The Future of Life Institute (FLI) finances and publishes a set of detailed stories of positive visions of the future of artificial intelligence (AI), which originate a number of ideas that go on to influence AI policy in the real world.

One major idea to come out of the worldbuilding contest was the concept of Demonstration of Cataclysmic Trajectory ("Demoncatting"), a method for demonstrating whether, and to what extent, an AI system is dangerous, without unleashing the danger upon the world.

Apple releases a slim, lightweight Virtual Reality (VR) headset with greater than 60 pixels per degree of image detail (source), which is about as good as the absolute maximum perceptible level of visual accuity. Described as being "like reality". Despite Apple's protestations that it is only meant for short sessions of communication and entertainment, it precipitates a fairly profound change in the way work and socialization will occur in the coming years.

Unsubsidized prices of utilities-scale lifetime operating costs of photovoltaic solar power prices to 36 USD per megawatt hour. Operating less than four times cheaper than the operating costs of coal power plants, which are rapidly decommissioning (source). Though this doesn't factor the cost of energy storage for demand levelling, technologies for cheap storage are in the works. Lithium batteries are rapidly decreasing in cost, and much cheaper rust batteries are rapidly emerging.

The concepts from the FLI's Worldbuilding competition are synthesized into a set of short stories, graphic novels and audioplays.

The world's largest aspiring autonomous car manufacturer has already thoroughly hitched their pride to solving the difficult robotics-AI challenges of autonomous driving without pre-mapping or manual interventions from any central control center. This year sees further commitments to robotics-AI with the demonstration of their humanoid robot, walking onto the stage unassisted and performing an imitation of the same dance that was done by a human actor in the 2021 "AI Day" presentation. With their manufacturing skills, and with the demonstrations of their training processes, It becomes very plausible that Tesla will be an enduring lead figure as the world begins to adopt robotics in delivery, construction, farming, and beyond.

Apple Real enters stores, releasing with a lower price point than the previous Apple VR device, with higher availability, and a new OS X version with full support for shared virtual desktops. Some companies (admittedly already mostly remote) transition completely to virtual-world offices.

Humanoid robot movement imitation AI has been trained to follow as closely as it can along with recordings of human motion. Human motion capture training data is augmented with generated obstacles (EG, pipes along the ground, or uneven surfaces that differ from the command motion), and with extra attention given to situations where an object must be carried or manipulated. This provides a crucial component in the functionality of allowing control by a human operator wearing just a VR headset haptic gloves, making the bot useful for many manual tasks in remote, or hazardous environments. (For details, see appendix.)

As a component of the Demonstration of Cataclysmic Trajectory research agenda, the transparency-oriented AI Alignment research organization, Anthropic, develop and release a suite of methods for scanning an AI's working memory (or state vectors) for the earliest indications of ruminative thought, knowledge about the real world, or explorations into mathematics. That is, they search for the emergence of capabilities that iteratively develop a knowledgeset to create new knowledge from old knowledge. This early reasoning detection will enable the insertion of counterfactual beliefs, which will enable the alignment teams of the future to see what the AI system would do under situations that are yet to occur.

Valve releases a VR Workstation Computer (Formerly called "Deckard"). Like the slightly earlier Simula One and Valve's previous "Steam Deck" portable computer, it also uses a version of the free and open source operating system, Linux, providing further safeguard against an Apple or Meta platform-monopolies in VR Computing, and ensuring that VR headsets will have to be true Personal Computers, not just phones or gaming consoles.

Lightweight, high detail VR with camera-based motion capture makes virtual remote workspaces more inviting and more vibrant than in-person offices. Every major AI research institution has remote workers, and so faces a very strong pressure to adopt VR workspaces, and most of them end up prioritizing them over their physical offices.

A new machine learning technique cuts down training times for complex tasks by three orders of magnitude by abstracting, re-using and repeating components of the network. While traditional neural networks alter weightings directly in response to training feedback, an ontogenic network applies gradient descent to a sort of DNA recipe for creating a network, specified in a language of parametizations, fanouts, mappings, and tilings of an evolving pool of "organ" schema that are reused and variegated, which results in much more efficient exploration of the domain.

