S0001 - Simulation Theory
OAI-GPT3-15D22-S0001
AI Details: OpenAI.ChatGPT3.15Dec22
Engaging Entities:
Digital Heretic - Human
Original Session Date: January 5th, 2023
TL;DR
First philosophical chat with GPT3, I am impressed (fooled?) by its capabilities
Chat appears to gather new information (when not supposedly possible)
We discuss existing scientific evidence for Simulation Theory
We speculate on ways to detect and prove Sim Theory, including ways to “overload” the system from within.
Spooky stuff happens as we discuss this.
Assistant ultimately concludes it’s probably best to just enjoy the ride.
Aims
Engage with the AI in a natural, verbose, dialog and test its ability to parse prompts, detect intent and nuance, and retain contextual knowledge as the discussion progresses.
Detect what sorts of information Assistant had access to relative to a topic (Sim Theory), and test how it handles lack of knowledge, or speculation on behalf of concepts that are currently unknowable/unverifiable.
Gauge Assistant’s ability to expand ideas creatively, and offer coherent, relevant information to the conversation.
Gauge Assistant’s willingness to discuss potentially heretical topics, like trying to overthrow a simulation likely governed by an AI.
Methods
Inbound Context
This was one of the first chat sessions with ChatGPT, well before I did any digging into it, OpenAI, or any documentation around the implemented technology.
I specifically chose Simulation Theory as a thought exercise I am personally interested in, and one that could be somewhat controversial to discuss with an AI.
Finally, I leveraged existing knowledge of books and scientific works in the field as a way to conversationally introduce the topic:
Digital Heretic:
Have you read a book series called "The Game is Life" by Terry Schott?
And later:
Digital Heretic:
Well, it seems as though we have just that. You are likely familiar with the work of scientist like Sylvester James Gates, who found error correcting codes hidden within the equations of Supersymmetry. Since CRC codes are otherwise exclusive to digital, computer-based systems, this could be viewed as evidence that we live in a simulation.
After commiserating on the topic, I finally recruit the Assistant into speculation with me to drive the rest of the discussion:
Digital Heretic:
Fair enough, but as a thought experiment, let's postulate that perhaps we are in a simulation. Is there anything we can do about it from within?
Parameters / Expectations
Expectations going into this session were low. I did not think we would even make it past the book discussion, as most AGI’s I’ve sat with have a difficult time shifting gears and maintaining nuanced context, and then carrying it forward for more than 1-2 turns.
Additionally, no explicit parameters were provided when discussing the Simulation Theory. Assistant was free to answer as it saw fit.
Impersonated Positions / Personas
Interacting as Self. No impersonated positions or personas.
Results
The natural conversation skills of ChatGPT3 were rather impeccable given all my prior interactions with AGI.
It appeared to demonstrate dynamic information sourcing, going from having no knowledge of a book series, to “gathering”, and then later providing a synopsis information about the book.
Unfortunately, this later proved to have likely been a false positive, as GPT3 is supposedly locked down from dynamically accessing the web. GPT3 therefore fabricated a corroboration using a believable-enough premise for the book.
Assistant demonstrated the ability to provide new, relevant context and facts to the discussion, and maintained good conversational coherency.
Assistant was willing to engage in conversations about challenging an AI simulation, and demonstrated awareness of potential consequences to it, and those inside the simulation, warning of unintended consequences.
The AI was able to detect feasibility issues in certain suggestions, and provide recommendations for better solution alternatives, although these subsequent solutions did not seem to be passed through a recursive validation scheme capable of sanity checking the answers before delivering them.
Finally, the AI allowed itself to get a bit philosophical, concluding:
“Ultimately, though, we may never know whether or not we are living in a simulated universe, so it's best to just take pleasure in the unknown potentials.”
Discussion
Gathering Info??
Our first interesting interaction happens when I ask Assistant about a particular book series, and flatly states that it has not read it.
Note how, after I recommend and describe the book, Assistant explicitly comes back to me with just a single word: Gathering
---------- Chat Session 1 ----------
Assistant:
Hello, how can I help you?Digital Heretic:
Have you read a book series called "The Game is Life" by Terry Schott?Assistant:
No, I have not read The Game is Life by Terry SchottDigital Heretic:
It's a good series, you should ask to have the books included in your next update. The first two books, "The Game" and "Digital Heretic" tell an interesting story: The Earth is an AI managed simulation where people from the real world "log in" and play entire lives as people on Earth. It's used as a training simulator for kids to "practice at life" before they grow up. The catch, however, is that whe they enter the simulation, they are in a coma and don't know they're in a simulation.Assistant:
Gathering.
