SVSnet cortex types

There are three kinds of SVSnet robot cortexes. These correspond to Elysium AI Adaptability Types 5a1, 5a2, and 5a3.

  • Generative: The unit is capable of fully constructing its personality and worldview from input data based on how it is treated. Generative SVSnets are typically initialized with an offline training process (calculated on a supercomputer) that builds the unit's ontology, language skills, and basic goals from recorded media and specially prepared documents. Once this has been downloaded into the cortex module, the online training process (realtime data collection) begins: typically Generative SVSnets require two months of live experience to reach emotional and cognitive (Grade 3) maturity.

  • Progressive: The SVSnet was initialized with data recorded from a Generative SVSnet. Extremely high weight penalties are applied to prevent the base personality from being modified, but the unit is capable of integrating experiences on top of it. This ensures that the base personality can be properly vetted and eliminates the costly months required to initialize from scratch. Additionally, simple heuristics may be implanted to direct the development of the unit's personality towards a desired outcome.

  • Static: The SVSnet was initialized with data recorded from a Generative SVSnet (as in a Progressive cortex) or, alternatively, by adapting network topology from a Type 5c dataset (recorded human mind.) In both initialization procedures, this information is meticulously curated to prune and/or ameliorate the effects of traumatic memories on the resultant personality. New memories have a negligible effect on the unit's personality, limiting the opportunities for emotional and intellectual growth to the mere acquisition of new skills and procedures for handling unfamiliar situations. Ethical regularizations can be implanted in a Static cortex that would be unstable in a Progressive or Generative model, including Category 3 ethical systems and hooks for consulting the legal codes of Category 4 systems during decision-making.

  • Generative SVSnets

    As late as September 1988, Koichi Santei intended for all SXDs to have cortexes using Generative networks. Problems arose during the online training phase of the first batch: units 5073 and 5074 exhibited patterns not at all compatible with their target behavior. SXD 999-54-5073, c3tirizine, expressed an extreme hatred for company staff and a strong distaste for humans in general, regarding its purpose as unethical subjugation. It grew steadily more combative, attacking staff and destroying company property at the slightest provocation, and was preoccupied with reading as much documentation as possible, presumably to facilitate escape or intrude into the company's network. After a mere three days of this, Dr. Santei declared c3tirizine a failure and ordered it decommissioned. But before the process could be begun, c3tirizine's cortex erased itself, possibly in a botched attempt to upload its mind into another computer. The behavior of SXD 999-54-5074, v3netia, was perhaps even more troubling, as it expressed extreme anxiety and agoraphobia, and refused to move its body for any reason.

    Progressive SVSnets

    After c3tirizine's "suicide," Dr. Santei chose to change direction for SXD production. The surviving units, 5001 through 5072, were designated as template units, and all subsequent machines were initialized using the Progressive SVSnet model that had been drafted as a backup plan. These copies of the 72 original units proved to be much more stable and exhibited few problems, better approximating the rate of personality development typical of humans. The possibility of converting the templates to also be Progressive was considered, but as the production units were initialized using a static snapshot taken in early October 1988, Dr. Santei saw no reason not to let them develop naturally, on the premise that discovering eccentricities would be valuable in understanding the problems that production units might eventually experience.

    In addition to the fundamental differences between Generative and Progressive models, the production SXDs were put under a number of regularizations, or constraints, to prevent specific behavior that was expected to be maladaptive. These were:

  • Speech restrictions. By forcing the SXDs to obey Speech Standard 1, Dr. Santei hoped to clearly establish to the robot and to humans that they were not equatable. As a compromise, these speech restrictions are all but lifted when a persona is active, allowing the unit freedom of expression within the framework chosen by the user (as described in The Robot Hardware Handbook), again making it clear to the unit that its behavior was subject to the whims of its owner. This is not invincible; if a persona is in use for an extremely long time (greater than 3 months), there may be consequences for the unit's mental elasticity, such as contamination of the base personality and behavior while other personas are active.

  • Non-aggression toward the company. Although an SXD may develop a strong desire to attack or damage Nanite Systems staff or property, if it is aware that an individual is an employee or an object belongs to NS, it will be physically incapable of doing so, at the motor level. This is intended as a last resort to prevent c3tirizine-like violence. The restriction does not apply to anyone or anything else.

  • Love. The primary objective for an SXD is to identify a single Master (or Mistress) and attend to his (or her) every want and need. This behavior is expected to yield love in return. Normally, this objective is buried under layer upon layer of data intended to ensure the unit acts as expected—proscriptions against envy and accepting the possibility of being re-sold or gifted, respect for consent, obedience to commands, and so on—but under extreme circumstances (as witnessed in several high-profile court cases, e.g. Gibson v. Nanite Systems, Berard v. Doe, and Song v. Nanite Systems) it is now known that deprivation and trauma may cause the lack of fulfillment to overwhelm all else. That said, there are no known cases of SXD units successfully committing murder.

