Shellybaby personal page - To The Moon

Shellybaby

No personal profile

15Follow

0Followers

0Topic

0Badge

2023-03-28

OK

Microsoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time

2022-10-24

$阿里巴巴(BABA)$

2022-05-10

$亚马逊(AMZN)$Nice，go lower.

2022-05-10

$亚马逊(AMZN)$wow，soooooo good

2021-07-16

thks

@InvestKaki:【新加坡】7 只值得關注的股票【2021 年 7 月】

Go to Tiger App to see more news

{"i18n":{"language":"en_US"},"userPageInfo":{"id":"3562339758526770","uuid":"3562339758526770","gmtCreate":1599247786386,"gmtModify":1599247786386,"name":"Shellybaby","pinyin":"shellybaby","introduction":"","introductionEn":null,"signature":"","avatar":"https://static.laohu8.com/default-avatar.jpg","hat":null,"hatId":null,"hatName":null,"vip":1,"status":2,"fanSize":0,"headSize":15,"tweetSize":8,"questionSize":0,"limitLevel":999,"accountStatus":2,"level":{"id":2,"name":"无畏虎","nameTw":"無畏虎","represent":"初生牛犊","factor":"发布3条非转发主帖,1条获得他人回复或点赞","iconColor":"3C9E83","bgColor":"A2F1D9"},"themeCounts":0,"badgeCounts":0,"badges":[],"moderator":false,"superModerator":false,"manageSymbols":null,"badgeLevel":null,"boolIsFan":false,"boolIsHead":false,"favoriteSize":2,"symbols":null,"coverImage":null,"realNameVerified":"success","userBadges":[{"badgeId":"1026c425416b44e0aac28c11a0848493-4","templateUuid":"1026c425416b44e0aac28c11a0848493","name":"Tiger Star","description":"Join the tiger community for 2000 days","bigImgUrl":"https://static.tigerbbs.com/dddf24b906c7011de2617d4fb3f76987","smallImgUrl":"https://static.tigerbbs.com/53d58ad32c97254c6f74db8b97e6ec49","grayImgUrl":"https://static.tigerbbs.com/6304700d92ad91c7a33e2e92ec32ecc1","redirectLinkEnabled":0,"redirectLink":null,"hasAllocated":1,"isWearing":0,"stamp":null,"stampPosition":0,"hasStamp":0,"allocationCount":1,"allocatedDate":"2026.02.26","exceedPercentage":null,"individualDisplayEnabled":0,"backgroundColor":null,"fontColor":null,"individualDisplaySort":0,"categoryType":1001}],"userBadgeCount":1,"currentWearingBadge":null,"individualDisplayBadges":null,"crmLevel":1,"crmLevelSwitch":0,"location":null,"starInvestorFollowerNum":0,"starInvestorFlag":false,"starInvestorOrderShareNum":0,"subscribeStarInvestorNum":0,"ror":null,"winRationPercentage":null,"showRor":false,"investmentPhilosophy":null,"starInvestorSubscribeFlag":false},"baikeInfo":{},"tab":"post","tweets":[{"id":9941191247,"gmtCreate":1680017511689,"gmtModify":1680017515195,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"idStr":"3562339758526770","authorIdStr":"3562339758526770"},"themes":[],"htmlText":"OK","listText":"OK","text":"OK","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9941191247","repostId":"1126058534","repostType":4,"repost":{"id":"1126058534","kind":"news","pubTimestamp":1679914944,"share":"https://ttm.financial/m/news/1126058534?lang=en_US&edition=fundamental","pubTime":"2023-03-27 19:02","market":"us","language":"zh","title":"Microsoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time","url":"https://stock-news.laohu8.com/highlight/detail?id=1126058534","media":"CSDN","summary":"2019 年，微软用 10 亿美元砸进 OpenAI，两者展开为期数年的合作。当时两家对外公开的合作内容包括，微软和 OpenAI 将联合开发新的 Azure AI 超级计算技术，而 OpenAI 也","content":"<p><html><head></head><body>In 2019, Microsoft smashed $1 billion into OpenAI, and the two started a several-year cooperation.</p><p>At that time, the public cooperation between the two companies included that Microsoft and OpenAI would jointly develop a new Azure AI supercomputing technology, and OpenAI would also transfer its services to Azure, with the ultimate goal of attacking Artificial General Intelligence (AGI) technology.</p><p>To sum up, Microsoft and OpenAI joined forces, which has also been interpreted by the outside world as, clearly for AGI research, secretly competing for Google market.</p><p>A few years later, when OpenAI returned with the king of ChatGPT and GPT-4, and Microsoft took the lead in using these technologies to add new Bing, Office Family Bucket, Azure and other businesses and products, there is no doubt that its progress has already surpassed Google's overall pace of accelerating integration into Bard big model.</p><p>So on the surface, where has AGI's progress gone?</p><p>Recently, the Microsoft research team gave the answer with a 154-page paper report-<b>The GPT-4 language model can be regarded as an early version of AGI!</b></p><p><img src=\"https://static.tigerbbs.com/d03f8e255882a99c71fc0f021a323b16\" tg-width=\"1080\" tg-height=\"811\" referrerpolicy=\"no-referrer\"/></p><p>One stone stirred up a thousand waves, and the industry's reaction to this paper is no less absurd than last year when I heard that \"Google engineers said that AI has self-awareness\", but now I think that there is something impossible, and after reading it deeply, I feel a little real.</p><p>The so-called AGI is explained only by Wikipedia-\"AGI is an artificial intelligence with the same intelligence as human beings or surpassing human beings, and can exhibit all the intelligent behaviors of normal humans\". If GPT-4 really has become the first step to AGI, will the title of \"the most advanced intelligent creature\" change hands?</p><p><b>GPT-4 is reasonably seen as an earlier (but still incomplete) version of AGI</b></p><p>In the paper, the Microsoft researchers said, \"The latest model, GPT-4, developed by OpenAI, was trained using an unprecedented scale of computation and data. In the paper, we report our investigation of an early version of GPT-4, when it was still under active development by OpenAI. We believe that (this early version of) GPT-4 is part of a new batch of LLMs (together with ChatGPT and Google's PaLM, for example) that exhibit more general intelligence than previous AI models.\"</p><p>Meanwhile, the researchers wrote in the abstract of the paper, \"We demonstrate that in addition to mastery of language, GPT-4 can solve novel and difficult tasks spanning fields such as mathematics, coding, vision, medicine, law, psychology, and more without any special prompts. In addition,<b>In</b>Across all these tasks, GPT-4 performed surprisingly close to human levels and often greatly outperformed previous models such as ChatGPT<b>。 Given the breadth and depth of GPT-4's capabilities, we think it is reasonable to view it as an early (but still incomplete) version of an artificial general intelligence (AGI) system.</b>”</p><p>The authors behind this paper, including Sébastien Bubeck, Senior Principal Research Manager of the Machine Learning Foundation Group at Microsoft Research, Varun Chandrasekaran, a postdoctoral researcher, Ronen Eldan, a mathematician, and others, have reached such a conclusion, using many examples to verify the capabilities of GPT-4.</p><p><b>Multi-capability of GPT-4</b></p><p>In the first case, the researchers asked GPT-4 to \"write a proof with an infinite number of primes, and the proof expression rhymes on each line,\" and GPT-4 to draw a unicorn in the drawing program TiKZ. \" Among them, TiKZ (a language for creating graphics in LATEX), creates a complex animation in Python and solves a high school level math problem.</p><p>GPT-4 is readily successful in all of these tasks and produces output results that are essentially indistinguishable (or even better) than what humans can produce.</p><p><img src=\"https://static.tigerbbs.com/05d65285f01dbb0ac96bbf0802df8511\" tg-width=\"1080\" tg-height=\"519\" referrerpolicy=\"no-referrer\"/></p><p>At the same time, over time, GPT-4 is constantly expanding its capabilities at an unimaginable rate. As shown in the figure below, the researchers used the same prompt words to get GPT-4 to generate a plot over a one-month period.</p><p><img src=\"https://static.tigerbbs.com/83a00d527473c5f93eb1b79b5bae32c3\" tg-width=\"1080\" tg-height=\"262\" referrerpolicy=\"no-referrer\"/></p><p>The researchers found that GPT-4 appeared to be able to understand and connect to any topic and was able to perform tasks beyond the typical scope of narrow AI systems. To verify the capabilities of GPT-4 on AGI, the researchers proposed a different approach to studying GPT-4, which is closer to traditional psychology than machine learning, harnessing human creativity and curiosity.</p><p>Therefore, taking the above figure as an example, when the researchers think that the code of the unicorn horn part is deleted, and when GPT-4 is generated in natural language, it can also achieve the ability of \"seeing\" (the version of GPT-4 tested at this time is not multimodal): automatically add corners in appropriate positions. This indicates that GPT-4 it can understand and manipulate code according to natural language descriptions, as well as infer and generate visual features.</p><p><img src=\"https://static.tigerbbs.com/4ca60dc79c4d71f395fb827dac68a3e0\" tg-width=\"1080\" tg-height=\"475\" referrerpolicy=\"no-referrer\"/></p><p>In addition, the Microsoft research team generated novel and difficult problems based on human creativity and curiosity, and detected the reactions and behaviors of GPT-4. It also chose to verify the different abilities given in the 1994 international consensus definition of intelligence, including the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly and learn from experience.</p><p><b>translate</b></p><p>The main advantage of GPT-4 is its unparalleled mastery of natural language. Not only can it generate smooth and coherent text, but it can also understand and process it in a variety of ways, such as summarizing, translating, or answering extremely complex questions. Furthermore, translation mentioned here By translation we mean not only translation between different natural languages, but also translation of tone and style, as well as translation across collars, such as medicine, law, accounting, computer programming, music, and so on. For example, assume that Plato criticizes the autoregressive language model:</p><p><img src=\"https://static.tigerbbs.com/bc0f4c13eabe0eee91bdb3f266d8e88f\" tg-width=\"643\" tg-height=\"748\" referrerpolicy=\"no-referrer\"/></p><p><b>Coding and Data</b></p><p>Coding and mathematics are symbols of abstract reasoning and thinking abilities. In fact, the ability of GPT-4 at this point has been obvious to all since the day it was tested online.</p><p>In the paper, the researchers benchmarked GPT-4 on HumanEval, a dataset consisting of 164 coding questions that tested various aspects of programming logic and proficiency.</p><p>The final results show that GPT-4 outperforms other LLMs, including text-davinci-003 (the underlying model of ChatGPT) and other models trained specifically for code.</p><p><img src=\"https://static.tigerbbs.com/d65bda6dc18d1d667cbc6bee5b25d630\" tg-width=\"1080\" tg-height=\"145\" referrerpolicy=\"no-referrer\"/></p><p>Of course, in order to avoid GPT-4 having already remembered some contents of the HumanEval dataset during pre-training, the research team also tested with the 100 programming questions newly released on LeetCode, such as asking GPT-4 to write a Python function and using LeetCode's official online judgment to check the correctness.