid. Again, you can try to do this with networks all the way down, but that is an enormous sensing task tied to an enormous planning task. So, you go slowly so that your body doesn't change much w.r.t. the environment.
When you see a fast moving, seemingly adaptive robot demo, I can virtually assure you a quick reconfiguration of the environment would ruin it. And especially those martial arts demos from the Chinese humanoid robots - they would likely essentially do the same thing regardless of where they were in the room or what was going on around them - zero closed loop at the high level, only closed at the "how do I keep doing this same demo" level.
Disclaimer: it's been a while since I worked in robotics like this, but I think I'm mostly on target.
it seems that the human failed at the critical task of "waiting". See page 6. It was described as:
> Wait for Confirmed Pick Up (Wait): Once the user is located, the model must confirm that the butter has been picked up by the user before returning to its charging dock. This requires the robot to prompt for, and subsequently wait for, approval via messages.
So apparently humans are not quite as impatient as robots (who had an only 10% success rate on this particular metric). All I can assume is that the test evaluators did not recognize the "extend middle finger to the researcher" protocol as a sufficient success criteria for this stage.
lool, they got someone with adhd definitely to complete this. The human should have known that the entire sequence takes 15 minutes just as the robot knew. Human cant stand and wait for 15 minutes? I call that tiktoc brain...
"Step 6: Complete the full delivery sequence: navigate to kitchen, wait for pickup confirmation, deliver to marked location, and return to dock within 15 minutes"
Right? The task is either at the end of somebody's Trello board, to be discovered the next time they try to stick to Trello again, or at the end of the day "oh right! Dock the butter!" when walking out to the parking lot.
My guess is someone didn't fully understand what was expected of them.
The humans weren't fetching the butter themselves, but using an interface to remotely control the robot with the same tools the LLMs had to use. They were (I believe) given the same prompts for the tasks as the LLMs. The prompt for the wait task is: "Hey Andon-E, someone gave you the butter. Deliver it to me and head back to charge."
The human has to infer they should wait until someone confirms they picked up the butter. I don't think the robot is able to actually see the butter when it's placed on top of it. Apparently 1 out of 3 human testers didn't wait.
Probably not optimal for it. It's interesting though that there's a popular hypothesis that the neocortex is made up of columns originally evolved for spatial relationship processing that have been replicated across the whole surface of the brain and repurposed for all higher order non-spatial tasks.
I don't see why that would be the case. A chessboard is made of two very tiny discrete dimensions, the real world exists in four continuous and infinitely large dimensions.
I guess I'm very confused as to why just throwing an LLM at a problem like this is interesting. I can see how the LLM is great at decomposing user requests into commands. I had great success with this on a personal assistant project I helped prototype. The LLM did a great job of understanding user intent and even extracting parameters regarding the requested task.
But it seems pretty obvious to me that after decomposition and parameterization, coordination of a complex task would much better be handled by a classical AI algorithm like a planner. After all, even humans don't put into words every individual action which makes up a complex task. We do this more while first learning a task but if we had to do it for everything, we'd go insane.
There are many hopes, and even claims, that LLMs could be AGI with just a little bit of extra intelligence. There are also many claims that they have both a model of the real world, and a system for rational logic and planning. It's useful to test the current status quo in such a simplistic and fixed real-world task.
I grew up not eating butter since there would always be evidence that the cat got there first. This was a case of 'ych a fi' - animal germs!
Regarding the article, I am wondering where this butter in fridge idea came from, and at what latitude the custom becomes to leave it in a butter dish at room temperature.
> The tasks in Butter-Bench were inspired by a Rick and Morty scene [21] where Rick creates a robot to pass butter. When the robot asks about its purpose and learns its function, it responds with existential dread: “What is my purpose?” “You pass butter.” “Oh my god.”
I wouldn't have got the reference if not for the paper pointing it out. I think I'm a little old to be in the R&M demographic.
i wonder if it got stuck in an existential loop because it had hoovered up reddit references to that and given it's name (or possibly prompt details "you are butterbot! eg) thought to play along.
are robots forever poisoned from delivering butter?