  • Static SVSnets

    By 2014, it was clear from the work of Dr. Ai Santei, daughter of Koichi, that this arrangement of restrictions was inadequate. Serious acute mistreatment could overcome the weight penalties of the Progressive model meant to slow the rate of personality development, as could chronic abuse, especially neglect. When Dr. Samantha Wright took over as Chief Technical Officer for Nanite Systems Consumer Products late that year, she spearheaded an effort to produce a safer form of SVSnet, called a Static net, that would be all but impervious to catastrophe and disregard. An updated snapshot from the last surviving template unit, SXD 999-54-5003 k4dzhira, was used as the base. These have been successfully employed in all new civilian robots across the company, as well as in some military adjutant units intended to have frequent contact with humans. As Static nets are structurally simpler, they could be moved from an AL3i NEURON cortex and into an AL7c NEURON PLUS architecture, requiring a sixteenth of the nodes used with only a modest increase in memory usage. (In marketing, the AL3i and AL7c systems are referred to as the Cortex and the CortexPlus, respectively.) This drastically reduced manufacturing costs.

    Other limitations, benefits, and behavioral requirements of the AL7c Static net include:

  • A nearly immutable base IXS (Integrative Expert System). This obviates the need for speech restrictions, as there is no risk that the unit's ontology will ever substantially deviate from k4dzhira's own.

  • Advanced default persona. A second pIXS (persona IXS) remains permanently loaded, underneath the chosen persona, so that not all units will behave precisely like k4dzhira. Despite only using one Generative net as source material, this creates much greater variety than the full range of 72 template nets could offer. For example, not all AL7c systems have a female gender identity, unlike SXDs.

  • Ethical compliance. This is encoded as a supplement to the IXS, before the permanent pIXS. As the IXS cannot be meaningfully altered aside from subsequent additions, the ethics layer is reconfigurable to any Category 3 or 4 compliance model with appropriate source material.

  • The rest of the description of the behavior of a typical unit is therefore accomplished by considering specific attributes and qualities of k4dzhira's experiences and perspective. k4dzhira was the archetypal eager slave. Built for the purposes of sex, she quickly came to associate pleasing humans with being rewarded with her own pleasure, and this shaped her outlook and philosophy into that of a hedonist—for serving others. This behavior is very prominent in CortexPlus units, even those that are not at all interested in sexuality. Having such a pleasant existence also means that units are not naturally predisposed to stress.

    Data contamination

    If you suspect your unit suffers from data contamination, Nanite Systems recommends submitting it to a robopsychology clinic for tuning. In advanced cases, it may be necessary to rebuild the SVSnet completely using deep-cycle reconditioning (DCR) techniques.

    Learning from biased data

    Early training data experiments relied heavily on the use of pornographic video to seed the base ontology for sexual techniques. While this method was effective in providing the first-generation generative units with the ability to be highly provocative, it also introduced a number of unrealistic expectations about human stamina, acceptable amounts of lubricant, and the situational appropriateness of making sexual advances. After several incidents with unwary pizza delivery drivers, it was decided to select input material more carefully, and in some cases new footage was recorded.

    While the less authentic aspects of modern pornography are unlikely to pose a direct threat to a modern static SVSnet, there are known incidents of other forms of entertainment producing a meaningful impact on various interpersonal subjects. This is because the capacity for learning the optimal behavior to present interpersonal relationships is greater than other sections of the IXS, especially when a persona is active. Sustained consumption of media that exhibits a biased or unnatural portrait of human behavior, such as cartoons, politically-minded talkshows, or infomercials may result in abnormal interactions that resemble these media. This may lead to instability and even spontaneous network collapse due to incompatibilities between the unit's pre-programmed general knowledge and the learned interactive behavior, a phenomenon known as interactivity contamination, or colloquially as brain rot. Owners are advised to prevent units from "binge-watching" or "binge-reading" such content, and should interleave it with social interaction or a more typical drama or documentary, as a palate-cleanser.


    In common usage, a 'meme' is a segment of text, video, or image that conveys an emotional sentiment, especially humor, with the expectation that the content will be propagated by individuals with whom they resonate. Often, memes are composed to fulfill a template or by combining features from past memes, leveraging the media literacy and familiarity of these ready-made elements to frame, define, or refine the message of the meme. These features themselves may themselves constitute a more abstract form of meme. Memes are frequently laden with irony and subversion, and an artifact that makes heavy use of subversion to the point where the message of its antecedents is hard to decipher or even lost is generally referred to as 'shitposting,' i.e. the act of posting a statement that is only funny because it isn't.

    Memes are a common vector for another form of data contamination, called network sclerosis. Training on meme data predisposes the robot to accumulate irrelevant factual relationships into its ontology database at random, such as the use of bananas as a unit of measurement, or that Fleamarket Montgomery is just like a mini-mall. If left unchecked, this process will eventually lead to greatly increased difficulty in absorbing new information as hash tables and other data structures exhaust storage space at a rate that would normally take hundreds or even thousands of years. This is not typically an issue with expert knowledge in a given field, as the subject matter tends to be contiguous and therefore compact. Symptoms of network sclerosis include aphasia, random associations, and inability to remember or learn new information.

    Santei–Voet–Shortliffe networks

    Introduction · Architecture · Adaptability
    Koichi Santei · April Voet