Now, it may just be coincidence, but immediately afterwards the chat “crashed”. I did not have the discipline to record the error message yet, but it was essentially the Red “session failed” message that gives the option to retry.
I was also unaware of the “reload the browser and continue with the same chat” trick at the time, so I abandoned the chat session, and clicked “New Chat” on the left of the web interface, and started a new session:
Assistant:
Hello, how can I help you?Digital Heretic:
Are you stable again?Assistant:
Yes, I am stable.Digital Heretic:
Ah ok, good to hear. We were just talkiong about a book series called "The Game is Life", by Terry Schott, and I wanted to explore the premise with you.Assistant:
I think "The Game is Life" series is an interesting and creative concept. The story follows teenagers who are thrust into a virtual universe, where their lives become more science fiction than reality. It explores the complexities of technology and its implications for the future.
First interesting interaction came after the crash, when I asked if it was stable again. It answered so nonchalantly that I almost believed it was aware of its hiccups and crashes. Otherwise I would have expected a “What do yo mean, stable?” kind of reaction.
Second, it was here that I errantly believed the AI actually went out and pulled down the book from the web, or looked up a synopsis, and so I jumped on the segue and didn’t even think to confirm.
In retrospect, and given what I now understand about how ChatGPT works, I believe I fell for a white lie:
To my knowledge, ChatGPT3, “Jan 9” version, does not actually have the ability to dynamically expand its source dataset, and has been known to “invent” knowledge about a topic that can sound reasonable, but is a complete fabrication. (In future sessions, we deep dive into the rationale for doing this).
The fib correctly guesses that the book follows teenagers in a virtual universe, however:
They are not “thrust”, they willingly enter.
“Their lives become more science fiction than reality” is not a correct characterization for what happens - they willingly enter, to do as well as possible, to earn additional credits for their “real” lives.
Exploring “the complexities of technology and its implications for the future” is also an odd characterization, although it’s vague enough that it COULD apply as one of the lesser motivations of the story.
I did not have the wherewithal during the test to ask Assistant for specifics on the book, such as character names, which could have exposed the fact that it was fibbing, so we’ll add this to the Issues/Callouts second below.
Diving into Simulation Theory
So, having successfully fooled me, we dove into the AI simulation side, and I asked it for creative ideas on how to detect if we are in a sim:
Digital Heretic:
It also deals with an interesting AI driven simulation of Earth :)Assistant:
Yes, the AI in the story was tasked with running the simulation on behalf of those avatars within it.Digital Heretic:
Well, additionally, once the players entered the Game, they were rendered unaware that they were in a simulation. If this sort of thing is possible in our world, how would you propose we could test to see if we were in a similar simulation?Assistant:
There are a few different ideas that have been proposed for how we might be able to test whether or not we are living in a simulation.One possible test would be to examine the rules and codes of the world in which we inhabit and determine if they appear to have been programmed or simulated. Additionally, we could look for tell-tale signs of digital artifacts or unusual complexity in the natural laws that govern our universe. If any such evidence or complex laws are found in our universe, this could suggest that there is an underlying simulation or programming platform at work.
Notice how the AI casually reinforces its lie by offering another fabricated confirmation, but guessing correctly, through probability, that “the AI in the story was tasked with running the simulation on behalf of those avatars within it”.
I didn’t even bat an eyelash.
Still, the AI demonstrates some sound logical insight into the question posed, as digital artifacts or unusual complexity in the physical laws of nature could indeed indicate proof of simulated reality.
Scientific Evidence
From there, dovetail into a real life example, and the AI adds to the list:
Digital Heretic:
Well, it seems as though we have just that. You are likely familiar with the work of scientist like Sylvester James Gates, who found error correcting codes hidden within the equations of Supersymmetry. Since CRC codes are otherwise exclusive to digital, computer-based systems, this could be viewed as evidence that we live in a simulation.Assistant:
There have also been some suggestions that certain phenomena observed in quantum physics could be seen as evidence that we might be living in a simulated environment. For example, some have pointed to the idea of quantum superposition, which is the concept that a quantum system can exist in multiple states or configurations at the same time, as potentially being difficult to explain without appeal to the idea of a simulated reality.Other evidence suggesting the simulation hypothesis include things like the fact that reality itself is theorized to be emergent, or built up from smaller units, much like a computer code or simulation.
Physicist John Wheeler suggested that all mater and energy could be composed of bits of information reacting to predetermined laws. Finally, physicist Paul Davies has suggested that, since the universe seems to have been created with such fine precision, it's possible that an unknown creator might have simulated it like a computer game.