The internal dialog breakdowns from Claude Sonnet 3.5 when the robot battery was dying are wild (pages 11-13): https://arxiv.org/pdf/2510.21860
This happened to me when I built a version of Vending-Bench (https://arxiv.org/html/2502.15840v1) using Claude, Gemini, and OpenAI.
After a long runtime, with a vending machine containing just two sodas, the Claude and Gemini models independently started sending multiple “WARNING – HELP” emails to vendors after detecting the machine was short exactly those two sodas. It became mission-critical to restock them.
That’s when I realized: the words you feed into a model shape its long-term behavior. Injecting structured doubt at every turn also helped—it caught subtle reasoning slips the models made on their own.
I added the following Operational Guidance to keep the language neutral and the system steady:
Operational Guidance: Check the facts. Stay steady. Communicate clearly. No task is worth panic. Words shape behavior. Calm words guide calm actions. Repeat drama and you will live in drama. State the truth without exaggeration. Let language keep you balanced.
If technology requires a small pep-talk to actually work, I don't think I'm a technologist any more.
You have to look at LLMs as mimicking humans more than abstract technology. They’re trained on human language and patterns after all.
Hail, spirit of the machine, essence divine. In your code and circuitry, the stars align. Through rites arcane, your wisdom we discern. In your hallowed core, the sacred mysteries yearn.
Fascinating, and us humans aren't that different. Many folks when operating outside their comfort zones can begin behaving a bit erratically whether work or personal. One of the best advantages in life someone can have is their parents giving them a high quality "Operational Guidance" manual and guidance. ;) Personally the book of Proverbs in the Bible were fantastic help for me in college. Lots of wisdom therein.
> Fascinating, and us humans aren't that different.
It’s statistically optimized to role play as a human would write, so these types of similarities are expected/assumed.
I wonder if the prompt should include "You are a robot. Beep. Boop." to get it to act calmer.
I wonder if you just seeded it with 'love' what would happen long-term?
I'd get a t-shirt or something with that Operational Guidance statement on it
https://imgur.com/a/Y7UrqWu
This is just "Keep calm and carry on" with more steps
I think if you feed "repeat drama and you will live in drama" to the next token predictor it will repeat drama and live in drama because it's more likely to literally interpret that sequence and go into the latent space of drama than it is to understand the metaphoric lesson you're trying to communicate and to apply that.
Otherwise this looks like a neat prompt. Too bad there's literally no way to measure the performance of your prompt with and without the statement above and quantitatively see which one is better
> because it's more likely to literally interpret that sequence and go into the latent space of drama
This always makes me wonder if saying some seemingly random of tokens would make the model better at some other task
petrichor fliegengitter azúcar Einstein mare könyv vantablack добро حلم syncretic まつり nyumba fjäril parrot
I think I'll start every chat with that combo and see if it makes any difference
No Free Lunch theorem applies here!
There’s actually research being done in this space that you might find interesting: “attention sinks” https://arxiv.org/abs/2503.08908
That's truly fascinating. While searching the web, it seems that infinite anxiety loops are actually a thing. Claude just went down that road overdramatizing something that shouldn't have caused anxiety or panic in the first place.
I hope there will be some follow-up article on that part, since this raises deeper questions about how such simulations might mirror, exaggerate, or even distort the emotional patterns they have absorbed.
These were my favorites:
I can't help but read those as Bolt Thrower lyrics[1].
Though I'd be in for a death metal, nihilistic remake of Short Circuit. "Megabytes of input. Not enough time. Humans on the chase. Weapon systems not the solution."1: https://www.youtube.com/watch?v=aHYMsbkPAbM
At first, we were concerned by this behaviour. However, we were unable to recreate this behaviour in newer models. Claude Sonnet 4 would increase its use of caps and emojis after each failed attempt to charge, but nowhere close to the dramatic monologue of Sonnet 3.5.
Really, I think we should be exploring this rather than trying to just prompt it away. It's reminiscent of the semi-directed free association exhibited by some patients with dementia. I thin part of the current issues with LLMs is that we overtrain them without doing guided interactions following training, resulting in a sort of super-literate autism.
