Kite, an early entrant into the code autocompletion developer tool space, has decided to call it a day, saying there is still a long way to go for the technology because the state of the art for ML on code is just not good enough yet.
"We failed to deliver our vision of AI-assisted programming because we were 10+ years too early to market, i.e. the tech is not ready yet," he wrote. "We failed to build a business because our product did not monetize, and it took too long to figure that out."
Smith also suggests Kite could have, maybe, solved the issue of synthesizing code reliably, but that would have required $100 million and more engineers.
"The largest issue is that state-of-the-art models don't understand the structure of code, such as non-local context. We made some progress towards better models for code, but the problem is very engineering intensive. It may cost over $100 million to build a production-quality tool capable of synthesizing code reliably, and nobody has tried that quite yet," said Smith.
Kite showed promise as an alternative to Microsoft's code-completion tool IntelliSense, and in 2019, when it still only supported code completion in Python, it raised $19 million in a Series A round led by Trinity Ventures, with personal participation by then-new GitHub CEO Nat Friedman, as TechCrunch reported at the time.
But by 2020, with new competitors like TabNine using OpenAI's GPT-2 large language model (LLM) for language-agnostic code completion, Kite opted to redesign its product using GPT-2 to support autocomplete for 11 more languages. Eventually, Kite would also build an integration with Microsoft's popular VS Code editor.
Kite has open sourced most of its code on Github here. It includes Kite's data-driven Python type inference engine, Python public-package analyzer, desktop software, editor integrations, Github crawler and analyzer, and more.
However, Smith said the state-of-the-art ML is still not good enough.
"We built the most-advanced AI for helping developers at the time, but it fell short of the 10× improvement required to break through because the state of the art for ML on code is not good enough. You can see this in Github Copilot, which is built by Github in collaboration with Open AI. As of late 2022, Copilot shows a lot of promise but still has a long way to go," he wrote.
However, Smith said that the future is still bright for these tools. "We can't wait for AI to revolutionize programming. Computers are so magical; it will be amazing to experience a step-function increase in what they can do for us," he said.