<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE article PUBLIC "-//NLM//DTD JATS (Z39.96) Journal Publishing DTD v1.3 20210610//EN" "JATS-journalpublishing1-3.dtd">
<article article-type="research-article" dtd-version="1.3" xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xml:lang="ru"><front><journal-meta><journal-id journal-id-type="publisher-id">cyberspace</journal-id><journal-title-group><journal-title xml:lang="ru">Философские проблемы информационных технологий и киберпространства</journal-title><trans-title-group xml:lang="en"><trans-title>Philosophical Problems of IT &amp; Cyberspace (PhilIT&amp;C)</trans-title></trans-title-group></journal-title-group><issn pub-type="epub">2305-3763</issn><publisher><publisher-name>Пятигорский государственный университет</publisher-name></publisher></journal-meta><article-meta><article-id pub-id-type="doi">10.17726/philIT.2024.2.6</article-id><article-id custom-type="elpub" pub-id-type="custom">cyberspace-327</article-id><article-categories><subj-group subj-group-type="heading"><subject>Research Article</subject></subj-group><subj-group subj-group-type="section-heading" xml:lang="ru"><subject>Статьи</subject></subj-group></article-categories><title-group><article-title>Почему трансформеры не мыслят, как люди?</article-title><trans-title-group xml:lang="en"><trans-title>Why don’t transformers think like humans?</trans-title></trans-title-group></title-group><contrib-group><contrib contrib-type="author" corresp="yes"><name-alternatives><name name-style="eastern" xml:lang="ru"><surname>Хомяков</surname><given-names>А. Б.</given-names></name><name name-style="western" xml:lang="en"><surname>Khomyakov</surname><given-names>A. B.</given-names></name></name-alternatives><bio xml:lang="ru"><p>Александр Б. Хомяков, магистр физических наук</p><p>Санкт-Петербург </p></bio><bio xml:lang="en"><p>Alexander B. Khomyakov, Master of Physical Sciences</p><p>Saint-Petersburg </p></bio><email xlink:type="simple">alexander.xom@gmail.com</email></contrib></contrib-group><pub-date pub-type="collection"><year>2024</year></pub-date><pub-date pub-type="epub"><day>14</day><month>01</month><year>2025</year></pub-date><volume>0</volume><issue>2</issue><fpage>87</fpage><lpage>98</lpage><permissions><copyright-statement>Copyright &amp;#x00A9; Хомяков А.Б., 2025</copyright-statement><copyright-year>2025</copyright-year><copyright-holder xml:lang="ru">Хомяков А.Б.</copyright-holder><copyright-holder xml:lang="en">Khomyakov A.B.</copyright-holder><license xml:lang="ru" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>Данная работа распространяется под лицензией Creative Commons Attribution 4.0.</license-p></license><license xml:lang="en" license-type="creative-commons-attribution" xlink:href="https://creativecommons.org/licenses/by/4.0/" xlink:type="simple"><license-p>This work is licensed under a Creative Commons Attribution 4.0 License.</license-p></license></permissions><self-uri xlink:href="https://cyberspace.pgu.ru/jour/article/view/327">https://cyberspace.pgu.ru/jour/article/view/327</self-uri><abstract><p>Большие языковые модели в виде чат-ботов очень правдоподобно имитируют диалог как всезнающий собеседник и поэтому получили широкое распространение. Но даже в чат-боте Google Gemini не советуют доверять тому, что напишет чат-бот, и просят проверять его ответы. В данном обзоре будут проанализированы различные типы ошибок LLM, такие как проклятие инверсии, обработка чисел и др., чтобы выявить их причины. Такой анализ привел к выводу об общих причинах ошибок, заключающихся в том, что трансформеры не обладают глубокой аналогией, абстракцией и избирательностью контента, учитываемого в вычислении ответа (inference). Но наиболее важным выводом является то, что трансформеры, как и другие нейросети, построены по концепции обработки входного сигнала, что создает сильную зависимость от нерелевантной информации, которую не может компенсировать слой внимания трансформера. Концепция нейросетей была заложена в 1950-х идеей перцептрона Ф. Розенблата и не учитывала тех достижений когнитивной психологии, которые появились позже. Согласно же конструктивистской парадигме, входной слой (или перцепция) является только способом проверки правильности сконструированной предиктивной модели для возможных ситуаций. Это же служит причиной самой большой проблемы трансформеров, называемой галлюцинациями. И устранение ее возможно только при изменении архитектуры нейросети, а не за счет большего количества данных в обучении.