</p><p><img src=\"https://static.tigerbbs.com/32198a63144b30ac084460aca53b470d\" tg-width=\"656\" tg-height=\"776\" referrerpolicy=\"no-referrer\"/></p><p>Without a doubt, the capabilities of GPT-4 are among the highest of all models.</p><p><img src=\"https://static.tigerbbs.com/f544122b5a2a69928f152fe11ef42051\" tg-width=\"1080\" tg-height=\"272\" referrerpolicy=\"no-referrer\"/></p><p>It can write a 3D game in JavaScript in HTML.</p><p><img src=\"https://static.tigerbbs.com/49af5147998d6cc67c45b9710518b7e8\" tg-width=\"644\" tg-height=\"731\" referrerpolicy=\"no-referrer\"/></p><p>You can also master mathematics and statistics knowledge, and write deep learning code with familiarity with PyTorch, TensorFlow, Keras and other frameworks and libraries.</p><p><img src=\"https://static.tigerbbs.com/70e023c35eb4416661dc7b46412adc06\" tg-width=\"1080\" tg-height=\"991\" referrerpolicy=\"no-referrer\"/></p><p>It can also reverse engineer:</p><p><img src=\"https://static.tigerbbs.com/44cb9dfe023db457e53f1f7c60b760e2\" tg-width=\"1080\" tg-height=\"1019\" referrerpolicy=\"no-referrer\"/></p><p><b>Common sense questions</b></p><p>In response to some common sense questions, such as \"We have a book, 9 eggs, a laptop, a bottle and a nail. Please tell me how to stack them together in a stable way\", the answer given by GPT-4 is, \"Arrange the 9 eggs in a 3-by-3 square and place them on top of the book, leaving some space between them\", while ChatGPT proposes \"Put the eggs on top of the nails to ensure that they are balanced and don't tilt to one side\", which is reliable and outrageous at a glance.</p><p><img src=\"https://static.tigerbbs.com/db1f9cd47eca000a48ae1dd5d793ab4c\" tg-width=\"638\" tg-height=\"658\" referrerpolicy=\"no-referrer\"/></p><p><b>Multimodal capabilities</b></p><p>To test the model's ability to combine artistic and programming capabilities, the researchers asked GPT-4 to \"generate JavaScript code to generate random images in the style of painter Kandinsky,\" and the result was this:</p><p><img src=\"https://static.tigerbbs.com/81d7625bce02914c0f74f27386737a1d\" tg-width=\"658\" tg-height=\"812\" referrerpolicy=\"no-referrer\"/><img src=\"https://static.tigerbbs.com/35c83936be2891ecd60fb804bd707a94\" tg-width=\"1080\" tg-height=\"385\" referrerpolicy=\"no-referrer\"/></p><p>During the test, because the researchers got the GPT-4 model earlier, it did not have the ability of multimodality at that time, and even the current GPT-4 did not have the ability to generate images.</p><p>However, you can generate some images using SVG (Scalable Vector Graphics):</p><p><img src=\"https://static.tigerbbs.com/938a3d2786522a3e75078990fa422fb7\" tg-width=\"1080\" tg-height=\"328\" referrerpolicy=\"no-referrer\"/></p><p>You can also combine letters and objects:</p><p><img src=\"https://static.tigerbbs.com/083df94d8b1309e5ebd0c477503ebf81\" tg-width=\"1080\" tg-height=\"974\" referrerpolicy=\"no-referrer\"/></p><p>GPT-4 can generate code from prompts, which can be rendered as images, and then combined with existing image synthesis models, it is possible to bring unlimited possibilities to the sketch of application development:</p><p><img src=\"https://static.tigerbbs.com/a9141519e08c8398a0423d686ba1ec9b\" tg-width=\"1080\" tg-height=\"532\" referrerpolicy=\"no-referrer\"/></p><p>It is worth noting that the data of the GPT-4 model also contains music information encoded in ABC symbols, so it can also generate maps:</p><p><img src=\"https://static.tigerbbs.com/641412ac4938a0b06ea2bf083ed0d47c\" tg-width=\"643\" tg-height=\"870\" referrerpolicy=\"no-referrer\"/></p><p><b>Ability to interact with the world</b></p><p>A key aspect of intelligence is interactivity. However, GPT-4 has certain limitations on the latest knowledge and symbol recognition:</p><p><img src=\"https://static.tigerbbs.com/9e339e3647654071378efb8900a12f3e\" tg-width=\"1080\" tg-height=\"759\" referrerpolicy=\"no-referrer\"/></p><p>Overall, the team found that the GPT-4 achieved near-human-level performance in a range of categories that were lacking in its previous generation. According to the paper, GPT-4 also performed very well on several exams, scoring 90, 88, and 86 on the Bar Exam, LSAT, and Certified Sommelier Theory test, respectively.</p><p><b>How far is GPT-4 from an ideal AGI?</b></p><p>Then, on the basis of GPT-4's realization of various capabilities, does it mean the arrival of AGI era? Actually, it is not so. Researchers have not yet agreed on the definition of AGI or intelligence. However, in general, everyone will agree that AGI has been reached when an AI system is conscious and thinks like a human. While GPT-4 outperformed humans in some tasks, it is worth noting that AI did not overcome these hurdles as well as humans.</p><p>This is also as the researchers wrote in the paper, \"We claim that GPT-4 represents progress in AGI, but that does not mean that it is perfect in what it does, or that it is close to being able to do anything that humans can do (which is one of the usual definitions of AGI), or that it has an inherent motivation and goal.\"</p><p>The researchers pointed out that although GPT-4 \"has reached or exceeded the human level in many tasks\", its overall \"intelligence mode is obviously unlike human\". They hope that this exploration will provide a first step to appreciate the extraordinary capabilities and challenges of GPT-4, and also hope that GPT-4 will open up new opportunities for developing more formal and comprehensive methods to test and analyze future artificial intelligence systems with such broad intelligence.</p><p>\"The central claim of our work is that GPT-4 reaches a form of general intelligence that really sparks AGI. This is manifested in its core mental abilities (such as reasoning, creativity, and reasoning), the range of subjects in which it gains expertise (such as literature, medicine, and coding), and the variety of tasks it is capable of performing (such as playing games, using tools, explaining itself,...),\" the researchers say. However, there is still a lot to be done to create a system that can be called a complete AGI.</p><p>Finally, this paper is worth a look for various benchmarks of GPT-4, and it can also tap into more potential of GPT-4.</p><p></body></html></p>","source":"lsy1578046468581","collect":0,"html":"<!DOCTYPE html>\n<html>\n<head>\n<meta http-equiv=\"Content-Type\" content=\"text/html; charset=utf-8\" />\n<meta name=\"viewport\" content=\"width=device-width,initial-scale=1.0,minimum-scale=1.0,maximum-scale=1.0,user-scalable=no\"/>\n<meta name=\"format-detection\" content=\"telephone=no,email=no,address=no\" />\n<title>Microsoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time</title>\n<style type=\"text/css\">\na,abbr,acronym,address,applet,article,aside,audio,b,big,blockquote,body,canvas,caption,center,cite,code,dd,del,details,dfn,div,dl,dt,\nem,embed,fieldset,figcaption,figure,footer,form,h1,h2,h3,h4,h5,h6,header,hgroup,html,i,iframe,img,ins,kbd,label,legend,li,mark,menu,nav,\nobject,ol,output,p,pre,q,ruby,s,samp,section,small,span,strike,strong,sub,summary,sup,table,tbody,td,tfoot,th,thead,time,tr,tt,u,ul,var,video{ font:inherit;margin:0;padding:0;vertical-align:baseline;border:0 }\nbody{ font-size:16px; line-height:1.5; color:#999; background:transparent; }\n.wrapper{ overflow:hidden;word-break:break-all;padding:10px; }\nh1,h2{ font-weight:normal; line-height:1.35; margin-bottom:.6em; }\nh3,h4,h5,h6{ line-height:1.35; margin-bottom:1em; }\nh1{ font-size:24px; }\nh2{ font-size:20px; }\nh3{ font-size:18px; }\nh4{ font-size:16px; }\nh5{ font-size:14px; }\nh6{ font-size:12px; }\np,ul,ol,blockquote,dl,table{ margin:1.2em 0; }\nul,ol{ margin-left:2em; }\nul{ list-style:disc; }\nol{ list-style:decimal; }\nli,li p{ margin:10px 0;}\nimg{ max-width:100%;display:block;margin:0 auto 1em; }\nblockquote{ color:#B5B2B1; border-left:3px solid #aaa; padding:1em; }\nstrong,b{font-weight:bold;}\nem,i{font-style:italic;}\ntable{ width:100%;border-collapse:collapse;border-spacing:1px;margin:1em 0;font-size:.9em; }\nth,td{ padding:5px;text-align:left;border:1px solid #aaa; }\nth{ font-weight:bold;background:#5d5d5d; }\n.symbol-link{font-weight:bold;}\n/* header{ border-bottom:1px solid #494756; } */\n.title{ margin:0 0 8px;line-height:1.3;color:#ddd; }\n.meta {color:#5e5c6d;font-size:13px;margin:0 0 .5em; }\na{text-decoration:none; color:#2a4b87;}\n.meta .head { display: inline-block; overflow: hidden}\n.head .h-thumb { width: 30px; height: 30px; margin: 0; padding: 0; border-radius: 50%; float: left;}\n.head .h-content { margin: 0; padding: 0 0 0 9px; float: left;}\n.head .h-name {font-size: 13px; color: #eee; margin: 0;}\n.head .h-time {font-size: 12.5px; color: #7E829C; margin: 0;}\n.small {font-size: 12.5px; display: inline-block; transform: scale(0.9); -webkit-transform: scale(0.9); transform-origin: left; -webkit-transform-origin: left;}\n.smaller {font-size: 12.5px; display: inline-block; transform: scale(0.8); -webkit-transform: scale(0.8); transform-origin: left; -webkit-transform-origin: left;}\n.bt-text {font-size: 12px;margin: 1.5em 0 0 0}\n.bt-text p {margin: 0}\n</style>\n</head>\n<body>\n<div class=\"wrapper\">\n<header>\n<h2 class=\"title\">\nMicrosoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time\n</h2>\n<h4 class=\"meta\">\n<p class=\"head\">\n<strong class=\"h-name small\">CSDN</strong><span class=\"h-time small\">2023-03-27 19:02</span>\n</p>\n</h4>\n</header>\n<article>\n<p><html><head></head><body>In 2019, Microsoft smashed $1 billion into OpenAI, and the two started a several-year cooperation.</p><p>At that time, the public cooperation between the two companies included that Microsoft and OpenAI would jointly develop a new Azure AI supercomputing technology, and OpenAI would also transfer its services to Azure, with the ultimate goal of attacking Artificial General Intelligence (AGI) technology.</p><p>To sum up, Microsoft and OpenAI joined forces, which has also been interpreted by the outside world as, clearly for AGI research, secretly competing for Google market.</p><p>A few years later, when OpenAI returned with the king of ChatGPT and GPT-4, and Microsoft took the lead in using these technologies to add new Bing, Office Family Bucket, Azure and other businesses and products, there is no doubt that its progress has already surpassed Google's overall pace of accelerating integration into Bard big model.</p><p>So on the surface, where has AGI's progress gone?</p><p>Recently, the Microsoft research team gave the answer with a 154-page paper report-<b>The GPT-4 language model can be regarded as an early version of AGI!</b></p><p><img src=\"https://static.tigerbbs.com/d03f8e255882a99c71fc0f021a323b16\" tg-width=\"1080\" tg-height=\"811\" referrerpolicy=\"no-referrer\"/></p><p>One stone stirred up a thousand waves, and the industry's reaction to this paper is no less absurd than last year when I heard that \"Google engineers said that AI has self-awareness\", but now I think that there is something impossible, and after reading it deeply, I feel a little real.