According to Cointelegraph, a new artificial intelligence chatbot named 'Reflection 70B' has been introduced by HyperWrite AI. CEO Matt Shumer announced the development on Sept. 5, claiming it to be the world's top open-source model. The AI was trained using a technique called 'Reflection-Tuning,' designed to enable large language models (LLMs) to correct their own mistakes.
Reflection Llama-3.1 70B, which is Meta's open-source AI launched in July, can compete with top closed-source models like Anthropic’s Claude 3.5 Sonnet and OpenAI’s GPT-4o in several benchmarks, according to Shumer. He explained that current AI models often hallucinate, meaning they generate inaccurate outputs by perceiving patterns or objects that do not exist. Reflection-Tuning aims to address this by allowing AI to recognize and correct its mistakes before finalizing an answer.
Reflection tuning involves feeding AI responses back into the model, asking it to evaluate its own outputs to identify strengths, weaknesses, and areas for improvement. This iterative process helps the AI continuously refine its capabilities, making it more self-aware and better at critiquing and improving its performance. Shumer noted that with the right prompting, the new model is highly effective for various use cases.
In 2023, Microsoft-backed OpenAI released a research paper discussing methods to prevent AI hallucinations. One proposed idea was 'process supervision,' which involves training AI models to reward themselves for each correct step of reasoning rather than just the final correct conclusion. Karl Cobbe, a researcher at OpenAI, emphasized the importance of detecting and mitigating logical mistakes or hallucinations as a critical step towards building aligned artificial general intelligence (AGI).