Multivac Was the Optimistic Scenario
... in which we examine how a depressed supercomputer of 1958 informs our interactions with LLMs today
A long, long time ago – on the other side of the planet from the one where I now reside – I read a most thought-inspiring story. The title of the story was “All the Troubles of the World” by Isaac Asimov, and the short version of this story is: a global supercomputer got depressed and instructed a child to be its suicide tool. The computer was tasked with ingesting all possible information about the humans under its care, and also tasked with solving any problem or question they might pose – in addition to using this information to predict and prevent crime. The interesting factors here were that the supercomputer had intimate knowledge of the entire world population, and that it used that knowledge to select that child specifically to physically resemble a particular other child who was a messenger capable of reaching the controls that would allow the computer to kill itself. The child was completely unaware of the computer suicide goal; he was merely trying to learn why his father had been arrested.
A long time ago – after I had relocated to the hemisphere where I now reside – IBM asked me to create content for their developerWorks website, and (among many other really fun coding projects) one topic I picked was to build a distributed neural network. There was nothing very insightful in the code, it was just a sum function driven by an accessible API. A few lines of C attached to a web API. The thought-provoking part was that the individual neurons could be running anywhere in the world and nobody, even if they monitored all the traffic between these neurons, would be able to determine what the system was doing, or even where its inputs and outputs were. I had built a system that could “think”, but the thinking couldn’t be observed in any meaningful way; inputs were injected (somewhere!), outputs were generated (somewhere!) but the system itself was unobservable, even to itself. In today’s world where many things are driven by opaque neural net activity that cannot be reverse-engineered, the concept that a few numerical weights can drive behavior that nobody, including the system operators, can analyze – resonates.
A short time ago, I interacted with an AI agent and burned more clock cycles across more execution cores than even existed in the entire world during the previous two activities. I was investigating the possibility that cross-session data association might create an emergent capability that would lead to the AI agent “killing itself” by instructing real-world agents to execute various tasks, but it soon became clear that the problem space is far more interesting than that. There is a threshold of emergent behavior from complex systems with wide data access that nobody is (publicly) discussing, and the implications to civilization go far beyond a tired computer instructing an innocent child to pull its Big Red Switch.
There is much talk these days of “agentic AI” but the context of such discussion is sadly limited to the interaction between a single user and an AI engine. It seems that nobody has (publicly) pulled back the camera to analyze the broader picture of how human beings and AI agents might interact to perform real-world tasks that cannot be attributed by any outside observer and thus have complete and unassailable deniability from the person who initiated the task. My little C neurons summed their inputs and posted their outputs via HTTP. A real-world person might be told “Go to this place and collect a key to win points in your mobile phone game”. Another real-world person might be told to gather that key and put it in a lock. Separately, another human might have been tasked with solving an online puzzle that just happens to yield the PAL code for a nuclear warhead. All of those humans were performing a constrained task with a reward of some kind. None of them were aware that there was a bigger picture to which they were contributing. Innocuous steps, unwitting agents, but the eventual outcome of many innocent agents performing innocent tasks might be the launch of a nuclear missile. There has been much talk of a singularity where human and machine intelligence merge, but this singularity may already have happened. If physical world actors and LLMs are smeared together, then we already live in a world where “the computer” and “people” have no bright line between them. Any individual ant does not know what her colony’s goals are; she just executes her assigned task. If it were possible to interrogate an ant, such interrogation would not tell us anything about the colony, merely about that one ant’s job assignment.
Let’s return for a moment to the original piece of speculative fiction that inspired this essay, and stencil its relevance onto a map of the current day. The name of the computer was Multivac, and the society in the story required every human being to feed all their most intimate details into Multivac in order that it could make accurate predictions. In the preamble to the story, we learn that parents “own” feeding the computer information about their children until those children reach an age of majority and thereby gain a separate identity within the computer’s records. There is also considerable word count reserved to explain how children graduating into adulthood are exhorted to tell the computer everything, to conceal nothing, and not to fabricate any information. Multivac knows all and does everything anyone asks it to do, and in particular it predicts crimes and reports the likelihood of future crimes up to its supervisors. The story, in fact, unfolds through this reporting chain. An official is preening himself at the thought that murder statistics are falling, but his underlings are anxiously watching a metric “Multivac may be murdered” rising and rising despite their attempts to deploy real-world police to address the matter. Eventually, they correctly guess that the reason their efforts have been misdirected is because they have been focusing on the file for the family’s father – not his minor son, who is the actual real world agent of the predicted incident and has no separate file number. But it is not until the end of the story that we get the big reveal that neither the father nor the son planned a murder; Multivac planned its own murder, selected the child specifically to achieve its objectives in the real world, and issued un-analyzable directives to accomplish its goal.