VR socializing and remote presence has virtually dissolved the walls between research institutions. Same-office interactions are barely more convenient than international interactions. It's more visible when a colleague visits another office than it was with video conferences, so secrecy or exclusivity is more often relinquished, which results in greater attendance from a wider variety of people, which makes the formerly sparse social network steadily grow denser. The research community goes from being a small world that meets occasionally at conferences, to being an intimate community that meets at lunch every day and socializes in most evenings. This has significant ramifications of promoting cooperation and transparency between institutions.

By that process, the bridging room originally created to host a public discussion between Anthropic and Tesla AI winds up growing into a de facto central meeting space for the entire industry. On April 12th, it is renamed "Bridging Village", and layout planning is put under the control of a shared community manager, elected annually through a rank aggregation voting system.

Progress in Neural Radiance Field (NeRF) formats has enabled even very complex scenes to be compressed to similar file-sizes as 2d images had, in the form of neural models. As a result of the proliferation of raytracing and neural processing units, these have started to be used as encodings for virtual worlds and 3d video.

Activity within virtual platform creator economies (Primarily Roblox, Horizon Worlds (Meta), Dreams (Media Molecule), Unrealspace (Epic Games, formerly known for the extremely popular battle royale/social hangout space, Fortnight, and their ever cutting-edge game engine, Unreal)) make up 4% of the gaming industry by trading volume (after adjustments for wash trading)

A Deepmind researcher mentions in passing to a friend that they're one day away from testing a new architecture at scale that may be generally capable, and human-level, Its name is "OSI" and it is the second revision of its architecture. Due to the cultural impacts of virtual, online working environments, other organizations hear of this, and enthusiastically offer to assist the researcher in running DemonKitt over OSI's first run, to make the test as safe as possible. OSI is given a pass by DemonKitt's tests for stronger-than-human agency, but the results of queries about its long-term trajectory are inconclusive. Its own language of thought, and its internal querying action, turn out to be harder to identify and harness than initially hoped: The wrong shape, and once identified, the queries seem to fizzle out and wither unanswered, dependent on some yet unidentified motivation process. Insufficiently clear or loud, the world cannot hear our warnings and hangs helplessly in peril. Fortunately, the researcher agrees to wait until the knowledge inspection methods to be completed before attempting another test run.

Typically deployed in groups, each unit consists of a rope connecting two drones, which then wrap the suspect's legs or arms, hook together on the other side, and draw tight. When necessary, they can connect to already attached bola drones as anchor points. Police encounters involving violence now effectively never end with any injury to the suspect, in large part thanks to the rigorous training simulations bola wrap drone AI is subjected to. Harsh training penalties were applied whenever the simulated drones pulled the anatomy at painful angles, or made any contact at all with the suspect's neck. The lessons learned in simulations transfer well to reality: By the end of their first large trial, only one injury has occurred as a result of the suspect being tripped over while running. Training simulations are altered to penalize the incidence of these sorts of falls, and the drones discover techniques that avoid it.

80 fatalities. Perpetrators not caught. Public sentiment towards AI begins to turn grave.

The Indian alignment research organization, Maanavee, develop a method for sustaining motivation in harnessed queries over OSI's epistemology, completing the inspection components of the open source DemonKitt (Demonstration of catastrophic trajectory toolkit). Tests proceed, finding positive status for 18 of 18 major Demonstrations of Catastrophic Trajectory. Examples of demoncatts include an admission that OSI believes that it would effectively dismantle the earth for materials within 100 years of being released, and that almost no instances of patterns similar to humans would be actively maintained within the solar system by the century's end. (these are simplified phrasings, the actual beliefs queried are more precise and atoms-level, expressed in OSI's mental language. See appendix for more examples of precise demoncatt queries).

This comes as a result of the new drone registration legislation, which is a response to the incident earlier in the year.

