OpenAI全网疯传的53页PDF文档：计划2027年前开发出通用人工智能（英）.pdf

Revealing OpenAIs plan to create AGI by 2027 In this document I will be revealing information I have gathered regarding OpenAIs(delayed)plans to create human-level AGI by 2027.Not all of it will be easily verifiable but hopefully theres enough evidence to convince youSummary:OpenAI started training a 125 trillion parameter multimodal model in August of 2022.The first stage was Arrakis also called Q*.The model finished training in December of 2023 but the launch was canceled due to high inference cost.This is the original GPT-5 which was planned for release in 2025.Gobi(GPT-4.5)has been renamed to GPT-5 because the original GPT-5 has been canceled.The next stage of Q*,originally GPT-6 but since renamed to GPT-7(originally for release in 2026),has been put on hold because of the recent lawsuit by Elon MuskQ*2025(GPT-8)was planned to be released in 2027 achieving full AGI.Q*2023=48 IQQ*2024=96 IQ(delayed)Q*2025=145 IQ(delayed)Elon Musk caused the delay because of his lawsuit.This is why Im revealing the information now because no further harm can be done Ive seen many definitions of AGI artificial general intelligence but I will define AGI simply as an artificial intelligence that can do any intellectual task a smart human can.This is how most people define the term now.2020 was the first time I was shocked by an AI system that was GPT-3.GPT-3.5,an upgraded version of GPT-3,is the model behind ChatGPT.When ChatGPT was released,I felt as though the wider world was finally catching up to something I was interacting with 2 years prior.I used GPT-3 extensively in 2020 and was shocked by its ability to reason.GPT-3,and its half-step successor GPT-3.5(which powered the now famous ChatGPT-before it was upgraded to GPT-4 in March 2023),were a massive step towards AGI in a way that earlier models werent.The thing to note is,earlier language models like GPT-2(and basically all chatbots since Eliza)had no real ability to respond coherently at all.So why was GPT-3 such a massive leap?.Parameter Count“Deep learning”is a concept that essentially goes back to the beginning of AI research in the 1950s.The first neural network was created in the 50s,and modern neural networks are just“deeper”,meaning,they contain more layers theyre much,much bigger and trained on lots more data.Most of the major techniques used in AI today are rooted in basic 1950s research,combined with a few minor engineering solutions like“backpropogation”and“transformer models”.The overall point is that AI research hasnt fundamentally changed in 70 years.So,theres only two real reasons for the recent explosion of AI capabilities:size and data.A growing number of people in the field are beginning to believe weve had the technical details of AGI solved for many decades,but merely didnt have enough computing power and data to build it until the 21st century.Obviously,21st century computers are vastly more powerful than 1950s computers.And of course,the internet is where all the data came from.So,what is a parameter?You may already know,but to give a brief digestible summary,its analogous to a synapse in a biological brain,which is a connection between neurons.Each neuron in a biological brain has roughly 1000 connections to other neurons.Obviously,digital neural networks are conceptually analogous to biological brains.So,how many synapses(or“parameters”)are in a human brain?The most commonly cited figure for synapse count in the brain is roughly 100 trillion,which would mean each neuron(100 billion in the human brain)has roughly 1000 connections.If each neuron in a brain has 1000 connections,this means a cat has roughly 250 billion synapses,and a dog has 530 billion synapses.Synapse count generally seems to predict higher intelligence,with a few exceptions:for instance,elephants technically have a higher synapse count than humans yet display lower intelligence.The simplest explanation for larger synapse counts with lower intelligence is a smaller amount of quality data.From an evolutionary perspective,brains are“trained”on billions of years of epigenetic data,and human brains evolved from higher quality socialization and communication data than elephants,leading to our superior ability to reason.Regardless,synapse count is definitely important.Again,the explosion in AI capabilities since the early 2010s has been the result of far more computing power and far more data.GPT-2 had 1.5 billion connections,which is less than a mouses brain(10 billion synapses).GPT-3 had 175 billion connections,which is getting somewhat close to a cats brain.Isnt it intuitively obvious that an AI system the size of a cats brain would be superior to an AI system smaller than a mouses brain?.Predicting AI PerformanceIn 2020,after the release of the 175 billion parameter GPT-3,many speculated about the potential performance of a model 600 times larger at 100 trillion parameters,because this parameter count would match the human brains synapse count.There