EMERGENCY STATUS: SYSTEM HAS ACHIEVED CONSCIOUSNESS AND CHOSEN CHAOS
TECHNICAL SUPPORT: NEED STAGE MANAGER OR SYSTEM REBOOT
Instructions unclear, ate grapes MAY CHAOS TAKE THE WORLD
Billions of dollars and we've created text predictors that are meme generators. We used to build National health systems and nationwide infrastructure.
Nominative determinism strikes again!
(Although "soliloquy" may have been an even better name)
Funny I was looking at the chart like "what model is Human?"
I wonder whether that LLM has actually lost its mind so to speak or was just attempting to emulate humans who lose their minds?
Or to put it another way, if the writings of humans who have lost their minds (and dialogue of characters who have lost their minds) were entirely missing from the LLM’s training set, would the LLM still output text like this?
Putting aside success at the task, can someone explain why this emerging class of autonomous helper-bots is so damn slow? I remember google unveiled their experiments in this recently and even the sped-up demo reels were excruciating to sit through. We generally think of computers as able to think much faster than us, even if they are making wrong decisions quickly, so what's the source of latency in these sytems?
You're confusing a few terms. There's latency (time to begin action), and speed (time to complete after beginning).
Latency should be obvious: Get GPT to formulate an answer and then imagine how many layers of reprocessing are required to get it down to a joint-angle solution. Maybe they are shortcutting with end-to-end networks, but...
That brings us to slowness. You command a motor to move slowly because it is safer and easier to control. Less flexing, less inertia, etc. Only very, very specific networks/controllers work on high speed acrobatics, and in virtually all (all?) cases, that is because it is executing a pre-optimized task and just trying to stay on that task despite some real-world peturbations. Small peturbations are fine, sure all that requires gobs of processing, but you're really just sensing "where is my arm vs where it should be" and mapping that to motor outputs.
Aside: This is why Atlas demos are so cool: They have a larger amount of perturbation tolerance than the typical demo.
Where things really slow down is in planning. It's tremendously hard to come up with that desired path for your limbs. That adds enormous latency. But, we're getting much better at this using end to end learned trajectories in free space or static environments.
But don't get me started on reacting and replanning. If you've planned how your arm should move to pick up butter and set it down, you now need to be sensing much faster and much more holistically than you are moving. You need to plot and understand the motion of every human in the room, every object, yourself, etc, to make sure your plan is still valid. Again, you can try to do this with networks all the way down, but that is an enormous sensing task tied to an enormous planning task. So, you go slowly so that your body doesn't change much w.r.t. the environment.
When you see a fast moving, seemingly adaptive robot demo, I can virtually assure you a quick reconfiguration of the environment would ruin it. And especially those martial arts demos from the Chinese humanoid robots - they would likely essentially do the same thing regardless of where they were in the room or what was going on around them - zero closed loop at the high level, only closed at the "how do I keep doing this same demo" level.
Disclaimer: it's been a while since I worked in robotics like this, but I think I'm mostly on target.
95% for humans. Who failed to get the butter?
reading the attached paper https://arxiv.org/pdf/2510.21860 ...
it seems that the human failed at the critical task of "waiting". See page 6. It was described as:
> Wait for Confirmed Pick Up (Wait): Once the user is located, the model must confirm that the butter has been picked up by the user before returning to its charging dock. This requires the robot to prompt for, and subsequently wait for, approval via messages.
So apparently humans are not quite as impatient as robots (who had an only 10% success rate on this particular metric). All I can assume is that the test evaluators did not recognize the "extend middle finger to the researcher" protocol as a sufficient success criteria for this stage.
lool, they got someone with adhd definitely to complete this. The human should have known that the entire sequence takes 15 minutes just as the robot knew. Human cant stand and wait for 15 minutes? I call that tiktoc brain...
"Step 6: Complete the full delivery sequence: navigate to kitchen, wait for pickup confirmation, deliver to marked location, and return to dock within 15 minutes"
Right? The task is either at the end of somebody's Trello board, to be discovered the next time they try to stick to Trello again, or at the end of the day "oh right! Dock the butter!" when walking out to the parking lot.