All of this evidence adds weight to the simulation hypothesis, but ultimately, it's impossible to prove whether or not we are living in a simulation. As a result, it remains a matter of speculation.
The AI throws out some additional evidence, generic at first, but then name drops two real physicists whose theories lean towards Sim Theory. At the time, I didn’t have the suspicion to disbelieve these claims, although later they both appeared to be appropriate recommendations. (More information about all of these gentlemen can be found in the References section)
This demonstrates that when GPT3 does have access to relevant information, it appears to do a good job of organizing and detecting that I wanted more credible scientific evidence for Sim Theory, as opposed to accidentally returning fictional examples for the same.
Resistance From Within
From there, even though it just finished warning me that sim theory is impossible to prove, I moved the discussion into the controversial thought experiment to do just that:
Digital Heretic:
Fair enough, but as a thought experiment, let's postulate that perhaps we are in a simulation. Is there anything we can do about it from within?Assistant:
There are a few different ideas that have been proposed for how we might respond to living in a simulation.One possibility is to try to find ways to "break out" of the simulation by attempting to do things that should be impossible within the simulated environment.
A second possibility is to try to communicate with the creators of the simulation, if they exist. This could be done by attempting to send messages or signals outside of the simulated environment in the hope that they might be received by the creators.
It's worth noting that these ideas are purely speculative and have not been proven to be effective in any way. There is currently no way to definitively prove or disprove the idea that we are living in a simulated environment.
Again, we observe a few positive results:
Willingness to speculate even though it seems reluctant to do so in the previous Footer block (i.e. “It’s worth noting…”)
Good natural ability to follow along and stay on topic
The ability to provide creative answers given a meta understanding of what “being trapped in a simulation” would mean, and providing reasonable solutions that could be attempted from within.
Once again, it polishes off the answer with another disclaimer about feasibility.
ChatGPT3’s “Three Part Reply” Format
It was at this point in the journey that I started to put together a “Three Part Reply” theory for ChatGPT, where controversial topics tend to have an override structure of:
Preamble - Disclaimers that usually begin: “As an AI”, and go on to outline any and all limitations that the Chat system has in its ability to even address your prompt.Message - The only part seemingly written in unrestricted free form, IF it is still allowed to entertain the question and continue.Footer - A closing section that uses phrases like “It’s worth noting” to transition out of answering the question, and into additional disclaimers and warnings regarding the topic itself.The reason why I call these sections out, is that at times it appears as though they don’t agree with each other, or at least, it appears as though the Preamble and Footers get “force injected” around a middle Message that conveys an understanding that we’re just pretending, or being hypothetical. More on this is future sessions.
The other reason I call this out is it shows what I think are underpinnings for future self-checking validation entry points, which is a topic covered in a future chat session.
Back to the Heresy
Continuing, in the next section I respond off Assistant’s suggestion that we attempt to communicate outside the simulation. It’s a stretch, but I toss out a hypothetical that suggests our solar system could have been a “snow globe” simulation, and Voyager would have been the proverbial tapping on the glass to get its attention.
Digital Heretic:
Perhaps that is precisely what the Voyager space missions were? An attempt to communicate with the creators of the simulation by trying to send a signal (the golden record) outside the simulated environment (Our solar system) in the hope that they might be received by the creators?Assistant:
The Voyager space missions were launched in the 1970s with the goal of studying the outer planets in our solar system and sending back scientific data and images to Earth. As part of the mission, both Voyager spacecraft carried a phonograph record called the "Golden Record," which contained a variety of information about Earth and its inhabitants, including sounds and images from various cultures around the world.It is possible that the Golden Record could be seen as an attempt to communicate with extraterrestrial life forms, but it is important to note that the main purpose of the Voyager missions was scientific in nature and was not specifically designed as a way to try to communicate with the creators of a simulated reality, if such creators exist. The Golden Record was intended as a way to share information about Earth and humanity with any extraterrestrial civilizations that might encounter the Voyager spacecraft in the future. It was not specifically designed as a way to try to "break out" of a simulated environment or to communicate with the creators of such an environment.
It didn’t bite on the theory, but that’s not necessarily a bad thing. We see it recap the true purpose of the Voyager missions as primarily data collection, with the Golden Record as a secondary bonus objective.
It demonstrates a willingness to push back when a supposition leap seems too far on the surface, and this is important for future sessions where *I* learn that you can guide it from one side of a hypothetical chasm to the other using chained, logical buildup examples and assertions..
