AI chatbots of the current generation are prone to making mistakes or misrepresenting concepts as real facts. That could become an issue as we become more and more dependent on them. However, a fresh open-source product might now provide a fix.
Popular AI companies such as Anthropic or OpenAI attract a lot of attention when they release new, improved models, primarily due to the global impact of AI on home and office users. AIs seem to be becoming more intelligent every day. However, a brand-new artificial intelligence (AI) system from New York-based startup HyperWrite is making waves for a different reason: it’s utilizing a brand-new open source error-trapping system to avoid many of the common “hallucination” problems that frequently befall chatbots like ChatGPT or Google Gemini, which made headlines earlier this year when it famously advised users to put glue on pizza.
According to reports, Meta’s open-source Llama model serves as the foundation for the new AI, known as Reflection 70B. The intention is to incorporate the new AI into the company’s flagship product, a writing assistant that assists users in crafting words and adjusts to the user’s needs. This kind of task is exactly the kind of “sparking” creative ideas that generative AI excels at.
But what’s most intriguing about Reflection 70B is that CEO and co-founder Matt Shumer is promoting it as the “world’s top open-source AI model,” and it includes a novel kind of error detection and correction known as “reflection-tuning.” Other generative AI models “have a tendency to hallucinate, and can’t recognize when they do so,” as Shumer noted in a post on X. With the help of the new correction system, LLMs are able to “identify their mistakes and correct them before committing to an answer.” In other words, the output from an AI is fed back into the system, which is asked to determine whether the output has any areas that need to be improved upon. This allows AIs to analyze their own outputs (thus the name “reflection”) in order to identify areas where they’ve gone wrong and learn from it.
The concept of AIs learning from their mistakes is not new; in April, Mark Zuckerberg of Meta proposed that the company’s Llama model should be able to train itself by approaching an issue in several ways, identifying which output is leading to the correct response, and then feeding that information back into the AI model to continue training it in a manner akin to a feedback loop. Rather than just reintroducing corrected data as training data, Reflection 70B appears to view this as a more direct approach to addressing the issue of AI hallucinations or false information. Shumer displayed an image of a conversion regarding the number of “Rs” in the word strawberry to demonstrate the type of “fix” that Reflection should be able to perform, according to the report. Recently, when leading AI models malfunctioned and stated that there were only two “Rs,” not three, this delightfully strange hallucination made headlines. This question is posed to Reflection in the model conversation, and it answers with “two” before identifying its own “reflection” error and reporting “I made a mistake.” I now see very clearly that the word “strawberry” actually contains three “r’s.”
The increasing number of people using AIs for news data searches, opinion polls on significant issues, and other purposes makes the business of AI accuracy, misinformation sharing, and other reliability matters crucial. With the EU, US, and UK signing a new agreement to guarantee AI safety, keeping future sophisticated AIs in line with humanity’s best interests is becoming more and more important.
The challenge with this kind of work is that, in order to pass laws that have real impact, regulators must grapple with extremely complex mathematical and logical problems that are fundamental to AI models such as ChatGPT and even more recent competitors like Reflection 70B. These problems begin with basic counting. California is set to implement laws pertaining to artificial intelligence that mandate disclosure when an AI model is trained on computers that can process between 10 and 26 floating point operations per second. That’s a lot of zeros to count—100 septillion math bits per second. All we can hope for is that lawmakers, who aren’t exactly known for their math skills, will be able to comprehend this better than ChatGPT did when it came to counting the “Rs” in strawberry.