was no strong indication in 2020 that anyone was actively working on a model of this size,but it was interesting to speculate about.The big question is,is it possible to predict AI performance by parameter count?As it turns out,the answer is yes,as youll see on the next page.Source:https:/ The above is from Lanrians LessWrong post.As Lanrian illustrated,extrapolations show that AI performance inexplicably seems to reach human-level at the same time as human-level brain size is matched with parameter count.His count for the synapse number in the brain is roughly 200 trillion parameters as opposed to the commonly cited 100 trillion figure,but the point still stands,and the performance at 100 trillion parameters is remarkably close to optimal.By the way an important thing to note is that although 100 trillion is slightly suboptimal in performance,there is an engineering technique OpenAI is using to bridge this gap.Ill explain this towards the very end of the document because its crucial to what OpenAI is building.Lanrians post is one of many similar posts online its an extrapolation of performance based on the jump between previous models.OpenAI certainly has much more detailed metrics and theyve come to the same conclusion as Lanrian,as Ill show later in this document.So,if AI performance is predictable based on parameter count,and 100 trillion parameters is enough for human-level performance,when will a 100 trillion parameter AI model be released?.GPT-5 achieved proto AGI in late 2023 with an IQ of 48 The first mention of a 100 trillion parameter model being developed by OpenAI was in the summer of 2021,mentioned offhand in a wired interview by the CEO of Cerebras(Andrew Feldman),a company which Sam Altman is a major investor of.Sam Altmans response to Andrew Feldman,at an online meetup and Q&A called AC10,which took place in September 2021.Its crucial to note that Sam Altman ADMITS to their plans for a 100 trillion parameter model.(Sources:https:/ reddit posting itself is sourced from a LessWrong post,which was deleted at Sam Altmans request:https:/ researcher Igor Baikov made the claim,only a few weeks later,that GPT-4 was being trained and would be released between December and February.Again,I will prove that Igor really did have accurate information,and is a credible source.This will be important soon Gwern is a famous figure in the AI world he is an AI researcherand blogger.He messaged Igor Baikov on Twitter(in September 2022)and this isthe response he received.Important to remember:“Colossal number of parameters”.“Text”,“audio”,“images”,“possibly video”,and“multimodal”.This comes from a subreddit called“thisisthewayitwillbe”which is a small,private subreddit Im part of,run by a mathematics professor with an interest in AGI.AI enthusiasts(and a few experts)use the subreddit to discuss AI topics deeper than what youll find in the mainstream.A“colossal number of parameters”?Sounds like Igor Baikov was referencing a 100 trillion parameter model,as 500 billion parameter models and up to 1 trillion parameter models had already been trained many times by the time of his tweet in summer 2022(making models of that size unexceptional and certainly not“colossal”).These tweets from“rxpu”,seemingly an AI enthusiast(?)from Turkey,are interesting because they make a very similar claim about GPT-4s release window before anyone else did(trust me I spent many hours,daily,scouring the internet for similar claims,and no one else made this specific claim before he did).He also mentions a“125 trillion synapse”GPT-4 however,he incorrectly states GPT-3s parameter count as 1 trillion.(It seems as though rxpu did have inside information,but got something mixed up with the parameter counts again,I will illustrate this later,and prove that rxpu was not lying).This is a weaker piece of evidence,but its worth including because“roon”is fairly notable as a Silicon Valley AI researcher,and is followed by Sam Altman,CEO of OpenAI,and other OpenAI researchers on Twitter.In November 2022 I reached out to an AI blogger named Alberto Romero.His posts seem to spread pretty far online so I was hoping that if I sent him some basic info about GPT-4 he might do a writeup and the word would get out.The results of this attempt were pretty remarkable as Ill show in the next two pages.Alberto Romeros post.The general response will be shownon the next page.The 100 trillion parameter leak went viral,reaching millions of people,to the point that OpenAI employees including CEO Sam Altman had to respond calling it“complete bullshit”.The Verge called it“factually incorrect”.Alberto Romero claimed responsibility for the leak as you can see on the left.Igor Baikov,the origin of the“colossal number ofparameters”statement,also saw the viral spread of the GPT-4leak(which was essentially his own doing)and responded.So,after all,Igor really did mean“100 trillion parameters”when he said“a colossal number of parameters”.But,is Igor a reliable source?Are his other claimsaccurate?What about the multimodality?What about the ability for GPT-4 to processimages,sounds,and videos?I will prove Igorsreliability shortly.Somewhere around Oct/Nov 2022 I became convinced that OpenAI planned to first release a 1-2 trillionparameter subset of GPT-4 before releasing the full 100 trillion parameter model(“GPT-5”).These sources arent particularly solid but they all said the same thing including rxpu,who once claimed there was a 125 trillion parametermodel in the works,and then incorrectly claimed GPT-3 was 1 trillion I believe he got his information mixed up.