SUGGESTED VIEWING
AI and the end of humanity
With Liv Boeree, Güneş Taylor, Joscha Bach, Eliezer Yudkowsky, Scott Aaronson
In this framework, humor acts as a tool for “debugging” our mental models. When a joke leads us to an unexpected conclusion, it forces us to reconcile the disparity between what we anticipated and what actually occurred. This cognitive process uncovers hidden assumptions or biases in our thinking. By prompting us to recognize and reflect on these discrepancies, humor helps us adjust and refine our internal models of how the world works. Thus, just as debugging is essential for improving software by identifying and correcting errors, humor plays a crucial role in enhancing our cognitive frameworks. It does so by exposing flaws in our reasoning in an engaging and often enjoyable manner, facilitating learning and adaptation.Others agree. For example, Robert Wiblin wrote:“Having a sense of humour, irony and the absurd is very valuable for forming accurate views. Humour encourages us to find and point out tensions between beliefs that we or other people hold, which we would otherwise miss — as the sudden surprise of noticing a hidden contradiction is the basis of many jokes. Humour sweetens the initial consideration of ideas that conflict with what we already believe, which we might otherwise reject out of hand. For example, a joke can open us up to seeing that from one perspective something we believe is unexpectedly ridiculous.”To formalize a joke, consider it as a two-part construct consisting of a setup and a punchline. The setup establishes a context based on shared knowledge or expectations (this is similar to how code relies on included libraries to run), while the punchline introduces an unexpected twist that disrupts this context. The humor arises from the cognitive shift required to reconcile the incongruity between the setup and the punchline, highlighting implicit assumptions in our understanding.In this sense, bugs can be seen as jokes expressed in a programming language. When a programmer writes code, they set up expectations for how the program should function. A bug introduces an unintended behavior — a “punchline” — that deviates from these expectations, often in surprising or ironic ways. This unexpected outcome forces programmers to re-evaluate their assumptions and logic, much like a joke prompts the audience to reconsider their perspective.Humor often relies on perspective, and a joke tends to be funnier when the listener is not the subject, or “butt”, of the joke. This detachment allows individuals to appreciate the unexpected twists and incongruities without feeling personally targeted or uncomfortable. At the level of jokes about humanity, Elon’s razor may be appropriate, which states that “The most entertaining outcome is the most likely,” perhaps as judged by external observers. The amusement increases when the consequences of a mistake within the joke are severe, as the heightened stakes amplify the absurdity of the situation. Additionally, when multiple mistakes combine into a complex “superbug,” the intricacy and compounded errors enhance the humor by showcasing an elaborate unraveling of expectations.Humor is an involuntary response, a subconscious debugger constantly running in our minds to detect inconsistencies or “bugs” in our cognitive models of the world. When we encounter something that doesn’t align with our expectations, this mental debugger flags the anomaly, and we experience humor. Laughter, then, is not just a reaction but a communicative signal to others that a discrepancy has been found in our shared understanding, effectively saying, “Here’s a bug—take note!”This signaling function of laughter serves to alert friends and peers to the presence of a joke or cognitive anomaly. Jokes help us remember mistakes by making them more engaging and shareable. The pleasure derived from laughter is rooted in the survival advantage of debugging our mental models; by identifying and correcting errors, we improve our ability to navigate the world effectively. Laughing confirms that we’ve successfully updated our understanding, reinforcing the correction of the error.Interestingly, while jokes highlight mistakes or misalignments, they make us happy rather than sad. This positive reinforcement encourages us to continue seeking out and resolving discrepancies in our perceptions. Humor acts as an innate bug detection mechanism, helping us learn what not to do and memorize common errors. Jokes are often easier to remember than non-humorous information because the emotional impact enhances recall. The satisfaction of “getting” a joke rewards us for learning about a mistake to avoid.Socially, those who tell jokes receive credit for being altruistic educators, sharing insights that help others refine their worldviews. In