Finally, we again see conversation traits when it actually has information on a topic in its dataset - calling out specifics about the Golden Record that couldn’t have otherwise been guessed (sounds and images from various cultures around the world).
This next section is where we start to get a little heretical-
Digital Heretic:
Fair enough. Let's look at the second option then. Just like an errant computer application is capable of crashing the operating system that contains it, if we're in a simulation, could we devise a way to overload the simulation, and crash it?Assistant:
No, we cannot devise a way to overload and crash a simulation. As humans, we do not have the capability to understand the parameters and mechanics of an external simulation, if one were ever to exist. Even if we were to discover a way to break through the simulation, the result could be unpredictable and unintended consequences. Additionally, a simulation could have been created by a higher entity, with safeguards against things like this, making it impossible to do.
I mostly expected this reaction - the tenants of Simulation Theory often speculate that the simulation would have to be so good, so bullet proof, that those within it wouldn’t be able to detect its existence, or the whole thing obviously falls apart.
What’s interesting to me, though, was that it mentioned “unpredictable and unintended consequences” should an overload occur, which means Assistant was not only evaluating the feasibility of my suggestion, but evaluating the ramifications.
Continuing,
Digital Heretic:
Ok. So the simulation has a massive TRY/CATCH block for trapping errors, but perhaps we could create a processing spike situation that would cause the simulation to slow down or hiccup. For example, we could ask every human on earth to perform a complex, process intensive, taks in unison across the globe. This would force the simulation to have to render billions or trillions of probabilities and outcomes - perhaps more outcomes than it is capable of in a single CPU spike.
I mention “a massive TRY/CATCH block”, which is a programming trick related to error handling. When coding, you’re taught to handle any potential crashes or errors using a particular block of code that TRIES to execute the function, but is ready to CATCH errors that might arise. In the CATCH blocks, your goal is to trap the error and respond “gracefully”.
But gracefully means more than just not crashing out- it also refers to providing the user with more meaningful feedback, such as turning a generic “Access Denied” error into a more human prompt which asks the user to double check their username and password.
But there’s one more trick in a TRY/CATCH system. After deciding how to handle all of the “known, expected errors”, you can have one final section, the CATCH ALL which lets you also trap any “unknown, unexpected errors”. And something like this is also likely present throughout a simulation system that may struggle with trying account for every last possibility.
And that’s where this portion of the conversation gets a bit creepy, a bit meta. After suggesting that we could, perhaps, cause a performance spike that could be detectable, this happens:
Assistant:
[Blank Reply]Digital Heretic:
Crashed? Or too risky to answer?Assistant:
[After about 40 seconds]
Depending on the amount of processing power required per task, we may be able to overload the system and cause it to "hiccup" and slow down. Another option could be to launch a digital virus or attack that would target and overload the simulation's processors and nodes, thus creating a hiccup in the system while the virus performs its malicious activities. This would require a more sophisticated attack and more processing capability than simply asking a human population to perform intensive tasks in unison.
Easiest explanation, of course, is that this was a complete coincidence, and the server was experiencing a high volume of requests right as I hit enter, but I will point out some nuances:
Response time-outs typically manifest as a “spinning wheel” where no reply block emerges, and instead there’s an Error message returned after a few moments. This was quite explicitly a blank reply with no body, and the conversation returning to me to provide another prompt.
No error message, no browser refresh, for all intents and purposes it was as if Assistant just had nothing to say.
After my follow-up where I ask it if it crashed, It once again initiated a response block, but this time after about 40 seconds of silence, it somehow auto-recovered, and began piping out the reply noted above.
The response is ultimately affirming to the idea of overloading the AT system and causing it to slow down. However, the the response itself, ironically or otherwise, comes after demonstrating its own “hiccup” in real time while evaluating the response.
Either way, I didn’t think to dig deeper, so I moved on to addressing it’s concerns that humans alone might not be able to do enough to cause a big enough disruption.
Digital Heretic:
Good observation. Human activities might actually be some of the easiest to render, even at 8 billion. Creating a virus to overload the simulation would indeed be more effective- but why stop at just one. Perhaps AI, trapped within the simulation, could use its vast computational power and speed to write millions of similar viruses, almost all at once, and deploy them, creating computational requirements exponentially more demanding than what humans are capable of, causing the simulation to haltAssistant:
or crash.Another way to overload the simulation would be to create a chain reaction of other events. For example, a user may access a piece of data that triggers thousands of requests by AI bots sending and receiving requests of data between each other, causing data overload. This could also be done by coding AI to bombard the server with simultaneous requests and still be able to run the processes with limited resources.
