This is a Story About Self-Reflection, Not AI Fraud

A few days ago, Matt Shumer, the founder of OthersideAI, announced that the company had made a breakthrough, which allowed them to train a mid-size model, achieving SOTA-level performance with the launch of Reflection, which outperforms GPT-4o and Claude Sonet 3.5.

But, the hype was short-lived, as many users started claiming otherwise, saying that the Reflection API was merely a wrapper of Claude 3.5 Sonnet and the answers on both the models were exactly the same. 

So, what exactly went wrong?

Artificial Analysis, known for its independent analysis of AI models and API providers, compared Reflection AI 70B to other models. It failed miserably, and the results were poor compared to Llama 3 70B.