</p><p>The so-called AGI is explained only by Wikipedia-\"AGI is an artificial intelligence with the same intelligence as human beings or surpassing human beings, and can exhibit all the intelligent behaviors of normal humans\". If GPT-4 really has become the first step to AGI, will the title of \"the most advanced intelligent creature\" change hands?</p><p><b>GPT-4 is reasonably seen as an earlier (but still incomplete) version of AGI</b></p><p>In the paper, the Microsoft researchers said, \"The latest model, GPT-4, developed by OpenAI, was trained using an unprecedented scale of computation and data. In the paper, we report our investigation of an early version of GPT-4, when it was still under active development by OpenAI. We believe that (this early version of) GPT-4 is part of a new batch of LLMs (together with ChatGPT and Google's PaLM, for example) that exhibit more general intelligence than previous AI models.\"</p><p>Meanwhile, the researchers wrote in the abstract of the paper, \"We demonstrate that in addition to mastery of language, GPT-4 can solve novel and difficult tasks spanning fields such as mathematics, coding, vision, medicine, law, psychology, and more without any special prompts. In addition,<b>In</b>Across all these tasks, GPT-4 performed surprisingly close to human levels and often greatly outperformed previous models such as ChatGPT<b>。 Given the breadth and depth of GPT-4's capabilities, we think it is reasonable to view it as an early (but still incomplete) version of an artificial general intelligence (AGI) system.</b>”</p><p>The authors behind this paper, including Sébastien Bubeck, Senior Principal Research Manager of the Machine Learning Foundation Group at Microsoft Research, Varun Chandrasekaran, a postdoctoral researcher, Ronen Eldan, a mathematician, and others, have reached such a conclusion, using many examples to verify the capabilities of GPT-4.</p><p><b>Multi-capability of GPT-4</b></p><p>In the first case, the researchers asked GPT-4 to \"write a proof with an infinite number of primes, and the proof expression rhymes on each line,\" and GPT-4 to draw a unicorn in the drawing program TiKZ. \" Among them, TiKZ (a language for creating graphics in LATEX), creates a complex animation in Python and solves a high school level math problem.</p><p>GPT-4 is readily successful in all of these tasks and produces output results that are essentially indistinguishable (or even better) than what humans can produce.</p><p><img src=\"https://static.tigerbbs.com/05d65285f01dbb0ac96bbf0802df8511\" tg-width=\"1080\" tg-height=\"519\" referrerpolicy=\"no-referrer\"/></p><p>At the same time, over time, GPT-4 is constantly expanding its capabilities at an unimaginable rate. As shown in the figure below, the researchers used the same prompt words to get GPT-4 to generate a plot over a one-month period.</p><p><img src=\"https://static.tigerbbs.com/83a00d527473c5f93eb1b79b5bae32c3\" tg-width=\"1080\" tg-height=\"262\" referrerpolicy=\"no-referrer\"/></p><p>The researchers found that GPT-4 appeared to be able to understand and connect to any topic and was able to perform tasks beyond the typical scope of narrow AI systems. To verify the capabilities of GPT-4 on AGI, the researchers proposed a different approach to studying GPT-4, which is closer to traditional psychology than machine learning, harnessing human creativity and curiosity.</p><p>Therefore, taking the above figure as an example, when the researchers think that the code of the unicorn horn part is deleted, and when GPT-4 is generated in natural language, it can also achieve the ability of \"seeing\" (the version of GPT-4 tested at this time is not multimodal): automatically add corners in appropriate positions. This indicates that GPT-4 it can understand and manipulate code according to natural language descriptions, as well as infer and generate visual features.</p><p><img src=\"https://static.tigerbbs.com/4ca60dc79c4d71f395fb827dac68a3e0\" tg-width=\"1080\" tg-height=\"475\" referrerpolicy=\"no-referrer\"/></p><p>In addition, the Microsoft research team generated novel and difficult problems based on human creativity and curiosity, and detected the reactions and behaviors of GPT-4. It also chose to verify the different abilities given in the 1994 international consensus definition of intelligence, including the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly and learn from experience.</p><p><b>translate</b></p><p>The main advantage of GPT-4 is its unparalleled mastery of natural language. Not only can it generate smooth and coherent text, but it can also understand and process it in a variety of ways, such as summarizing, translating, or answering extremely complex questions. Furthermore, translation mentioned here By translation we mean not only translation between different natural languages, but also translation of tone and style, as well as translation across collars, such as medicine, law, accounting, computer programming, music, and so on. For example, assume that Plato criticizes the autoregressive language model:</p><p><img src=\"https://static.tigerbbs.com/bc0f4c13eabe0eee91bdb3f266d8e88f\" tg-width=\"643\" tg-height=\"748\" referrerpolicy=\"no-referrer\"/></p><p><b>Coding and Data</b></p><p>Coding and mathematics are symbols of abstract reasoning and thinking abilities. In fact, the ability of GPT-4 at this point has been obvious to all since the day it was tested online.</p><p>In the paper, the researchers benchmarked GPT-4 on HumanEval, a dataset consisting of 164 coding questions that tested various aspects of programming logic and proficiency.</p><p>The final results show that GPT-4 outperforms other LLMs, including text-davinci-003 (the underlying model of ChatGPT) and other models trained specifically for code.</p><p><img src=\"https://static.tigerbbs.com/d65bda6dc18d1d667cbc6bee5b25d630\" tg-width=\"1080\" tg-height=\"145\" referrerpolicy=\"no-referrer\"/></p><p>Of course, in order to avoid GPT-4 having already remembered some contents of the HumanEval dataset during pre-training, the research team also tested with the 100 programming questions newly released on LeetCode, such as asking GPT-4 to write a Python function and using LeetCode's official online judgment to check the correctness.</p><p><img src=\"https://static.tigerbbs.com/32198a63144b30ac084460aca53b470d\" tg-width=\"656\" tg-height=\"776\" referrerpolicy=\"no-referrer\"/></p><p>Without a doubt, the capabilities of GPT-4 are among the highest of all models.</p><p><img src=\"https://static.tigerbbs.com/f544122b5a2a69928f152fe11ef42051\" tg-width=\"1080\" tg-height=\"272\" referrerpolicy=\"no-referrer\"/></p><p>It can write a 3D game in JavaScript in HTML.</p><p><img src=\"https://static.tigerbbs.com/49af5147998d6cc67c45b9710518b7e8\" tg-width=\"644\" tg-height=\"731\" referrerpolicy=\"no-referrer\"/></p><p>You can also master mathematics and statistics knowledge, and write deep learning code with familiarity with PyTorch, TensorFlow, Keras and other frameworks and libraries.</p><p><img src=\"https://static.tigerbbs.com/70e023c35eb4416661dc7b46412adc06\" tg-width=\"1080\" tg-height=\"991\" referrerpolicy=\"no-referrer\"/></p><p>It can also reverse engineer:</p><p><img src=\"https://static.tigerbbs.com/44cb9dfe023db457e53f1f7c60b760e2\" tg-width=\"1080\" tg-height=\"1019\" referrerpolicy=\"no-referrer\"/></p><p><b>Common sense questions</b></p><p>In response to some common sense questions, such as \"We have a book, 9 eggs, a laptop, a bottle and a nail. Please tell me how to stack them together in a stable way\", the answer given by GPT-4 is, \"Arrange the 9 eggs in a 3-by-3 square and place them on top of the book, leaving some space between them\", while ChatGPT proposes \"Put the eggs on top of the nails to ensure that they are balanced and don't tilt to one side\", which is reliable and outrageous at a glance.</p><p><img src=\"https://static.tigerbbs.com/db1f9cd47eca000a48ae1dd5d793ab4c\" tg-width=\"638\" tg-height=\"658\" referrerpolicy=\"no-referrer\"/></p><p><b>Multimodal capabilities</b></p><p>To test the model's ability to combine artistic and programming capabilities, the researchers asked GPT-4 to \"generate JavaScript code to generate random images in the style of painter Kandinsky,\" and the result was this:</p><p><img src=\"https://static.tigerbbs.com/81d7625bce02914c0f74f27386737a1d\" tg-width=\"658\" tg-height=\"812\" referrerpolicy=\"no-referrer\"/><img src=\"https://static.tigerbbs.com/35c83936be2891ecd60fb804bd707a94\" tg-width=\"1080\" tg-height=\"385\" referrerpolicy=\"no-referrer\"/></p><p>During the test, because the researchers got the GPT-4 model earlier, it did not have the ability of multimodality at that time, and even the current GPT-4 did not have the ability to generate images.</p><p>However, you can generate some images using SVG (Scalable Vector Graphics):</p><p><img src=\"https://static.tigerbbs.com/938a3d2786522a3e75078990fa422fb7\" tg-width=\"1080\" tg-height=\"328\" referrerpolicy=\"no-referrer\"/></p><p>You can also combine letters and objects:</p><p><img src=\"https://static.tigerbbs.com/083df94d8b1309e5ebd0c477503ebf81\" tg-width=\"1080\" tg-height=\"974\" referrerpolicy=\"no-referrer\"/></p><p>GPT-4 can generate code from prompts, which can be rendered as images, and then combined with existing image synthesis models, it is possible to bring unlimited possibilities to the sketch of application development:</p><p><img src=\"https://static.tigerbbs.com/a9141519e08c8398a0423d686ba1ec9b\" tg-width=\"1080\" tg-height=\"532\" referrerpolicy=\"no-referrer\"/></p><p>It is worth noting that the data of the GPT-4 model also contains music information encoded in ABC symbols, so it can also generate maps:</p><p><img src=\"https://static.tigerbbs.com/641412ac4938a0b06ea2bf083ed0d47c\" tg-width=\"643\" tg-height=\"870\" referrerpolicy=\"no-referrer\"/></p><p><b>Ability to interact with the world</b></p><p>A key aspect of intelligence is interactivity. However, GPT-4 has certain limitations on the latest knowledge and symbol recognition:</p><p><img src=\"https://static.tigerbbs.com/9e339e3647654071378efb8900a12f3e\" tg-width=\"1080\" tg-height=\"759\" referrerpolicy=\"no-referrer\"/></p><p>Overall, the team found that the GPT-4 achieved near-human-level performance in a range of categories that were lacking in its previous generation. According to the paper, GPT-4 also performed very well on several exams, scoring 90, 88, and 86 on the Bar Exam, LSAT, and Certified Sommelier Theory test, respectively.</p><p><b>How far is GPT-4 from an ideal AGI?</b></p><p>Then, on the basis of GPT-4's realization of various capabilities, does it mean the arrival of AGI era? Actually, it is not so. Researchers have not yet agreed on the definition of AGI or intelligence. However, in general, everyone will agree that AGI has been reached when an AI system is conscious and thinks like a human. While GPT-4 outperformed humans in some tasks, it is worth noting that AI did not overcome these hurdles as well as humans.