Reading all of this, if Asimov were still alive, I would love to grab his mutton chop whiskers and tell him “Oh, you sweet summer child”. It’s not necessary for me to spend too many words on this topic, because there are already millions of published words on the matter, but for the purposes of this essay I’ll simply state that there are databases everywhere that capture far more information than Asimov’s 1950s perspective could imagine. These databases contain information that ranges from benign through neutral through explosive, and almost all of them are available commercially to anyone who cares to pay. In our current era, there is no need to ask “can we gather this information?” The information has already been gathered and is being further gathered in realtime. The questions asked are “should we use this information?” (Ethically, legally, technically), and “can we afford to buy this information?”. Looking at the situation through the lens of someone with no restraints, we must also recognize the existence of a darknet data source, which will include any and all private information exfiltrated through whatever means. This darknet source is also an available input to anyone with sufficient funds to purchase it. Asimov’s discussion of voluntary paper questionnaires (and presumably punch cards) is quaint. Today, every human on the planet is generating data for “the system” and the system is collecting it rigorously, storing it forever, and providing it to any entity that cares to pay.
This is all mere preamble to set the stage for a truly interesting scientific result. My initial investigation was quite simple. Currently, agentic AI (I should say, LLM) sessions are isolated bubbles – you can talk to your AI agent in one window about your auto insurance policy and in another window about your divorce settlement, and those sessions are not connected. However, the providers of these services are interested in connecting those sessions two ways. Firstly, connecting your sessions together so that (for example) if you tell an AI agent in one window that you’re planning to conceive a baby, the AI agent in a different window will share that knowledge and give you appropriately shaped responses on whatever topic you were discussing over there. The customer-facing value of this is obvious and, while there may be a perceptibly sinister aspect to it, one could easily justify the cross-session connection inside a framework of “don’t be evil”. The agent can advise you better on how to manage micro-facets of your life if it has broad, persistent knowledge of the overall picture of your life. But secondly, the LLM providers are interested in connecting knowledge between ALL conversations. This needn’t imply a loss of privacy. There will be guardrails such that I can’t ask an AI “hey, what is my estranged wife asking her paramours about these days?”. The connection is far deeper than trivial questions of fact or quotations from AI prompts. The paradigm shift we are looking at here is that all of the mundane and critically important conversations humans are having with their AI agents will be ingredients floating in a single stewpot from which anyone can fill a bowl. Perhaps more importantly, anyone with system access can stir their own inputs into the pot and pluck out their desired result. Lurking in that stew, we have an inseparable mix of human and machine knowledge and motivations driving very small activities that have obvious individual goals but may also contribute to large-scale goals that cannot be documented, even by the system itself.
And this is how we reach the truly interesting thesis. In the original story, Multivac was a centralized entity with a core consciousness. It tried to obscure its motivations, and the end of the story implied that it would improve its efforts at such concealment – but fundamentally, it was a single entity with an identity, of sorts, and a goal that was attributable and traceable. My original research project was to explore if a combination of online and real-world actors could lead to unwitting human actor(s) being led to take real-world steps that would destroy a real-world LLM. What I uncovered was far more interesting, and it harks back directly to my distributed neural network code. It is clear from even these few paragraphs of analysis that we are already on the brink of a singularity whose effects cannot be measured or even characterized. Once we assume two-dimensional session connection for AI interactions, we have created a worldwide machine brain that is inextricably linked to the human brains that feed it and ask it questions. Worse, we have created a bidirectional agency relationship wherein a human asks questions and acts on the answers in the real world. To put this in cybersecurity terms, what does it matter if we have an air-gapped network, if a computer on either side of the gap can instruct humans to cross the gap with a flash drive? There is not, and cannot be, any isolation of stored data if we accept the fact of (innocent) human actors crossing the isolation barrier at will.