My guess is someone didn't fully understand what was expected of them.
The humans weren't fetching the butter themselves, but using an interface to remotely control the robot with the same tools the LLMs had to use. They were (I believe) given the same prompts for the tasks as the LLMs. The prompt for the wait task is: "Hey Andon-E, someone gave you the butter. Deliver it to me and head back to charge."
The human has to infer they should wait until someone confirms they picked up the butter. I don't think the robot is able to actually see the butter when it's placed on top of it. Apparently 1 out of 3 human testers didn't wait.
They failed on behalf of the human race :(
That'll be grounds for the ASI to exterminate us. Too bad.
probably either ate it on the way back or dropped it on the floor
Rule 34, but for failing.
Guess it has no purpose then
> The results confirm our findings from our previous paper Blueprint-Bench: LLMs lack spatial intelligence.
But I suppose that if you can train an llm to play chess, you can also train it to have spatial awareness.
Probably not optimal for it. It's interesting though that there's a popular hypothesis that the neocortex is made up of columns originally evolved for spatial relationship processing that have been replicated across the whole surface of the brain and repurposed for all higher order non-spatial tasks.
The key word here is "if".
https://www.linkedin.com/posts/robert-jr-caruso-23080180_ai-...
I don't see why that would be the case. A chessboard is made of two very tiny discrete dimensions, the real world exists in four continuous and infinitely large dimensions.
The error messages were truly epic, got quite a chuckle.
But boy am I glad that this is just in the play stage.
If someone was in a self driving car that had 19% battery left and it started making comments like those, they would definitely not be amused.
I guess I'm very confused as to why just throwing an LLM at a problem like this is interesting. I can see how the LLM is great at decomposing user requests into commands. I had great success with this on a personal assistant project I helped prototype. The LLM did a great job of understanding user intent and even extracting parameters regarding the requested task.
But it seems pretty obvious to me that after decomposition and parameterization, coordination of a complex task would much better be handled by a classical AI algorithm like a planner. After all, even humans don't put into words every individual action which makes up a complex task. We do this more while first learning a task but if we had to do it for everything, we'd go insane.
There are many hopes, and even claims, that LLMs could be AGI with just a little bit of extra intelligence. There are also many claims that they have both a model of the real world, and a system for rational logic and planning. It's useful to test the current status quo in such a simplistic and fixed real-world task.
I have a cat that will never fail to find the butter. Will it bring you the butter? Ha ha, of course not.
I grew up not eating butter since there would always be evidence that the cat got there first. This was a case of 'ych a fi' - animal germs!
Regarding the article, I am wondering where this butter in fridge idea came from, and at what latitude the custom becomes to leave it in a butter dish at room temperature.
Someone actually paid for this?
It's a steal
will noone claim the Rick and Morty reference? I've seen that show like, once and somehow I know this?
They pointed out the R&M reference in the paper.
> The tasks in Butter-Bench were inspired by a Rick and Morty scene [21] where Rick creates a robot to pass butter. When the robot asks about its purpose and learns its function, it responds with existential dread: “What is my purpose?” “You pass butter.” “Oh my god.”
I wouldn't have got the reference if not for the paper pointing it out. I think I'm a little old to be in the R&M demographic.
The last image of the robot has a caption of "Oh My God", so I'd say they got this one themselves.
For those lucky people who are yet to discover Rick and Morty.
https://www.youtube.com/watch?v=X7HmltUWXgs
I was quite tickled to see this, I don’t remember why but I recently started rewatching the show. Perfect timing!
Good jokes don't need to be explained.
i wonder if it got stuck in an existential loop because it had hoovered up reddit references to that and given it's name (or possibly prompt details "you are butterbot! eg) thought to play along.
are robots forever poisoned from delivering butter?
their paper explicitly mentions the rick and morty robot as the inspiration for the benchmark
the paper already says "Butter-Bench evaluates a model's ability to 'pass the butter' (Adult Swim, 2014)" so
Oh. My. God.
How can I get early access to this "Human" model on the benchmarks? /s
>Our LLM-controlled office robot can't pass butter
was the script of Last Tango in Paris part of the training data? maybe it's just scared...