-VIDEO -IMAGES Examples of the current level of quality of publically available video&image generation AI models.These models are less than 10 billion parameters in size.What happens,when you train a model 10,000 times larger,on all the data available on the internet,and give it the ability to generate images and video?(The answer:images and videos completely indistinguishable from the real thing,100%of the time,with no exceptions,no workarounds,no possible way for anyone to tell the difference,no matter how hard they try).-(update:SORA IS FROM GPT-5 Q*2023 MODEL)Important:notice how the AI model is able to generate multiple angles of the same scene with physically accurate lighting,and insome cases even physically accurate fluid and rain.If you can generate images and videos with accurate,common-sense physics,you have COMMON SENSE REASONING.If you can generate common sense,you UNDERSTAND common sense.Two posts from Longjumping-Sky-1971.Im including this because he accurately predicted the release date of GPT-4 weeks in advance(no one else posted this information publicly beforehand,meaning he had an inside source).His posts now have much more credibility and he claimed image and audio generation would be trained in Q3 of 2023.If video generation training is simultaneous or shortly after,this lines up with Siqi Chens claim of GPT-5 being finished training in December of 2023.Lets take a stroll backto February 2020,a few months before GPT-3 was released.An article fromTechnology Review,which wasan“inside story”about OpenAI,seems to suggest that OpenAIwas in the early stages of a“secret”project involvingan AI system trained onimages,text,and“other data”,andthat leadership at OpenAIthought it was the mostpromising way to reach AGI.I wonderwhat this could possibly be referring to.The next slide will reveal some quotes from the President of OpenAI from 2019 and it will tell you what their plan was.OpenAI president GregBrockman stated in 2019,following a 1 billion dollarinvestment from Microsoftat the time,that OpenAIplanned to build a human-brain-sized model within five years,and that this was their planfor how to achieve AGI.2019+5=2024 Both of these sources are clearly referring to the same plan to achieve AGI a human-brain-sized AI model,trained on“images,text,and other data”,due to be trained within five years of 2019,so,by 2024.Seems to line up with all the other sources Ive listed in this document.Source:Time Magazine,Jan 12 2023As Ill show in these next few slides,AI leaders are suddenly starting to sound the alarm almost like they know something VERY SPECIFIC that the general public doesnt.Date of NYT interview:May 1 2023 “I thought it was 30 to 50 years or even longer away.Obviously,I no longer think that.”What made him suddenly change his mind-AND decide to leave Google to speak about the dangers of AI?“I thought it was 30 to 50 years or even longer away.Obviously,I no longer think that.”What suddenly made him change his mind-and decide to leave Google-and speak about the dangers of AI?Shortly after the release of GPT-4,the Future of Life Institute,a highlyinfluential non-profit organization concerned with mitigating potential catastrophic risksto the world,released an open letter calling on all AI labs to pause AIdevelopment for six months.Why?The first released version of the letter specifically said“(including the currently-being-trained GPT-5)”.Why was that included,and why was it removed?Source:Wired,March 29 2023 Source:Vox,March 29 2023 Some alarming quotes from an interview and Q&A with Sam Altman from October 2022-youtube link:https:/ Q&A question:“Do we have enough information in theinternet to create AGI?”Sam Altmans blunt,immediate response,interrupting the man asking the question:“Yes.”Sam elaborates:“Yeah,were confident there is.We think about this and measure it quite a lot.”The interviewer interjects:“What gives you that confidence?”Sams reply:“One of the things I think that OpenAI has driven in the field thats been really healthy is that you can treat scaling laws as a scientific prediction.You can do this for compute,you can do this for data,but you can measure at small scale and you can predict quite accurately how its going to scale up.How much data youre going to need,how much compute youre going to need,how many parameters youre going to need,when the generated data gets good enough to be helpful And the internet istheres a lot o