On August 14, AI generative company Cosine announced the closure of a $2.5 million seed funding round led by UpHonest Capital and SOMA Capital, with participation from Lakestar, Focal, and others.

Cosine has developed an autonomous AI engineer named Genie, capable of code refactoring, feature development, and bug fixing. Alistair Pullen, Co-founder and CEO of Cosine, said, "Genie was initially trained to think and act like a human software engineer (SWE)."

"Cosine is not just providing AI capabilities; they are fundamentally teaching AI how to reason, providing companies with a true AI colleague," said Ellen Ma, Partner at lead investor UpHonest Capital.

Ben Tossell, founder of Ben's Bites, also praised the company, saying, "I've seen thousands of AI startups, but none have focused on human reasoning like Cosine. Genie proves their vision, strategy, and team are correct, bringing us closer to artificial general intelligence (AGI)."

Cosine plans to expand Genie's model portfolio in the future, including small models for simple tasks and larger models capable of handling more complex challenges.

01

Achieving the Highest Score

Autonomous AI Engineer Genie Proficient in 15 Programming Languages

Recently, Genie achieved the highest score of 30.08% on the SWE-Bench testing platform, significantly outperforming other tools. SWE-Bench is an evaluation framework specifically designed to assess whether language models can automatically solve real-world GitHub issues. This means Genie independently solved 30% of GitHub open-source project issues in the SWE-Bench coding benchmark test.

Genie's performance surpassed competitors like AWS's Amazon Q Developer and Cognition's Devin, which scored below 20% on the same benchmark. This result is better than the 13.8% previously demonstrated by autonomous AI engineer Devin and 22 times the performance of OpenAI's GPT4 (1.31%).

Image Source: Cosine

To achieve this, Cosine spent nearly a year compiling a dataset from real-world software development by engineers.

Currently, Genie can program in 15 languages, including JavaScript, Python, Java, C++, etc. The code generated by Genie will be stored in users' GitHub repositories, meaning Cosine does not retain copies or pose any related security risks.

Cosine's software platform has integrated Slack and system notifications, allowing it to alert users of its status, ask questions, or flag issues, similar to an excellent human colleague. It can even respond to questions raised by colleagues.

Technical Report Link: https://cosine.sh/blog/genie-technical-report

02

AI as a Colleague

Cosine Receives Support from OpenAI Resources

Currently, Cosine plans to offer two pricing models for Genie at its early stages.

The first model offers Genie for approximately $20 but with limited functionality and usage, targeted at individuals and small teams.

The second model positions Genie as an enterprise product with nearly unlimited usage, providing a perfect AI colleague for businesses at a higher price point.

Cosine hopes Genie will transform how engineering resources are allocated, enabling teams to focus on more strategic initiatives. "The value of an AI colleague capable of diving into unknown codebases and solving unknown issues orders of magnitude faster than humans is undeniable and has a significant global impact," said Pullen.

Image Source: Cosine

"We're transforming the way developers work. We've invested a fraction of the time and money compared to similar products, yet our product outperforms OpenAI and others in complex software tasks," said Yang Li, COO of Cosine.

Cosine's CIO Sam Stenner also mentioned that the team views Genie as a colleague rather than an assistant. "We understand how to generate datasets encoding human reasoning and use them to train large language models," he said. "We'll collaborate with OpenAI's fine-tuning team to access their long context window capabilities. We believe we can continuously surpass our best results in the future."

Currently, Genie is available to select users, but general access requires an application (Application Link: https://cosine.sh/register). Cosine plans to regularly update Genie's features based on customer feedback.

"The latest SWE-Bench submission requirements mention disclosing the complete workflow of AI models, which could be a challenge for us. Currently, we keep these internal processes confidential but plan to make Genie's achievements public and independently verifiable on GitHub," Pullen mentioned in his blog.

03

Software Engineering is Just the Most Intuitive Starting Point

Founded in 2022, Cosine was selected for the Y Combinator accelerator. Cosine defines itself as a lab for human reasoning, focused on enabling large language models to mimic human software engineers' behavior to perform complex coding tasks.

Cosine aims to create truly resilient AI engineers capable of solving problems across various domains. "We've been chasing a dream of creating a highly reliable AI colleague that can automatically execute end-to-end programming tasks without intervention. Genie is the first step towards realizing this goal," said Pullen.

Currently, the company has three co-founders: CEO Alistair Pullen, CIO Sam Stenner, and COO Yang Li.

From left to right: Alistair Pullen, Sam Stenner, and Yang Li | Image Source: UTKN

Alistair Pullen has been interested in programming since childhood, releasing and commercializing his first software application at age 9.

Another co-founder, Yang Li, graduated from Oxford University's Department of Sociology and experienced one IPO, two acquisitions, and the growth of three unicorns during his six-year career.

Currently in its early stages, Cosine has a team of only five members with offices in London and San Francisco. They are also seeking AI developers to join their team and create the world's most human-like autonomous AI software developer (Recruitment Team: https://app.dover.com/jobs/cosine).

"We firmly believe we can bring human reasoning to any job and industry," Alistair Pullen stated in a blog post. "Software engineering is just the most intuitive starting point. We can't wait to show you everything else we're working on."