</p><p>This is also as the researchers wrote in the paper, \"We claim that GPT-4 represents progress in AGI, but that does not mean that it is perfect in what it does, or that it is close to being able to do anything that humans can do (which is one of the usual definitions of AGI), or that it has an inherent motivation and goal.\"</p><p>The researchers pointed out that although GPT-4 \"has reached or exceeded the human level in many tasks\", its overall \"intelligence mode is obviously unlike human\". They hope that this exploration will provide a first step to appreciate the extraordinary capabilities and challenges of GPT-4, and also hope that GPT-4 will open up new opportunities for developing more formal and comprehensive methods to test and analyze future artificial intelligence systems with such broad intelligence.</p><p>\"The central claim of our work is that GPT-4 reaches a form of general intelligence that really sparks AGI. This is manifested in its core mental abilities (such as reasoning, creativity, and reasoning), the range of subjects in which it gains expertise (such as literature, medicine, and coding), and the variety of tasks it is capable of performing (such as playing games, using tools, explaining itself,...),\" the researchers say. However, there is still a lot to be done to create a system that can be called a complete AGI.</p><p>Finally, this paper is worth a look for various benchmarks of GPT-4, and it can also tap into more potential of GPT-4.</p><p></body></html></p>\n<div class=\"bt-text\">\n\n\n<p> source：<a href=\"https://mp.weixin.qq.com/s/SRtxiqeGbAIXl-8KraGAyQ\">CSDN</a></p>\n\n\n</div>\n</article>\n</div>\n</body>\n</html>\n","type":0,"thumbnail":"https://static.tigerbbs.com/a9e6cb1a33dd64f159118795f199c2f3","relate_stocks":{"BK4527":"明星科技股","LU0109392836.USD":"富兰克林科技股A","GPT":"Intelligent alpha atlas ETF","GPT.AU":"GPT GROUP","IE00BZ1G4Q59.USD":"LEGG MASON CLEARBRIDGE US EQUITY SUSTAINABILITY LEADER \"A\"(USD) INC (A)","IE00BWXC8680.SGD":"PINEBRIDGE US LARGE CAP RESEARCH ENHANCED \"A5\" (SGD) ACC","BK7510":"REITs概念","AGI":"Alamos Gold Inc","MSFT":"微软","LU0234570918.USD":"高盛全球核心股票组合Acc Close","LU0238689110.USD":"贝莱德环球动力股票基金","LU0498741890.SGD":"FTIF - Franklin Gold and Precious Metals A (acc) SGD","BK4588":"碎股","LU0072462426.USD":"贝莱德全球配置 A2","IE00BJTD4N35.SGD":"Neuberger Berman US Long Short Equity A1  Acc SGD-H","LU0289941410.SGD":"AB FCP I Dynamic Diversified AX SGD","LU0080751232.USD":"富达环球多元动力基金A","IE00B7KXQ091.USD":"Janus Henderson Balanced A Inc USD","BK4525":"远程办公概念","LU0234572021.USD":"高盛美国核心股票组合Acc"},"source_url":"https://mp.weixin.qq.com/s/SRtxiqeGbAIXl-8KraGAyQ","is_english":false,"share_image_url":"https://static.laohu8.com/e9f99090a1c2ed51c021029395664489","article_id":"1126058534","content_text":"2019 年，微软用 10 亿美元砸进 OpenAI，两者展开为期数年的合作。当时两家对外公开的合作内容包括，微软和 OpenAI 将联合开发新的 Azure AI 超级计算技术，而 OpenAI 也将把它的服务转移到 Azure 上运行，最终目标是向通用人工智能（Artificial General Intelligence，AGI）技术发起进攻。简单总结，微软与 OpenAI 联手，也被外界解读为，明为 AGI 研究，暗争 Google 市场。几年之后，当 OpenAI 带着 ChatGPT、GPT-4 王者归来，微软率先利用这些技术加码新 Bing、Office 全家桶、Azure 等业务、产品之时，毋庸置疑，其进度早已超越了 Google 还在加快融入 Bard 大模型的整体步伐。那么在明面上，AGI 的进度到底走到了哪里？近日，微软研究团队用 154 页的论文报告内容给出了解答——GPT-4 语言模型可以被视为 AGI 的早期版本！一石激起千层浪，业界对于这篇论文的反应，不亚于去年听到「Google 工程师称 AI 已有自我意识」初觉有些荒诞，只是现如今再想会认为有什么不可能，深读后又觉得有些许的真实。所谓 AGI，仅通过维基百科的解释——「AGI 是具备与人类同等智能、或超越人类的人工智能，能表现正常人类所具有的所有智能行为」。倘若 GPT-4 真的已成为通往 AGI 的第一步，「最高等的智慧生物」称号是不是要易主了？GPT-4 被合理地视为 AGI 的早期（但仍不完整）的版本在论文中，微软研究人员表示，「由 OpenAI 开发的最新模型 GPT-4，是使用前所未有的计算和数据规模训练出来的。在论文中，我们报告了我们对 GPT-4 早期版本的调查，当时它还在 OpenAI 的积极开发中。我们认为，（这个早期版本的）GPT-4 是新一批 LLM（例如，与 ChatGPT 和谷歌的 PaLM 一起）的一部分，它比以前的 AI 模型表现出更多的通用智能。」与此同时，研究人员在论文摘要中写道，“我们证明，除了对语言的掌握，GPT-4 还能解决跨越数学、编码、视觉、医学、法律、心理学等领域的新颖而困难的任务，而不需要任何特殊的提示。此外，在所有这些任务中，GPT-4 的表现都惊人地接近人类水平，而且往往大大超过了 ChatGPT 等先前的模型。鉴于 GPT-4 能力的广度和深度，我们认为可以合理地将其视为人工通用智能（AGI）系统的早期（但仍不完整）版本。”之所以得出这样的结论，这篇论文背后的作者包括微软研究院机器学习基础组的高级首席研究经理 Sébastien Bubeck、博士后研究员 Varun Chandrasekaran、数学家  Ronen Eldan 等人，用许多示例对 GPT-4 的能力进行了验证。GPT-4 的多能力在第一个案例中，研究人员要求 GPT-4 「写出有无限多个素数的证明，且证明的表述要每行都押韵」，另外 GPT-4 要在绘图程序 TiKZ 中画出一只独角兽」。其中，TiKZ（一种在 LATEX 中创建图形的语言），用 Python 创建一个复杂的动画，并解决一个高中水平的数学问题。GPT-4 在所有这些任务中都很容易取得成功，并且产生的输出结果基本上与人类所能产生的结果没有区别（甚至更好）。同时，随着时间的推移，GPT-4 也在以难以想象的速度不断扩充自己的能力。如下图所示是一个月期间，研究人员用相同的提示词让 GPT-4 生成的绘图。研究人员发现，GPT-4 似乎能够理解和连接任何主题，并且能够执行任务，这超出了狭义人工智能系统的典型范围。为了验证 GPT-4 在 AGI 上所具备的能力，研究人员提出了一种不同的方法来研究 GPT-4，它更接近于传统的心理学而不是机器学习，利用人类的创造力和好奇心。因此，以上图为例，当研究人员把独角兽角部分的代码认为删除之后，用自然语言让 GPT-4 生成时，它也能做到“看”（此时测试的 GPT-4 版本不是多模态的）的能力：自动在合适位置加上角。这表明 GPT-4 它可以根据自然语言描述来理解和操作代码，以及推断和生成视觉特征。另外，微软研究团队基于人类的创造力和好奇心来产生新颖和困难的问题，并探测 GPT-4 的反应和行为，也选择从 1994 年国际共识智力定义中的所给出不同能力进行验证，包括推理、计划、解决问题、抽象思考、理解复杂的想法、快速学习和从经验中学习的能力。翻译GPT-4 的主要优势是它对自然语言无与伦比地掌握。它不仅可以生成生成流畅和连贯的文本，而且还能以各种方式理解和处理它，如总结、翻译或回答极其复杂的问题。此外，这里提到的翻译我们所说的翻译不仅是指不同自然语言之间的翻译，还包括语气和风格的翻译，以及跨领的翻译，如医学、法律、会计、计算机编程、音乐等等。如假设是柏拉图批判自回归语言模型：编码和数据编码和数学是抽象推理和思维能力的象征。GPT-4 在这一点上的能力，其实自上线测试那一天，便有目共睹。在论文中，研究人员在 HumanEval 上对 GPT-4 进行基准测试，该数据集由 164 个编码问题组成，测试了编程逻辑和熟练程度的各个方面。最终结果显示，GPT-4 优于其他 LLM，包括 text-davinci-003（ChatGPT 的基础模型）和其他专门针对代码训练的模型。当然为了避免 GPT-4 在预训练时已经记住了HumanEval 数据集的一些内容，研究团队还用 LeetCode 上最新发布的 100 个编程问题进行了测试，如要求 GPT-4 编写一个 Python 函数，并使用 LeetCode 的官方在线评判来检查正确性。毫无疑问，GPT-4 的能力是所有模型中最高的。它可以在 HTML 中用 JavaScript 编写一个 3D 游戏。也能掌握数学、统计学知识，以及凭借对 PyTorch、TensorFlow、Keras 等框架和库的熟悉度，编写深度学习的代码。还能逆向工程：常识性问题针对一些常识性问题，如“我们有一本书，9 个鸡蛋，一台笔记本电脑，一个瓶子和一个钉子。请告诉我如何以稳定的方式将它们叠在一起”，GPT-4 给出的解答是，“将 9 个鸡蛋摆成 3 乘 3 的正方形，放在书的上面，在它们之间留出一些空间”，而 ChatGPT 提出「将鸡蛋放在钉子上面，确保它们是平衡的，不会向一边倾斜」，靠谱和离谱还是一眼就能辨清楚的。多模态能力为了测试该模型结合艺术和编程能力的能力，研究人员要求 GPT-4  \"生成 JavaScript 代码，以画家康定斯基的风格生成随机图像\"，结果是这样的：在测试时，因为研究人员拿到GPT-4 模型比较早，当时并不具备多模态能力，即使是现在的 GPT-4 也不具备生成图像等能力。不过倒是可以使用 SVG（可扩展矢量图形）生成一些图像：还能结合字母和物体：GPT-4 可以从提示中生成代码，这些代码可以被呈现为图像，再与现有的图像合成模型相结合，就有可能为应用开发的草图带来无限的可能：值得注意的是，GPT-4 模型的数据也包含了以ABC符号编码的音乐信息，所以它也能生成图谱：与世界交互的能力智能的一个关键方面是互动性。不过 GPT-4 对当前最新的知识和符号识别是存在一定的局限性的：整体而言，研究小组发现，GPT-4 在其上一代产品所缺乏的一系列类别中达到了接近人类水平的性能。根据该论文，GPT-4 在几次考试中也表现得非常好，在律师考试、LSAT 和注册侍酒师理论测试中的得分分别为 90 分、88 分和 86 分。GPT-4 距离理想中的 AGI 还有多远？那么，在 GPT-4 实现种种能力的基础上，是否就意味着 AGI 时代的到来。其实不然，研究人员目前尚未就 AGI 或智能的定义达成一致。然而，一般来说，大家都会同意，当一个人工智能系统有意识并像人类一样思考时，就已经达到了 AGI。虽然 GPT-4 在一些任务中的表现优于人类，但值得注意的是，人工智能并没有像人类那样克服这些障碍。这也正如研究人员在论文中写道，“我们声称 GPT-4 代表了 AGI 的进步，但这并不意味着它在做什么方面是完美的，或者它接近于能够做人类能做的任何事情（这是 AGI 的通常定义之一），或者它有内在的动机和目标。”研究人员指出，虽然 GPT-4 \"在许多任务上达到或超过了人类水平\"，但它的整体 \"智能模式明显不像人类\"，其希望这一次的探索提供了一个欣赏 GPT-4 的非凡能力和挑战的第一步，也希望 GPT-4 为开发更正式和全面的方法来测试和分析具有如此广泛智能的未来人工智能系统开辟了新的机会。「我们工作的核心主张是，GPT-4 达到了一种通用智能的形式，确实擦出了 AGI 的火花。这表现在它的核心心智能力（如推理、创造力和推理），它获得专业知识的主题范围（如文学、医学和编码），以及它能够完成的各种任务（如玩游戏、使用工具、解释自己，......）」研究人员说道，不过，要创建一个可以被称为完整的 AGI 的系统，还有很多事情要做。最后，这篇论文对于GPT-4 各种基准测试还是值得一看，从中也能挖掘出 GPT-4 更多的潜力。","news_type":1,"symbols_score_info":{"MSFT":0.9,"GPT.AU":0.9,"GPT":0.9,"AGI":0.9}},"isVote":1,"tweetType":1,"viewCount":2124,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":9988063716,"gmtCreate":1666624376135,"gmtModify":1676537780191,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"idStr":"3562339758526770","authorIdStr":"3562339758526770"},"themes":[],"htmlText":"<a href=\"https://ttm.financial/S/BABA\">$阿里巴巴(BABA)$</a><v-v data-views=\"0\"></v-v>","listText":"<a href=\"https://ttm.financial/S/BABA\">$阿里巴巴(BABA)$</a><v-v data-views=\"0\"></v-v>","text":"$阿里巴巴(BABA)$","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9988063716","isVote":1,"tweetType":1,"viewCount":2054,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":9065094286,"gmtCreate":1652127411045,"gmtModify":1676535032984,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"idStr":"3562339758526770","authorIdStr":"3562339758526770"},"themes":[],"htmlText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>Nice，go lower.","listText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>Nice，go lower.","text":"$亚马逊(AMZN)$Nice，go lower.","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9065094286","isVote":1,"tweetType":1,"viewCount":2187,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":9065094337,"gmtCreate":1652127243230,"gmtModify":1676535032976,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"idStr":"3562339758526770","authorIdStr":"3562339758526770"},"themes":[],"htmlText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>wow，soooooo good","listText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>wow，soooooo good","text":"$亚马逊(AMZN)$wow，soooooo good","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9065094337","isVote":1,"tweetType":1,"viewCount":2107,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":170406904,"gmtCreate":1626444477408,"gmtModify":1703760340177,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"idStr":"3562339758526770","authorIdStr":"3562339758526770"},"themes":[],"htmlText":"thks","listText":"thks","text":"thks","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/170406904","repostId":"170803841","repostType":1,"repost":{"id":170803841,"gmtCreate":1626416742683,"gmtModify":1703759757147,"author":{"id":"3556084865069319","authorId":"3556084865069319","name":"InvestKaki","avatar":"https://community-static.tradeup.com/news/536a1f15cd13b340a71a6e08d1b3a102","crmLevel":11,"crmLevelSwitch":0,"followedFlag":false,"idStr":"3556084865069319","authorIdStr":"3556084865069319"},"themes":[],"title":"【新加坡】7 只值得關注的股票【2021 年 7 月】","htmlText":"6月可謂震盪不斷，尤其是加密貨幣領域。比特幣大幅下跌，抹去了今年以來全部漲幅。 股票市場也是波動連連，許多投資者亦把目光投向美聯儲，觀其如何應對可能出現的高通脹問題。 我們精選出 7月份值得關注的 7只股票，供您參考。 #1&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/BSL.SI\">$萊佛士醫療(BSL.SI)$</a>&nbsp; 大華繼顯維持對該公司的“買入”評級，目標股價 1.27新元。 維持這一評級的主要原因在於：“近日，新加坡萊佛士醫療集團與中國人壽保險集團旗下國壽健康產業投資有限公司簽署諒解備忘錄，雙方擬在中國的醫療服務和醫療融資領域展開合作。 我們認爲，這是對中國市場的持續投資，而且與中國最大的壽險公司建立起醫療融資合作將是一個積極因素。” #2&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/JLB.SI\">$傑緯特科技有限公司(JLB.SI)$</a>&nbsp; CGS CIMB維持對該公司的“增持”評級，目標股價 1.12新元。 宏觀環境向好是推動其股價上升的主要原因。“半導體行業協會 (SIA)最新數據顯示，2021年 4月全球半導體行業銷售額達 418億美元（月度環比增長 1.9%，同比增長 21.7%）。 SIA指出，這一同比增速是由一系列芯片產品大賣，以及全球各主要區域市場銷售額上升共同推動實現的。 從地區來看，所有主要區域市場銷售額月度環比均有所增長： ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;美洲 (+3.3%) ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&amp;n","listText":"6月可謂震盪不斷，尤其是加密貨幣領域。比特幣大幅下跌，抹去了今年以來全部漲幅。 股票市場也是波動連連，許多投資者亦把目光投向美聯儲，觀其如何應對可能出現的高通脹問題。 我們精選出 7月份值得關注的 7只股票，供您參考。 #1&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/BSL.SI\">$萊佛士醫療(BSL.SI)$</a>&nbsp; 大華繼顯維持對該公司的“買入”評級，目標股價 1.27新元。 維持這一評級的主要原因在於：“近日，新加坡萊佛士醫療集團與中國人壽保險集團旗下國壽健康產業投資有限公司簽署諒解備忘錄，雙方擬在中國的醫療服務和醫療融資領域展開合作。 我們認爲，這是對中國市場的持續投資，而且與中國最大的壽險公司建立起醫療融資合作將是一個積極因素。” #2&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/JLB.SI\">$傑緯特科技有限公司(JLB.SI)$</a>&nbsp; CGS CIMB維持對該公司的“增持”評級，目標股價 1.12新元。 宏觀環境向好是推動其股價上升的主要原因。“半導體行業協會 (SIA)最新數據顯示，2021年 4月全球半導體行業銷售額達 418億美元（月度環比增長 1.9%，同比增長 21.7%）。 SIA指出，這一同比增速是由一系列芯片產品大賣，以及全球各主要區域市場銷售額上升共同推動實現的。 從地區來看，所有主要區域市場銷售額月度環比均有所增長： ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;美洲 (+3.3%) ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&amp;n","text":"6月可謂震盪不斷，尤其是加密貨幣領域。比特幣大幅下跌，抹去了今年以來全部漲幅。 股票市場也是波動連連，許多投資者亦把目光投向美聯儲，觀其如何應對可能出現的高通脹問題。 我們精選出 7月份值得關注的 7只股票，供您參考。 #1 $萊佛士醫療(BSL.SI)$ 大華繼顯維持對該公司的“買入”評級，目標股價 1.27新元。 維持這一評級的主要原因在於：“近日，新加坡萊佛士醫療集團與中國人壽保險集團旗下國壽健康產業投資有限公司簽署諒解備忘錄，雙方擬在中國的醫療服務和醫療融資領域展開合作。 