Links table
Abstract
Related work
Media, liquidation bubbles and echo rooms
Network effects and consequently information
The collapse of the model
The well -known biases in llms
A model for the collapse of knowledge
results
Discussion and references
Excessive
Comparison of tails
Determine the collapse of knowledge
A model for the collapse of knowledge
Determine the collapse of knowledge
However, in other areas, it is less clear, especially within the regions. Historically, knowledge did not advance in a monitor, as it is clear from the fall of the Western Roman Empire, the destruction of the House of Wisdom in Baghdad and the subsequent decline in the Abbasid Empire after 1258, or the collapse of the Maya civilization in the eighth or ninth century. Or, to martyrdom with specific examples, the ancient Romans had a recipe for concrete that was lost later, and despite the progress that we did not re -discover after the secrets of its duct (Seymour et al . Culturally, there are many cultural and artistic languages and practices, and religious beliefs that were once by human societies that have now lost that they are not present between any known sources (Nettle and Romaine, 2000).
Knowledge distribution vary across individuals over time. For example, traditional fishermen can identify thousands of different plants and know their medicinal uses, while most people today only know dozens of plants and whether they can be purchased in a grocery store. This can be considered a more efficient form in the specialization of information across individuals, but it may also affect our beliefs about the value of these species or walk through a forest, or affect scientific or policy -related provisions.
Unofficially,[2] We define the collapse of knowledge as gradual restrictions over time (or technological procedures) of the information group available to humans, as well as narrowing the accompanying availability and benefit of different information groups. The latter is important because for many purposes it is not enough to have a capacity, for example, on an archive to find some information. If all members consider that it is very expensive or is not worth searching for some information, then the theoretically available information will be neglected.
Model overview
The main focus of the model is whether individuals decide to invest in innovation or learning (we treat it as an interchangeable) in the “traditional” way, through a process perhaps cheaper than AI, or not at all. The idea is to capture, for example, the difference between a person who does intense research in an archive instead of relying on easily available materials, or someone who takes the full time to read a full book instead of reading paragraphs consisting of paragraphs- born a summary.
Humans, unlike LLMS, who have been trained by researchers, have an agency to make a decision between possible inputs. Consequently, one of the main dynamics of the model is to allow the possibility that rational factors are able to prevent or correct distortion from excessive dependence on “medium” information. If the previous samples neglect the “tail” areas, the returns from this knowledge should be relatively higher. To the extent they notice this, individuals will be ready to pay more (putting more workers) to profit from these additional gains. Consequently, we achieve under the circumstances of such a modernization between individuals, it is sufficient to maintain an accurate vision of the truth of society as a whole.
The cost of cost and the return on investing in new information depends on the expected value of that information. Anyone who experiments with artificial intelligence, for example SUM2 for more discussion and a more accurate definition, see the appendix. Preparing, develops an intuitive feeling when artificial intelligence provides the main idea well enough for a specific purpose and when it is worth going directly to the source. We assume that individuals cannot predict the future, but they notice that the rewards achieved from the previous rounds. The decision also depends on the type of individual. Specifically, individuals N have θN types derived from the distribution of Lognormal with µ = 1, σ = 0.5. Depending on how to calculate its usefulness (not the objective focus here), it can be explained as various expected returns from innovation (EGTECHNOPTISTS versus pessimists), or their relative ability or desire to engage in innovation.
We represent knowledge as a process to bring up the distribution of possibilities (students T).[3] This is simply borrowing, although it looks like, for example, Shuailov et al. (2023), but we do not claim that the “truth” was distributed in a deep way from Gaousi 1D. This is an assumption of modeling to work with a process with well -known properties, where there is a large and long -standing central mass, which we take in a general way that reflects the nature of knowledge (and the distribution of training data for LLMS.)
The group of individuals who decide to invest in information receives a sample of real distribution, while those who invest in the sample created from artificial intelligence. To change the range of mass in tails, we put the real distribution design as a student T distribution with 10 degrees of freedom. The results are similar to the standard natural distribution, and as expected, the problem of collapsing knowledge is more clear to the broader tail (CF APPENDIX 7).
While individuals choose whether to invest in innovation according to their personal reward, when they invest, they also contribute to their knowledge of the public. This means, the function of distributing the possibility of general knowledge (‘pdf’ public) is created by collecting NSAMP = 100 modern samples[4] And generate an estimate of the truth using the density of the nucleus density. The distance between the general PDF and the truth provides an abbreviation of the general welfare of society. We define the collapse of knowledge as it occurs when there is a great and increasing distance between the general PDF and the real reality as a result of the collapse of the tail areas and the increase of the mass near the center.
