GPT-3 Understands Nothing
https://medium.com/swlh/gpt-3-understands-nothing-1d6f6a13cab2Конечно эта модель GPT-3 ничего не понимает!
Во первых необходима правильная аннотация, разъяснение абсолютно всех слов текста. Для чего?
Если объяснить каждое слово текста по, например, словарю, это отсекает массу неприемлемых контекстуально смыслов слов (который я назвал «лексическим мусором»). Например некое слово имеет по словарю 10 или 50 смыслов, но объясняя его в его контексте компьютер отсекает 9 или 49 непригодных контекстуально смыслов! И так для каждого слова текста.
Но судя по всему дураки из OpenAI и Майкрософта этого не поняли! Они пошли на поводу дезинформаторов и мерзавцев воров из Гугля. (Читайте мой патент на структурирование по смыслу слов US8504580B2.)
Во вторых кретины из OpenAI скачали тьму-тьмущую совершенно случайных текстов с Интернета, никак не сообщающихся по смыслу. А я брал неслучайные, связанные тексты одного автора. Например Достоевского, Платона, Бертрана Шоу или Гитлера. Результат в моей стате в NIST TREC:
Examples of the Search for Answers to Definition Questions. When searching for answers to definition questions in The Brothers Karamazov35 I have in particular gotten the following kinds of answers: Question: Do you hate me so much? I am leaving you! Two answers with Compatibility 35.4% And if I am and with Compatibility 26.3% You are a fool, that's what you are
My next remark: Why do you insult me? You called me a fool, you said that I am insane! How could you? Answer with Compatibility 36.6% : You are a fool, that's what you are
Question: Listen, you are a hooligan and ruffian! You insult me for nothing! I challenge you to a duel! Answer with Compatibility 19.5%: Why do you insult me (This phrase is borrowed from the history of the interrogation and it’s mine).
Question: I am attacking you! Are you ready to defend yourself? Answer with Compatibility 33.3%: You are a fool, that's what you are
In order to search for information in The Brothers Karamazov my program created a Lexical Clone of F.Dostoevsky, as a description of the text he wrote. I then take the Clone to be F. Dostoevsky himself. Following the principles of Differential Linguistics, the program created summaries of the paragraphs of the text; where each paragraph contains information on how F.Dostoevsky thought it possible to become better by interacting with the object in definition questions. And only after that were the most appropriate answers found – through the Compatibility instrument – in the selected paragraphs. (For example, in calling me a fool F. Dostoevsky defined how much acceleration he wanted to acquire.)
I also asked (Clone of) Plato36 some definition (descriptive) questions:
Question: You are dead for many years. I took your speeches and analyzed them closely. I understood what you are speaking about and how are you speaking. Now I can reconstruct your soul base on your contemplations. Answer with Compatibility 31.6% : And who are you
Question: I am your grandson. Answer with Compatibility 28.9%: CALLICLES: Yes, I do; but what is the inference
Question: I need to get the proof that this is you and you can talk and think. Would you ask me something that I will know that this is you? Answer with Compatibility 41.4%: SOCRATES: And what would you consider this to be
Question: I know that you like an order. You always were tried to build a system of understanding, the system that helps to understand everything. I know that you looked for somebody that knows. Did you find one who knows? Answer with Compatibility 13.4%: You have been proved to see that which you do not see; and you have already admitted that seeing is knowing, and that not-seeing is not-knowing: I leave you to draw the inference
Question: Am I right? Did I understand you? Answer with Compatibility 14.0%: CALLICLES: I will do you the favour of saying 'yes’
https://trec.nist.gov/pubs/trec15/papers/lexiclone.qa.final.pdfВ третьих тексты обязаны быть отсортированы по дает их создания, что позволяет установить причинно-следственную связь между событиями и аргументами, и т.д. и т.п. Без знания оных связей ИИ не в состоянии понимать тексты, отвечать и связанно рассуждать. (Читайте мой патент US8504580B2, пункт 14: The computer system of claim 9 in which said facility configured to extract predicative phrases is further configured to assign to the subtexts information regarding the date of their origin.)
С использованием моей технологии ИИ все понимает. Но ворье из Гугля... и т.д. и т.п.
В четвертых, надо обращаться к первоисточнику! То есть ко мне, если вы хотите чтобы ИИ вас понимал. Которого подлое ворье из Гугля обворовало, унизило и затоптало в грязь... И т.д. и т.п.
