OpenAI, in Duel With Anthropic, Doubles Down on AI That Writes Software
The Takeaway
• Anthropic’s AI recently showed it had an edge in coding over OpenAI’s by some measures
• OpenAI’s coding products could threaten GitHub Copilot, Cursor and other startups
• OpenAI has developed a tool to help its AI researchers, including by generating code
OpenAI’s ChatGPT has become a multibillion-dollar business in large part because programmers use it to write and check their code, fix bugs and translate code into different programming languages.
Now, facing competition from rival artificial intelligence startup Anthropic, OpenAI is putting more effort into improving the tools it offers for software programming. Some products or features under development aim to make it easier to use OpenAI’s AI for coding tasks inside major code-editing programs like Microsoft’s Visual Studio Code, while others aim to take on bigger software development tasks.
For instance, OpenAI has worked on a product to handle software engineering tasks that might take a human hours or days, and to automatically write and execute code for complex applications based on customers’ instructions, said a person who has heard from OpenAI leaders about it.
It isn’t clear when OpenAI would launch such a product.
Coding tasks became an early application of large language models developed by OpenAI, in part because AI-generated code can quickly be tested to see if it works or not. Microsoft’s GitHub unit used OpenAI’s LLMs to power an AI Copilot product, starting in 2021, that gives code suggestions to programmers while they type.
But the launch of ChatGPT in late 2022 provided a more widely accessible and free alternative that quickly gained popularity. OpenAI then convinced millions of programmers to pay for an upgraded version of ChatGPT—and get access to upgraded LLMs before GitHub Copilot did—that could respond to conversational requests for handling such tasks. Those capabilities powered what is now a subscription product on pace to generate about $3 billion annually.
AI firms also believe models that can generate good code also can generate better answers to other types of queries.
Now OpenAI is somewhat on its back foot in AI coding because rival AI startup Anthropic has become a formidable competitor. On Tuesday, Anthropic announced new software that can use computers the way humans do to take actions on behalf of people, such as moving a cursor, clicking buttons and typing text.
The Anthropic software can theoretically help programmers with tasks like building a website and improving the way it looks. OpenAI has been developing a similar product, known as a computer-using agent, for months but hasn’t launched it.
By some metrics OpenAI uses to evaluate models’ coding abilities, Anthropic’s models recently had an edge, according to a person who has talked to senior OpenAI leaders about the issue.
Anthropic co-founder Daniela Amodei recently told The Information that AI coding can’t fully replace engineers yet and needs “some coaching,” but its Claude chatbot has boosted its developers’ productivity to the point where Anthropic might alter its hiring plans. Anthropic has projected that by the end of 2024, it would generate revenue at a pace of $1 billion annually, or $83 million per month. That’s 25% higher than its original projection at the start of the year, though Anthropic shares a bigger portion of its revenue with sales and cloud partners like Amazon Web Services than OpenAI does with its cloud partner Microsoft.
Internal Research Assistant
OpenAI’s internal tools show the broad potential of its AI for coding-related tasks. The company has developed an internal research assistant tool to help AI researchers work faster, including by generating code for their experiments related to AI models, according to an employee. The product is popular internally, this person said, but it’s unclear whether the company would ever release it publicly.
The internal tool appears to be a step toward the development of AI that can conduct AI research on its own—a skill that requires far more than coding, including generating ideas for new experiments, and that OpenAI leaders have publicly said was likely to happen in the coming years.
A spokesperson for OpenAI declined to comment.
OpenAI employees have also considered developing a version of the company’s Canvas product—a new ChatGPT interface making it easier for users to draft documents and code—that can integrate with developers’ code repositories, according to a person who has spoken to OpenAI managers about the idea.
Such a feature would provide the AI with a better understanding of a user’s existing code base and the types of projects they’re working on, but users would likely still need to undergo the arduous task of copying and pasting their code in and out of the chatbot.
Eventually, OpenAI hopes to offer more general-purpose computer-using agents that can control customers’ computers to handle a broader range of tasks beyond writing code, similar to the one Anthropic announced on Tuesday. OpenAI has internally demonstrated a preliminary version of an agent that would use the user’s computer to do tasks like order food for delivery, according to a person who has seen the demonstration. An autonomous agent tasked with a coding or app development assignment might be able to consult web-based sources for help if it runs into issues.
Competing With Customers
The OpenAI products under development could threaten customers of its application programming interface business who sell coding assistants. Among them is Cursor, which uses OpenAI models for its product and has taken venture capital from OpenAI.
OpenAI’s future coding products may overlap with open-source coding assistants such as Continue or SWE-agent, which use LLMs to help users write code and fix bugs.
The new products also could compete with GitHub Copilot. Microsoft’s financial disclosures in July implied that GitHub Copilot was on track to generate $300 million annually from subscriptions. The company has used the tool’s popularity to sell customers on its other AI products, such as its suite of Copilot features for Office 365 apps like Word, Outlook, Excel and PowerPoint. In cases where customers buy multiple products, Microsoft has discounted the price of GitHub Copilot and other products.
Spokespeople for Microsoft and Cursor didn’t immediately respond to a request for comment.