HyperWrite Launches the World's Most Powerful Open AI Model: Reflection 70B
HyperWrite launches Reflection 70B, a groundbreaking open AI model that offers unique self-correction capabilities and outperforms both open and closed models on several benchmarks.
HyperWrite, an AI writing startup led by co-founder and CEO Matt Shumer, has launched Reflection 70B, a new large language model (LLM) based on Meta's open Llama 3.1-70B Instruct. The model stands out for a unique self-correction technique and performs better than other models on independent benchmark tests. According to Shumer, who unveiled the model on the social network X, Reflection 70B is now "the world's best open AI model."
Reflection 70B has been thoroughly tested on several benchmarks, including MMLU and HumanEval, with results verified using LMSys's LLM Decontaminator to avoid data contamination. The model consistently outperforms Meta's Llama series and can hold its own against the best commercial models.
Unique Features That Set Reflection 70B Apart
One of the most distinctive features of Reflection 70B is its ability to identify and self-correct errors. Shumer explained in an interview with VentureBeat that the model uses a technique called "reflection tuning" to identify and correct its own mistakes before delivering final answers to the user. The model introduces special tokens for reasoning and error correction, making it easier for users to interact with the model in a more structured way.
On the demo site, users can test the model with suggested prompts that showcase its ability to solve problems other AI models often struggle with, like counting letters in words or comparing decimals correctly. Although answers can take time to generate, the model ultimately delivers correct answers by breaking its reasoning down into distinct steps.
Bigger Models on the Way
Reflection 70B is just the beginning of the Reflection series. An even bigger model, Reflection 405B, is planned for launch next week and is expected to outperform even the best closed models on the market. Shumer also mentioned that HyperWrite is working on integrating Reflection 70B into its main AI writing assistant.
The underlying model for Reflection 70B is built on Meta's Llama 3.1 70B Instruct and uses the standard Llama chat format, ensuring compatibility with existing tools and pipelines.
Partnership With Glaive for Rapid Progress
A key factor behind Reflection 70B's success is the synthetic data generation company Glaive, which provides custom-tailored datasets. Glaive's platform makes it possible to quickly train small, highly focused language models, helping to democratize access to AI tools. Glaive has managed to create synthetic datasets tailored to specific needs, making rapid, cost-effective fine-tuning of models easier.
By using Glaive's technology, the Reflection team was able to quickly generate high-quality synthetic data to train Reflection 70B. According to Shumer, the training process took three weeks across five iterations of the model.
HyperWrite's Journey and Future Plans
HyperWrite, founded as Otherside AI by Matt Shumer and Jason Kuperberg in 2020 on Long Island, has grown its platform from a simple Chrome extension into a powerful AI assistant with over two million users. The company recently secured funding to further develop its AI-powered features, including tools that turn browsers into virtual butlers.
Going forward, Shumer plans to continue expanding the Reflection series and expects the upcoming Reflection 405B to outperform market-leading models like OpenAI's GPT-4. This development could shift the balance of power in the fast-growing AI industry and give open models a competitive edge over proprietary alternatives.