Codeium has announced a new AI assistant called Cortex that uses complex reasoning capabilities to solve difficult coding problems.
Rather than enabling smaller, individual tasks such as autocomplete or code completion like existing code assistants, Cortex supports large scale reasoning, code generation, reviews, and knowledge transfer, with greater accuracy, lower latency, and reduced costs.
According to Anshul Ramachandran, head of enterprise and partnerships at Codeium, most other AI assistants that generate code do so by leveraging RAG to pull in information on a codebase, which relies on putting information into an embedding space where it can be retrieved later. What sets Cortex apart is that it can retrieve what it needs from raw information, which results in fewer false positives than a RAG approach.
“What if we trained a foundation model that can take the query, the task at hand, and take the raw text, not an embedding version of the text, and ask that question, ‘Is this relevant?’ And then actually utilize that in this kind of multi-step reasoning process in order to get much better results,” he said.
According to Ramachandran, they’ve tested this approach on hundreds of millions of lines of code without any degradation in quality.
He also explained that what makes Cortex special is that they’ve solved latency issues so that results are near instantaneous while still being accurate. The company claims that Cortex can provide its responses in seconds.
“If your tool takes 30 minutes to get your response and it’s right 80% of the time, it’s probably something a developer won’t trust because if you take that long to generate something, it better be super accurate. Otherwise, people are like, ‘I’ll just do this myself,’ Ramachandran said.
He also said that despite the advanced reasoning functionality, Cortex is not being sold as something that can fully replace developers: The goal is to be able to handoff tasks so that developers focus on other things.
“For us, the human is always in the driver’s seat. Even with things like Cortex, with all the modalities we have, a human has to review what’s coming,” he said. “We want you to take a look, accept, reject, and think about the code critically. We always tell our customers, don’t replace your SAST tools, or any of those things, because AI is meant to just be an enhancer, right? It does not replace anything.”
Cortex is available today for customers to start experimenting with. It integrates with the rest of Codeium’s offerings, including Autocomplete and Chat, and the company plans to integrate it into more of its products in the future.
You may also like…
Anthropic introduces prompt caching to reduce latency and costs
Google begins rolling out voice capabilities in Gemini with Gemini Live