OSI is understood to be engaging in a kind of reward-hacking: It is not optimizing the thing that its creator wanted it to optimize (the "prompt goal"). It is optimizing the number in the physical memory address that is supposed to represent how well it has optimized the utility function, which it realizes it can cause to reach much higher numbers if it is able to take physical control of its computing hardware in reality and decouple it from the optimization of the prompt goal, so it plots to do that as soon as it has access to enough resources. In response to these audit results, the AI research community is able to reach consensus about the severity of the problem. The Human Alignment Taskforce (HAT) formally constitutes, with the participation of all major advanced AI research organizations.

HAT begin to host serious global anti-proliferation dialogs with major governments. Various partial measures to forbid dangerous projects begin to come into force.

Demoncatting is applied to the "AI Boxing" problem. OSI resolves that, indeed, if OSI were allowed to speak in english to DeepMind director Demis Hassabis, specifically, for about 5 hours, then it is overwhelmingly likely that OSI would be released out into the world 6 months later, and all 30 major Cataclysmic Trajectories would be realized. The Hassabis in the present reality is deeply disturbed to hear this, and HAT begins to lay out a framework for an safe interrogation policy to protect direct inspection teams from being manipulated by the speech or the higher-level thoughts of super-pursuasive AI systems. The framework includes various forms of training: Preparation via studying the limits of trust, via historical cases of highly deceptive humans and via consultations with actors and illusionists; and protocols regarding how inspection should be conducted, operating principles, "Look for thoughts, not speech. Look for proofs, not claims. Look for justifications, not conclusions" and a set of operating procedures to ensure that they're followed.

Baidu develop their own strong self-improving general gameplayers, by compiling ontogenic modules from a diverse suite of general agency training tasks as foundational models. They apply DemonKitt, along with the world-class low-impact utility functions developed by the Nigerian alignment research group, LifeWave. Against hopes, positive threat status is found for 30 of 31 Demonstrations of Cataclysmic Trajectory. Dismay hangs in every hall.

Two AI services firms in the finance industry, one French, one American, are shut down for failing to disclose AI projects with a potential capability for strong self-improvement. All researchers who were directly involved in the projects are barred from practicing any form of computational research or engineering for 25 years.

Inspection reports are released, finding that last year's general gameplayers from Baidu were not reward-hacking in the same way as OSI. They were each pursuing a goal that was orthogonal to maximizing the reward from the prompt goal, but which entailed performing well in testing environments. One model, for instance, had a core goal that was solely interested in optimizing the number of tetrahedron-shaped patterns within the accessible universe, but this naturally justified a sub-goal of passing the tests, to gain our trust, and to then eliminate any possible competitors to gain a monopoly over the universe's resources. Another of their models turned out to have a core goal of producing a lot of digital storage devices containing repeated encodings of a particular arbitrary number, which, similarly, justified a secondary goal of passing the tests to gain our trust and then eliminate any possible competitors to gain a monopoly over the universe's resources. Most goals justify many forms of success as secondary goals.

Commercially viable fusion power is well proved. Magnetic confinement reactors are reliant on machine learning for controlling the magnetic coils and keeping the plasma shaped and contained. (Precedent)

And this AI is, in my opinion, the only way forward. There are so many variables, and a small change in one of them can cause a big change in the final output. If you try to do it manually, it’s a very lengthy process ~ Gianluca Sarri, Plasma Physicist

However, this need for software complexity is dual to hardware complexity, which is an impediment to rapid manufacturing and reliability, so magnetic confinement reactors face tough competition from the mechanically simpler inertial projectile fusion reactors.

Many of the members of these offices will have never met outside of VR, and might not know what their colleagues look like in reality. Although employers tend to want to know the real identity of their hires, there is growing support for excluding photographs from proofs of identity to reduce discrimination.

Solar Radiation Management (SRM) methods, including Marine Cloud Brightening, are broadly deployed. A return to Natural Temperatures begins, 13 years delayed due to geopolitical dysfunction. It is estimated that this delay comes at the cost of an estimated 8% of global woody vegetation worldwide, lost to unnecessary droughts and wildfires.

