Zencoder has hired a bunch of search engine veterans to help it build a tool that can analyze large codebases and figure out what is and isn’t relevant. This detailed context reduces hallucinations and improves the quality of code that large language models can produce, says Filev: “We call it repo grokking.”
Cosine also thinks context is key. But it draws on that context to create a new kind of data set. The company has asked dozens of coders to record what they were doing as they worked through hundreds of different programming tasks. “We asked them to write down everything,” says Pullen: “Why did you open that file? Why did you scroll halfway through? Why did you close it?” They also asked coders to annotate finished pieces of code, marking up sections that would have required knowledge of other pieces of code or specific documentation to write.
Cosine then takes all that information and generates a large synthetic data set that maps the typical steps coders take, and the sources of information they draw on, to finished pieces of code. They use this data set to train a model to figure out what breadcrumb trail it might need to follow to produce a particular program, and then how to follow it.
Poolside, based in San Francisco, is also creating a synthetic data set that captures the process of coding, but it leans more on a technique called RLCE—reinforcement learning from code execution. (Cosine uses this too, but to a lesser degree.)
RLCE is analogous to the technique used to make chatbots like ChatGPT slick conversationalists, known as RLHF—reinforcement learning from human feedback. With RLHF, a model is trained to produce text that’s more like the kind human testers say they favor. With RLCE, a model is trained to produce code that’s more like the kind that does what it is supposed to do when it is run (or executed).
Gaming the system
Cosine and Poolside both say they are inspired by the approach DeepMind took with its game-playing model AlphaZero. AlphaZero was given the steps it could take—the moves in a game—and then left to play against itself over and over again, figuring out via trial and error what sequence of moves were winning moves and which were not.
“They let it explore moves at every possible turn, simulate as many games as you can throw compute at—that led all the way to beating Lee Sedol,” says Pengming Wang, a founding scientist at Poolside, referring to the Korean Go grandmaster that AlphaZero beat in 2016. Before Poolside, Wang worked at Google DeepMind on applications of AlphaZero beyond board games, including FunSearch, a version trained to solve advanced math problems.
When that AlphaZero approach is applied to coding, the steps involved in producing a piece of code—the breadcrumbs—become the available moves in a game, and a correct program becomes winning that game. Left to play by itself, a model can improve far faster than a human could. “A human coder tries and fails one failure at a time,” says Kant. “Models can try things 100 times at once.”
