OpenAI的幽默与执着:从ChatGPT Pro到GPT-4O

in hive-180932 •  7 months ago 

image.png

今天随便聊聊天,今天早上一起来看到半夜里open AI又出了新产品,叫GPT四o.当然我们这里有半夜,正好是他们的工作时间嘛,时差原因可能就不想是价魔搭致行吧,行品种么?请你认真,对不对?这个GPT四o中的o指的是omini,就是所谓的原模态,不光可以看图,还可以音频,其实还可以音频对话实时翻译等等。其实这些应用模式都不算什么新鲜的,主要是open AI的GPT四o模验比现成的一些要好很多,延迟更低。比较有意思的是,在open AI暂时推出GPT四o之前,曾经在著名的语言模型竞技场网站上推出了一个叫做chat board GPTR的的聊天机器人。这个语言模型竞技场,用户可以自己提问,对用户的提问可以返回由不同模型产生的答案。然后由用户对这些答案进行打分。不过用户并不知道哪个答案是由哪个模型产生的。这就非常类似于医学上的双盲实验得出的结果是非常客观的。在这个榜单上,GPT四系列模型一直处于领先状态。虽然国内和国外很多AI公型,包括谷歌都宣称自己的模型达到甚至超过GPT四的水平。但在这个榜单上面,这些模型都只是保持颈椎的状态,虽然差距越来越小,但从这个榜单上看,还没有哪一个模型真正超越过GPT四唯一的例外就是这个神秘的chat board GPT二。当时人们就怀疑这个就是open AIA推出的下一代模型的测试版。没有。他昨天发布会也证实了这一个猜测,其中也可以看出open AI的一些幽默搞笑的精神。因为GPT二是open AI早期开源的一款参数量比较小的GVT语言模型效果并不好。所以即使开源呢也没有引起大家的广泛兴趣,直到他的下一代GVT三恰TGBT出现才引起轰动。而open AI将他们的GVT四升级版命名成GPT二,似乎是在说他们能够点石成金的技术,能够将GPT二改造成超越GPT四的聊天机器人。这也不是open AI第一次恶搞了。其实open AI刚创立的时候,世界首富马斯克也是他的投资人之一。后来因为理念不同,撤资了,在看到chad GP大火之后,不知道这位世界首富是不是对自己过早抽身感到后悔,反正他一边呼吁政府管制要求暂停停止比GTT四更大的的i模模的开发,一边自己购买大创显卡,采始训练自己的聊天机器人grorook要时还指责open AI违反了当初创立的复兴,就是研究和发布开源的人工智能模型,让每个人都能享受AI带来的好处,而不是让这项技术仅仅集中在少数大公司的手里。为了表达自己言行一致,马斯克还扬言要将自己开发的grorook聊天机器人所用的语言模型进行开源。结果open AI并没有回应马斯克的他们底是谁言论,而是幽默了一把抢先在github上以open enem名义开源了一个叫grock的项目。当然,其中并没有达何AI模型或者训练数据等等。不过话说回来,这次新发布的chat新发布的VTTO虽然在GPT四的能力基础上有所提升,但是并没有达到质的飞跃的水平,主要的还是在使用体验上下的功夫。比如会察言观色、了解判断使用者的情绪,快速实时的翻译等等。不是不要你那还是仔细的这似乎说明在大语言模型刚刚推出的时候,很多人乐观的认为在语言模型的基础上,就可以实现人类梦寐以求的agi了,或者在近近半模的的发展来在语言模型似乎已经到了它发展的瓶颈期,或者说已经很接近它的天花板了,或者说GPT四呃能力已经接近了语言模型的天天板向向上提升的发展已经不大了。而且根据语言模型的原理来看,它主要是学习语言符号之间的关系。虽过在大大部分情况下,已类的思考过程,思思想交流流是通过语言完成的,但是在有一些时间的关系,那些对智力水平要求比较高的领域,比如数学、光靠语言,比如数学、物理、光靠语言作为思考工具是远远不够的。这在我辅导小宝贝数学学习的时候就表现的很明显。比如一到对小学生来说还是很有难度的盈亏问题。一群小朋友分苹果,每个人分四个,则多出了七个苹果。如果每个人分五个又少了两个苹果,请问有多少个苹果和多少个小朋友?这样题目高年级的学生来说是没有难度的,他可以用二元一次方程组轻松的解出答案。但是小学二三年级的学生来说,往往不知道不下手,这时候就可以用结合图形将这两种情况画出来,然后分析他们的转化过程也能够很容易的得出答案。说明图形的动态思维,在数数学、物理这些需要高度智能的学科中有着非常重要的作用。光靠对语言的学习是无法掌握的。还有爱因斯坦发现广义相对论的过程,也是通过在大脑中不断的做思维实验。比如思考一个人从高楼上掉下来是什么感受?这样的问题得出了引力,是一种几何效应的结论。而这个过程仅仅通过语言思考,通过语言工具来思考,是很难得出爱因斯坦的洞见的。那么真正的AGI肯定不能只掌握语言工具。嗯,oh my god嗯,反抗尔尔的得出,这就需要我们将其他形态的AI模型结合起来才行。当然现在已经有具有视觉能力的语言模型,但它的视觉模块主要是识别功能,并没有思考的作用,也许会思考的视觉模型才是实现AGI的关键吧。这是我的一点点想法,希望能够启发到有关的人士。