</p></abstract><trans-abstract xml:lang="en"><p>Large language models in the form of chatbots very realistically imitate a dialogue as an omniscient interlocutor and therefore have become widespread. But even Google in its Gemini chatbot does not recommend trusting what the chatbot will write and asks to check its answers. In this review, various types of LLM errors such as the curse of inversion, number processing, etc. will be analyzed to identify their causes. Such an analysis led to the conclusion about the common causes of all errors, which is that transformers do not have deep analogy, hierarchy of schemes and selectivity of content taken into account in the inference. But the most important conclusion is that transformers, like other neural networks, are built on the concept of processing the input signal, which creates a strong dependence on superficial noise and irrelevant information that the transformer’s attention layer cannot compensate for. The concept of neural networks was laid down in the 1950s by the idea of F. Rosenblatt’s perceptron and did not take into account the achievements of cognitive psychology that appeared later. According to the constructivist paradigm, the input word (or perception) is only a way to check the correctness of the constructed predictive model for possible situations. This is the cause of the biggest problem of transformers, called hallucinations. And its elimination is possible only by changing the architecture of the neural network, but not by increasing the amount of data in training.</p></trans-abstract><kwd-group xml:lang="ru"><kwd>LLM</kwd><kwd>трансформеры</kwd><kwd>мышление</kwd><kwd>аналогия</kwd><kwd>когнитивная психология</kwd><kwd>перцептивный цикл</kwd><kwd>галлюцинации</kwd></kwd-group><kwd-group xml:lang="en"><kwd>LLM</kwd><kwd>transformers</kwd><kwd>thinking</kwd><kwd>analogy</kwd><kwd>cognitive psychology</kwd><kwd>perceptual cycle</kwd></kwd-group></article-meta></front><back><ref-list><title>References</title><ref id="cit1"><label>1</label><citation-alternatives><mixed-citation xml:lang="ru">Robison K. OpenAI cofounder Ilya Sutskever says the way AI is built is about to change // The Verge, Dec 14, 2024. https://www.theverge.com/2024/12/13/24320811/what-ilya-sutskever-sees-openai-model-datatraining.</mixed-citation><mixed-citation xml:lang="en">Robison K. OpenAI cofounder Ilya Sutskever says the way AI is built is about to change // The Verge, Dec 14, 2024. https://www.theverge.com/2024/12/13/24320811/what-ilya-sutskever-sees-openai-model-datatraining.</mixed-citation></citation-alternatives></ref><ref id="cit2"><label>2</label><citation-alternatives><mixed-citation xml:lang="ru">Найссер У. Познание и реальность. М.: Прогресс, 1981. С. 42-43. 230 с.</mixed-citation><mixed-citation xml:lang="en">Neisser W. Cognition and Reality. M.: Progress, 1981. P. 42-43. 230 p.</mixed-citation></citation-alternatives></ref><ref id="cit3"><label>3</label><citation-alternatives><mixed-citation xml:lang="ru">Mitchell M. How do we know how smart ai systems are? // Science, 381(6654). https://www.science.org/doi/10.1126/science.adj5957.</mixed-citation><mixed-citation xml:lang="en">Mitchell M. How do we know how smart ai systems are? // Science, 381(6654). https://www.science.org/doi/10.1126/science.adj5957.</mixed-citation></citation-alternatives></ref><ref id="cit4"><label>4</label><citation-alternatives><mixed-citation xml:lang="ru">Nezhurina M. Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models. https://arxiv.org/html/2406.02061v1#bib.bib12, 04 Jun 2024.</mixed-citation><mixed-citation xml:lang="en">Nezhurina M. Alice in Wonderland: Simple Tasks Showing Complete Reasoning Breakdown in State-Of-the-Art Large Language Models. https://arxiv.org/html/2406.02061v1#bib.bib12, 04 Jun 2024.</mixed-citation></citation-alternatives></ref><ref id="cit5"><label>5</label><citation-alternatives><mixed-citation xml:lang="ru">Лурия А. Р. Основы нейропсихологии: учеб. пособие. М.:Издательский центр «Академия», 2003. 384 с., с. 123-126. (</mixed-citation><mixed-citation xml:lang="en">Luria A. R. Fundamentals of Neuropsychology: textbook. M.: Publishing Center “Academy”, 2003. 384 p., p. 123-126.</mixed-citation></citation-alternatives></ref><ref id="cit6"><label>6</label><citation-alternatives><mixed-citation xml:lang="ru">Berglund L. and etc. The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A”. https://arxiv.org/abs/2309.12288, 26 May 2024.</mixed-citation><mixed-citation xml:lang="en">Berglund L. and etc. The Reversal Curse: LLMs trained on “A is B” fail to learn “B is A”. https://arxiv.org/abs/2309.12288, 26 May 2024.</mixed-citation></citation-alternatives></ref><ref id="cit7"><label>7</label><citation-alternatives><mixed-citation xml:lang="ru">Hagag B. Discover What Every Neuron in the Llama Model Does // Towards Data Science. https://towardsdatascience.com/-0927524e4807, Oct 25, 2024.</mixed-citation><mixed-citation xml:lang="en">Hagag B. Discover What Every Neuron in the Llama Model Does // Towards Data Science. https://towardsdatascience.com/-0927524e4807, Oct 25, 2024.</mixed-citation></citation-alternatives></ref><ref id="cit8"><label>8</label><citation-alternatives><mixed-citation xml:lang="ru">Хомяков А. Б., Чижик П. Новый способ нахождения аналогов как возможность исследования языка, мышления и построения систем искусственного интеллекта // Философские проблемы информационных технологий и киберпространства. 2024. № 1. С. 77-88. https://doi.org/10.17726/philIT.2024.1.5.</mixed-citation><mixed-citation xml:lang="en">Khomyakov A. B., Chizhik P. A new way of finding analogues as an opportunity to study language, thinking and build artificial intelligence systems // Philosophical Problems of IT &amp; Cyberspace (PhilIT&amp;C). 2024. № 1. Р. 77-88. (In Russ.). https://doi.org/10.17726/philIT.2024.1.5.</mixed-citation></citation-alternatives></ref><ref id="cit9"><label>9</label><citation-alternatives><mixed-citation xml:lang="ru">De La Cruz R. Frank Rosenblatt’s Perceptron, Birth of The Neural Network // Medium. https://medium.com/@robdelacruz/frank-rosenblattsperceptron-19fcce9d627f, 1 November 2023.</mixed-citation><mixed-citation xml:lang="en">De La Cruz R. Frank Rosenblatt’s Perceptron, Birth of The Neural Network // Medium. https://medium.com/@robdelacruz/frank-rosenblattsperceptron-19fcce9d627f, 1 November 2023.</mixed-citation></citation-alternatives></ref><ref id="cit10"><label>10</label><citation-alternatives><mixed-citation xml:lang="ru">Tiehen J. Perception as controlled hallucination // ResearchGate.org. https://www.researchgate.net/publication/359601789_Perception_as_controlled_hallucination march 2022.</mixed-citation><mixed-citation xml:lang="en">Tiehen J. Perception as controlled hallucination // ResearchGate.org. https://www.researchgate.net/publication/359601789_Perception_as_controlled_hallucination march 2022.</mixed-citation></citation-alternatives></ref><ref id="cit11"><label>11</label><citation-alternatives><mixed-citation xml:lang="ru">Хоффман Д. Как нас обманывают органы чувств. М.: АСТ, 2022.</mixed-citation><mixed-citation xml:lang="en">Hoffman D. How Our Senses Deceive Us. M.: AST, 2022.</mixed-citation></citation-alternatives></ref></ref-list><fn-group><fn fn-type="conflict"><p>The authors declare that there are no conflicts of interest present.</p></fn></fn-group></back></article>