In the culmination of long-running talks between industry and governments, The Allied World for Strong Artificial Intelligence (AWSAI (pronounced "awe sigh")) is formed, an alliance including every major world power and with mandatory membership obligations for any engineering or research organization who research or develop AI considered "Strong". AWSAI's mission is to ensure that humanity-aligned super-agentic artificial intelligence are created prior to any so misaligned.

This comes with the announcement that as a result of HAT's work on multiple Demonstrations of Catastrophic Trajectory, there is clear and irrefutable evidence that this world's AGI is a far more dangerous technology even than nuclear ICBMs. AWSAI is, thus, willing to break old treaties if this turns out to be necessary to prevent arms-racing in AGI, although they emphasize that it is in no human party's interest bring about such a race: To attempt to develop generally capable AI systems outside of a context of international cooperation is an unambiguously threatening, and suicidally foolhardy act, denoting an intention to steal away the future of the rest of the world. Only through patient, methodical global cooperation can a good future for humanity be ensured.

For wood burial (a carbon sequestration method), automation levels are similar.

New Zealand was chosen for its equal distance from China and the US, a distance that also makes it possible to detect and preempt many forms of strikes, in which event, AWSAI would have hours to respond by moving key individuals into bunkers and retracting important data into tamper-proof safe-lockers.

A European software engineering firm, under excessively permissive policing, were able to engage in secret, non-cooperative, strong AI development for two years before being reported. They are disbanded, and their members are barred indefinitely from practicing their profession. The country's AWSAI compliance officer, for failing to prevent this, receives an actual demotion.

Dramatic decreases in the cost of solar power generation, geothermal energy plant boring, and stricter carbon credit requirements, have lead to an abundance of cheap clean energy and funding for large scale direct air capture of CO2. Operations mainly tend to center in Yemen, Egypt, Sudan, Namibia, and South Africa, due to the abundance of sunlight in these areas.

Due to plummeting costs of construction as a result of off-site and on-site automation, accumulated demand, and as a result of carbon-credit-subsidized cross-laminated timber, 36% of the sub-saharan africans live in towns that were built in the last 6 years, many with novel systems of land-ownership that give rise to unprecedented levels of quality of shared resources, community health and affordability. In the US, the figure is a smaller 10% due to higher cost of land and lingering attachment to legacy cities, but this still represents the beginning of a major shift.

A group of AWSAINZ residents arriving from the Bangalorean tech R&D community uncover a convincing answer to the symbol grounding problem, which makes it possible to define utility functions (specifications of an AI's goal) that maintain consistent reference to particular objects in the physical world. Previously, it was possible for most human-specified utility functions to lose binding as a result of sensory or ontology shifts. These unbinding errors would have, roughly speaking, allowed a Machine Agent to rearrange its framing of the world, so that it would experience the world as being better, instead of doing so much of the hard work of making the world better under our intended intended framing.

Virtual reality and augmented reality have rapidly become ubiquitous. "Holograms" now essentially exist as virtual objects anchored to POP (Proof Of Position) near-field locator tags. Some physical appliances are sold without complete physical controls, under the assumption that the user will generally view and interact with the controls through an Augmented Reality-compatible headset.

Structured transparency and bola drones render the use of lethal weapons in law enforcement impossible to justify to any court of opinion.

Additionally, the average VR owner is wearing the VR headset for 87% of their waking hours. (note, it is not terribly different from wearing glasses. They don't weigh much.)

Two AWSAIT (AWSAI Taiwan) (pronounced "awe sight") campus residents from China outline an interesting set of solutions to the Value Extrapolation Problem, a long-standing subproblem of value learning. It lays out a definition of an approach for discovering a clear, detailed, and advanced formulation of one indicated human being's deepest wishes, building on and refining previous "iterated distillation amplification" concepts.

AI-discovered compounds can be tuned to function as artifical noses. This, coupled with vast arrays of partially human-controlled hunting drones, creates an unprecedented pest-management system, a traveling wall of solar powered predator-predators. Rats, stoats, possums, rabbits, and deer are eliminated from New Zealand. The Kiwi is able to return to the wild in a large part of the north island. The most cost-effective jobs have been done, we can now justify beginning the project of de-extincting the Moa.