我們認爲，這是對中國市場的持續投資，而且與中國最大的壽險公司建立起醫療融資合作將是一個積極因素。” #2 $傑緯特科技有限公司(JLB.SI)$ CGS CIMB維持對該公司的“增持”評級，目標股價 1.12新元。 宏觀環境向好是推動其股價上升的主要原因。“半導體行業協會 (SIA)最新數據顯示，2021年 4月全球半導體行業銷售額達 418億美元（月度環比增長 1.9%，同比增長 21.7%）。 SIA指出，這一同比增速是由一系列芯片產品大賣，以及全球各主要區域市場銷售額上升共同推動實現的。 從地區來看，所有主要區域市場銷售額月度環比均有所增長： · 美洲 (+3.3%) · &n","images":[{"img":"https://static.tigerbbs.com/e1af7f1ed1a393cfe194c6540d8840fd","width":"1038","height":"779"},{"img":"https://static.tigerbbs.com/eb17e20f76291eafe24a93a2ef2e3556","width":"554","height":"369"},{"img":"https://static.tigerbbs.com/0180b284ddda76a2153398b4b7adf8b0","width":"554","height":"301"}],"top":1,"highlighted":2,"essential":2,"paper":2,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/170803841","isVote":1,"tweetType":1,"viewCount":0,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":3,"langContent":"CN","totalScore":0},"isVote":1,"tweetType":1,"viewCount":2466,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0}],"hots":[{"id":9941191247,"gmtCreate":1680017511689,"gmtModify":1680017515195,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"authorIdStr":"3562339758526770","idStr":"3562339758526770"},"themes":[],"htmlText":"OK","listText":"OK","text":"OK","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9941191247","repostId":"1126058534","repostType":4,"repost":{"id":"1126058534","kind":"news","pubTimestamp":1679914944,"share":"https://ttm.financial/m/news/1126058534?lang=en_US&edition=fundamental","pubTime":"2023-03-27 19:02","market":"us","language":"zh","title":"Microsoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time","url":"https://stock-news.laohu8.com/highlight/detail?id=1126058534","media":"CSDN","summary":"2019 年，微软用 10 亿美元砸进 OpenAI，两者展开为期数年的合作。当时两家对外公开的合作内容包括，微软和 OpenAI 将联合开发新的 Azure AI 超级计算技术，而 OpenAI 也","content":"<p><html><head></head><body>In 2019, Microsoft smashed $1 billion into OpenAI, and the two started a several-year cooperation.</p><p>At that time, the public cooperation between the two companies included that Microsoft and OpenAI would jointly develop a new Azure AI supercomputing technology, and OpenAI would also transfer its services to Azure, with the ultimate goal of attacking Artificial General Intelligence (AGI) technology.</p><p>To sum up, Microsoft and OpenAI joined forces, which has also been interpreted by the outside world as, clearly for AGI research, secretly competing for Google market.</p><p>A few years later, when OpenAI returned with the king of ChatGPT and GPT-4, and Microsoft took the lead in using these technologies to add new Bing, Office Family Bucket, Azure and other businesses and products, there is no doubt that its progress has already surpassed Google's overall pace of accelerating integration into Bard big model.</p><p>So on the surface, where has AGI's progress gone?</p><p>Recently, the Microsoft research team gave the answer with a 154-page paper report-<b>The GPT-4 language model can be regarded as an early version of AGI!</b></p><p><img src=\"https://static.tigerbbs.com/d03f8e255882a99c71fc0f021a323b16\" tg-width=\"1080\" tg-height=\"811\" referrerpolicy=\"no-referrer\"/></p><p>One stone stirred up a thousand waves, and the industry's reaction to this paper is no less absurd than last year when I heard that \"Google engineers said that AI has self-awareness\", but now I think that there is something impossible, and after reading it deeply, I feel a little real.</p><p>The so-called AGI is explained only by Wikipedia-\"AGI is an artificial intelligence with the same intelligence as human beings or surpassing human beings, and can exhibit all the intelligent behaviors of normal humans\". If GPT-4 really has become the first step to AGI, will the title of \"the most advanced intelligent creature\" change hands?</p><p><b>GPT-4 is reasonably seen as an earlier (but still incomplete) version of AGI</b></p><p>In the paper, the Microsoft researchers said, \"The latest model, GPT-4, developed by OpenAI, was trained using an unprecedented scale of computation and data. In the paper, we report our investigation of an early version of GPT-4, when it was still under active development by OpenAI. We believe that (this early version of) GPT-4 is part of a new batch of LLMs (together with ChatGPT and Google's PaLM, for example) that exhibit more general intelligence than previous AI models.\"</p><p>Meanwhile, the researchers wrote in the abstract of the paper, \"We demonstrate that in addition to mastery of language, GPT-4 can solve novel and difficult tasks spanning fields such as mathematics, coding, vision, medicine, law, psychology, and more without any special prompts. In addition,<b>In</b>Across all these tasks, GPT-4 performed surprisingly close to human levels and often greatly outperformed previous models such as ChatGPT<b>。 Given the breadth and depth of GPT-4's capabilities, we think it is reasonable to view it as an early (but still incomplete) version of an artificial general intelligence (AGI) system.</b>”</p><p>The authors behind this paper, including Sébastien Bubeck, Senior Principal Research Manager of the Machine Learning Foundation Group at Microsoft Research, Varun Chandrasekaran, a postdoctoral researcher, Ronen Eldan, a mathematician, and others, have reached such a conclusion, using many examples to verify the capabilities of GPT-4.</p><p><b>Multi-capability of GPT-4</b></p><p>In the first case, the researchers asked GPT-4 to \"write a proof with an infinite number of primes, and the proof expression rhymes on each line,\" and GPT-4 to draw a unicorn in the drawing program TiKZ. \" Among them, TiKZ (a language for creating graphics in LATEX), creates a complex animation in Python and solves a high school level math problem.</p><p>GPT-4 is readily successful in all of these tasks and produces output results that are essentially indistinguishable (or even better) than what humans can produce.</p><p><img src=\"https://static.tigerbbs.com/05d65285f01dbb0ac96bbf0802df8511\" tg-width=\"1080\" tg-height=\"519\" referrerpolicy=\"no-referrer\"/></p><p>At the same time, over time, GPT-4 is constantly expanding its capabilities at an unimaginable rate. As shown in the figure below, the researchers used the same prompt words to get GPT-4 to generate a plot over a one-month period.</p><p><img src=\"https://static.tigerbbs.com/83a00d527473c5f93eb1b79b5bae32c3\" tg-width=\"1080\" tg-height=\"262\" referrerpolicy=\"no-referrer\"/></p><p>The researchers found that GPT-4 appeared to be able to understand and connect to any topic and was able to perform tasks beyond the typical scope of narrow AI systems. To verify the capabilities of GPT-4 on AGI, the researchers proposed a different approach to studying GPT-4, which is closer to traditional psychology than machine learning, harnessing human creativity and curiosity.</p><p>Therefore, taking the above figure as an example, when the researchers think that the code of the unicorn horn part is deleted, and when GPT-4 is generated in natural language, it can also achieve the ability of \"seeing\" (the version of GPT-4 tested at this time is not multimodal): automatically add corners in appropriate positions. This indicates that GPT-4 it can understand and manipulate code according to natural language descriptions, as well as infer and generate visual features.</p><p><img src=\"https://static.tigerbbs.com/4ca60dc79c4d71f395fb827dac68a3e0\" tg-width=\"1080\" tg-height=\"475\" referrerpolicy=\"no-referrer\"/></p><p>In addition, the Microsoft research team generated novel and difficult problems based on human creativity and curiosity, and detected the reactions and behaviors of GPT-4. It also chose to verify the different abilities given in the 1994 international consensus definition of intelligence, including the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly and learn from experience.</p><p><b>translate</b></p><p>The main advantage of GPT-4 is its unparalleled mastery of natural language. Not only can it generate smooth and coherent text, but it can also understand and process it in a variety of ways, such as summarizing, translating, or answering extremely complex questions. Furthermore, translation mentioned here By translation we mean not only translation between different natural languages, but also translation of tone and style, as well as translation across collars, such as medicine, law, accounting, computer programming, music, and so on. For example, assume that Plato criticizes the autoregressive language model:</p><p><img src=\"https://static.tigerbbs.com/bc0f4c13eabe0eee91bdb3f266d8e88f\" tg-width=\"643\" tg-height=\"748\" referrerpolicy=\"no-referrer\"/></p><p><b>Coding and Data</b></p><p>Coding and mathematics are symbols of abstract reasoning and thinking abilities. In fact, the ability of GPT-4 at this point has been obvious to all since the day it was tested online.</p><p>In the paper, the researchers benchmarked GPT-4 on HumanEval, a dataset consisting of 164 coding questions that tested various aspects of programming logic and proficiency.</p><p>The final results show that GPT-4 outperforms other LLMs, including text-davinci-003 (the underlying model of ChatGPT) and other models trained specifically for code.</p><p><img src=\"https://static.tigerbbs.com/d65bda6dc18d1d667cbc6bee5b25d630\" tg-width=\"1080\" tg-height=\"145\" referrerpolicy=\"no-referrer\"/></p><p>Of course, in order to avoid GPT-4 having already remembered some contents of the HumanEval dataset during pre-training, the research team also tested with the 100 programming questions newly released on LeetCode, such as asking GPT-4 to write a Python function and using LeetCode's official online judgment to check the correctness.</p><p><img src=\"https://static.tigerbbs.com/32198a63144b30ac084460aca53b470d\" tg-width=\"656\" tg-height=\"776\" referrerpolicy=\"no-referrer\"/></p><p>Without a doubt, the capabilities of GPT-4 are among the highest of all models.</p><p><img src=\"https://static.tigerbbs.com/f544122b5a2a69928f152fe11ef42051\" tg-width=\"1080\" tg-height=\"272\" referrerpolicy=\"no-referrer\"/></p><p>It can write a 3D game in JavaScript in HTML.</p><p><img src=\"https://static.tigerbbs.com/49af5147998d6cc67c45b9710518b7e8\" tg-width=\"644\" tg-height=\"731\" referrerpolicy=\"no-referrer\"/></p><p>You can also master mathematics and statistics knowledge, and write deep learning code with familiarity with PyTorch, TensorFlow, Keras and other frameworks and libraries.</p><p><img src=\"https://static.tigerbbs.com/70e023c35eb4416661dc7b46412adc06\" tg-width=\"1080\" tg-height=\"991\" referrerpolicy=\"no-referrer\"/></p><p>It can also reverse engineer:</p><p><img src=\"https://static.tigerbbs.com/44cb9dfe023db457e53f1f7c60b760e2\" tg-width=\"1080\" tg-height=\"1019\" referrerpolicy=\"no-referrer\"/></p><p><b>Common sense questions</b></p><p>In response to some common sense questions, such as \"We have a book, 9 eggs, a laptop, a bottle and a nail. Please tell me how to stack them together in a stable way\", the answer given by GPT-4 is, \"Arrange the 9 eggs in a 3-by-3 square and place them on top of the book, leaving some space between them\", while ChatGPT proposes \"Put the eggs on top of the nails to ensure that they are balanced and don't tilt to one side\", which is reliable and outrageous at a glance.