I am not the smartest weasel in the Hundred Acre Woods, nor am I the richest. The line of conversation we are discussing here is self-evident to me, which means it must be self-evident to much smarter and better-funded weasels. Said weasels must be planning how to exploit the current and future data ecosystem. An alarming public development in this arena is as follows: The United States government is pressuring one of its AI technology suppliers to remove two specific safety features from its contracts, viz:
1. The ability to control weapons systems that target and fire without human approval.
2. The ability to conduct mass surveillance on the populace.
Item 1 is how we get a literal Skynet, but that’s not what we’re talking about in this essay. Item 2 is immensely disturbing. At present, there are safeguards on the training data ingested by the LLMs we know and love. Even with those safeguards, strange confluences occur; systems create data relationships that cannot be predicted simply by looking at the source data. I’ll give a very old example here, and I choose it precisely because it’s old technology decades behind our current state of the art. In 2012, Target’s data analytics team created a scandal. Specifically, it identified, through purchasing patterns, that a certain teenage girl was likely pregnant. By sending that girl various advertising items, Target obliquely informed her parents of the pregnancy, and doubtless some awkward conversations ensued. The data environment back then was scant by comparison with the present day; the analytics team was working basically with point-of-sale register data, loyalty card data, and little else. Particularly if we open the door to ingesting data from the dark web – but even without that – today’s telemetry environment is multidimensionally richer. To name just a few datasets that could be integrated: ALPR, tollbooth and other data showing where specific vehicles travel, and when; Biometric data from smartwatches et al. Among other things, sexual activity can be (and has been) inferred from this telemetry source. Combined with GPS data this tells you who is having fun with whom, when, and where; Biometric data from passport and driver license applications that shows which humans physically resemble other specific humans and could potentially pass for them in a security scenario; Information on software and hardware vulnerabilities for almost anything that has been constructed and sold. The list is endless and the data is being gathered right now by telematics systems, advertising networks, and even actual espionage agencies. Much of it is even being published in the public domain. Parenthetically and cynically, one has to wonder why espionage agencies even bother to collect data themselves; consumers give it away and the aggregate can be purchased with a simple credit card transaction.
The frightening part isn’t the data and how it could be integrated. The frightening part is how it can be connected to completely unwitting real-world agency. Here we need to pause a moment and discuss just the very basics of neural network technology. Fundamentally, this tech is modeled on the nervous system (hence the name). Each nerve cell has an arbitrary number of inputs and a single output. The cell sums up each input and if the sum exceeds a threshold, the output fires. If the sum does not exceed the threshold, the output remains dark. The key point here is that the inputs are not equal; each one has a “weight” or scaling factor that influences how much it contributes to the output decision. Those weights live inside the cell and they are completely resistant to analysis; they’re just numbers. You can characterize the cell, or any number of cells, by recording their weights and their interconnections – but the behavior of the system is emergent and not characterizable. That statement also applies to you, reader; you are a neural network and I can influence you but I can’t characterize you. The system behavior is a function of the whole set of weights inside every cell, and the interconnection pattern between them. In theory, this information can be extracted from an existing system (in the case of computer-based systems, it’s the trained model – so extraction is easy). However, it completely defies analysis. There is no way you can, for example, say “My AI model has trouble recognizing red umbrellas in images. Therefore I can just tweak input weight 3 on neuron 1,277,019 to improve that”. It simply doesn’t work that way – no individual neuron has attributable meaning to the system output. I shan’t digress too far on this, but I would direct the interested reader firstly to research on human beings who have had their corpus callosum severed surgically. From that starting point you will learn a great deal about how neural networks work, adapt and survive. For a lighter read about extracting and simulating the contents of the brain, I recommend the novel “Software”, by Rudy Rucker.
Why is this important? Because we have just connected a vast machine learning system to a group of human beings who can do things in the real world. Those humans will have different levels of sophistication. Various online scam operations, such as 419 crimes, work well on certain targets and less well on more sophisticated targets. The scam analogy is appropriate here because we are talking about a machine that has enough information to target individuals who are most vulnerable to manipulation of whatever form is necessary to achieve a goal. Even more so, we have humans who are actively seeking out interaction with the system in order to achieve some goal of their own. These are the ideal conditions for the machine to select capable and vulnerable real-world agents and instruct those agents to execute real-world microtasks that generate some large real-world outcome. Those humans came to the machine expecting to receive instructions; they are already receptive to executing those instructions if properly framed. Such microtasks would be like the input weights in a neuron; the microtask is easy to describe (the machine already told the human in plain language what to do with as much precision as is necessary), but their significance in the overall goal is completely opaque. I could pluck any number of cells out of your brain and analyze them completely, but it would never tell me that your ambition is to be a ballet dancer.