You'll never see anything go into a black hole. Imagine that you trap your enemy in a rocket ship and launch him into a black hole. He looks back at you and shakes his fist at a constant rate. As he accelerates closer, gravity becomes stronger, so you would expect him to accelerate. But that's not what you see. Instead, the rocket ship appears to be slowing down. Not only that, but his fist movements seemed to be getting slower and slower. This is because from your perspective, his time is slowing down. At the moment he was supposed to cross the event horizon, the point from which not even light can escape, he and his rocket ship did not disappear. Instead, they seemed to stop, frozen in time. The light from the ship became dimmer and redder until it was completely out of view. This is what any object looks like when it crosses the event horizon. The light was still coming from the point he passed through, it was just too redshifted to be seen. But if you could see that light, then in theory you would see everything that ever fell into the black hole frozen at its event horizon, including the stars that formed it. But in reality, photons are emitted at discrete intervals, so there will be a last photon emitted outside the event horizon, so these images will gradually fade after a while. This is just one of the strange results of general relativity, our current best theory of gravity. The first solution to Einstein's equations predicted not only black holes, but also their opposite, white holes. It also hints at the existence of parallel universes, perhaps even a way to travel between them. This is a video about the real science of black holes, white holes and wormholes. General relativity stems at least in part from a fundamental flaw in Newtonian gravity. In the 17th century, Isaac Newton pondered how apples fall to the ground, how the moon orbits the Earth, and how the Earth orbits the sun, and concluded that every object with mass must attract every other object. But Newton was troubled by his own theory. How do objects so far apart exert forces on each other? He wrote that an object could exert a force on another object at a distance through a vacuum without anything else mediating it. One person who certainly had the ability to think competently, he writes, was Albert Einstein, who, more than 200 years later, figured out how gravity is mediated. Objects do not exert forces directly on each other. Instead, a mass like the sun bends space-time around it. This in turn bends the space-time around it, and so on, all the way to Earth. So the Earth goes around the sun because the space-time it travels through is curved. Mass is affected by the local curvature of space-time and therefore does not need to be acted upon at a distance. Mathematically, this is described by Einstein's field equations. Can you write down Einstein's field equations? It was the result of Einstein's hard work in the decade after his special theory of relativity. Essentially, what we get in the field equation, on the one hand, tells me the distribution of matter and energy. The other side tells you what the curvature of space-time is that arises from this distribution of matter and energy. It's a line. It looks like, oh, this is a simple equation, right? But it's not really an equation. It's a set of equations. To make life more difficult, they are coupled equations, so they depend on each other, they are differential equations. That means you have to integrate, and so on. So you need to do a lot of steps to solve the field equation. To see what the solutions to these equations would look like, we need a tool to understand spacetime. So imagine you're floating in empty space. A light flashed above your head and spread in all directions. Now your entire future, anything that can and will happen to you, will happen inside this bubble. Because the only way out is to travel faster than light. In two dimensions, the bubble is just an expanding circle. If we allow time to run up the screen and take snapshots at regular intervals, then this bubble of light depicts a cone, your future cone of light. By convention, the axis is scaled so that the light always travels at a 45 degree Angle. This cone reveals the only region of space-time you can hope to explore and influence. Now imagine that instead of a flash of light flashing over your head, those photons actually traveled from all corners of the universe, met at that moment, and then continued in their respective directions. Well, in this case, going back in time, those photons also reveal a light cone, your past light cone. Only the events that take place within this cone can affect you up to the present moment. We can simplify this diagram further by plotting a space and a time dimension. This is a space-time map of empty space. If you want to measure the distance between two events in spacetime, you use something called a spacetime interval. The square of the interval is equal to minus dt squared plus dx squared. Since spacetime is flat and the geometry is the same everywhere, this formula holds true throughout the diagram, making it very easy to measure the distance between any two events. But around mass, spacetime is curved, so you need to modify the equation to account for geometry. This is what the solution to Einstein's equation looks like. They tell you how space-time curves and how to measure the distance between two events in that curved geometry. Einstein published his equations in 1915, in the midst of World War I, but he could not find an exact solution. Fortunately, one copy of his paper made its way to the Eastern Front, where Germany was at war with Russia.

