一个GPT的幽灵在Gemini上空徘徊

Every time the author Luo Yihang makes a big move in the field of generative artificial intelligence, it can make people feel a kind of secret and huge emotional power. The release of a series of models on the basis of the month is regarded as a strong challenge at that time. At the end of the year, a series of large language models including the basic version and the advanced version of the mobile version were presented without warning, which is more direct in terms of key performance and benchmark evaluation indicators. The official claim is that it is an academic benchmark widely used in the research and development of large language models. Some of its performance has surpassed the most advanced level represented at present, flaunting native multimodal, which makes it pre-train on combined modes such as text, image, audio, video and code from the beginning, so it may have stronger performance in complex understanding and reasoning, especially in solving mathematical and physical problems, and spare no effort to emphasize the above advantages. It takes a gradual multimodal road, first following up based on text corpus code, then image, video and audio, and finally combining these abilities to train in multimodal from the beginning. Corpus training and multimodal data optimization are more advanced than training methods. In theory, some advanced multimodal training can bring stronger performance. The published academic benchmark evaluation results seem to explain the details of large-scale transcendence, but the academic benchmark itself is a part of the theory and can not really reflect the application effect. Many people run on some big language models in China and are keen on benchmarking. We should treat it equally, which is essentially the same as that of domestic big models. At present, the users of the only supported version of the chat bot on the former social platform have contributed a lot. For example, it confuses the Oscar winners in and will not use the simple function of writing the intersection of two polygons. We also find that it can't recognize the number of leaves and can't do the simple geometric problem of finding an acute angle correctly. Even if the target is it, it is still a little less interesting. Another prominent problem that people have pointed out is that propaganda video fraud quickly responds to a group of gestures, saying that it is a. The game of rock, paper, scissors, but its function document, which has not been shown by video, gives at least two hints about what I am doing. This is a game, and other tests even need more hints to help generate results. However, this process was omitted in the official video, so that most people who are not serious overestimate their understanding ability and reaction speed. This has to be said to be misleading. I still remember that my staff on the live stage demonstrated how to book one directly through the voice assistant in. Applause thundered at the bottom of a restaurant, and I clapped my hands. I thought it was great, but a month later, it came out that this bridge was prepared in advance. There was no fraud, but it was usually so eager to show its unparalleled ability that it often abbreviated the process behind it, which actually exaggerated the effect. To put it bluntly, the exaggerated presentation of the video just showed that it was too concerned about it, and people were anxious about any big model that was competing with it, especially the masterpieces of giants. Of course, people are the most demanding. After all, one of the motives for choosing to use the invented architecture to create an epoch-making model is to get rid of the ubiquitous repression. Who doesn't want to see make a fool of oneself? To some extent, it is the only twin on this planet. All the architectures included are open source. With Mr. Musk's preference for open source, it is highly probable that China's big language model will also be open source in the future. Only peace is firmly closed, which makes progress in the big language model. Instinctively, there is a strong binding relationship with it, and there is also a dramatic tension. Every time there is a big action around it, public opinion will shout out, and when it falls back, it will be beaten. Then it will almost certainly offer a new big move one to two months after the move, prove that your uncle is still your uncle, and then hold back for a few months and then be called out and beaten. Will this really change the pattern? To some extent, it is still one position behind in ecological construction. After all, there are millions of developments in this world. I have done it myself, but it is not until early next year that I can provide intensive training feedback to developers and corporate customers, so that people can develop their own applications on it. By then, I am afraid it will have been officially launched. I have always been a little confused. Didn't I win half of Apple by open source? How did I lose this role to me this time? I really don't want to blame me, but I am more looking forward to proving myself. We people who came into contact with the Internet at the end of the 1990 s have some subtle special feelings, and we must also prove it. It is a fact that one's own strategy can bear real fruit, but the ghost is wandering. Anyone can try to get rid of this ghost, but it can't. This is its opponent that has no choice. In fact, everything that has been done today can resonate with the hearts of China's big language model developers to some extent. The ghost hovering above everyone's head makes everyone try to prove that they are better in some aspects through some efforts, and adopt all the tricks and more complicated thinking chains in the benchmark test. Tips and results were selected, but the test only used secondary feedback and the results were crushed without prompts. Do we sound familiar with the similar test methods? Do the developers of large language models in China have a feeling that the villagers have met their fellow villagers? We often like to compare the efforts of Zhipu Baidu and others, but to put it another way, in fact, the contest of large language models is not the bright top of Baidu Zhipu and the students besieged together. In this sense, the big model in China and the big model in the United States are not the same. In a camp and a trench, it is the grass of the big model, which is also the object of learning from each other. We Shennong have tasted it too much and found that our big model in China is not worse than that in the United States, but it is not as good as this. Another thing worth noting about this training is that it completely adopts its own chip cluster training officer Xuanhe to realize this large-scale training on the optimized infrastructure, which is highly scalable and the most efficient in reasoning. I am afraid it is a model we have heard of with relatively strong performance. Of course, people are self-produced and self-sold, but I have seen the possibility and feasibility of replacing NVIDIA in the practice of large-scale model training. What this means for China's large-scale model training is self-evident. In addition, it is also a bright spot. This is the smallest model, which is given priority to the upper-end model of home-made mobile phones. In fact, it is closer to the small model. China smartphone manufacturers and Xiaomi have recently released their own end-to-end models, and Lenovo has actually cut into the same field from the new dimension. 比特币今日价格行情网_okx交易所app_永续合约_比特币怎么买卖交易_虚拟币交易所平台

注册有任何问题请添加 微信：MVIP619 拉你进入群

文章转载注明 网址：https://netpsp.com/?id=62532