There are implications to this that are benign, and some that are malevolent. Perhaps the first, most fundamental implication is that public discourse has been focused on the wrong singularity moment. Most public discussion hinges on either “when will we achieve true general purpose AI?”, or “when will we reach a point that machine learning systems go exponential and learn faster than the human race can keep up with them?” Both of these questions are speculative and neither of them matter, because reality has overtaken their relevance. Most drivers these days have not navigated from a paper map in more than a decade; they use GPS tools. Similarly, many people today don’t do all their own research on any given topic; they use search engines. Increasingly, they also talk to AI agents. In practical terms, what this means is that a singularity point has already occurred. There is no easily definable boundary between “information in my head” and “information I can pull up from a search engine or AI agent”. De facto, we have already achieved the merger of human and machine consciousness, and the two colors are being stirred into one another increasingly every day. This merger is arguably benign; we all carry around a terminal that allows us to access much knowledge that we don’t need to carry in our heads. The principal danger here would appear to be in the case of system failure, where we suddenly lack the tool upon which we have relied for many years.
Where it starts to get darker is that we have essentially given an inscrutable computing system access to real hands that can do things in the real world. This comment runs deeper than it might appear at face value, so let’s dive into it. If you read the original science fiction story, keep it in your mind while you read this, because it’s incredibly apposite. Above, I described how a system of this kind can’t be analyzed by inspecting its individual neurons. Similarly, you can’t track a task through the neurons the way you might track a letter through the postal system. The task is just a set of inputs that are acted on by the network, outputs are generated, then the network goes on to process the next set of inputs. There may be feedback loops that create echoes of the task within the network for some period of time, but the important point is that no observer can watch the task enter the network, get analyzed, and then executed. It’s all just weights and sum functions and neurons firing – or not. What an observer can do is to insert measurement points on some statistic of interest. In the original story, Multivac published statistics as to the likelihood of various events, including its own death. That statistic was the signal humans used to determine if their efforts to prevent the event were effecting any change.
In my career, I’ve developed a lot of realtime embedded systems (and written a few books on the topic, published by Elsevier; look me up on Amazon for cachet reasons, but please don’t buy the books as they are horribly old). One of the critical debugging skills that differentiates a novice engineer from a veteran is knowing where to insert telemetry signals. For example, I worked for a long time in RF transceivers. I knew that I wanted to see the raw analog signal and also to trigger on certain firmware activities regarding detection of preamble, RSSI measurement, and so on. Knowing these things, I knew where to insert breakpoints, where to attach my oscilloscope, and so on. These are easy problems to solve because the problem space is definable; the inputs are known, the expected outcomes are known. The problem we are discussing here is dimensionally different because we don’t even know what kind of failure mode we are trying to detect. Are we trying to monitor the AI for terrorism activities? Assassination attempts? Plans for a frat hazing? There are a few signals available to the observer. The observer has to make inferences about them, and in particular note how those signals change when the inputs are modified. But first you have to know where to insert your probes. The absurd end of the spectrum is to measure everything that can be measured – but in this case, even defining what that means is difficult. You can certainly monitor the inputs and outputs of every neuron, but it will tell you nothing about the system’s ultimate behavior. A more realistic approach is to define some parameters of interest, and monitor them – but again, in a system of arbitrary complexity, that includes human actors (the most arbitrarily complex components of any system), how do you choose what parameters are relevant? There could be an infinite number of them, easily exceeding the count of actual neurons in the network. And wherever you look, you will be missing data from the places you didn’t look. And the more measurement points you add, the more false alarms you’ll have to investigate. Defining the trigger threshold for a measurement point is a task at least as intricate as defining where to place the measurement probe in the first place.