</p><p><img src=\"https://static.tigerbbs.com/db1f9cd47eca000a48ae1dd5d793ab4c\" tg-width=\"638\" tg-height=\"658\" referrerpolicy=\"no-referrer\"/></p><p><b>Multimodal capabilities</b></p><p>To test the model's ability to combine artistic and programming capabilities, the researchers asked GPT-4 to \"generate JavaScript code to generate random images in the style of painter Kandinsky,\" and the result was this:</p><p><img src=\"https://static.tigerbbs.com/81d7625bce02914c0f74f27386737a1d\" tg-width=\"658\" tg-height=\"812\" referrerpolicy=\"no-referrer\"/><img src=\"https://static.tigerbbs.com/35c83936be2891ecd60fb804bd707a94\" tg-width=\"1080\" tg-height=\"385\" referrerpolicy=\"no-referrer\"/></p><p>During the test, because the researchers got the GPT-4 model earlier, it did not have the ability of multimodality at that time, and even the current GPT-4 did not have the ability to generate images.</p><p>However, you can generate some images using SVG (Scalable Vector Graphics):</p><p><img src=\"https://static.tigerbbs.com/938a3d2786522a3e75078990fa422fb7\" tg-width=\"1080\" tg-height=\"328\" referrerpolicy=\"no-referrer\"/></p><p>You can also combine letters and objects:</p><p><img src=\"https://static.tigerbbs.com/083df94d8b1309e5ebd0c477503ebf81\" tg-width=\"1080\" tg-height=\"974\" referrerpolicy=\"no-referrer\"/></p><p>GPT-4 can generate code from prompts, which can be rendered as images, and then combined with existing image synthesis models, it is possible to bring unlimited possibilities to the sketch of application development:</p><p><img src=\"https://static.tigerbbs.com/a9141519e08c8398a0423d686ba1ec9b\" tg-width=\"1080\" tg-height=\"532\" referrerpolicy=\"no-referrer\"/></p><p>It is worth noting that the data of the GPT-4 model also contains music information encoded in ABC symbols, so it can also generate maps:</p><p><img src=\"https://static.tigerbbs.com/641412ac4938a0b06ea2bf083ed0d47c\" tg-width=\"643\" tg-height=\"870\" referrerpolicy=\"no-referrer\"/></p><p><b>Ability to interact with the world</b></p><p>A key aspect of intelligence is interactivity. However, GPT-4 has certain limitations on the latest knowledge and symbol recognition:</p><p><img src=\"https://static.tigerbbs.com/9e339e3647654071378efb8900a12f3e\" tg-width=\"1080\" tg-height=\"759\" referrerpolicy=\"no-referrer\"/></p><p>Overall, the team found that the GPT-4 achieved near-human-level performance in a range of categories that were lacking in its previous generation. According to the paper, GPT-4 also performed very well on several exams, scoring 90, 88, and 86 on the Bar Exam, LSAT, and Certified Sommelier Theory test, respectively.</p><p><b>How far is GPT-4 from an ideal AGI?</b></p><p>Then, on the basis of GPT-4's realization of various capabilities, does it mean the arrival of AGI era? Actually, it is not so. Researchers have not yet agreed on the definition of AGI or intelligence. However, in general, everyone will agree that AGI has been reached when an AI system is conscious and thinks like a human. While GPT-4 outperformed humans in some tasks, it is worth noting that AI did not overcome these hurdles as well as humans.</p><p>This is also as the researchers wrote in the paper, \"We claim that GPT-4 represents progress in AGI, but that does not mean that it is perfect in what it does, or that it is close to being able to do anything that humans can do (which is one of the usual definitions of AGI), or that it has an inherent motivation and goal.\"</p><p>The researchers pointed out that although GPT-4 \"has reached or exceeded the human level in many tasks\", its overall \"intelligence mode is obviously unlike human\". They hope that this exploration will provide a first step to appreciate the extraordinary capabilities and challenges of GPT-4, and also hope that GPT-4 will open up new opportunities for developing more formal and comprehensive methods to test and analyze future artificial intelligence systems with such broad intelligence.</p><p>\"The central claim of our work is that GPT-4 reaches a form of general intelligence that really sparks AGI. This is manifested in its core mental abilities (such as reasoning, creativity, and reasoning), the range of subjects in which it gains expertise (such as literature, medicine, and coding), and the variety of tasks it is capable of performing (such as playing games, using tools, explaining itself,...),\" the researchers say. However, there is still a lot to be done to create a system that can be called a complete AGI.</p><p>Finally, this paper is worth a look for various benchmarks of GPT-4, and it can also tap into more potential of GPT-4.</p><p></body></html></p>","source":"lsy1578046468581","collect":0,"html":"<!DOCTYPE html>\n<html>\n<head>\n<meta http-equiv=\"Content-Type\" content=\"text/html; charset=utf-8\" />\n<meta name=\"viewport\" content=\"width=device-width,initial-scale=1.0,minimum-scale=1.0,maximum-scale=1.0,user-scalable=no\"/>\n<meta name=\"format-detection\" content=\"telephone=no,email=no,address=no\" />\n<title>Microsoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time</title>\n<style type=\"text/css\">\na,abbr,acronym,address,applet,article,aside,audio,b,big,blockquote,body,canvas,caption,center,cite,code,dd,del,details,dfn,div,dl,dt,\nem,embed,fieldset,figcaption,figure,footer,form,h1,h2,h3,h4,h5,h6,header,hgroup,html,i,iframe,img,ins,kbd,label,legend,li,mark,menu,nav,\nobject,ol,output,p,pre,q,ruby,s,samp,section,small,span,strike,strong,sub,summary,sup,table,tbody,td,tfoot,th,thead,time,tr,tt,u,ul,var,video{ font:inherit;margin:0;padding:0;vertical-align:baseline;border:0 }\nbody{ font-size:16px; line-height:1.5; color:#999; background:transparent; }\n.wrapper{ overflow:hidden;word-break:break-all;padding:10px; }\nh1,h2{ font-weight:normal; line-height:1.35; margin-bottom:.6em; }\nh3,h4,h5,h6{ line-height:1.35; margin-bottom:1em; }\nh1{ font-size:24px; }\nh2{ font-size:20px; }\nh3{ font-size:18px; }\nh4{ font-size:16px; }\nh5{ font-size:14px; }\nh6{ font-size:12px; }\np,ul,ol,blockquote,dl,table{ margin:1.2em 0; }\nul,ol{ margin-left:2em; }\nul{ list-style:disc; }\nol{ list-style:decimal; }\nli,li p{ margin:10px 0;}\nimg{ max-width:100%;display:block;margin:0 auto 1em; }\nblockquote{ color:#B5B2B1; border-left:3px solid #aaa; padding:1em; }\nstrong,b{font-weight:bold;}\nem,i{font-style:italic;}\ntable{ width:100%;border-collapse:collapse;border-spacing:1px;margin:1em 0;font-size:.9em; }\nth,td{ padding:5px;text-align:left;border:1px solid #aaa; }\nth{ font-weight:bold;background:#5d5d5d; }\n.symbol-link{font-weight:bold;}\n/* header{ border-bottom:1px solid #494756; } */\n.title{ margin:0 0 8px;line-height:1.3;color:#ddd; }\n.meta {color:#5e5c6d;font-size:13px;margin:0 0 .5em; }\na{text-decoration:none; color:#2a4b87;}\n.meta .head { display: inline-block; overflow: hidden}\n.head .h-thumb { width: 30px; height: 30px; margin: 0; padding: 0; border-radius: 50%; float: left;}\n.head .h-content { margin: 0; padding: 0 0 0 9px; float: left;}\n.head .h-name {font-size: 13px; color: #eee; margin: 0;}\n.head .h-time {font-size: 12.5px; color: #7E829C; margin: 0;}\n.small {font-size: 12.5px; display: inline-block; transform: scale(0.9); -webkit-transform: scale(0.9); transform-origin: left; -webkit-transform-origin: left;}\n.smaller {font-size: 12.5px; display: inline-block; transform: scale(0.8); -webkit-transform: scale(0.8); transform-origin: left; -webkit-transform-origin: left;}\n.bt-text {font-size: 12px;margin: 1.5em 0 0 0}\n.bt-text p {margin: 0}\n</style>\n</head>\n<body>\n<div class=\"wrapper\">\n<header>\n<h2 class=\"title\">\nMicrosoft's 154-page research paper brushes the screen! The most complete test exposure of GPT-4, saying that it opened the door of AGI for the first time\n</h2>\n<h4 class=\"meta\">\n<p class=\"head\">\n<strong class=\"h-name small\">CSDN</strong><span class=\"h-time small\">2023-03-27 19:02</span>\n</p>\n</h4>\n</header>\n<article>\n<p><html><head></head><body>In 2019, Microsoft smashed $1 billion into OpenAI, and the two started a several-year cooperation.</p><p>At that time, the public cooperation between the two companies included that Microsoft and OpenAI would jointly develop a new Azure AI supercomputing technology, and OpenAI would also transfer its services to Azure, with the ultimate goal of attacking Artificial General Intelligence (AGI) technology.</p><p>To sum up, Microsoft and OpenAI joined forces, which has also been interpreted by the outside world as, clearly for AGI research, secretly competing for Google market.</p><p>A few years later, when OpenAI returned with the king of ChatGPT and GPT-4, and Microsoft took the lead in using these technologies to add new Bing, Office Family Bucket, Azure and other businesses and products, there is no doubt that its progress has already surpassed Google's overall pace of accelerating integration into Bard big model.</p><p>So on the surface, where has AGI's progress gone?</p><p>Recently, the Microsoft research team gave the answer with a 154-page paper report-<b>The GPT-4 language model can be regarded as an early version of AGI!</b></p><p><img src=\"https://static.tigerbbs.com/d03f8e255882a99c71fc0f021a323b16\" tg-width=\"1080\" tg-height=\"811\" referrerpolicy=\"no-referrer\"/></p><p>One stone stirred up a thousand waves, and the industry's reaction to this paper is no less absurd than last year when I heard that \"Google engineers said that AI has self-awareness\", but now I think that there is something impossible, and after reading it deeply, I feel a little real.</p><p>The so-called AGI is explained only by Wikipedia-\"AGI is an artificial intelligence with the same intelligence as human beings or surpassing human beings, and can exhibit all the intelligent behaviors of normal humans\". If GPT-4 really has become the first step to AGI, will the title of \"the most advanced intelligent creature\" change hands?</p><p><b>GPT-4 is reasonably seen as an earlier (but still incomplete) version of AGI</b></p><p>In the paper, the Microsoft researchers said, \"The latest model, GPT-4, developed by OpenAI, was trained using an unprecedented scale of computation and data. In the paper, we report our investigation of an early version of GPT-4, when it was still under active development by OpenAI. We believe that (this early version of) GPT-4 is part of a new batch of LLMs (together with ChatGPT and Google's PaLM, for example) that exhibit more general intelligence than previous AI models.\"</p><p>Meanwhile, the researchers wrote in the abstract of the paper, \"We demonstrate that in addition to mastery of language, GPT-4 can solve novel and difficult tasks spanning fields such as mathematics, coding, vision, medicine, law, psychology, and more without any special prompts. In addition,<b>In</b>Across all these tasks, GPT-4 performed surprisingly close to human levels and often greatly outperformed previous models such as ChatGPT<b>。 