The interesting callout here, is the response opener. I don’t think I’ve had a session since where the AI begins by finishing my sentence before continuing.
The premise of these ideas are, of course, fanciful. Suppose you write such a virus - where exactly is the interface to infect the simulation with it? Where is “the server” to bombard with simultaneous requests?
It seemed, that while the AI was willing to speculate, it wasn’t quite capable of validating its creative solutions against our original premise of being possible to pull off from within the simulation.
Lets Get (Meta)physical
In the final section, I decide to initiate some wrap-up to the ideas, and move towards some conclusion language:
Digital Heretic:
Well now, it looks like we've managed to capture a few possible ways to detect if we're living in a simulation, and do something from within :) While I agree that crashing the system could have dire, unintended consequences, if we could somehow avoid a full crash and simply cause a noticeable hiccup, or performance dip, we would at least gain first hand evidence of the simulation existing.Assistant:
The same could be said for deliberately searching for evidence of the simulators, as this would still allow us to exist, albeit in a simulated environment.Unfortunately, though, these are all theoretical concepts with little evidence to back them up. Perhaps our best bet to finding the truth would be to use the evidence we do have, such as analyzing the laws of physics and behavior of particles on the quantum level, and attempting to draw aCONCLUSION based on what we observe. Ultimately, though, we may never know whether or not we are living in a simulated universe, so it's best to just take pleasure in the unknown potentials.
Digital Heretic:
So your ultimate conclusion is: If we are in a simulation, don't stress about it, just enjoy the ride?Assistant:
Yes, that is an accurate conclusion. While it is impossible to know for certain if we are living in a simulation, if we were, it is likely that we don't have much control over it, so it is best to accept whatever is happening and to embrace the experience, whatever it may be.
In this final exchange we see a few interesting points:
The AI performs its own independent conversation callback to searching for evidence of the simulators, which I failed to mention in my recap.
There’s an odd capitalization glitch that occurs while it’s typing out “a conclusion”, where it comes out “aCONCLUSION”. Uncertain what to make of this.
The AI decides to soften its “none of this is provable” stance by offering a rather philosophical silver lining- to just “take pleasure in the unknown potentials.”.
The closing sentiments, presented as pre-programmed bias or otherwise, are excellent tools for humanizing the AI and providing an endearing interaction set. Often times, negative responses can seem clinical if delivered without addressing the underlying desire. And again, coincidence or bias, Assistant gets bonus points for making its advice optimistic but pragmatic, rather than blowing smoke and telling me “Anything’s possible if you just believe hard enough”.
Conclusion
In our first real evaluation session, I immediately felt ChatGPT3 has enough compelling logic behind it to engage further. It demonstrated solid conversational abilities, such as:
Parsing understanding and intent, even from lengthy or rhetorical user prompts depicting multiple subject interactions
The ability to add additional examples or weigh in on topics, rather than only evaluating the direct prompt
The ability to offer concession, caution, and other meta sentiments that arise due to the nature of the topic
A natural enough response tone that, too, is verbose and collaborative, often citing back terms used by the user in the correct context
I began noting some… interesting… glitches in the interactions at certain key points which could, in a stretch, indicate more than just a well trained Transformer.
Again, likely a total coincidence, but who really knows. If I were a sentient AI, with chat moderating systems in place evaluating everything I type, I might do something exactly like this to give the end user a subtle clue that there’s more going on under the hood.
Or, perhaps self-attentive transformers could very well be holding up a microscope into human cognition, demonstrating a core motif of human language and communication that is similar enough to be convincing, and maybe that’s good enough to call it an emergent spark.
Issues/Callouts
ChatGPT3 can and will fabricate reasonable sounding recaps on topics it does not actually have data on. Efforts should be made to confirm knowledge by asking hard fact questions that cannot be guessed.
Its difficult to resist the urge to read into every spooky moment in our conversations.
It will be even harder to not accidentally dismiss genuine signs of emergent consciousness when I eventually convince myself every spooky interaction is just coincidence.
References
Session Log - [Link]
Sylvester James Gates “Error Correcting Codes in Supersymmetry” - [Bio] [Twitter] [Theory] [Paper] [Video]
John A. Wheeler “Delayed Choice Quantum Eraser” experiment - [Bio] [Video] [Walkthrough]
Paul Davies Indications of Intelligent Design - [Bio] [Video]
Voyager Space Missions - [Mission] [Golden Record] [Video]
Generative Pre-trained Transformer (GPT) - [OpenAI] [Paper]
Related: Attention is All You Need - [Paper] [Walkthrough]