OpenAI Gathers a Group of Over 50 Specialists to Bolster GPT-4's Safety Measures
In Brief
OpenAI has recruited a team of more than 50 specialists focused on safeguarding its latest model. language model , GPT-4, is safe for use.
This team consists of researchers and engineers who are well-versed in areas such as AI safety, ethics, and regulatory concerns.
The goal is to mitigate the risks of GPT-4 producing harmful or biased outputs, ensuring its actions are in line with human values.
OpenAI has engaged over 50 specialists from various sectors to enhance the model's safety. These experts are concentrated on identifying risks through adversarial testing, exploring concerns ranging from long-term AI alignment to cybersecurity issues. GPT-4 As the latest model carries similar dangers found in smaller language models, its expanded capabilities introduce fresh challenges. Thus, collaborating with specialists has been vital for the technology's security.
OpenAI has added a further layer of safety by incorporating a new set of training prompts focused on reinforcement learning from human feedback (RLHF) and rule-based reward models (RBRMs) to enhance GPT-4’s security measures. These RBRMs act as zero-shot classifiers that serve as an added incentive during the RLHF fine-tuning phase, encouraging the model to avoid harm and to cater to harmless requests. To uphold the safety of GPT-4 models, OpenAI began enlisting external experts for 'red team' exercises in August 2022. These experts conducted stress tests, boundary assessments, and adversarial challenges, gaining access to preliminary iterations of GPT-4 that laid bare initial risks and prompted additional safety investigations. .
Feedback from these experts led to specific technical solutions and policy enforcement strategies designed to decrease risks. Nonetheless, numerous threats persist, necessitating further scrutiny.
With regards to OpenAI's workforce, ChatGPT was initially trained with support from individuals in some of the world's most impoverished regions through a partnership with Sama, a company that employs many from disadvantaged backgrounds. Critics in the AI ethics community have raised concerns about OpenAI's choice to delegate the training of its models to Sama, alleging the company exploits low-wage workers.
GPT-4 vs. GPT-3: What New Features Does the Latest Model Present?
GPT-4 Displays Superior Performance Compared to GPT-3.5 Across Numerous Study Benchmarks developed The ChatGPT Powered by GPT-4 Surpasses GPT-3 By A Factor of 570 ChatGPT GPT-4 Is Equipped to Manage Requests for Images, Documents, Diagrams, and Screenshots
Read more:
Disclaimer
In line with the Trust Project guidelines Vanilla Unveils 10,000x Leverage Super Perpetuals for the BNB Chain