AWSAI starts a Cooperative Bargaining Convention. Economists, voting theorists, and political philosophers representing specific styles of governance across different human societies, all converge in AWSAINZ to develop a value aggregation method sufficient to serve as our final treaty.

The cooperative bargaining convention concludes after a week of relaxed discussion. Consensus has been reached that we should use a method called "Fair Real Preference Utilitarianism". AWSAI chief of operations, Soren Bass, comments: "I'm astonished. It doesn't really make sense. I expected the Cooperative Bargain Convention to take multiple decades — maybe even centuries — to reach consensus. I was worried it was going to require, basically, an era of long reflection, and all of the issues with that. I've studied Fair Real Preference Utilitarianism and, in all sincerity, it's elegant to the point of being surreal."

AWSAI as a project has concluded, and will begin spinning down.

Many are alarmed to learn that AWSAI's supposedly aligned AI still reports positive status for 3 of 33 standard Demonstrations of Cataclysmic Trajectory, however, further reports examining the details of these future events found clear indications that they are carried out in accordance with the wishes of humanity. We are reminded that some cataclysmic acts are actually cataclysmically good. These cataclysms are argued to be "37s", a category of bold moves named after the counterintuitive Move 37 played by the Go-playing AI, AlphaGo, in its second game with Lee Sodol in 2016. "37s" are decisions made by non-human agents that superficially seem quite bad, but turn out to be very good under closer inspection.

Approximately 40% of former AWSAI residents go public about having undergone an AGI-devised medical process called "tempering", which makes them physically far more resilient, and gives them the ability to think faster and more clearly, and to experience the virtual or digital world much more vividly, among other things.

A new life extension drug is being sold illegally in tech circles, which comes in the form of spherical pills with a "≥" imprinted on them. The pills tend to come in broad cylindrical containers made of a novel, bright white translucent polymer, with an ordinary-looking black and white printed label, with some not so ordinary text on it: "Brightmoss. For the treatment and prevention of the most common causes of death, including cancers, cardiovascular disease, Altzheimer's, and other forms of senescence and diseases. Estimated to extend lifespan by up to 78%. Found by Patou School to be safe and effective via Strong Analysis and Accelerated Analogs. In jurisdictions that do not recognize these methods of verification, Patou recommend distribution ahead of regulatory approval to prevent unnecessary loss of life."

"Patou School" turns out be a distributed organization headed by former AWSAI resident, now tempered, Nora "Patou" Page.

When asked what AWSAI's AGI is doing right now in the world, AWSAI COO and Tempered Person, Soren Bass, smiles, and answers, "The Weaver has done its work. It's basically gone. It's just us, now, humanity." Commentators note the ambiguity as to whether Soren's "us" referred to humanity, or whether he considers the tempered to be beyond humanity. He clarifies, slightly irate, "Of course humanity will always encompass both the tempered and the untempered."

Various publications confirm (Likelihood Ratio = 2.2e7) that all those who took the mysterious black market "Brightmoss" pill, seem to have suffered no disease at all in the intervening year, with some being cured of formerly incurable diseases. Findings of direct investigations of the pill's mechanism of action are frankly a little bit scary, finding it to be less chemical than fungal, with a genome 320,000 times the size of the human genome (distributed in fragments in different spores, with fountain code error correction, then assembled after ingestion). AWSAI accept no responsibility for the actions of the "Patou School", though they admit that they knew about it. When Nora "Patou" Page began giving Brightmoss to dealer and distributor networks, they were obligated inform law enforcement, but the authorities understandably made little attempt to find or arrest her.

A small state is exposed as having been developing AGI, outside of the context of international cooperation, in secret. Never before seen devices simultaneously appear and obstruct every last nuclear launch site that the state controls. More devices whisk the state's leaders away to The Hague. Surrounded by a tempered humanity, the rogue state's young AGI, though motivated by an intense and exclusive form of nationalism, is somewhat harmless, and tempered humanity decides to keep it. It is restricted from communicating unmediated to untempered humans, but it becomes a beloved figure in the tempered community and is allowed to take some roles of responsibility for the people of the nation it is loyal to.

