Powered by GPT-4 technology

GPT-3.5 and GPT-4 are similar in casual conversations, but GPT-4 is better at complex tasks and can understand more detailed instructions. This was tested on different exams made for humans.

GPT-4 can process text and images together to perform various tasks. It can generate text outputs based on mixed inputs of text and images, similar to how it handles text-only inputs. It can also be enhanced with techniques like few-shot and chain-of-thought prompting. However, image inputs are still being researched and not yet accessible to the public.

Work is progressing on defining AI behavior, including steerability. Instead of a fixed personality, developers can customize the AI's style and task using system messages. This allows users to customize their experience within limits. While there may be some imperfections, we encourage you to try it out and give us feedback. Ask your SalesMachines account manager about Guided vs Raw AI LLM.

Biases remain in GPT-4 outputs and still needs improvement. It lacks knowledge of recent events and doesn't learn from experience, yet. It can make reasoning errors and be easily fooled by false statements. It sometimes fails at difficult problems and can introduce vulnerabilities. It can also make incorrect predictions without double-checking its work. Although the base model is usually accurate, the calibration is decreased during post-training.

GPT-4 has limitations like earlier models, including unreliability and errors. Caution is advised when using language model outputs in important situations, with specific protocols tailored to the use-case. GPT-4 reduces hallucinations compared to previous models and performs 40% better than GPT-3.5. We made progress on TruthfulQA, a benchmark that tests the model's ability to distinguish fact from incorrect statements. It includes questions with appealing but incorrect answers.

GPT-4 has been made safer and more aligned through various measures, such as filtering data, engaging experts, improving model safety, and monitoring. It still carries risks, but we have tested it extensively with the help of over 50 experts to understand and mitigate those risks. Feedback from experts has led to improvements, including collecting more data to prevent the generation of dangerous advice. GPT-4 also incorporates a safety reward system during training to reduce harmful outputs. Our mitigations have significantly improved GPT-4's safety compared to GPT-3.5.

Please give our product management team more time to answer AI technology questions as they are currently prioritizing work for members. We enjoy collaborating and having strategic discussions.