GPT-4: Overview

GPT-4: Overview


OpenAI (the makers of chatgpt) just released their latest large language model GPT-4.

What is GPT? Generative Pre-trained Transformer?is a autoregressive language model?built by OpenAI which can produce human-like text


Brief overview video: https://www.youtube.com/watch?v=oc6RV5c1yd0

Developer livestream: https://www.youtube.com/watch?v=outcGtbnMuQ&ab_channel=OpenAI


Major highlights of GPT-4 include:

  • Longer context: GPT-4 is now capable of handling 32,768 token context. (about 50 pages of text)
  • Steerability: You can now prescribe the AI’s style and task by putting directions in the "system" message box. This enables customizing and finetune the user experience.?

No alt text provided for this image
Example from OpenAI asking GPT-4 to answer as a Shakespearean pirate

  • Multilingual:?In the 24 of 26 languages tested, GPT-4 outperforms the English-language performance of it previous model GPT-3.5???

No alt text provided for this image

  • Multimodel: GPT-4 allows for both text and image input, outputting text. In the developer livestream the presenter hand drew a UX mock and asked GPT-4 to "write brief HTML/JS to turn this mock-up into a colorful website, where jokes are replaced by two real jokes". Note that image inputs are still a research preview and not publicly available.

No alt text provided for this image
Hand drawn UX mock & text input


No alt text provided for this image
Generated HTML/JS code

  • Increased safety: Improvements were made to significantly improve the safety when compared to GPT-3.5. GPT-4 decreased response to disallowed content requests by 82% compared to GPT-3.5.

No alt text provided for this image
Example of disallowed content


Other points you might have missed:

- Pretrained data: Pretrained data is still the same as GPT-3.5 and still cuts off in September 2021. However during the developer livestream the presenter was testing GPT-4 against the discord API and got around this by providing the latest API information when prompting the model.

- Hallucinations: GPT-4 still can hallucinate i.e. “produce content that is nonsensical or untruthful in relation to certain sources. However through internal adversarial factuality evaluations GPT-4 scores 40% higher than GPT-3.5

- Companies are already using GPT-4 in their products. For example Stripe leverages GPT-4 to streamline user experience and combat fraud. https://openai.com/customer-stories/stripe

- OpenAI has open sourced a software framework to evaluate the performance of its AI models called Evals, https://github.com/openai/evals

- GPT-4 is accessible if you have ChatGPT plus access https://chat.openai.com/chat. There is still a waitlist for API access.???


Further links:

GPT-4 product page - https://openai.com/product/gpt-4

GPT-4 research page with mode details: https://openai.com/research/gpt-4

GPT-4 technical report - https://cdn.openai.com/papers/gpt-4.pdf

Andrea Rigotti

Senior Director Digital Technology - Architecture of Tomorrow

2 年

Love your work Kevin ??

回复
Tony Ta

On a mission to humanize enterprise systems with human centric AI

2 年

Great summary KC!

回复

要查看或添加评论,请登录

Kevin Chan的更多文章

社区洞察

其他会员也浏览了