Recently, HyperWrite co-founder and CEO Matt Shumer announced the launch of Reflection 70B, a large-scale language model (LLM) based on the Meta open-source Llama 3.1-70B Instruct. Shumer stated that Reflection 70B utilizes new error self-correction technology and has performed exceptionally well in third-party benchmark tests.
Shumer referred to Reflection 70B as the "world's top open-source AI model." However, subsequent third-party evaluations failed to reproduce the claimed performance metrics, leading to accusations of fraud against Shumer on social media.
The distinguishing feature of this model is its reflection tuning technology, which allows the model to detect and correct errors before generating text. This application of technology has enabled Reflection 70B to excel in certain benchmark tests, particularly in MMLU and HumanEval testing.
To verify the performance of Reflection 70B, the public can try a demo version on the "Playground" website. However, due to increased traffic, HyperWrite is actively seeking more GPU resources to support the operation of this model.
Shumer mentioned that the success of Reflection 70B is partly attributed to Glaive AI, a startup focused on creating specific-use case datasets. Glaive's rapid AI model training technology has helped accelerate the development process of Reflection 70B for HyperWrite.
Originally founded by Shumer and Jason Kuperberg, HyperWrite is located in Melville, Long Island, New York. The company gained attention for developing the HyperWrite Chrome extension, which assists users in writing emails and other documents. As of November 2023, HyperWrite has amassed two million users and was included in Forbes' annual "30 Under 30" list for its influence.
In March 2023, HyperWrite completed a new funding round led by Madrona Venture Group, securing a $2.8 million investment. Following this funding round, the company added new AI features such as virtual assistant services.
Shumer plans to release an even larger model, Reflection 405B, in the future, which is expected to surpass the current top closed-source models in the market. Additionally, HyperWrite will publish a detailed report outlining the training process and benchmark test results of the Reflection models.
