Last November, the company behind Facebook released an intriguing chatbot called Galactica, but it quickly faced criticism for inventing historical events and sharing false information. As a result, Meta, the company behind Facebook, removed it from the internet. However, just two weeks later, another San Francisco start-up called OpenAI unveiled its own chatbot called ChatGPT, which became a global sensation. What set ChatGPT apart was the technique used to train it, which was revolutionizing the field of artificial intelligence.
In the months leading up to ChatGPT’s release, OpenAI employed hundreds of people to use an early version of the bot and provide precise suggestions to improve its skills. These individuals acted as tutors, teaching the bot how to respond to specific questions, rating its responses, and highlighting its mistakes. By analyzing this valuable feedback, ChatGPT became a better chatbot.
The technique behind ChatGPT, known as “reinforcement learning from human feedback,” has had a profound impact on the development of AI in the industry. It has transformed chatbots from mere curiosities into mainstream technology. These chatbots are built on new AI systems capable of learning skills by analyzing data. A significant portion of this data is crafted, refined, and sometimes created by large teams of low-paid workers worldwide.
For years, companies like Google and OpenAI have relied on these workers to curate data used to train AI technologies. They have helped identify objects in photos for self-driving cars and detect signs of diseases in medical videos. In the case of chatbots, workers play a crucial role in refining the AI system’s responses. They provide precise feedback to improve its behavior and reduce the production of misinformation and biased responses.
OpenAI, Anthropic, Hugging Face, and other leading labs utilize freelance workers, including those hired through platforms like Upwork, Scale AI, and Surge. These workers have varying educational backgrounds and are evenly split between genders; some do not identify as either male or female. While U.S.-based workers earn between $15 and $30 per hour, their international counterparts earn considerably less.
The nature of this work demands meticulous writing, editing, and rating, often requiring significant hours to perfect responses. The valuable human feedback allows today’s chatbots to engage in more natural turn-by-turn conversations rather than providing static responses. It also helps mitigate misinformation and biases generated by these systems.
However, researchers caution that the technique is not fully understood and can have unintended consequences. Recent studies have shown that OpenAI’s technology’s accuracy has dropped in certain situations, such as solving math problems or coding, possibly due to continuously applying human feedback. Fine-tuning the system in one area may introduce biases and side effects, causing it to perform poorly in other areas.
The conundrum of AI development lies in the potential for machines to exhibit unexpected and potentially harmful behavior as they learn through data analysis. To combat this, OpenAI has developed algorithms that allow the AI system to receive regular guidance from human teachers while learning tasks through data analysis.
Companies have built large language models by training them on vast amounts of digital text, resulting in advanced systems capable of generating articles, solving math problems, and annotating images. However, these systems, like Meta’s Galactica, can produce untruthful and biased information. To address this, labs have fine-tuned large language models with human feedback, similar to how OpenAI trained its AI system to play a video game.
Workers utilize various methods to train chatbots, including providing specific prompts or editing bot-generated responses. They also rate the bot’s responses for helpfulness, truthfulness, and harmlessness. However, biases may inadvertently influence these judgments, as workers’ perspectives shape the feedback.
It is essential to note that human feedback cannot solve all chatbot problems, as their responses rely on mathematical probabilities. AI systems must learn patterns of behavior through human feedback and apply them in various situations. Although human feedback works reasonably well in preventing negative outcomes, it is not a flawless solution.
Yann LeCun, chief AI scientist at Meta, believes it is crucial to develop new techniques to enhance chatbots’ reliability further. While human feedback can prevent problematic behavior, it cannot guarantee perfection. Chatbots remain an exciting and evolving field, and the careful implementation of human feedback is an important aspect of their development.
Denial of responsibility! Vigour Times is an automatic aggregator of Global media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, and all materials to their authors. For any complaint, please reach us at – [email protected]. We will take necessary action within 24 hours.
