The English writer Percy Bysshe Shelley used to bang on about poets being “the unacknowledged legislators of the world.” More recently, and more prosaically, the Harvard scholar Lawrence Lessig has argued that computer code is the new law.
But computers, as well as humans, are now starting to write that code. Will that make algorithms the unacknowledged legislators of the world?
For the moment, the question is premature. But the rapid advances of OpenAI’s Codex code-writing program, released by the San Francisco-based research company over the summer, highlights the direction of travel. An offshoot of OpenAI’s powerful GPT-3 text generation model, Codex can instantly translate natural language into computer code in 12 different computer languages — and can even translate between them.
GitHub, the open source software resource owned by Microsoft, is already using the technology to power Copilot, a software tool that suggests the next line of code, like an auto-complete function in a search engine. When Codex emerges from its current beta testing, OpenAI argues it will help democratise access to smart software.
“By removing the barriers to entry, everyone is going to do more,” says Greg Brockman, OpenAI’s co-founder and chief technology officer. But he suggests there is still a long way to go to improve Codex’s current limitations and ensure it is used appropriately. Although startling in its speed and flexibility, Codex currently generates the “right” code in 37 per cent of use cases, meaning it still requires close supervision. “Codex is a point on the journey but it is not the end point,” Brockman says.
One of the areas in which Codex is likely to be used first is in digital education. Sana Labs, a Swedish start-up run by Joel Hellermark, is exploring how Codex can help provide more personalised and dynamic learning courses, teaching users, among other things, the principles of coding. Hellermark’s view is that Codex will initially be used by experienced software developers as a tool to write faster code. But over time it will empower non-technical users to do all kinds of things they currently deem impossible.
“You will still need to have an intuition of how software works. But it will put more value on computational thinking rather than the exact syntax of a particular computer language,” Hellermark says. “Intelligence is becoming increasingly commoditised.”
In a launch paper published in July, OpenAI researchers trumpeted Codex’s technological capabilities but also flagged many of the concerns about automated code-writing programs. As with GPT-3 itself, the data sets and code resources used to train the Codex model are partial and imperfect and need to be scrutinised for errors and baked-in biases. “Codex can generate code with structure that reflects stereotypes about gender, race, emotion, class,” the paper warns.
As with all software, there is also a significant risk of misalignment, when programmers fail to specify the desired outputs clearly enough. Think of the Disney film Fantasia in which Mickey Mouse as the sorcerer’s apprentice misuses magic with disastrous effects. By replicating buggy software, Codex might inject security vulnerabilities into systems. It could also be used to create malware or empower phishing attacks. Aware of the dangers, OpenAI says it is already focused on ways of mitigating such potential harms.
Codex also forms part of a far broader transformation of software, a cognitive revolution that might be comparable in its impact with the Industrial Revolution, according to Kevin Scott, chief technology officer of Microsoft, which has invested in OpenAI. This Software 2.0 revolution, as Scott calls it, will help programmers move from an era of artisanal handwork into one of mass production. “It changes the process of coding from you very explicitly telling a machine to accomplish a task in terms that are convenient for the machine to you teaching a computer how to accomplish a task in terms that are convenient for a human,” Scott says.
The possibilities for enhancing human creativity and productivity are enticing. Like Shelley’s conception of poetry, software can awaken and enlarge the mind and help “lift the veil from the hidden beauty of the world.” But the recent history of flawed technological rollouts suggests we should make haste slowly and ensure that humans remain the acknowledged legislators of the world. “Human oversight and vigilance is required for safe use of code generation systems like Codex,” as the academic launch paper rightly concludes.