Chapter 160: Go to the basketball team and show off (First update)
"Orange...hehehei..." Youzi giggled a few times, and her laughter was full of emotion.
Fang Yu glanced at Youzi. In a sense, this big model could also be considered Youzi’s child.
I just don’t know how far this child can grow?
The underlying layer of the Orange model is not only composed of multiple neural networks, but also contains a simplified version of Orange's own architectural principles. It has more than 300 million parameters, and the scale of the model itself exceeds 10G.
The parameter of 300 million is a very terrifying scale in the current year of 3061 in the lunar calendar.
The Deep Q-Network, which Deepmind just announced a few months ago, has only 1.68 million parameters.
Although the number of parameters of DeepFace, a deep learning facial learning system released by Feisibue in the middle of the year, has not been announced, it is speculated that it should be at the level of more than 10 million parameters.
The Google Brain Project, launched by Google three years ago, used 16,000 CPUs for training and claimed to have 1 billion parameters, but the proportion of invalid and negative parameters exceeded 70%.
Although unsupervised learning on the video side was also achieved, the training effect was not good.
But the Orange model is different.
Since Yuzu completed the framework construction of the Orange model in his own body, with the assistance of arcane magic, the invalid parameters and negative parameters among the 300 million parameters of the Orange model can be basically controlled within 10%!
It can be said that the newly born Orange model is currently the most powerful AI model in the world!
The artificial intelligence parameters under the neural network are equivalent to the synapses in the human brain.
The number of parameters is one of the most important factors affecting the capabilities of artificial intelligence models, and even the decisive factor.
More parameters generally mean that the model has higher representational power and can capture and express more complex patterns and relationships.
In simple terms, the more parameters there are, the more human-like the artificial intelligence becomes.
Moreover, models with more parameters can better fit the training data and reduce the training error.
In simple terms, the more parameters there are, the stronger the artificial intelligence's ability to understand.
Generally speaking, the more parameters there are, the stronger the capabilities of artificial intelligence will be. This statement is correct.
Although there is only 40G of training data at present, the Orange model has demonstrated a considerable level of intelligence.
This also shows that the deep learning training framework created by YouZi is much more efficient than the TensorFlow training framework version 0.5 released by Google just a month ago.
It is worth noting that the AI training framework and the model framework of the AI large model are two different things.
For example, the Orange model uses a multi-layer neural network and the hierarchical structure and connection method of the neural network, which is the Orange model framework.
The training framework is a software platform that provides tools and interfaces for building, training, evaluating, and deploying deep learning models.
To put it simply, if the large model framework that has not been trained with data is a brand new brain, then the training framework is the school, the teacher, and the entire education system.
The hierarchy and structure of the AI big model framework itself is the IQ of this new brain.
Training data is the knowledge that the education system teaches to this new brain using various methods.
Teachers have different levels, different education systems, and different knowledge taught, so the efficiency and accuracy of students' mastery of knowledge will naturally be different.
Whether a student's academic performance is good or not depends on his or her personal IQ and efforts on the one hand, and on whether the education method and education system are scientific and the teaching level of the teacher on the other hand.
On the other hand, this knowledge should be correct. Teaching students incorrect knowledge will be of no use in exams and practical applications.
Similarly, contaminated erroneous data cannot be used to train a usable AI large model. Using contaminated data to train a large model will result in the trained large model having almost no practicality.
The three complement each other and none of them can be missing.
Otherwise, how could school district housing be sold at such a high price?
Otherwise, why would tutoring classes be so expensive?
"Yuzu, use your Yuzu Technology account to upload the training framework's prerequisite technologies to GitHub in batches, in the order they were prepared, every three days. Use the Apache 2.0 license."
"Then I wrote three papers on multi-head attention mechanisms and published them on arXiv, also weekly."
"Also, we're looking for high-tech talent within the Great Zhou on Github, arXiv, and LinkedIn. The requirements are as follows..."
Fang Yu gave Youzi three clear instructions.
It’s time to find a technical team for Youzi Technology. Otherwise, no one would believe that a small company with only three employees could suddenly come up with a training framework and a mature AI model.
As a startup, how can you attract high-level technical talent?
It’s very simple, you just need to be a high-level technical talent first.
Genius has a clustering effect.
These things put on GitHub are bait.
Both the pomelo and orange models will definitely be hidden. Fang Yu plans to strip the orange model down to its most basic framework and then hand it over to these geniuses to fill in. If the filled model is not as efficient as the one made by pomelo, he will modify it himself.
In short, just keep your abilities at the level of a top genius and make sure that what you produce is not suspected by others.
In fact, the core members of a large model team and training architecture team are often not large in number, perhaps only a dozen or even a few people.
Therefore, Fang Yu only needs to recruit three to five algorithm scientists, five to ten engineers, three data processing personnel, and a dozen or so clerical staff to support this large model team.
The total number of people on the product side can be controlled within 30 people.
Moreover, on the product side, Fang Yu does not plan to hire any foreigners.
It’s not that Fang Yu is particularly nationalistic, it’s mainly out of consideration for confidentiality.
Since he is in the Great Zhou, he can deal with any accidents as soon as possible, but if he is abroad, it will be more troublesome.
If it were other companies, they might still have concerns that it would be difficult to recruit top talents in Dazhou.
But Youzi Technology does not need to worry about this. Fang Yu is looking for high-level talents, not top talents.
If it weren't for the practical issues, he alone, with a finance and operations team, could build the entire product side by relying on Youzi without the help of anyone else, and the efficiency would be even higher.
At that time, the only department that may need a large number of manpower is the AI alignment department. To put it bluntly, it is to align the ethics of AI with the ethics of human society.
This part of the staff cannot be laid off. We need full-time social science experts and a large number of testers to discover the ethical issues of AI through various strange conversations with AI and prevent them from happening.
No matter where you save, the auditors can't save.
However, this is all later.
Before that, Fang Yu had to find an HR for Youzi Technology.
Oh, no, I have to go to the basketball team and show off first.
I have tried my best to write this chapter in a simple and easy-to-understand way. I have revised it many times, but I still retained this part of the content.
Because there are so many things surrounding artificial intelligence in the future, we should first try to make everyone understand what the artificial intelligence model is, what the principles are, and how an artificial intelligence is born.
These are not the author's way of showing off or increasing the word count, but to illustrate that in real life, if the protagonist really comes up with a separate training framework and model framework, how can he release this model without arousing suspicion, and how can he maximize his own interests from a professional perspective.
In this way, the subsequent plot can be exciting.
(End of this chapter)
Continue read on readnovelmtl.com
