Despite the massive layoffs taking place in the technology sector, Open AI commitment to the recruitment of personnel for the creation of new artificial intelligence models that allow them to improve their products, among them, ChatGPT. The objective of the company co-founded by Elon Musk would be incorporate 1,000 new workers in different countries around the world to work on creating improved code in their AI tools.
Of those thousand new employees, about 600 (40%) are programmers whose work will focus on AI performing software engineering tasks. The remaining 60% will focus on the ‘data labeling’, that is, in the creation of raw data (text, audio, images and videos) that will then be labeled to specify the context in which the AI is developed. The profile currently most in demand by OpenAI would be that of python developera highly qualified programming language.
This study, included in a Semafor report, shows that new hires will have to create training data that will encompass from lines of code to explanations of it in natural language. OpenAI has a tool with these characteristics launched in 2021, OpenAI Codexwhich received training on data pulled from GitHub (Microsoft’s code repository). Similarly, Codex is used by Microsoft to power GitHub Copilot, a service that will help programmers write code.
Open AI Codex masters more than a dozen programming languages and is capable of interpreting simple commands in natural language and creating an interface for existing applications. Regardless, coding could improve ChatGPT’s ability to converse with humans and thus overcome barriers such as those of the Stack Overflow website, which accuses the OpenAI AI tool of not offering reliable answers.
The key to being able to establish improvements in a coding assistant will be to build a system capable of anticipating the effect of its own actions and set up an internal system. Meta’s chief artificial intelligence scientist, Yann LeCunclaimed that there was a great similarity between cruise control systems in cars and coding assistants like Copilot. And it is that if the human does not establish a constant supervision on Copilot, it will end up making almost negligible code errors. The problem resides when we are dealing with a code model that is sometimes executed and sometimes not.
The OpenAI Selection Process
For the election of qualified personnel regarding the programming and control of ChatGPT, OpenAI subjected the candidates to a five hour unpaid coding test having to undertake different tasks. One of them was to explain in writing how they would solve a coding problem and, in the event of detecting errors in the process, explain what they were and how they should be corrected. In the end, the objective is to feed the tool with very specific training data that the human himself marks for it.