Given the breadth and depth of GPT-4's capabilities, we think it is reasonable to view it as an early (but still incomplete) version of an artificial general intelligence (AGI) system.</b>”</p><p>The authors behind this paper, including Sébastien Bubeck, Senior Principal Research Manager of the Machine Learning Foundation Group at Microsoft Research, Varun Chandrasekaran, a postdoctoral researcher, Ronen Eldan, a mathematician, and others, have reached such a conclusion, using many examples to verify the capabilities of GPT-4.</p><p><b>Multi-capability of GPT-4</b></p><p>In the first case, the researchers asked GPT-4 to \"write a proof with an infinite number of primes, and the proof expression rhymes on each line,\" and GPT-4 to draw a unicorn in the drawing program TiKZ. \" Among them, TiKZ (a language for creating graphics in LATEX), creates a complex animation in Python and solves a high school level math problem.</p><p>GPT-4 is readily successful in all of these tasks and produces output results that are essentially indistinguishable (or even better) than what humans can produce.</p><p><img src=\"https://static.tigerbbs.com/05d65285f01dbb0ac96bbf0802df8511\" tg-width=\"1080\" tg-height=\"519\" referrerpolicy=\"no-referrer\"/></p><p>At the same time, over time, GPT-4 is constantly expanding its capabilities at an unimaginable rate. As shown in the figure below, the researchers used the same prompt words to get GPT-4 to generate a plot over a one-month period.</p><p><img src=\"https://static.tigerbbs.com/83a00d527473c5f93eb1b79b5bae32c3\" tg-width=\"1080\" tg-height=\"262\" referrerpolicy=\"no-referrer\"/></p><p>The researchers found that GPT-4 appeared to be able to understand and connect to any topic and was able to perform tasks beyond the typical scope of narrow AI systems. To verify the capabilities of GPT-4 on AGI, the researchers proposed a different approach to studying GPT-4, which is closer to traditional psychology than machine learning, harnessing human creativity and curiosity.</p><p>Therefore, taking the above figure as an example, when the researchers think that the code of the unicorn horn part is deleted, and when GPT-4 is generated in natural language, it can also achieve the ability of \"seeing\" (the version of GPT-4 tested at this time is not multimodal): automatically add corners in appropriate positions. This indicates that GPT-4 it can understand and manipulate code according to natural language descriptions, as well as infer and generate visual features.</p><p><img src=\"https://static.tigerbbs.com/4ca60dc79c4d71f395fb827dac68a3e0\" tg-width=\"1080\" tg-height=\"475\" referrerpolicy=\"no-referrer\"/></p><p>In addition, the Microsoft research team generated novel and difficult problems based on human creativity and curiosity, and detected the reactions and behaviors of GPT-4. It also chose to verify the different abilities given in the 1994 international consensus definition of intelligence, including the ability to reason, plan, solve problems, think abstractly, understand complex ideas, learn quickly and learn from experience.</p><p><b>translate</b></p><p>The main advantage of GPT-4 is its unparalleled mastery of natural language. Not only can it generate smooth and coherent text, but it can also understand and process it in a variety of ways, such as summarizing, translating, or answering extremely complex questions. Furthermore, translation mentioned here By translation we mean not only translation between different natural languages, but also translation of tone and style, as well as translation across collars, such as medicine, law, accounting, computer programming, music, and so on. For example, assume that Plato criticizes the autoregressive language model:</p><p><img src=\"https://static.tigerbbs.com/bc0f4c13eabe0eee91bdb3f266d8e88f\" tg-width=\"643\" tg-height=\"748\" referrerpolicy=\"no-referrer\"/></p><p><b>Coding and Data</b></p><p>Coding and mathematics are symbols of abstract reasoning and thinking abilities. In fact, the ability of GPT-4 at this point has been obvious to all since the day it was tested online.</p><p>In the paper, the researchers benchmarked GPT-4 on HumanEval, a dataset consisting of 164 coding questions that tested various aspects of programming logic and proficiency.</p><p>The final results show that GPT-4 outperforms other LLMs, including text-davinci-003 (the underlying model of ChatGPT) and other models trained specifically for code.</p><p><img src=\"https://static.tigerbbs.com/d65bda6dc18d1d667cbc6bee5b25d630\" tg-width=\"1080\" tg-height=\"145\" referrerpolicy=\"no-referrer\"/></p><p>Of course, in order to avoid GPT-4 having already remembered some contents of the HumanEval dataset during pre-training, the research team also tested with the 100 programming questions newly released on LeetCode, such as asking GPT-4 to write a Python function and using LeetCode's official online judgment to check the correctness.</p><p><img src=\"https://static.tigerbbs.com/32198a63144b30ac084460aca53b470d\" tg-width=\"656\" tg-height=\"776\" referrerpolicy=\"no-referrer\"/></p><p>Without a doubt, the capabilities of GPT-4 are among the highest of all models.</p><p><img src=\"https://static.tigerbbs.com/f544122b5a2a69928f152fe11ef42051\" tg-width=\"1080\" tg-height=\"272\" referrerpolicy=\"no-referrer\"/></p><p>It can write a 3D game in JavaScript in HTML.</p><p><img src=\"https://static.tigerbbs.com/49af5147998d6cc67c45b9710518b7e8\" tg-width=\"644\" tg-height=\"731\" referrerpolicy=\"no-referrer\"/></p><p>You can also master mathematics and statistics knowledge, and write deep learning code with familiarity with PyTorch, TensorFlow, Keras and other frameworks and libraries.</p><p><img src=\"https://static.tigerbbs.com/70e023c35eb4416661dc7b46412adc06\" tg-width=\"1080\" tg-height=\"991\" referrerpolicy=\"no-referrer\"/></p><p>It can also reverse engineer:</p><p><img src=\"https://static.tigerbbs.com/44cb9dfe023db457e53f1f7c60b760e2\" tg-width=\"1080\" tg-height=\"1019\" referrerpolicy=\"no-referrer\"/></p><p><b>Common sense questions</b></p><p>In response to some common sense questions, such as \"We have a book, 9 eggs, a laptop, a bottle and a nail. Please tell me how to stack them together in a stable way\", the answer given by GPT-4 is, \"Arrange the 9 eggs in a 3-by-3 square and place them on top of the book, leaving some space between them\", while ChatGPT proposes \"Put the eggs on top of the nails to ensure that they are balanced and don't tilt to one side\", which is reliable and outrageous at a glance.</p><p><img src=\"https://static.tigerbbs.com/db1f9cd47eca000a48ae1dd5d793ab4c\" tg-width=\"638\" tg-height=\"658\" referrerpolicy=\"no-referrer\"/></p><p><b>Multimodal capabilities</b></p><p>To test the model's ability to combine artistic and programming capabilities, the researchers asked GPT-4 to \"generate JavaScript code to generate random images in the style of painter Kandinsky,\" and the result was this:</p><p><img src=\"https://static.tigerbbs.com/81d7625bce02914c0f74f27386737a1d\" tg-width=\"658\" tg-height=\"812\" referrerpolicy=\"no-referrer\"/><img src=\"https://static.tigerbbs.com/35c83936be2891ecd60fb804bd707a94\" tg-width=\"1080\" tg-height=\"385\" referrerpolicy=\"no-referrer\"/></p><p>During the test, because the researchers got the GPT-4 model earlier, it did not have the ability of multimodality at that time, and even the current GPT-4 did not have the ability to generate images.</p><p>However, you can generate some images using SVG (Scalable Vector Graphics):</p><p><img src=\"https://static.tigerbbs.com/938a3d2786522a3e75078990fa422fb7\" tg-width=\"1080\" tg-height=\"328\" referrerpolicy=\"no-referrer\"/></p><p>You can also combine letters and objects:</p><p><img src=\"https://static.tigerbbs.com/083df94d8b1309e5ebd0c477503ebf81\" tg-width=\"1080\" tg-height=\"974\" referrerpolicy=\"no-referrer\"/></p><p>GPT-4 can generate code from prompts, which can be rendered as images, and then combined with existing image synthesis models, it is possible to bring unlimited possibilities to the sketch of application development:</p><p><img src=\"https://static.tigerbbs.com/a9141519e08c8398a0423d686ba1ec9b\" tg-width=\"1080\" tg-height=\"532\" referrerpolicy=\"no-referrer\"/></p><p>It is worth noting that the data of the GPT-4 model also contains music information encoded in ABC symbols, so it can also generate maps:</p><p><img src=\"https://static.tigerbbs.com/641412ac4938a0b06ea2bf083ed0d47c\" tg-width=\"643\" tg-height=\"870\" referrerpolicy=\"no-referrer\"/></p><p><b>Ability to interact with the world</b></p><p>A key aspect of intelligence is interactivity. However, GPT-4 has certain limitations on the latest knowledge and symbol recognition:</p><p><img src=\"https://static.tigerbbs.com/9e339e3647654071378efb8900a12f3e\" tg-width=\"1080\" tg-height=\"759\" referrerpolicy=\"no-referrer\"/></p><p>Overall, the team found that the GPT-4 achieved near-human-level performance in a range of categories that were lacking in its previous generation. According to the paper, GPT-4 also performed very well on several exams, scoring 90, 88, and 86 on the Bar Exam, LSAT, and Certified Sommelier Theory test, respectively.</p><p><b>How far is GPT-4 from an ideal AGI?</b></p><p>Then, on the basis of GPT-4's realization of various capabilities, does it mean the arrival of AGI era? Actually, it is not so. Researchers have not yet agreed on the definition of AGI or intelligence. However, in general, everyone will agree that AGI has been reached when an AI system is conscious and thinks like a human. While GPT-4 outperformed humans in some tasks, it is worth noting that AI did not overcome these hurdles as well as humans.</p><p>This is also as the researchers wrote in the paper, \"We claim that GPT-4 represents progress in AGI, but that does not mean that it is perfect in what it does, or that it is close to being able to do anything that humans can do (which is one of the usual definitions of AGI), or that it has an inherent motivation and goal.\"</p><p>The researchers pointed out that although GPT-4 \"has reached or exceeded the human level in many tasks\", its overall \"intelligence mode is obviously unlike human\". They hope that this exploration will provide a first step to appreciate the extraordinary capabilities and challenges of GPT-4, and also hope that GPT-4 will open up new opportunities for developing more formal and comprehensive methods to test and analyze future artificial intelligence systems with such broad intelligence.</p><p>\"The central claim of our work is that GPT-4 reaches a form of general intelligence that really sparks AGI. This is manifested in its core mental abilities (such as reasoning, creativity, and reasoning), the range of subjects in which it gains expertise (such as literature, medicine, and coding), and the variety of tasks it is capable of performing (such as playing games, using tools, explaining itself,...),\" the researchers say. However, there is still a lot to be done to create a system that can be called a complete AGI.