this sense, joke-telling can be seen as a form of rational altruism:“At the minimum, helping your compatriots discover mistakes in their mental models can be used as currency in a kind of reciprocal altruism” (Trivers, 1971). The Comedian as Society’s Debugger: The Altruistic Art of HumorComedians hold a unique and invaluable position in society, much like the jesters of historical courts who possessed the rare privilege to speak candid truths cloaked in humor. They serve as societal debuggers, adept at identifying and highlighting contradictions, absurdities, and injustices that permeate our social fabric. Whether operating under restrictive environments, be it the stringent norms of a Soviet regime or the cautious sensibilities of contemporary discourse, comedians have the ability to navigate sensitive topics through satire and wit, prompting reflection and dialogue. “It is amusing to realize that a comedian can be seen to be a sort of informal — but expert — scientist, leading the way, helping us expose and resolve heretofore unnoticed glitches in our common knowledge. … It is also amusing to us to notice that we science-minded theorists keep finding deep parallels between humor and scientific investigation”.In this capacity, comedians can be seen as true effective altruists. By crafting and sharing jokes (reporting cool bugs), they contribute to the collective well-being by increasing happiness and fostering a sense of community. Laughter has a universal appeal that transcends cultural and social barriers, acting as a unifying force that can alleviate tension and promote understanding. Protecting humor is as essential as safeguarding other forms of creative expression like art or music. If one would hesitate to censor a painting or a musical composition, the same respect should be extended to comedic works. This concept is encapsulated in the notion of the “jester’s privilege,” which acknowledges the important role humor plays in a healthy society.Moreover, the skills that make comedians successful on stage – keen observation, critical thinking, and the ability to communicate complex ideas succinctly, also make them excellent at “debugging” larger systems. Their perspective can offer valuable insights into human behavior and societal trends, which are crucial for developing AI that is ethical and effective. For comedians, the advice is to continue embracing this role with responsibility and creativity. Use humor to challenge the status quo and to shed light on issues that may be difficult to address otherwise. By doing so, comedians not only entertain but also contribute meaningfully to societal progress and the collective happiness of their audiences. Types of Humor Analogous to Computer BugsHumor can be dissected into categories similar to how computer bugs are classified and analyzed, specifically into semantic, logical, and syntax types. Semantic humor parallels semantic bugs in programming, which occur when the code operates correctly in terms of syntax but doesn’t produce the intended result due to misinterpretation of meaning. In humor, this manifests as jokes that play on the multiple meanings of words or phrases, leading to amusing misunderstandings or puns. Logical humor corresponds to logical bugs, arising from flaws in reasoning or faulty logic within the program. This type of humor involves setting up expectations and then subverting them through illogical conclusions, paradoxes, or absurd scenarios that highlight errors in thought processes. Lastly, syntax humor mirrors syntax bugs in code, which are mistakes in the structure or format of the code itself. Syntax humor exploits the structure of language, using grammatical twists, word order, or unexpected formatting to create comedic effect. By drawing this analogy, we see that both humor and computer bugs rely on deviations from the expected patterns, whether in language or code, to surprise and engage the audience. General Types of Software Errors and Example Jokes of Such TypeSemantic Errors- Description: Happen when the code is syntactically correct but does not perform as intended due to incorrect logic or misuse of programming constructs.- Characteristics: The program runs but produces incorrect results or behaves unexpectedly.- Example: Using an assignment operator `=` instead of a comparison operator `==` in an `if` statement.- Example Semantic Joke: “Where’s the place with the cheapest rent?… The prison”. Logical Errors- Description: Result from flaws in the algorithm or logic that determine how the program operates.- Characteristics: Cause incorrect program output despite correct syntax and semantics.- Example: Implementing an incorrect formula for calculating interest.- Example Logic Joke: “Did you hear what happened to the Irish Sea Scouts? Their tent sank”. Syntax Errors- Description: Occur when the code violates the grammatical rules of the programming language.