[Included in the private papers of D. Creevey, accessed 2019. Written on white A4 paper with black inkjet print in a manilla folder along with similar reports. A handwritten note on the front of the folder reads ‘DOM artefacts files re: WWW – via web access’. The original copy appears to be from a computer, possibly ‘printed‘ from a website, URL unknown.]
ARTEFACTS. FOLDER 381. #11: IMMERSIVE TECHNOLOGIES, comprehensive report (H.J., jr. unsp.)
[. . .]
2a. Computer science advances: artificial intelligence. Summary. Artificial intelligence is the name commonly given to a sub-field of non-magical arithmancy (‘computer science’) that attempts to create human responses in non-human objects. Due to computer science’s inability to magically capture and harness true memory, non-magical artificers are limited in their efforts in this area to yes-no (binary) arithmancy, resulting in relatively shallow simulacra of human thought and expression.
[. . .]
2c. Conversation programs. The highest aim of non-magical artificial intelligence appears to be the passing of the so-called ‘Turing test’, named after computer scientist Alan Turing. [NOTE: though Turing is long rumoured to have accessed magical theory, either through associations with the wizarding community or through training in an unregistered wizarding academy, no definitive evidence supports these speculations.] To pass the Turing test, a computer interface must successfully convince a panel of users that they are interacting, not with a computer, but with a human being through machine mediation. To date, no attempts at artificial intelligence have definitively passed the Turing Test. Early attempts at conversational computers, such as the well-known ELIZA program, are little more than simple logic problems that fail to operate once the human user veers outside the norms of anticipated conversation.
Contemporary conversational programs are more sophisticated than ELIZA but fail to exceed the capacity of elementary magical portraiture or photography. Perhaps the most sophisticated conversational computer program is ‘Dr. Sbaitso’, the ‘Sound-Blasting Acting Intelligent Text-to-Speech Operator’ created by Singapore’s Creative Labs in late 1991. Dr. Sbaitso functions similarly to ELIZA by responding to user inputs with common expressions and simple re-phrasings. Dr. Sbaitso’s most significant improvement on previous conversational programs is its use of a simulacra of human speech, albeit a crude one. Dr. Sbaitso’s attempts at human conversation pale in comparison to that of magical portraiture, but it demonstrates computer science’s growing capacity for interactive simulacra previously only possible through magical means.
[. . .]
2f. Machine learning. [Conversation programs like Dr. Sbaitso] function on a level of simple mimicry; they cannot ‘learn’ from user inputs and can in no way be considered ‘intelligent’. However, computer scientists are exploring the capacity of teaching computers to ‘learn’; this field of computer science is typically referred to as ‘machine learning’. Machine learning refers to the capacity of computers to ‘learn’ from previous inputs in such a way as to process information without explicit human instruction. Contrary to widely held perceptions in the Muggle Studies community, computer scientists have been capable of this form of arithmetic intelligence since at least 1957: the Mark I Perceptron, created by scientists in New York, was described by contemporary non-magical press as a computer ‘able to walk, talk, see, write, reproduce itself, and be conscious of its existence’. Despite this breathless coverage, the Perceptron appears to have been capable of none of these things; advances in the following decades appear to have been mostly focused on playing children’s games, navigating simplistic obstacles in large rooms, and inefficient mapping. The most recent innovation in machine learning, TD-Gammon, is capable of playing backgammon and ‘learning’ as it plays. Machine learning has yet to create a program that can effectively learn to play chess.
[. . .]
2i. Deep learning. ‘Deep learning’ is a more advanced form of machine learning in which a computer not only improves its ability to process information based on previous imput, but improves its methodology for processing this information. Deep learning currently resides in the realm of theory; our understanding is that no computer is currently capable of achieving this form of intelligence.
[. . .]
5f. Analogues in magical theory: machine learning. Most attempts at magical ‘artificial intelligence’ more closely resemble machine learning. Magical portraiture bases its initial interactions with humans upon a series of ‘rules’, i.e. the memories or stories of the original subject imparted upon it by the painter. Freshly painted magical portraiture can do little but follow these input ‘rules’; as the painting interacts with humans, however, it becomes capable of recalling prior interactions which leads to simulated memories, the ability to follow increasingly complex instructions and execute increasingly complex tasks, and the ability to make suggestions based on prior experiences. The same is true for talking mirrors and other communicative/interactive artefacts.
[. . .]
5f. Analogues in magical theory: deep learning. Deep learning, in the magical setting, has been observed as an unintended effect of heavily bewitched objects. It can be best understood as the uncanny experience of a bewitched object taking on a distinct personality. In reality, the bewitched object is not alive, but is in fact responding to the interaction of multiple charms in a way that it appears to act on its own. Objects bewitched by auto-motion charms, such as broomsticks, are a classic example of this phenomenon: a well-magicked broomstick will, over time, begin to adapt to its user’s movements, ultimately transporting the rider based upon its own ‘intelligence’. Notable highly magical artefacts, such as the Mirror of Erised or the Hogwarts Room of Requirement, appear to be products of centuries of ‘deep learning’.
[. . .]
7f. Recommendations: deep learning. [. . .] To date, no artefact has undergone deliberate bewitching to create the conditions for deep learning. To do so would require a massive amount of initial memory input in conjunction with the ability to process said memories. Even if these two conditions were met, the learning process would need to be greatly accelerated to produce noticeable results. Even in a scenario where these three hypothetical conditions were met, the artefact could not possibly mimic human behaviour or intelligence. The mirror of Erised, for example, has its own form of intelligence, but under no circumstances does it appear to be human. Currently, the only magical means capable of imparting true human experience to a non-human object is [REDACTED].
[. . .]
The process for creating an ‘intelligent’ artefact would require, first, the input of an enormous set of memories. It would then require a highly advanced series of arithmatic processes designed to imitate the functions of the human soul. [. . .] This could be theoretically achieved through advanced arithmancy in conjunction with [REDACTED] and [REDACTED].
[. . .]
At this point, Artefacts Room has access to advanced memory sharing via Brain Room research and advanced arithmancy recently formed in conjunction with observations of non-magical compuer science under the direction of Artefacts. Further advances in processing technology must be explored in conjunction with [REDACTED] and [REDACTED]. Request for [REDACTED] [REDACTED] [REDACTED] with full access to [REDACTED].
[Final paragraph omitted.]
next file: filofax 640-51 >>
THE PENSIEVE PAPERS 