Matt Shumer, CEO of AI startup HyperWrite, has announced a significant milestone: the launch of the Reflection70B model, based on Meta's open-source Llama3.1-70B Instruct. This new model not only excels in third-party benchmark tests but has also been hailed as "the world's most powerful open-source AI model."
Shumer shared the news on social network X, showcasing charts where Reflection70B outperformed the Meta Llama series in multiple tests, with performance rivaling top commercial models. Following the announcement, HyperWrite's demo website saw a surge in traffic, and the team is currently seeking more GPU resources to meet user demand.
What sets Reflection70B apart is its self-correction technology. Shumer noted that existing large language models (LLMs) can produce errors but often lack self-correction abilities. Reflection70B reviews and assesses the accuracy of its generated text post-production, ensuring the reliability of answers provided to users.
The model introduces several new special symbols, enabling users to interact with it in a more structured manner. During reasoning, the model outputs its thought process within special tags, allowing for real-time error correction. On the demo website, users can try simple questions like "How many letters 'r' are in the word 'Strawberry'?" or "Which is larger, 9.11 or 9.9?" Reflection70B processes these questions and delivers correct answers after a brief period.
This feature makes Reflection70B ideal for tasks requiring high precision. It improves accuracy by breaking down the reasoning process into clear steps. Now, Reflection70B is available for download via AI code repository Hugging Face, with API access soon to be provided by GPU service provider Hyperbolic Labs.
Additionally, HyperWrite is working to integrate the Reflection70B model into its main AI writing assistant product, with a more powerful version—Reflection405B—set to launch next week. Shumer revealed they are continuously exploring how to better apply this technology to practical products.
The success of Reflection70B also owes to the contributions of Glaive, which specializes in generating datasets for specific use cases, significantly accelerating model training. Glaive's platform quickly produces high-quality synthetic data, enabling companies to fine-tune models quickly and cost-effectively, addressing a major bottleneck in AI development regarding the availability of high-quality, task-specific data.
Based in Long Island, HyperWrite, founded by Shumer and Jason Kuperberg in 2020, has been actively exploring the AI field. Initially a Chrome extension, it has evolved into an AI assistant capable of handling more complex tasks, boasting a user base of 2 million and earning recognition on Forbes' annual "30 Under 30" list.
HyperWrite's latest funding round, disclosed in March 2023, raised $2.8 million from investors including Madrona Venture Group. These funds have supported the development of new AI-driven features, transforming web browsers into virtual assistants capable of tasks ranging from booking flights to finding job candidates on LinkedIn.
Shumer emphasized that accuracy and safety are always HyperWrite's top priorities, especially when exploring complex automated tasks. The platform continuously monitors and improves through user feedback to refine its personal assistant tools. This cautious approach, akin to the structured reasoning and reflection embedded in Reflection70B, reflects Shumer's commitment to precision and responsibility.
Model download link: https://huggingface.co/mattshumer/Reflection-Llama-3.1-70B