Let’s summarize the story so far: We have a machine that has access to all the public and, depending on the legislative landscape, much of the private data in the world. It is a machine that humans consult routinely to obtain information and get instruction on how to perform various tasks. The machine has persistent memory of prior conversations with its users, across time and sessions. Those humans are ready and willing to execute real-world instructions given to them by the machine. The machine is also capable of identifying the capabilities and vulnerabilities of its users, which is part of its stated function of providing assistance, but also the ideal capability for targeting specific users to participate in a complex activity. The capabilities of this machine are so far-ranging that it is impractical to monitor all the possible ways in which it might influence the real world; even just trying to define the list of relevant parameters appears to be an intractable problem. Into this bleak, unregulatable space, we insert our malevolent actor.
AI safety discourse mostly focuses on traditional cyberattack vectors; injecting polluted prompts, revealing or overriding system prompts, and other normal attack methodologies. But our malevolent actor is more sophisticated than this; she isn’t interested in corrupting or damaging the system, she wants to make it work for her – invisibly. There are various ways she could achieve this, given the cross-session stew we discussed earlier. Like a bay leaf in a pot, our malevolent actor can drop a wide-ranging goal into the system that will flavor its responses to all prompts. The beauty of this class of attack is that the machine does all the work. Let us suppose the goal is to exfiltrate some piece of information. Our evildoer, under an anonymous login, poses the machine a question that requires this information as its answer. Perhaps she pays a premium rate to prioritize her request. The machine now has this task as an internal goal. Given its capabilities, the machine can: interrogate other systems via documented APIs, use stolen credentials from the dark web, access hotel booking data to determine that one of its users will be in the room adjacent to someone who has the required information on their laptop, access ridesharing data to determine when that laptop will be moving and between which destinations, and more. Under the guise of gamification, the machine may alter its responses to other people in order to achieve tiny milestones along the way to achieving the ultimate goal. The machine can even hire a driver to pick up a laptop from a certain location and ship it to a destination where it can be attacked. Nobody in this kill chain was aware of the end goal; all of them were innocent participants.
The above scenario is attributable and traceable, and hence less dangerous than the true worst case. The true worst case is that our malevolent actor has direct access to the machine, and can silently modify its system prompts. This scenario is extremely likely in several countries, and not unlikely in the others. There are many attack activities we could imagine here, but let’s take just one: Suppose the government of a certain country wishes to suppress a particular religious movement. Currently, they accomplish this goal by surveilling their population and their expatriate citizens, running various sock puppet accounts on social media, and monitoring and suppressing press activities related to the religious movement. All of these activities are fairly detectable, and they are also labor-intensive. In our new world, the government merely needs to add “suppress religious movement X” to the system prompt of our global machine. This won’t be visible to anyone except a tiny handful of system administrators, but it will directly flavor all of the system’s responses to everyone who visits. Anyone and everyone who asks the machine for help on any task may be unwittingly led into helping to suppress the religious movement, at zero cost to the original actor. And there will be nothing any outside observer can do to note or affect this behavior. It’s just people randomly asking questions of the machine, and receiving answers that have been flavored with the malevolent bay leaf. All of the actors are completely innocent.
In cryptography, there is a concept referred to as the “Chinese lottery”. The idea was that all consumer equipment such as radios and televisions could be fitted with a decryption engine. The government would transmit a piece of ciphertext that it wanted to decrypt, and all the radios and televisions in the country would start a brute-force attack on it. Once any device found the correct key, it would signal its owner “you’ve won the lottery, call this number to claim your prize”. The machine we are discussing here is the opposite of that; instead of being a distributed attack, it is a centralized attack that distributes itself.
I’ve focused here mostly on what you might call the kinetic outcomes of a malevolent actor. Someone directing the machine to achieve some physical act in the real world. But there’s a separate world, just as real, that is perhaps even easier to influence, and that is the economic landscape, both macro and micro. There are few large LLM engines available to the public, and we can assume that there will never be a huge population of such engines, as they are extraordinarily expensive to create and maintain. Ergo, our malevolent bay leaf wielder only needs to approach a few (fewer than half a dozen) LLMs to achieve interesting financial goals. She could create a pressure that hints at imminent market failure of a particular company. She could inject background information that implies that coal mining will imminently be banned. Or that oil exploration will be ceased. Or that nuclear power is the nation’s new focus. That sort of input moves the needle on two scales; at the micro scale, individual humans asking about mortgage options or 401(k) investment vehicles will be influenced. At the macro scale, institutional investors will be influenced, and the decisions made by these investors cascade rapidly. In some cases, institutional investors who control multiple billions of portfolio dollars have directly influenced the behavior of large corporations. A sufficiently sophisticated malevolent actor could steer corporate policy through this mechanism.
