Table Of Contents
It was announced by “Cognition,” a newly established company in the field of artificial intelligence, backed by the Founders Fund owned by Peter Thiel and prominent tech figures including Elad Gil, former CEO of Twitter, and Tony Shue, co-founder of “Dordash,” about the launch of a completely independent AI program called “Devine.”
Despite the availability of many software assistants in the market, including the brilliant assistant GitHub Copilot, Devine stands out in managing programming projects comprehensively from start to finish – from code writing to error correction to final execution. This service, the first of its kind, can even handle projects on platforms like Upwork, as demonstrated by the startup company.
Devine’s announcement signifies a significant shift in AI-supported development as it equips developers with a fully integrated AI tool to use in their projects, going beyond just writing initial code or suggesting code snippets.
However, Devine is still not open to the public. The company has granted the opportunity to use it to a limited group of clients, including Bloomberg journalist Ashley Vance, who shared details of his experience in this regard.
Devine’s Mission:
An article on Cognition’s website explained by Scott Wu, founder and CEO of Cognition and a decorated figure in the sports software field, stated that Devine is capable of leveraging well-known programming tools like command-line interface, code editor, and browser within a specialized computer environment to design and implement complex engineering tasks that require thousands of decisions.
The human user formulates commands in a conversational language within Devine’s platform. Then, the AI engineer takes responsibility for preparing a detailed execution plan to deal with the presented case. Subsequently, it proceeds to apply the project using its development tools in the same manner as humans do – writing code, solving problems, conducting tests, and providing instant reports on progress, allowing the user to monitor all aspects of the process during execution.
If the human user notices an error, they can go to the conversation platform and provide instructions for the AI to make necessary adjustments. Cognition explains that this process enables engineering teams to delegate some of their tasks to AI, allowing them to focus on more creative tasks that require higher human cognitive abilities.
In this manner, Devine introduces a modern model that may provide us with a futuristic glimpse of how computer programs – and computer-based work in general – might evolve in the near future, through the use of AI software agents under human supervision and management.
Devine: Capable of Efficiently Handling Various Diverse Development Tasks
As explained by Wu, Devine has the ability to efficiently perform various types of tasks in its current state. This includes engaging in typical engineering projects like app launches and website enhancements from A to Z, finding and resolving errors in databases, and even tackling more complex tasks such as making modifications.
Take a minute to delve into big language database rules by utilizing links to research repositories on GitHub or exploring innovative techniques that may be unconventional.
In a specific scenario, he benefited from an article on a blog to learn how to execute code to create images that include hidden messages. Meanwhile, in another situation, he worked on a project through the Hub platform related to running a machine learning model by programming codes and fixing resulting programming errors.
In the well-known “SWE-bench” test that evaluates the performance of AI assistants by analyzing problems extracted from the “GitHub” platform linked to open-source software projects, a software engineer working with AI succeeded in addressing 13.86% of the cases accurately from A to Z without the need for human intervention. On the other hand, the “Cloud 2” system only managed to solve 4.80% of these cases, while the “SWE-Llama-13b” and “GPT-4” systems were able to handle 3.97% and 1.74% of these issues, respectively. However, all these systems ultimately required some support, as they were guided to the specific file that needed to be fixed.
The underlying technology remains undescribed
The use of artificial intelligence in software development is not new; various tools have been used in this context for a long time. These tools include many well-known programs, such as GitHub Copilot and StarCoder, extending to include Replit, which features a library of smart code examples within the Hugging Face. Moreover, Codeium, which recently received $65 million in funding in a Series B round, estimated to reach $500 million.
However, most of these offerings have primarily focused on using artificial intelligence to assist in programming. This AI can generate basic code based on textual instructions, tailor it to fit the context of the integrated development environment, or retrieve pieces of code, speeding up team workflow. It appears that the AI company utilizing “Devin” is taking a step (or several steps) forward by providing us with an all-encompassing AI to manage projects entirely.
Although the tool still needs further testing, its ability to handle multiple steps and keep on the right path to complete software engineering projects stands out. The company “Cognitio” did not disclose how it achieved this feat, whether by using its proprietary model or relying on another model, but it hinted that the work is a result of its “long-term progress in inference and planning.”
The company is currently expanding its capabilities and providing early access to the “Devin” service for a limited group of users. Interested parties looking to enhance their engineering work can contact via email to join the access list. It is expected that access to the service will become more widely available in the next stage.
Cognitio also mentions on its website that programming is just a starting point, hinting at the potential to leverage it for technical advancement in AI fields or employ it in various other professional domains. So far, the company has managed to raise funding of up to $21 million.
