Faced with that reality, some startups in Silicon Valley are looking for ways to develop technology to verify the accuracy of code created by AI.
AI chatbots such as ChatGPT or current programming support systems can create computer code in just a few seconds. However, experts say these systems sometimes still make mistakes or even create inaccurate information.
A study published in January by researchers at Carnegie Mellon University (USA) shows that AI code-generating tools can help speed up software development. However, they can also reduce code quality, causing projects to have long-term problems.
Faced with this challenge, many startups in Silicon Valley are developing technology to verify computer code.
Technology startups in the field of artificial intelligence such as Axiom Math and Harmonic (in Palo Alto, California) or Logical Intelligence (San Francisco) are building AI systems capable of verifying codes in a similar way to mathematicians proving a problem.
Carina Hong, founder and CEO of Axiom Math, believes that code verification could become the next step in the AI technology development process.
Recently, this company announced the successful raising of 200 million USD from venture capital funds, including Menlo Ventures, Greycroft and Madrona.
Although it was only established about a year ago and only had about 20 employees, Axiom Math was valued at 1.6 billion USD. Investors believe that the company's technology can help improve the quality of code created by AI systems such as Codex or Claude Code.
According to Matt Kraning (a venture capitalist in the technology field and a partner at Menlo Ventures), the biggest problem when using AI to write code today is that users cannot be sure if the code is faulty or not.
Mr. Matt Kraning believes that code verification technology like Axiom's can help partially solve this problem.
Initially, Axiom developed technology to solve complex mathematical problems. The company's system named AxiomProver achieved absolute scores in the Putnam Math exam (one of the best university exams in the US and Canada).
This technology works by using Lean programming language, which was developed more than a decade ago to demonstrate mathematical propositions.
Thanks to the ability to clearly identify right and wrong in mathematics, the system can eliminate logical errors in the problem solving process.
Researchers hope that the same method can be applied to verify computer code. This is an example of "transfer learning", when an AI system learns skills in one field and successfully applies them to another task.
However, some experts warn that this method still has limitations. According to Bogdan Vasilescu, professor of computer science at Carnegie Mellon University, it is not always possible to clearly define what a "true" computer code is.
In reality, many software, especially online services serving millions of users, will have a complex and unpredictable operating environment.
Therefore, although AI can help check for some errors, this technology can hardly completely eliminate all problems in computer code.