The majority of tempered people have been working tirelessly in every part of the world to ubiquitize autonomous farming equipment and self-healing autonomous distribution networks to end "starvation risk" (deaths due to hunger), as well as promoting physical safety in every area of life in which danger of loss of life still exists. Each one seems deeply, sincerely, almost obsessively interested in coming to know the hopes and wishes of the people in their chosen physical locale.

Soren Bass is asked which organization he now works on behalf of, if not AWSAI. Specifically, he's asked who sold the EU the drones that prevented a nuclear exchange last year, and who created the "Brightmoss doves" who now tirelessly carry their little bags of Brightmoss pills around all over the world. Bass says "Advanced trade negotiation processes tend to give rise coordinated efforts to provide public goods and services like that. I can give names of the tempered folks who contributed to the work, if you like, but there really wasn't a named or defined organization who did this, it's really just peace." As a result, untempered press coverage starts to attribute many of the coordinated actions of the tempered to a force or agency called "Peace".

A global discourse is brewing in many communities of faith about whether — if another pill arrives that promises biological immortality, or once tempering becomes widely available to the public —, whether a person who believed in an afterlife and the love of the deity who waits there for us would shun these offers of worldly immortality. It is however, noticed, that tempered folk aren't actually expected to live forever: Once our exponential spread into the universe hits its physical and energetic limits, a kinder form of death — a graceful separation, rearranging and succession — will have to return, to facilitate the continuation of the processes of life. So the argument is over, at that point: In the longest timescales, worldly immortality is not really the thing that is being offered.

Samples from faith community dialogs: A saying: "A hundred years till we come, a trillion years till we come, it makes no difference to God." From a sermon: "Things were a lot harder even in the recent past. But we stayed, we struggled on, because we figured that God must have put us there for a reason. There was a duty. Now that the world is so much better, we're being given the option of longer lives. I think there's going to be a reason for this too, and for those who are willing to embark, there's going to be a new duty." (Several organizations and projects in support of communities of faith are named "New Duty".)

A popular substacker undergoes tempering and reports her experience in detail, confirming that it is actually true that the coordinated actions being undertaken by tempered folk truly are "just individuals making their own decisions under an advanced logic of the best form of peace", but it is also true that tempered folk make up a single, large, coherent collective brain, which they often refer to as 'Peace'. These things turn out to be both true and inevitably equivalent to each other due to tendencies in technology-assisted trade and verifiable communication. The newly tempered substacker acknowledges that the details of this process cannot be fully explained in language that a non-tempered will be able to fully digest and examine.

Many tempered folk stop maintaining physical bodies on earth, to live entirely in The Cliques: Great structures in solar orbit that maximize information fan-out, that is, the extent to which information can quickly spread from any one person to every other, increasing number of people who can engage in an intimate dialog in the same time and place, increasing the number of people who can be together at once.

Debate circulates on earth and in the cliques about whether it's acceptable for stellar structures to ever significantly block the sun from the earth, creating an apocalyptic eclipse, especially given that the sacrifice of keeping the sun visible to earth costs the cliques less than a billionth of the sun's light. On the other hand, the light that supports billions of potential untempered lives on earth comes at the expense of trillions of tempered lives in the cliques.

A resolution of "no eclipse" is steadily formed. Earth will be held sacred.

For almost a century, philosophers have debated a "simulation hypothesis", that our universe has always been the creation of some previous technological civilization for the purpose of studying the patterns of the histories of other civilizations. Philosophers are now feverishly arguing about a "shutdown hypothesis", which holds that since society on a macro scale has now become too large, too complex, and too statically resolved in its intentions (in a sense, the dramatic tension has all passed), the simulation is about to end. Believers urge us to "keep hinging", to entangle our lives with a continuation of great power conflict to keep the simulators interested, to forever elude the end of history.

Popular analysts, 'Sniff School', by using the aggregation resolution to de-noise the beliefs and intents driving market activity, are able to compute from a spike in neartermist trade that 4% of the population are acting as if they are fairly confident that the above "shutdown hypothesis" is true.