The individual’s return is calculated according to the distance he transmits to the general PDF towards the real PDF. This means that innovation (individual returns) created by an individual is calculated an additional sample (N + 1) with regard to the real PDF distance (x) and PDF PPPublic (x), based on the distance of Hellen H (P (x), Q (X (X) )))[5]As follows:
Innovation = previous distance – a new distance
In Figure 2, we show the innovation account for the virtual example where the distance between the current year PDF and the real PDF is 0.5, while the n + 1 sample reduces the distance to 0.4, thus generating the innovation of 0.1.
This can be considered closer to the patent process, as the individual receives a patented patent rental (to the extent to which you are really innovative) in exchange for contributing to the general knowledge that benefits others.
As mentioned above, individuals cannot predict the true future value of their innovation options (they do not know the sample that they will receive or the amount of value that they will add. Instead, they can only assess the relative values of innovation based on the previous rounds. In particular, they update their belief in the options Depending on the full and severed samples of the previous round (and at least three), according to the learning rate (η) as follows.
By changing the learning rate, we can evaluate the effect of obtaining more or less information about the value of different sources It was relatively more expensive.
While the individual reward depends on the real movement of the general PDF towards the real PDF, the general PDF is updated based on all samples. This reflects that public awareness is steeped in the claims of knowledge and cannot evaluate each of them, so that a consensus is formed on the sum of all votes. Unlike the individual innovator who has a narrow focus and notes whether its patent is generated in the end, the public sphere has limited attention and is forced to accept the total contributions of the ideas market
As a result, individuals’ investments in innovation have positive and positive operations to the extent that they can transfer general knowledge towards the truth. However, if a very large number of people invested in “popular” or “central” knowledge by taking samples from the broken distribution, this may have a negative external, by distorting general knowledge towards the center and alleviating its disclosure.[6]
We also offer the possibility of generations in some models to explore the impact on the collapse of knowledge. This can be taken either as the literal generations of humans, as in the models of the “overlapping” economic “generation” (for example, Will, 2008), or instead reflects the luar nature of relying on the interlocking AI systems, which can generate the same The result is within the fast time frame
In the version of the model with the change of generations, the new generation PDF takes the current year to be a representative, and thus the samples of the distribution begins with the same contrast (perhaps the smaller) (and in return the limits of deduction are updated). This was explained in terms of human generations, and this can be understood as the new generation that determines the “cognitive horizon” based on the previous generation. This means that the new generation may reduce the breadth of possible knowledge and then depend on these perceived limits to restrict their research.[7] The information series model can justify such a position if individuals assume that the previous actors would have invested in the knowledge of the removal if they are of value, and thus take the absence of information that it indicates that they must be of little value.[8]
The second interpretation sees these “generations” not in terms of human population, but as a result of the linguistic dynamics between the systems of artificial intelligence, such as when the user reads a summary created from artificial intelligence to the artificial research article written from artificial intelligence, which was created in itself from the Wikipedia articles edited with Artificial intelligence, etc., a great version of the phone game.
[2] For more discussion and more accurate definition, see the appendix.
[3] Fullly identical copy code is available on: https://github.com/aristotle-tek/knowndLedge-collapse
[4] Changing this has a commercial effect on the model, although the higher values can distort general knowledge.
[6] If individuals knew that they were taking samples from the broken distribution, they can use the expected algorithm to restore the full distribution, but again this process aims to be metaphorical, and there is no realistic method known to restore knowledge of the source from the source the content created by artificial intelligence.
[7] Zamora-Bonilla (2010, p. 328) proposes a scientific process for “Verisimiltude”, where we judge the evidence with reference to the objective truth through “the perceived rapprochement of what we know experimentally about the truth, weighed in the quantity of the information this experimental knowledge contains.” To obtain a more modern review of models and experiences on the transmission of human culture, see (Mesoudi and Whiten, 2008), especially the Henrich model (2004), which tries to explain how Tasmania has a number of useful technologies over time.
[8] For example, Christian societies sometimes promoted “ecclesiastical” texts actively with the neglect or ban of others, which led to the exclusion of those who were excluded from reproduction by scribes to have little value. Perhaps it was possible that the opinion made by Aristarkos from Samos in the third century BC is easier (repeated) if his work was not neglected (Rousseau and others, 2003, Chapter 3). A number of authors, such as Basilidis, are known today only through the texts that condemn (and sometimes distort) their views (Layton, 1989).
