A few months ago, AIM observed many Indian founders making their way to the world’s largest startup accelerator Y Combinator. Many have since secured backing from prominent investors to scale their ideas.
One such founder is Cyril Gorlla, a 23-year-old prodigy from Hyderabad who mastered coding at the age of 11. He has now built CTGT, an AI startup that has attracted funds from prominent investors, such as Mark Cuban and Mike Knoop (co-founder of Zapier).
Gorlla and Trevor Tuttle founded CTGT, a name interspersed with their initials. The company was selected for Y-Combinator’s Fall 2024 batch and also made it into TechCrunch Disrupt’s Startup Battlefield Top 20. It aims to bring greater efficiency and interpretability to AI, with its model already being piloted by Fortune 10 companies.
The startup claims to eliminate AI hallucinations and make ML nearly 10x more efficient. “CTGT’s vision is to make AI more transparent and accessible without sacrificing on performance,” said Gorlla in an exclusive interaction with AIM.
Its platform integrates with both open-source local models and API-based models, allowing enterprise clients to train customised, high-performance AI systems that deploy 10 times faster than traditional models.
Co-founders Cyril Gorlla and Trevor Tuttle
CTGT and Hallucinations
AI hallucinations have been a huge concern for everyone, including big-tech companies like Google. Recently, answer engine Perplexity got into trouble for hallucinating fake news and citing them under real publications. Dow Jones and the New York Post filed a copyright lawsuit against the company.
Interestingly, Microsoft has filed a patent for a method that claims to reduce and eventually eliminate hallucinations.
Gorlla believes that existing solutions for LLM interpretability involve training hundreds of other models to identify the concepts in a model and then modify each concept individually.
Since hallucinations are a universal concern for AI models, many research papers have attempted to address the problem. CTGT addresses the issue by using a novel platform that bypasses traditional deep learning techniques, which often require massive compute resources and extensive fine-tuning to mitigate inaccuracies.
“These models often aren’t suited for the business needs of enterprises, especially those in critical applications like finance and healthcare. The ‘last mile’ layer to ensure that these models are at a level of trustworthiness and reliability with regard to hallucinations and brand alignment doesn’t exist – that’s what we provide,” said Gorlla.
However, the platform instead examines the internal structures of models directly, enabling companies to “steer” AI behaviour without introducing additional layers of complexity.
Surpasses Benchmark
Gorlla highlighted that many existing methods are computationally inefficient, “with one leading foundation model provider’s state-of-the-art LLM interpretability taking more compute to use than the foundation model itself”, rendering such methods inaccessible to most companies.
“By focusing on understanding the foundational mechanisms of learning, we’re building models from the ground up to be both efficient and interpretable,” said Gorlla.
CTGT’s AI models have been evaluated on several metrics. In a benchmark of 121 classification datasets, neural nets took five hours to train, whereas CTGT’s method took only 40 minutes.
“In a benchmark of large (>50,000 training samples) classification and regression datasets, our method provides accuracy ≥ current transformer, tree-based, and MLP models,” said Gorlla. He also mentioned that CTGT’s method used 3,600 compute hours in total, while other methods used 20,000 hours for tuning.
Gorlla said that the next version of their training algorithm will be 500 times faster than the existing one.
What’s the Plan?
With a $500k YC deal, $125k from Character Labs, and an undisclosed amount from Cuban, CTGT is using its funds to refine its technology stack, expand research, and grow its customer base.
“Mark thinks interpretability and the “black box” of AI is a huge issue. We spoke to him about what we were doing and he was immediately on board with our vision for truly intelligent AI and that the current status quo of deep learning will not lead to true AI,” said Gorlla.
His passion for coding and building started at a very young age. “Growing up as an Indian-origin immigrant to the US, I taught myself how to code and aced my mother’s community college programming course at 11 amidst periodically disconnected household utilities,” he said.
It was in high school that he saw AI’s potential to drive global change—and recognised the importance of ensuring that everyone had access to the resources to benefit from it.
The post How This Hyderabadi Founder Won Over Mark Cuban To Solve AI Hallucination appeared first on Analytics India Magazine.