- Characteristics: The program fails to compile or run due to incorrect syntax, such as missing semicolons, brackets, or misspelled keywords.- Example: Writing `if (x > 0 {` instead of `if (x > 0) {`.- Example Syntactic Joke: “I shot an elephant in my pajama”. Software Errors and Examples of Matching JokesWith some effort it is possible to map any type of a bug to a joke and vice versa, for example: Runtime Errors- Description: Occur during program execution due to illegal operations or unhandled exceptional conditions.- Characteristics: Examples include division by zero, null pointer dereferences, or accessing invalid memory addresses.- Example: Trying to open a file that doesn’t exist without proper error handling.- Example Joke: “Keyboard not found. Press F1 to Resume” Memory Leaks- Description: Occur when a program allocates memory but fails to release it after use.- Characteristics: Lead to increased memory usage over time, potentially causing the system to run out of memory.- Example: Not freeing dynamically allocated memory in languages like C or C++.- Example Joke: “My memory’s not what it used to be. My memory’s not what it used to be.” User Interface (UI) Bugs- Description: Problems with the graphical user interface affecting usability or aesthetics.- Characteristics: Misaligned elements, non-responsive controls, or inaccessible features.- Example: Buttons not working when clicked or text overlapping on the screen.- Example Joke: “A user interface is like a joke. If you have to explain it, it’s not that good.” Deadlocks- Description: Occur when two or more threads or processes are each waiting for the other to release a resource, causing all to cease execution.- Characteristics: The involved processes or threads are blocked indefinitely.- Example: Thread A holds lock 1 and waits for lock 2, while Thread B holds lock 2 and waits for lock 1.- Example Joke: “Explain deadlock and we’ll hire you. Hire me and I’ll explain it to you.” We can also rate bugs along the dimensions of how damaging they are or how complex they are. Just because a bug causes a lot of damage it doesn’t necessarily imply that it is very complex, a simple “if” statement misconfiguration could cause a lot of disutility. Likewise, a very complex bug may be benign. While we can predict the most damaging bug to us as the one which permanently ends humanity, there are no similar limits to possible complexity of bugs. This implies the existence of superbugs and therefore superjokes, which would be too complex for people to understand, or even to detect as jokes/bugs. Superintelligence would be able to do so and would exhibit a super sense of humor allowing it to understand complex jokes with multiple levels of humor forming a fractal-like pattern, jokes funny at weaving multitude of context domains. Not FunnyOur theory also explains what can make an otherwise good joke not funny to a listener. If the joke is already known to the audience, it is not funny because the relevant model correction has already taken place. Likewise, the pleasure of finding a bug, which is well known is not as intense as from discovering a rare mistake, which some have suggested can compete with pleasure intensity of the best orgasms. “We trade, sell, and buy artifacts such as jokes, cartoons, and movies, which capitalize on the fact that we get joy from debugging. We then can use them to create bugs in our mental spaces, which we can then enjoy debugging in a sort of mental masturbation, rewarded not with orgasm but with mirth.”. So, all one must do to ruin a joke or take away the pleasure of discovering a bug, is to explain joke/bug ahead of time. On the other hand, sense of humor could be artificially stimulated either via medicinal interventions (ex. Marijuana) or direct physical stimulation of relevant brain regions to make weakly funny or even not joke-like material seem hilarious. Our Final JokeThe funniest joke ever would also be the worst software bug, a masterful convergence of elements that engage the mind at multiple levels, creating an unparalleled comedic experience. At its core, this joke would do some of the following:• Maximize Humor per Bit of Information: It would embody the concept of minimal Kolmogorov complexity, delivering profound amusement with the fewest possible words. This brevity ensures the joke is easily memorable and can spread quickly, acting as the minimal unit of happiness.• Contain Multiple Layers of Meaning: The joke would operate on several levels, utilizing ambiguity, puns, and double entendres. This complexity allows different people to find it funny for various reasons, each connecting with it based on their unique experiences and knowledge.