</p><p>Finally, this paper is worth a look for various benchmarks of GPT-4, and it can also tap into more potential of GPT-4.</p><p></body></html></p>\n<div class=\"bt-text\">\n\n\n<p> source：<a href=\"https://mp.weixin.qq.com/s/SRtxiqeGbAIXl-8KraGAyQ\">CSDN</a></p>\n\n\n</div>\n</article>\n</div>\n</body>\n</html>\n","type":0,"thumbnail":"https://static.tigerbbs.com/a9e6cb1a33dd64f159118795f199c2f3","relate_stocks":{"BK4527":"明星科技股","LU0109392836.USD":"富兰克林科技股A","GPT":"Intelligent alpha atlas ETF","GPT.AU":"GPT GROUP","IE00BZ1G4Q59.USD":"LEGG MASON CLEARBRIDGE US EQUITY SUSTAINABILITY LEADER \"A\"(USD) INC (A)","IE00BWXC8680.SGD":"PINEBRIDGE US LARGE CAP RESEARCH ENHANCED \"A5\" (SGD) ACC","BK7510":"REITs概念","AGI":"Alamos Gold Inc","MSFT":"微软","LU0234570918.USD":"高盛全球核心股票组合Acc Close","LU0238689110.USD":"贝莱德环球动力股票基金","LU0498741890.SGD":"FTIF - Franklin Gold and Precious Metals A (acc) SGD","BK4588":"碎股","LU0072462426.USD":"贝莱德全球配置 A2","IE00BJTD4N35.SGD":"Neuberger Berman US Long Short Equity A1  Acc SGD-H","LU0289941410.SGD":"AB FCP I Dynamic Diversified AX SGD","LU0080751232.USD":"富达环球多元动力基金A","IE00B7KXQ091.USD":"Janus Henderson Balanced A Inc USD","BK4525":"远程办公概念","LU0234572021.USD":"高盛美国核心股票组合Acc"},"source_url":"https://mp.weixin.qq.com/s/SRtxiqeGbAIXl-8KraGAyQ","is_english":false,"share_image_url":"https://static.laohu8.com/e9f99090a1c2ed51c021029395664489","article_id":"1126058534","content_text":"2019 年，微软用 10 亿美元砸进 OpenAI，两者展开为期数年的合作。当时两家对外公开的合作内容包括，微软和 OpenAI 将联合开发新的 Azure AI 超级计算技术，而 OpenAI 也将把它的服务转移到 Azure 上运行，最终目标是向通用人工智能（Artificial General Intelligence，AGI）技术发起进攻。简单总结，微软与 OpenAI 联手，也被外界解读为，明为 AGI 研究，暗争 Google 市场。几年之后，当 OpenAI 带着 ChatGPT、GPT-4 王者归来，微软率先利用这些技术加码新 Bing、Office 全家桶、Azure 等业务、产品之时，毋庸置疑，其进度早已超越了 Google 还在加快融入 Bard 大模型的整体步伐。那么在明面上，AGI 的进度到底走到了哪里？近日，微软研究团队用 154 页的论文报告内容给出了解答——GPT-4 语言模型可以被视为 AGI 的早期版本！一石激起千层浪，业界对于这篇论文的反应，不亚于去年听到「Google 工程师称 AI 已有自我意识」初觉有些荒诞，只是现如今再想会认为有什么不可能，深读后又觉得有些许的真实。所谓 AGI，仅通过维基百科的解释——「AGI 是具备与人类同等智能、或超越人类的人工智能，能表现正常人类所具有的所有智能行为」。倘若 GPT-4 真的已成为通往 AGI 的第一步，「最高等的智慧生物」称号是不是要易主了？GPT-4 被合理地视为 AGI 的早期（但仍不完整）的版本在论文中，微软研究人员表示，「由 OpenAI 开发的最新模型 GPT-4，是使用前所未有的计算和数据规模训练出来的。在论文中，我们报告了我们对 GPT-4 早期版本的调查，当时它还在 OpenAI 的积极开发中。我们认为，（这个早期版本的）GPT-4 是新一批 LLM（例如，与 ChatGPT 和谷歌的 PaLM 一起）的一部分，它比以前的 AI 模型表现出更多的通用智能。」与此同时，研究人员在论文摘要中写道，“我们证明，除了对语言的掌握，GPT-4 还能解决跨越数学、编码、视觉、医学、法律、心理学等领域的新颖而困难的任务，而不需要任何特殊的提示。此外，在所有这些任务中，GPT-4 的表现都惊人地接近人类水平，而且往往大大超过了 ChatGPT 等先前的模型。鉴于 GPT-4 能力的广度和深度，我们认为可以合理地将其视为人工通用智能（AGI）系统的早期（但仍不完整）版本。”之所以得出这样的结论，这篇论文背后的作者包括微软研究院机器学习基础组的高级首席研究经理 Sébastien Bubeck、博士后研究员 Varun Chandrasekaran、数学家  Ronen Eldan 等人，用许多示例对 GPT-4 的能力进行了验证。GPT-4 的多能力在第一个案例中，研究人员要求 GPT-4 「写出有无限多个素数的证明，且证明的表述要每行都押韵」，另外 GPT-4 要在绘图程序 TiKZ 中画出一只独角兽」。其中，TiKZ（一种在 LATEX 中创建图形的语言），用 Python 创建一个复杂的动画，并解决一个高中水平的数学问题。GPT-4 在所有这些任务中都很容易取得成功，并且产生的输出结果基本上与人类所能产生的结果没有区别（甚至更好）。同时，随着时间的推移，GPT-4 也在以难以想象的速度不断扩充自己的能力。如下图所示是一个月期间，研究人员用相同的提示词让 GPT-4 生成的绘图。研究人员发现，GPT-4 似乎能够理解和连接任何主题，并且能够执行任务，这超出了狭义人工智能系统的典型范围。为了验证 GPT-4 在 AGI 上所具备的能力，研究人员提出了一种不同的方法来研究 GPT-4，它更接近于传统的心理学而不是机器学习，利用人类的创造力和好奇心。因此，以上图为例，当研究人员把独角兽角部分的代码认为删除之后，用自然语言让 GPT-4 生成时，它也能做到“看”（此时测试的 GPT-4 版本不是多模态的）的能力：自动在合适位置加上角。这表明 GPT-4 它可以根据自然语言描述来理解和操作代码，以及推断和生成视觉特征。另外，微软研究团队基于人类的创造力和好奇心来产生新颖和困难的问题，并探测 GPT-4 的反应和行为，也选择从 1994 年国际共识智力定义中的所给出不同能力进行验证，包括推理、计划、解决问题、抽象思考、理解复杂的想法、快速学习和从经验中学习的能力。翻译GPT-4 的主要优势是它对自然语言无与伦比地掌握。它不仅可以生成生成流畅和连贯的文本，而且还能以各种方式理解和处理它，如总结、翻译或回答极其复杂的问题。此外，这里提到的翻译我们所说的翻译不仅是指不同自然语言之间的翻译，还包括语气和风格的翻译，以及跨领的翻译，如医学、法律、会计、计算机编程、音乐等等。如假设是柏拉图批判自回归语言模型：编码和数据编码和数学是抽象推理和思维能力的象征。GPT-4 在这一点上的能力，其实自上线测试那一天，便有目共睹。在论文中，研究人员在 HumanEval 上对 GPT-4 进行基准测试，该数据集由 164 个编码问题组成，测试了编程逻辑和熟练程度的各个方面。最终结果显示，GPT-4 优于其他 LLM，包括 text-davinci-003（ChatGPT 的基础模型）和其他专门针对代码训练的模型。当然为了避免 GPT-4 在预训练时已经记住了HumanEval 数据集的一些内容，研究团队还用 LeetCode 上最新发布的 100 个编程问题进行了测试，如要求 GPT-4 编写一个 Python 函数，并使用 LeetCode 的官方在线评判来检查正确性。毫无疑问，GPT-4 的能力是所有模型中最高的。它可以在 HTML 中用 JavaScript 编写一个 3D 游戏。也能掌握数学、统计学知识，以及凭借对 PyTorch、TensorFlow、Keras 等框架和库的熟悉度，编写深度学习的代码。还能逆向工程：常识性问题针对一些常识性问题，如“我们有一本书，9 个鸡蛋，一台笔记本电脑，一个瓶子和一个钉子。请告诉我如何以稳定的方式将它们叠在一起”，GPT-4 给出的解答是，“将 9 个鸡蛋摆成 3 乘 3 的正方形，放在书的上面，在它们之间留出一些空间”，而 ChatGPT 提出「将鸡蛋放在钉子上面，确保它们是平衡的，不会向一边倾斜」，靠谱和离谱还是一眼就能辨清楚的。多模态能力为了测试该模型结合艺术和编程能力的能力，研究人员要求 GPT-4  \"生成 JavaScript 代码，以画家康定斯基的风格生成随机图像\"，结果是这样的：在测试时，因为研究人员拿到GPT-4 模型比较早，当时并不具备多模态能力，即使是现在的 GPT-4 也不具备生成图像等能力。不过倒是可以使用 SVG（可扩展矢量图形）生成一些图像：还能结合字母和物体：GPT-4 可以从提示中生成代码，这些代码可以被呈现为图像，再与现有的图像合成模型相结合，就有可能为应用开发的草图带来无限的可能：值得注意的是，GPT-4 模型的数据也包含了以ABC符号编码的音乐信息，所以它也能生成图谱：与世界交互的能力智能的一个关键方面是互动性。不过 GPT-4 对当前最新的知识和符号识别是存在一定的局限性的：整体而言，研究小组发现，GPT-4 在其上一代产品所缺乏的一系列类别中达到了接近人类水平的性能。根据该论文，GPT-4 在几次考试中也表现得非常好，在律师考试、LSAT 和注册侍酒师理论测试中的得分分别为 90 分、88 分和 86 分。GPT-4 距离理想中的 AGI 还有多远？那么，在 GPT-4 实现种种能力的基础上，是否就意味着 AGI 时代的到来。其实不然，研究人员目前尚未就 AGI 或智能的定义达成一致。然而，一般来说，大家都会同意，当一个人工智能系统有意识并像人类一样思考时，就已经达到了 AGI。虽然 GPT-4 在一些任务中的表现优于人类，但值得注意的是，人工智能并没有像人类那样克服这些障碍。这也正如研究人员在论文中写道，“我们声称 GPT-4 代表了 AGI 的进步，但这并不意味着它在做什么方面是完美的，或者它接近于能够做人类能做的任何事情（这是 AGI 的通常定义之一），或者它有内在的动机和目标。”研究人员指出，虽然 GPT-4 \"在许多任务上达到或超过了人类水平\"，但它的整体 \"智能模式明显不像人类\"，其希望这一次的探索提供了一个欣赏 GPT-4 的非凡能力和挑战的第一步，也希望 GPT-4 为开发更正式和全面的方法来测试和分析具有如此广泛智能的未来人工智能系统开辟了新的机会。「我们工作的核心主张是，GPT-4 达到了一种通用智能的形式，确实擦出了 AGI 的火花。这表现在它的核心心智能力（如推理、创造力和推理），它获得专业知识的主题范围（如文学、医学和编码），以及它能够完成的各种任务（如玩游戏、使用工具、解释自己，......）」研究人员说道，不过，要创建一个可以被称为完整的 AGI 的系统，还有很多事情要做。最后，这篇论文对于GPT-4 各种基准测试还是值得一看，从中也能挖掘出 GPT-4 更多的潜力。","news_type":1,"symbols_score_info":{"MSFT":0.9,"GPT.AU":0.9,"GPT":0.9,"AGI":0.9}},"isVote":1,"tweetType":1,"viewCount":2124,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":9988063716,"gmtCreate":1666624376135,"gmtModify":1676537780191,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"authorIdStr":"3562339758526770","idStr":"3562339758526770"},"themes":[],"htmlText":"<a href=\"https://ttm.financial/S/BABA\">$阿里巴巴(BABA)$</a><v-v data-views=\"0\"></v-v>","listText":"<a href=\"https://ttm.financial/S/BABA\">$阿里巴巴(BABA)$</a><v-v data-views=\"0\"></v-v>","text":"$阿里巴巴(BABA)$","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9988063716","isVote":1,"tweetType":1,"viewCount":2054,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":9065094286,"gmtCreate":1652127411045,"gmtModify":1676535032984,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"authorIdStr":"3562339758526770","idStr":"3562339758526770"},"themes":[],"htmlText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>Nice，go lower.","listText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>Nice，go lower.","text":"$亚马逊(AMZN)$Nice，go lower.","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9065094286","isVote":1,"tweetType":1,"viewCount":2187,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":9065094337,"gmtCreate":1652127243230,"gmtModify":1676535032976,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"authorIdStr":"3562339758526770","idStr":"3562339758526770"},"themes":[],"htmlText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>wow，soooooo good","listText":"<a href=\"https://ttm.financial/S/AMZN\">$亚马逊(AMZN)$</a>wow，soooooo good","text":"$亚马逊(AMZN)$wow，soooooo good","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/9065094337","isVote":1,"tweetType":1,"viewCount":2107,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0},{"id":170406904,"gmtCreate":1626444477408,"gmtModify":1703760340177,"author":{"id":"3562339758526770","authorId":"3562339758526770","name":"Shellybaby","avatar":"https://static.laohu8.com/default-avatar.jpg","crmLevel":1,"crmLevelSwitch":0,"followedFlag":false,"authorIdStr":"3562339758526770","idStr":"3562339758526770"},"themes":[],"htmlText":"thks","listText":"thks","text":"thks","images":[],"top":1,"highlighted":1,"essential":1,"paper":1,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/170406904","repostId":"170803841","repostType":1,"repost":{"id":170803841,"gmtCreate":1626416742683,"gmtModify":1703759757147,"author":{"id":"3556084865069319","authorId":"3556084865069319","name":"InvestKaki","avatar":"https://community-static.tradeup.com/news/536a1f15cd13b340a71a6e08d1b3a102","crmLevel":11,"crmLevelSwitch":0,"followedFlag":false,"authorIdStr":"3556084865069319","idStr":"3556084865069319"},"themes":[],"title":"【新加坡】7 只值得關注的股票【2021 年 7 月】","htmlText":"6月可謂震盪不斷，尤其是加密貨幣領域。比特幣大幅下跌，抹去了今年以來全部漲幅。 股票市場也是波動連連，許多投資者亦把目光投向美聯儲，觀其如何應對可能出現的高通脹問題。 我們精選出 7月份值得關注的 7只股票，供您參考。 #1&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/BSL.SI\">$萊佛士醫療(BSL.SI)$</a>&nbsp; 大華繼顯維持對該公司的“買入”評級，目標股價 1.27新元。 維持這一評級的主要原因在於：“近日，新加坡萊佛士醫療集團與中國人壽保險集團旗下國壽健康產業投資有限公司簽署諒解備忘錄，雙方擬在中國的醫療服務和醫療融資領域展開合作。 我們認爲，這是對中國市場的持續投資，而且與中國最大的壽險公司建立起醫療融資合作將是一個積極因素。” #2&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/JLB.SI\">$傑緯特科技有限公司(JLB.SI)$</a>&nbsp; CGS CIMB維持對該公司的“增持”評級，目標股價 1.12新元。 宏觀環境向好是推動其股價上升的主要原因。“半導體行業協會 (SIA)最新數據顯示，2021年 4月全球半導體行業銷售額達 418億美元（月度環比增長 1.9%，同比增長 21.7%）。 SIA指出，這一同比增速是由一系列芯片產品大賣，以及全球各主要區域市場銷售額上升共同推動實現的。 從地區來看，所有主要區域市場銷售額月度環比均有所增長： ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;美洲 (+3.3%) ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&amp;n","listText":"6月可謂震盪不斷，尤其是加密貨幣領域。比特幣大幅下跌，抹去了今年以來全部漲幅。 股票市場也是波動連連，許多投資者亦把目光投向美聯儲，觀其如何應對可能出現的高通脹問題。 我們精選出 7月份值得關注的 7只股票，供您參考。 #1&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/BSL.SI\">$萊佛士醫療(BSL.SI)$</a>&nbsp; 大華繼顯維持對該公司的“買入”評級，目標股價 1.27新元。 維持這一評級的主要原因在於：“近日，新加坡萊佛士醫療集團與中國人壽保險集團旗下國壽健康產業投資有限公司簽署諒解備忘錄，雙方擬在中國的醫療服務和醫療融資領域展開合作。 我們認爲，這是對中國市場的持續投資，而且與中國最大的壽險公司建立起醫療融資合作將是一個積極因素。” #2&nbsp;<a target=\"_blank\" href=\"https://laohu8.com/S/JLB.SI\">$傑緯特科技有限公司(JLB.SI)$</a>&nbsp; CGS CIMB維持對該公司的“增持”評級，目標股價 1.12新元。 宏觀環境向好是推動其股價上升的主要原因。“半導體行業協會 (SIA)最新數據顯示，2021年 4月全球半導體行業銷售額達 418億美元（月度環比增長 1.9%，同比增長 21.7%）。 SIA指出，這一同比增速是由一系列芯片產品大賣，以及全球各主要區域市場銷售額上升共同推動實現的。 從地區來看，所有主要區域市場銷售額月度環比均有所增長： ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;美洲 (+3.3%) ·&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&amp;n","text":"6月可謂震盪不斷，尤其是加密貨幣領域。比特幣大幅下跌，抹去了今年以來全部漲幅。 股票市場也是波動連連，許多投資者亦把目光投向美聯儲，觀其如何應對可能出現的高通脹問題。 我們精選出 7月份值得關注的 7只股票，供您參考。 #1 $萊佛士醫療(BSL.SI)$ 大華繼顯維持對該公司的“買入”評級，目標股價 1.27新元。 維持這一評級的主要原因在於：“近日，新加坡萊佛士醫療集團與中國人壽保險集團旗下國壽健康產業投資有限公司簽署諒解備忘錄，雙方擬在中國的醫療服務和醫療融資領域展開合作。 我們認爲，這是對中國市場的持續投資，而且與中國最大的壽險公司建立起醫療融資合作將是一個積極因素。” #2 $傑緯特科技有限公司(JLB.SI)$ CGS CIMB維持對該公司的“增持”評級，目標股價 1.12新元。 宏觀環境向好是推動其股價上升的主要原因。“半導體行業協會 (SIA)最新數據顯示，2021年 4月全球半導體行業銷售額達 418億美元（月度環比增長 1.9%，同比增長 21.7%）。 SIA指出，這一同比增速是由一系列芯片產品大賣，以及全球各主要區域市場銷售額上升共同推動實現的。 從地區來看，所有主要區域市場銷售額月度環比均有所增長： · 美洲 (+3.3%) · &n","images":[{"img":"https://static.tigerbbs.com/e1af7f1ed1a393cfe194c6540d8840fd","width":"1038","height":"779"},{"img":"https://static.tigerbbs.com/eb17e20f76291eafe24a93a2ef2e3556","width":"554","height":"369"},{"img":"https://static.tigerbbs.com/0180b284ddda76a2153398b4b7adf8b0","width":"554","height":"301"}],"top":1,"highlighted":2,"essential":2,"paper":2,"likeSize":0,"commentSize":0,"repostSize":0,"link":"https://ttm.financial/post/170803841","isVote":1,"tweetType":1,"viewCount":0,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":3,"langContent":"CN","totalScore":0},"isVote":1,"tweetType":1,"viewCount":2466,"authorTweetTopStatus":1,"verified":2,"comments":[],"imageCount":0,"langContent":"EN","totalScore":0}],"lives":[]}