Google wants to optimize and rewrite code with AI. The tech giant is also said to be planning a major breakthrough in generative AI. But there’s also skepticism about code-writing AI.
In 2018, Google put artificial intelligence at the center of its operations and even rebranded its research department as Google AI. Since then, more and more AI applications are found in Google products, such as Pixel smartphones. According to a Business Insider report, Google will focus even more on generative AI in the future.
A fork to clean up the code
According to Business Insider, Google’s Pitchfork project is working on an AI system that learns to write, fix and update code on its own. In February, Google’s sister company Deepmind unveiled Alphacode, an AI system that aims to program at a human level, but is primarily intended as a support system for humans.
The AI software developed under the Pitchfork project, which is part of Google’s AI Developer Assistance Program, is intended to act more autonomously than currently known code AIs such as Alphacode or Copilot. from Microsoft.
Pitchfork will be able to learn, write, and rewrite code from scratch, learn different programming styles, and generate new code based on them, according to an internal project description.
Pitchfork began as a Moonshot project in Alphabet’s “X” research unit. Since the summer, the project has been located at Google Labs, which should underline its importance. At Google Labs, the company is banking on long-term technological transformations.
The code project is said to be part of a larger move by Google towards generative AI. Google recently made its own powerful image AI, Imagen, available in a test environment for some users. A wider deployment in the near future is possible.
Google also showed text-to-content search systems Imagen Video for AI-generated videos and Dreamfusion for 3D objects.
Towards a new era of code – or maybe not
Only time will tell if and to what extent code AIs catch on. Github CEO Thomas Dohmke, which sells Copilot, an AI code tool, predicts that up to 80% of code will be written by machines in five years.
Amjad Masad, founder of co-coding platform Replit, sees similarly drastic changes in the software market: AI is a 100x increase in productivity for developers, he said. Currently known AI tools only manage a 30-50% increase.
The next generation of code AI goes beyond auto-completion and will lead to “rapid change in the way we build software,” writes Masad.
Replit works on its own AI code that allows even non-programmers to create complex software. “Everyone in the world will at least be able to use John Carmack-level software,” Masad writes.
Put these things together, a developer will have the power of an entire network of AIs, people, and services at their fingertips. I believe a 100x increase in productivity is the lower limit here.
— Amjad Massad ⠕ (@amasad) November 23, 2022
Kite founder Adam Smith has a totally different perspective. Since 2014, his startup had been working on AI-powered programming tools with autocomplete. Now Kite is giving up, releasing its code as open source on Github.
Smith’s reasoning is interesting: his startup was at least ten years too early to market. “The technology isn’t ready yet,” Smith wrote in his farewell remarks on his company’s website. Smith explicitly includes Github Copilot in his review.
The biggest problem is that state-of-the-art models don’t understand code structure, like non-local context. We have made progress towards better code models, but the problem is very engineering intensive. It can cost upwards of $100 million to create a production-grade tool that can reliably synthesize code, and no one has tried that yet.
Smith, however, also admits mismanagement in Kite’s software monetization and product development that would have contributed significantly to his startup’s bankruptcy.
The question of copyright is also an open question in the context of code AIs. A lawsuit is currently pending in the United States against Github Copilot for possible copyright infringement involving sample code that was adopted during AI training.