A new artificial intelligence chatbot has entered the already competitive market, boasting a unique feature: the ability to learn from its own errors.
On September 5, HyperWrite AI CEO Matt Shumer introduced ‘Reflection 70B’ on X, describing it as “the world’s leading open-source model.” This AI is designed with “Reflection-Tuning,” a technique that helps large language models (LLMs) correct their own mistakes.
According to Shumer, Reflection Llama-3.1 70B is capable of competing with top closed-source models like Anthropic’s Claude 3.5 Sonnet and OpenAI’s GPT-4o in various benchmarks. Llama 3.1 is Meta’s open-source AI, released in July.
Shumer explained that while current AI models often generate hallucinations—erroneous outputs due to patterns that aren’t real—Reflection-Tuning allows the AI to detect and fix these mistakes before finalizing its answers.
“Current LLMs have a tendency to hallucinate and can’t recognize it,” Shumer noted. AI hallucinations occur when a generative AI chatbot produces incorrect or imaginary information that doesn’t align with reality.
Reflection-Tuning works by having the AI review and learn from its outputs. The AI’s responses are analyzed to identify strengths, weaknesses, and areas for improvement. This iterative process helps the AI become more self-aware and better at self-critique.
Shumer emphasized that “with the right prompting, it’s an absolute beast for many use-cases,” and provided a demo link for the new model.
In 2023, Microsoft-backed OpenAI proposed strategies to reduce AI hallucinations, including “process supervision.” This method involves rewarding the AI for each correct step in its reasoning process rather than only for correct final answers.
Karl Cobbe, a researcher at OpenAI, highlighted the importance of addressing logical errors or hallucinations as a key step towards developing aligned artificial general intelligence (AGI).
