AI system content recommendation issue 24

1、DeepMind Publish comparable to ordinary programmers AlphaCode

 

DeepMind And OpenAI Two well-known AI Research institutions publish important research results respectively ：DeepMind Released based on Transformer Model AlphaCode, Can write computer programs comparable to human beings ; meanwhile ,OpenAI The developed neural theorem prover has successfully solved two International Mathematical Olympiad problems .

 

link ：https://mp.weixin.qq.com/s/evIZTPGHDJlo8Xee20j56A

 

2、 Ali BladeDISC The deep learning compiler is officially open source

 

With the continuous development of deep learning ,AI The model structure is evolving rapidly , The underlying computing hardware technology is emerging in endlessly , For the majority of developers, we should not only consider how to effectively exert our computing power in complex and changeable scenes , Also deal with the continuous iteration of the computing framework . Deep compiler has become the technical direction to deal with the above problems , Let users only focus on the development of the upper model , Reduce human development costs for manually optimizing performance , Further squeeze hardware performance space . Alibaba cloud machine learning PAI Open source has the dynamic trend of putting into practical business applications earlier in the industry shape Deep learning compiler BladeDISC, This article will explain in detail BladeDISC Design principle and application of .

 

link ：https://zhuanlan.zhihu.com/p/462641670

 

3、TensorFlow 2.8.0 The official launch

 

In recent days, TensorFlow It's official 2.8.0 Official version , The last time distance  2.7 edition It hasn't been long since the update of , The new version provides more bug Repair and function improvement , In addition, the new version has also released patches for vulnerabilities .

 

link ：https://mp.weixin.qq.com/s/FwrqPED66Nd-a2akMtrJBA

 

4、AI Thinking and Prospect of system full stack technology

 

This paper discusses how to think about the whole stack of software and hardware AI The necessity of system technology , Included in performance , Ease of use and TCO The benefits of . Combined with some specific examples , Hardware and software are discussed co-design Yes AI Potential of system performance improvement . Through some actual use cases , Discussed AI The interaction between different modules of the software stack , It highlights the core point of the whole stack design of the system . And MLPerf For example , The relationship and contradiction between extreme performance optimization for specific computing evaluation benchmark and automatic optimization for user out of the box are discussed . Beyond these discussions , We also discussed the impact of organizational structure design at the company and industry levels on AI The impact of the evolution of system technology architecture , as well as AI Core software stack and MLOps The connection of .

 

link ：https://zhuanlan.zhihu.com/p/462810001

 

5、 Dedicated architecture and AI Software stack （5）

 

With auto tensorize As the core of decoupling new ideas , By solving this problem DSA Decoupling of hardware and upper application , The final desired scenario is that the front-end user writes serial scalar code to describe business logic , Back end operator developers / Chip manufacturers write operators / Function description of the instruction （ Serial scalar form ） The actual implementation of the and operator / The actual call mode of the instruction is two parts , What the compiler wants to solve is to try to use operators / The function description form of the instruction is the target , Find the equivalent transformation of user code , So as to achieve front end 、 Back end and compiler The decoupling of .

 

link ：https://zhuanlan.zhihu.com/p/458544969

 

6、Ion Stoica： Make it Spark and Ray Secrets of two star projects

 

Spark and Ray It is not only an open source project with great influence in the industry , They have developed into commercially successful companies based on open source projects . They must have done a series of difficult and correct things to get today , What the hell did they do right ？ In the machine learning program 《Gradient Dissent》 in , host Lukas Biewald And Ion Stoica Conducted an in-depth interview , From this, we can learn from the first-hand information about the initiation of Spark and Ray、 Start up companies 、 Pay attention to open source 、 How to make a series of key decisions to embrace the cloud . Through this article , I hope my friends can find Spark and Ray The secret of success .

 

link ：https://mp.weixin.qq.com/s/pfl5Kp29ECWsxsYAV-PR9g

 

7、 How does the deep learning framework do operator alignment tasks gracefully ？

 

This article introduces OneFlow The operator of AutoTest frame , It provides a deep learning method to do operator alignment gracefully , So that developers and users can write PyTorch That makes it easy to write test programs .

 

link ：https://mp.weixin.qq.com/s/Vb9zHOwnPn-19ZF-sKC3hg

 

8、PyTorch Analysis of video memory mechanism

 

Two years on , There is no good article to summarize PyTorch All aspects of the video memory mechanism , This time for PyTorch Make a small summary of the video memory mechanism of .

 

link ：https://zhuanlan.zhihu.com/p/424512257