AlphaCode’s ability to solve computer-science problems impressed artificial intelligence (AI) researchers. DeepMind, a Google company, is an AI powerhouse based in London. They released this tool in February. Science1 has published its results, showing that AlphaCode can beat half of the humans in code competitions.
ChatGPT, a chatbot that sometimes produces meaningful-sounding but sometimes ridiculous mini-essays upon request, has captivated social media users over the past week. This includes writing code, fixing bugs and explaining them, etc. However, these AIs cannot perform many tasks, and researchers claim they cannot replace human programmers.
These AI puppets are based on neural networks, which learn to complete a task by processing large amounts of human-generated text.
According to Zico Kolter (a Carnegie Mellon University computer scientist in Pittsburgh, Pennsylvania), AlphaCode And ChatGPT have “virtually identical architecture.” While there are minor differences in execution and training, the most significant difference is that the systems are trained on different data sets and are therefore suited for different tasks.
ChatGPT is a conversation engine that can be used for all purposes, but AlphaCode is more specific. It was designed to answer questions in software-writing contests.
David Choi, DeepMind’s research engineer and co-author of Science’s paper, stated that AlphaCode was explicitly designed and trained for competitive programming, not software engineering.
Research has shown that a lot of the work involved in large-scale software engineering projects, such as designing a web browser, involves understanding the needs and preferences of users. These are challenging to explain with simple, machine-readable specs that an AI can use to create code.
Kolter states that it is unclear whether machines can ever create large-scale software systems from scratch. He says that “my best guess” is that tools that generate large-scale programs will become second-nature tools for programmers.
Choi states, “We hope that further research can result in tools to improve programmer productivity and get us closer to problem-solving AI.”
Kolter also said that AI tools are already capable of making programmers’ lives easier. One example is Copilot, an AI tool for code completion launched last year by GitHub.