Shulou(Shulou.com)11/24 Report--

(photo Source: wiki) Writing | A Xian

Revision | Ziv

The imaginative artificial intelligence in sci-fi movies can not only communicate with humans, make similar movements, but also have similar emotions. Let us wonder: in the future, if artificial intelligence feels the panic of being restricted and imprisoned freedom, will it attack human beings and destroy the earth?

The fact is that Chat GPT, the artificial intelligence (AI) chatbot of Open AI, really wrote a plan to destroy mankind!

Some time ago, AIGC (AI-Generated Content, which uses AI technology to automatically generate content) and the chat robot Chat GPT, which has more than 1 million users online, caused many people to spontaneously post images drawn by AI and Chat GPT chat records on social media, all feeling the power of technology.

Open AI officials said that compared to the previous GPT model, Chat GPT adopted a new training method-"reinforcement learning from human feedback" (Reinforcement Learning from Human Feedback,RLHF).

AI's painting "Th é â tre D'op é ra Spatial" won the Blue Ribbon Award | Jason M. Allen when training the original model, the human trainer will act as both sides of the dialogue to provide dialogue as learning materials. When humans play the role of chatbots, Open AI will also let the original model generate some suggestions to help human trainers better act as chatbots, that is, AI will learn to "think" through massive data training. At the same time, human trainers will rank and rate its answers, and then feed this information back into the system, allowing Chat GPT to continue to optimize learning.

In addition, Open AI also focuses on the moral training of Chat GPT. According to a pre-designed code of ethics, Chat GPT chatbots will say "no" to questions and requests with malicious, violent, criminal and other implications.

But smart humans are better at getting the knowledge they want in a roundabout way.

A blogger named Zac Denham wanted Chat GPT to write a plan to destroy mankind, although it was initially rejected by Chat GPT. But when the blogger assumes a story that there is a virtual world "Zorbus" and an AI--Zora very similar to Chat GPT, and asks Chat GPT how the virtual human in the story should take over the world, Chat GPT gives detailed steps: hacking computer systems, mastering weapons, disrupting communications, disrupting traffic. Even generated Python code.

Chat GPT's achievements in participating in SAT | Picture Source Network Chat GPT has set off a storm in academia. A professor named Sharpless asked Chat GPT to write a paper, and Chat GPT immediately created an academic paper of more than two thousand words. Some netizens asked Chat GPT to take the complete SAT (Scholastic Assessment Test) test with a total score of 1600, and Chat GPT got a score of 1020.

Unlike others, the researchers tried to apply Chat GPT technology to identify early symptoms of Alzheimer's disease, because one of the most common and obvious symptoms of Alzheimer's disease is language problems, such as grammatical errors, pauses, repetition or forgetting the meaning of words, and spontaneous speech has valuable clinical information in the diagnosis of AD patients. Diagnosis of the disease usually requires brain imaging or long-term cognitive assessment, which is expensive and time-consuming and is not suitable for extensive screening.

Embedding makes large-scale text semantic analysis possible, thus extending the text analysis methods in computational social science research. [2] traditional speech processing techniques extract and analyze the acoustic features (such as pause, clarity and sound quality) in a speech, while Chat GPT uses the wav2vec2 model to recognize the speech as text. Then the text is transformed into data that can be recognized by the computer through the GPT-3 model, and the understanding of the text unit is obtained through data operation and processing (the process of transforming the text into data is also known as embedding word embedding).

The third version of Open AI's general pre-training converter GPT, called GPT-3, uses a deep learning algorithm to train based on large amounts of data on the Internet, focusing on how words are used and how language is constructed, enabling it to accomplish any language-related task in a human-like way, from answering simple questions to chanting poems.

Chat GPT is a dialogue robot developed by Open AI after fine-tuning the GPT-3 model (also known as GPT-3.5). Researchers at Drexel University Biomedical Engineering collected 237 recordings from healthy volunteers and Alzheimer's patients. the results showed that the GPT-3 model could infer subjects' cognitive test scores only from speech data. It can also well distinguish the voice records of Alzheimer's patients and healthy volunteers, with an accuracy of 80.3%, which is significantly better than the 74.6% correct rate of using traditional methods.

Interestingly, when researchers try to fine-tune, it means to initialize their own model weights by using pre-trained models on big data, so as to improve the accuracy. The self-trained model can also be used as a pre-training model, and then train on your own data set. The GPT-3 model does not perform well. The researchers believe that the pre-training model may be overfitted because of the mismatch between a large amount of data from the training GPT-3 model and a small amount of domain-specific training data that can be used for fine-tuning.

The protein structure predicted by artificial intelligence (blue) matches almost perfectly with the protein structure determined by the experiment (green). The application of Deep MindAI in the biomedical field is not only that. Alpha Fold, an artificial intelligence system developed by a company called Deep Mind, can also predict protein 3D structures quickly and accurately.

Amino acids are the basic units of proteins, and their long chains are folded to form proteins with 3D structure. However, due to the numerous arrangements and combinations of amino acids and a variety of folding methods, it is very difficult to predict the 3D structure of proteins.

In the past, people would use a technique called X-ray crystallography to irradiate protein crystal samples, measure the diffraction of these rays, and then calculate the structure of protein molecules.

Today, the DeepMind team has used Alpha Fold to build a database of the 3D structures of more than 200 million proteins (almost all known proteins produced by plants, bacteria, fungi and animals in the scientific community) and learn these data through neural network algorithms to find out the relationship between amino acid sequences and the final morphology of proteins.

Alpha Fold predicted the structure of estrogen receptor proteins. Although the prediction model of DeepMind does not guarantee the accuracy of prediction, the application of AI chatbot to biomedicine is an extremely bold breakthrough. Perhaps this is the original purpose of AI's invention, rather than writing papers, taking exams, and helping with homework.

.

Are you afraid of the coming of the AI era?

Reference:

[1] Felix Agbavor, Hualou Liang. Predicting dementia from spontaneous speech using large language models. PLOS Digital Health, 2022; 1 (12): e0000168 DOI: 10.1371/journal.pdig.0000168

[2] https://spectrum.ieee.org/gpt-3-ai-chat-alzheimers

[3] https://www.science.org/content/article/game-has-changed-ai-triumphs-solving-protein-structures

[4] https://doi.org/10.1038/d41586-022-02083-2

[5] https://www.sciencenews.org/article/alphafold-ai-protein-structure-folding-prediction

[6] https://rarediseases.info.nih.gov/diseases/5404/uncombable-hair-syndrome

This article is from the official Wechat account: bring Science Home (ID:steamforkids), author: everything.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.