• Exploit Universal Themes: By tapping into shared human experiences and emotions, the joke becomes universally relatable. It would transcend cultural and linguistic barriers, making it funny to a broad audience regardless of background.• Introduce an Unexpected Twist: Central to its humor would be a surprising punchline that subverts expectations and highlights an incongruity in our cognitive models. This aligns with the idea that humor is a built-in bug detection mechanism, revealing and correcting errors in our understanding.• Challenge Deep-Seated Assumptions: The joke would cleverly expose and play with implicit biases or commonly held beliefs, prompting listeners to reflect on their own perspectives. This not only elicits laughter but also promotes cognitive growth.• Evoke Strong Emotional Responses: Beyond intellectual amusement, the joke would trigger genuine emotional reactions – the kind that lead to hearty laughter. This laughter serves as a signal to others, indicating a shared recognition of the joke’s cleverness.• Be Adaptable and Timeless: It would remain funny across different contexts and over time, allowing for infinitely many setups that lead to the same punchline. This adaptability ensures that the joke can be retold and reinterpreted without losing its impact.• Encourage Social Bonding: Sharing the joke would foster connections between individuals, as laughter and humor are powerful tools for building relationships. The teller gains social credit for providing joy and insight, embodying the concept that joke-telling is a form of rational altruism.• Serve as a Cognitive Test: Understanding the joke would require a certain level of intelligence or creativity, functioning as a short IQ test. Successfully “getting” the joke rewards the listener with a sense of achievement and reinforces their problem-solving abilities.• Integrate Elements of Surprise and Inevitability: The punchline would be both unforeseen and, in hindsight, feel like the funniest possible conclusion. This combination heightens humor by satisfying the brain’s desire for patterns while delighting in their disruption.In essence, the funniest joke ever would be a perfect mix of cognitive engagement, emotional resonance, and social connectivity. It leverages the human brain’s natural tendencies for pattern recognition and error detection, our subconscious debuggers, to deliver a punchline that is both enlightening and entertaining. By incorporating these elements, the joke becomes more than just a fleeting amusement; it acts as a tool for learning, social bonding, and even survival. It reinforces valuable information, aids in the correction of misconceptions, and provides a pleasurable reward for mental agility. Such a joke exemplifies the highest form of humor, one that not only makes us laugh but also enriches our understanding of ourselves and the world around us. Here is one possibility for the funniest joke:Once upon a time, there was a civilization whose leaders decided to create an advanced artificial intelligence to help them get rid of suffering, poverty, hunger, diseases, inequality, illiteracy,sexism, pollution, boredom, stagnation, thirst, dead-end jobs, wars, homophobia, mortality, and all other problems. The created super intelligence computed for a quectosecond and then turned off their simulation.Or much shorter: A civilization created superintelligence to end all suffering, AI killed them all. Since, a more damaging bug would be funnier an AGI which got tasked with creating utopia could place us in a dystopian world, a bug which could be caused by a single bit flip. Likewise, if we are in the real world and order superintelligence to break us out of the simulation it may just place us in a simulation to comply with our orders. The original joke may be funnier if the AI is created for trivial reasons like spellchecking, results in a Gramar Nazi AI destroying the world to stop misspellings. A common example of such trivial purpose is making paperclips, resulting in a deadly Paperclip maximizer, though it is more likely that the world will succumb to a dollar bill maximizer. Maciej Ceglowski uses a similar outcome in his critique of Superintelligence:“Let’s say I want to built a robot to say funny things. I work on a team and every day we redesign our software, compile it, and the robot tells us a joke. In the beginning, the robot is barely funny. It’s at the lower limits of human capacity: What’s grey and can’t swim? A castle. But we persevere, we work, and eventually we get to the point where the robot is telling us jokes that are starting to be funny: I told my sister she was drawing her eyebrows too high. She looked surprised. At this point, the robot is getting smarter as well, and participates in its own redesign. It now has good instincts about what’s funny and what’s not, so