今天随便聊聊天,今天找到一起来看到半夜里又出了新产品,叫吉皮T四欧。当然我们这里半夜正好是他们的工作时间嘛,时差原因上去是不�通的,欧指的是欧米尼,就是所谓的人模态,不光可以看图,还可以音频,其实还可以音频对话,实时翻译等等。其实这些应用模式都不算什么新鲜的,主要是欧盆的吉布提斯,欧比现成的一些要好很多,延迟更低。比较有意思的是,在欧盆正式推出。对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对对。比现成的一些要好,很多,延迟更低。比较有意思的是,在欧盆证实推出之前,曾经在著名的语言模型竞技场网站上推出了一个叫做七七二的的聊天机器人。这个语言模型竞技场,用户可以自己提问,对用户的提问可以返回由不同模型产生的答案。然后由用户对这些答案进行打分。不过用户并不知道哪个答案是由哪个模型产生的。这就非常类似于医学上的双盲实验得出的结果是非常客观的。在这个榜单上,七四系列模型一直处于领先状态。虽然国内和国外很多公司,包括谷歌都宣称自己的模型达到甚至超过七的水平。但在这个榜单上面,这些模型都只是保持颈椎的状态,虽然差距越来越小,但从这个榜单上看,还没有哪一个模型真正超越过七唯一的例外就是这个神秘的加特波特七批T二。当时人们就怀疑这个就是欧盆艾艾推出的下一代模型的测试版啊。你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你你把你其他有的都两个。我们是唉呀啊昨天发布会也证实了这一个猜测,其中也可以看出欧盆的的一些幽默搞笑的精神。因为GPT二是欧盆的早期开源的一款参数量比较小的一比七,语言模型效果并不好,所以即使开源了,也没有引起大家的广泛兴趣,直到它的下一代七三下的出现才引起轰动啊,将他们的四升级版命名成七二,似乎是在说他们能够点石成金的技术,能够将七二改造成超越吉皮四的聊天机器人,这也不是欧盖第一次恶搞了。其实欧盖刚创立的时候,世界首富马斯克也是他的投资人之一。后来因为理念不同,撤资了,在看到恰的七批大火之后,不知道就世界首富,是不是对自己过早抽身感到后悔,反正他一边呼吁政府管制要求暂停停止以及第四更强大的模型的开发,一边自己购买大量显卡,开始训练自己的聊天机器人国,同时还指责欧格维反了当初创立的不新,就是研究了这个新的新的、新的、新的、新的、新的、新的、新的、新的、新的、新的、新的、新的新的这项技术仅仅集中在少数大公司的手里,为了表达自己言行一致。马斯克还扬言要将自己开发的谷歌聊天机器人所用的语言模型进行开源,结果并没有回应马斯克的言论,而是幽默了一把抢先在上比开开开开的一个叫格洛克的上有所提升,但是并没有达到质的飞跃的水平。主要的还是在使用体验上下的功夫。比如会察言观色,了解判断使用者的情绪,快速实时的翻译等等啊,然后然后然后然后然后然后然后然后语言模型刚刚推出的时候,很多人乐观地认为,在语言模型的基础上,就可以实现人类梦寐以求的AGI了。不过最近半年来的发展来看,语言模型似乎已经到了它发展的瓶颈期,或者说已经很接近它的天花板了,或者说七七四的能力已经接近了大圆模型的天花板向上提升的幅度已经不大了。而且根据语言模型的原理来看,它主要是学习语言符号时间的关系。虽然在绝大部分情况下,人类的思考过程和思想的交流都是通过语言完成的。但是在有些领域,特别是那些对智力水平要求比较高的领域,比如数学、光靠语言,比如数学、物理、光靠语言作为思考工具是远远不够的。这在我辅导小宝贝数学学习的时候就表现得很明显。比如一道对小学生来说还是很有难度的盈亏问题。一群小朋友分苹果,每个人分四个,则多出了七个苹果。如果每个人分五个又少了两个苹果,请问有多少个苹果和多少个小朋友?这样题目高年级的学生来说是没有懒惰的,他可以用二元一次方程组轻松地解出答案。但是小学二三年级的学生来说,往往不知道和下手,这时候就可以用结合符琴将这两种情况画出来,然后分析他们的转化过程也能够很容易的得出答案。说明图形的动态思维,在数学、物理这些需要高度智能的学科中有着非常重要的作用。光靠对语言的学习是无法掌握的。还有爱因斯坦发现广义相对论的过程,也是通过在大脑中不断的做思维实验。比如思考一个人从高楼上掉下来是什么感受?这样的问题得出了引力,是一种几何效应的结论,把这个过程仅仅通过语言思考,通过语言工具来思考,是很难得出爱因斯坦的洞见的。那么真正的A肯定不能只找我语言工具。嗯,哦买嘎啊啊,有房卡你来了。嗯,这就需要我们将其他形态的模型结合起来才行。当然现在已经有具有视觉的模型了嘛。啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊啊的视觉模块主要是识别功能并没有思考的作用。也议学模型才是实现给借爱的关销织织学能力的语学的想法,希望能够企发到有关的人。

Authors get paid when people like you upvote their post.
If you enjoyed what you read here, create your account today and start earning FREE STEEM!
Sort Order:  

Upvoted! Thank you for supporting witness @jswit.

888888888

后半段有点敷衍了哦

ai是造福人类呢,还是毁灭人类?