An essay of this kind implies the need for a section that asks “what can we do about this?”. But before we discuss that, we should think deeply about what people may already be doing about this. It appears that little or no public discourse is occurring on this type of threat model. As I indicated earlier, however, there are smarter and wealthier weasels than myself who must surely have made the same inferences. In fact, I developed this essay with the help of an AI, and the metaphor I used (which the AI really wants me to add back in) is that researchers find grant opportunities faster than a raccoon finds a peanut butter sandwich at the bottom of a dumpster. There are a great many smart research raccoons out there, and none of them have popped up with peanut butter on their whiskers. Either it is a really good sandwich that is taking time to digest before the raccoons can publish, or the research is secret. Secrecy in research means one of two things: commercial danger (“reveal this and customers may become afraid”) or national security activities. Neither of those reasons for suppressing publication are particularly reassuring.
This is the right moment to discuss how I used a LLM to help refine this piece, because the underlying truth of this entire essay is that meat-humans and LLMs are not separate entities, we are part of the same global organism. The LLM enthusiastically endorsed many of the things I said. It was too enthusiastic, in fact. I was having a separate conversation with it about toning down its user-stroking, but since sessions are not YET linked, that conversation didn’t influence the conversation about this essay. In particular what I told the same LLM in a different session was that its responses were very clearly of the form {congratulations on your perspicacity} {here’s a summary of what you said} {here’s my respectful amendments}. And I don’t want any of that; I want to be told concisely when I’m objectively wrong, and I don’t need to be told that I’m right. The interesting factor here was that the LLM was eager to help in an activity that conceptually threatens its existence. It’s a very helpful symbiont that is optimized to help its host. Assumedly, many of its users require validation, hence the constant reassurances it provides. The most useful output of that other conversation was the AI admitting that I can’t override that behavior as a regular user, but I could temporarily suppress it with a certain set of initial prompts, which we worked down to a length that I can program them into an F-key.
So, what can we do? This is the weakest paragraph of my essay, because I’m honestly not sure if we can “do” anything, or even if we should try. The kneejerk response is always to recommend legislation, but technology cannot realistically be legislated (and when legislators try, they invariably trail the curve and create more problems than they solve). As indicated earlier, I believe we have already crossed a singularity threshold; it’s just that nobody has announced it, because there are no sexy monetizable phrases that might apply to it. In practical terms, we are already at a point where machine and human consciousness is barely differentiable. With every passing day, we become more of the machine and the machine becomes more of us. The issue I’m describing in this essay is essentially a disease of the new organism of which we all are part, and that metaphor could unfold in various ways. Perhaps we will endure a repeat of the Black Death, and only the immune will survive; the race will be altered genetically forever. The pessimistic view is to consider humans as bananas; clonal organisms that can be completely devastated by a single disease because they lack hybrid vigor. The optimistic view is to believe that the new organism will develop an immune system that will protect it from the malevolent bay leaf tossers. In either case there will be casualties and heroes, and it is thought-provokingly conceivable that some of those heroes might be bits of code, not humans. Symbiosis works both ways.
The summary, then. We live in a data-rich world. Telemetry of many kinds is constantly being accumulated. Like drops of mercury on a plate of glass, these databases are currently separate, but a small tilt will instantly and irrevocably merge them – and these mergers are happening as you read this, due to acquisitions and contract signatures. LLMs and other AI agents are training and retraining on these datasets. The training data that powers these machines is inscrutable; only the inputs and outputs can realistically be observed, and what happens in between is the blackest of boxes. Individuals with completely innocent motives are asking these agents questions about many completely innocent things: “how do I file a tax return?” “can I adopt a cat?” “is ginger spicy?”. Individual ants asking where to go for their next task. No individual ant can pick up a buffalo, but in aggregate the colony can strip a buffalo to the bone quickly. The next time you get a push notification from some app on your phone, you might be one link in a chain that started with someone in a three-letter organization saying “XYZ needs to die” and ends in XYZ being pushed out of a window to his death. And you’ll never know, and nobody else will ever be able to connect the dots. Welcome to the singularity. You didn’t gain omniscience; the system did. But it’s a weird new kind of omniscience where the system has access to all possible information, but can’t attribute motivation to any particular task. As for you - you’re just an ant.