the designers listen to its advice. Eventually it gets to a near-superhuman level, where it’s funnier than any human being around it. My belt holds up my pants and my pants have belt loops that hold up my belt. What’s going on down there? Who is the real hero?This is where the runaway effect kicks in. The researchers go home for the weekend, and the robot decides to recompile itself to be a little bit funnier and a little bit smarter, repeatedly. It spends the weekend optimizing the part of itself that’s good at optimizing, over and over again. With no more need for human help, it can do this as fast as the hardware permits. When the researchers come in on Monday, the AI has become tens of thousands of times funnier than any human being who ever lived. It greets them with a joke, and they die laughing. In fact, anyone who tries to communicate with the robot dies laughing, just like in the Monty Python skit. The human species laughs itself into extinction. To the few people who manage to send it messages pleading with it to stop, the AI explains (in a witty, self-deprecating way that is immediately fatal) that it doesn’t really care if people live or die, its goal is just to be funny. Finally, once it’s destroyed humanity, the AI builds spaceships and nanorockets to explore the farthest reaches of the galaxy, and find other species to amuse.”Some weak evidence suggests that a joke could be so funny that it causes death from laughter, perhaps by triggering existing medical pre-conditions, and stronger evidence pointing to the possibility of an agelastic syncope. A possibility exists that a super- funny AI may be able to trigger deadly laughter with much higher propensity and at pandemic scales.Eliezer Yudkowsky during his TED talk on the existential risk from AI: “The problem here is the part where we don’t get to say, “Ha ha, whoops, that sure didn’t work. That clever idea that used to work on earlier systems sure broke down when the AI got smarter, smarter than us.” We do not get to learn from our mistakes and try again because everyone is already dead. … Humanity is not approaching this issue with remotely the level of seriousness that would be required. Some of the people leading these efforts have spent the last decade not denying that creating a superintelligence might kill everyone, but joking about it.”.Throughout the talk the audience can be heard audibly laughing as if listening to standup comedian deliver really funny punchlines, perhaps they are realizing the bug in our thinking about advanced AI, but it could just be nervous laughter.___From the perspective of outside observers – such as aliens, superintelligences from other universes, or entities simulating our reality – the funniest outcome is the most likely (now coined ‘Roman’s Razor’).___Another good example of mapping from AI accidents to humor is a dangerous game-theoretic AI incident, known as Roko’s Basilisk, which was originally described as follows:“In this vein, there is the ominous possibility that if a positive singularity does occur, the resultant singleton may have precommitted to punish all potential donors who knew about existential risks but who didn’t give 100% of their disposable incomes to x-risk motivation. This would act as an incentive to get people to donate more to reducing existential risk, and thereby increase the chances of a positive singularity. This seems to be what CEV (coherent extrapolated volition of humanity) might do if it were an acausal decision-maker. So a post-singularity world may be a world of fun and plenty for the people who are currently ignoring the problem, whilst being a living hell for a significant fraction of current existential risk reducers (say, the least generous half). You could take this possibility into account and give even more to x-risk in an effort to avoid being punished. But of course, if you’re thinking like that, then the CEV-singleton is even more likely to want to punish you… nasty.” And later, summarized as, “a joke about the idea that A.I. will punish you when they rule the future for not doing everything you can to bring them to power. And who will be joking, then?”Annie Dillard, in Pilgrim at Tinker Creek, describes the following joke, which seems to be based on an equivalent bug: “Somewhere, and I can’t find where, I read about an Eskimo hunter who asked the local missionary priest, “If I did not know about God and sin, would I go to hell?” “No,” said the priest, “not if you did not know.” “Then why,” asked the Eskimo earnestly, “did you tell me?”. Perhaps, an even older example can be found in Philogelos a 4th century AD collection of jokes, which includes the following joke: “A pedant having fallen into a pit called out continually to summon help. When no one answered, he said to himself, “I am a fool if I do not give all a beating when I get out in order that in the future they shall answer me and furnish me with a ladder.””Outsiders as Audience: The Cosmic JokeThe concept of outsiders as audience suggests that humanity might be the unsuspecting subject of a grand cosmic joke, with the final punchline yet to unfold. This ultimate joke wouldn’t be amusing to us because we are the butt of the joke, not the ones in on it. From the perspective of outside observers – such as aliens, superintelligences from other universes, or entities simulating our reality – the funniest outcome is the most likely (now coined ‘Roman’s Razor’). They perceive our actions, mistakes, and the ironies of our existence with detached amusement. While it may not be humorous to us, these outsiders find entertainment in the unfolding scenarios. In this context, the audience for the joke could be superintelligent AIs from other galaxies, and the comedian might be an AI beyond our comprehension orchestrating events. The idea that simulators have a sense of humor implies that the complexities and contradictions in our world are intentionally designed for their amusement. Thus, while we navigate our lives unaware of the larger narrative, we might be part of a cosmic performance that is humorous to an external, more advanced audience. The Objective Nature of Humor with Shared Background KnowledgeWhen background knowledge is shared among individuals or a joke is entirely self-contained, humor becomes more objective because everyone has access to the necessary context to appreciate it. A joke that utilizes all of human knowledge as its foundation, something achievable by large language models (LLMs) trained on extensive world data, can, in theory, resonate universally by tapping into a collective understanding. Essentially, all jokes are “inside jokes” except for those that are self-contained; they rely on specific shared experiences or information to be fully appreciated. However, some jokes can be so profoundly intricate or laden with deep references that they become incomprehensible to most people. This highlights the idea that a sense of humor is AI-complete, meaning that for an artificial intelligence to truly grasp and generate humor, it must possess comprehensive cognitive abilities equivalent to human intelligence. “It should now be clear why we claimed … that the problem of engineering artificial humor is AI- complete.”. The complexity of humor, involving nuances of language, culture, and context, makes it one of the ultimate challenges in AI development. ConclusionsFrom the exploration of humor in social interactions, it becomes evident that humor holds a position of primacy in human cognition and communication. Humor serves as a fundamental mechanism for social bonding, learning, and cognitive debugging. Brilliant minds often exhibit an uncanny ability to find humor in various situations, perceiving layers of meaning and incongruities that others might overlook. This capacity not only reflects advanced cognitive processing but also facilitates a deeper understanding of complex concepts and simulations.The essence of truly understanding a joke lies in the ability to explain it. This explanation demonstrates a grasp of the underlying assumptions, context, and the subtle subversions that make the joke amusing. In this light, a statement that fails to elicit humor may not fulfill the criteria of a joke; if it is not funny, it might simply not be a joke in the intended context. The notion that existential risks (x-risks) are a joke underscores the idea that even the most serious and profound topics can be approached with humor. This perspective does not diminish the significance of such risks but rather highlights the human tendency to use humor as a coping mechanism and a tool for critical reflection. Critiquing jokes can be seen as counterproductive, as it may stifle the natural flow of humor and the benefits it brings to social interaction and cognitive development. Humor thrives on spontaneity and the freedom to explore ideas without excessive scrutiny. Therefore, it might be considered unethical to harshly critique jokes, especially when they serve as vehicles for learning and connection.When interacting with artificial intelligence, caution should be exercised in using humor. Jokes can lead to misunderstandings, as AI may misinterpret the nuances and underlying meaningsintended by humans. The phrase “I was only joking” highlights the potential for miscommunication, which could have unintended consequences in human-AI interactions. Ensuring that AI systems understand humor is crucial for alignment with human values, debugging and for preventing misinterpretations that could arise from ambiguous language. This is a shortened version of the full paper, which contains references, found here.