10 things you need to know about GPT-4

On March 14, 2023, OpenAI officially released a new large-scale multi-modal model GPT-4.

Just a few months after OpenAI shocked the world with ChatGPT, it quickly launched a new generation of milestone GPT-4. Compared with the previous generation of ChatGPT and GPT-3.5, what updates does GPT-4 bring, this article will give you a quick inventory:

1. ChatGPT has upgraded to GPT4.0

ChatGPT and the new Bing have been upgraded to the GPT-4 version. Friends who have ChatGPT or new Bing accounts can already try the powerful functions of GPT-4 directly through the original account. At the same time, GPT-4 has opened the API interface, which is so powerful and unreasonable.

2. Iteration of professional academic ability

The test performance of GPT-4 in law, calculus, chemistry, physics and other subjects has been greatly improved, which is close to the test performance of top human students. Taking the simulated bar exam as an example, the score of GPT-4 can be ranked in the top 10% of test takers, while the previous generation of GPT-3.5 intelligence can achieve the bottom 10%. The lawyers were shocked and started emo.

3. Text input limit increased

In terms of text input, GPT-4 has achieved 8 times the input limit of the previous generation ChatGPT, reaching 25,000 words/times. According to this trend, in the near future, all the independent book texts of human beings can be input into gpt as a prompt at one time. Quantum reading papers have been realized, and quantum reading is just around the corner.

4. Support visual image input

GPT-4 has begun to support image input. This feature is currently not open to the public, and the official has only released some dialogue previews.

GPT-4 can accept both text and image cues, which parallel the text-only setup, allowing users to specify any visual or language task.

Specifically, it generates textual output (natural language, code, etc.) given an input consisting of interspersed text and images. In a range of domains—including documents with text and photos, diagrams, or screenshots—GPT-4 demonstrated similar capabilities to plain text input. Additionally, it can be enhanced with test-time techniques developed for plain text language models, including few-shot and thought-chain cues.

5. Improved answer accuracy

After 6 months of intensive learning, GPT-4’s ability to deal with complex and subtle questions is significantly higher than that of GPT3.5, with more reliable answers and more creative answers.

Comparing GPT-4 with several other large models based on several mainstream machine learning model evaluation benchmarks, it is not difficult to see that GPT-4 is currently the top large-scale model in the fields of multi-disciplinary professional question answering, trial reasoning, code writing and arithmetic.

6. Leap forward in cross-lingual ability

To get a first look at the capabilities of other languages, OpenAI used Azure Translate (see Appendix) to translate the MMLU benchmark — a set of 14,000 multiple-choice questions covering 57 topics — into multiple languages. GPT-4 outperforms GPT-3.5 and other LLMs (Chinchilla, PaLM) in English language performance in 24 of the 26 languages ​​tested, including low-resource language tests such as Latvian, Welsh, and Swahili.

7. Customizable style of GPT

OpenAI has been working on optimizing ChatGPT’s answer level, including stylistic operability. Unlike classic ChatGPT personalities with fixed verbosity, tone and style, developers (and soon ChatGPT users) can now dictate the style and tasks of their AI by describing these directions in “system” messages. System messages allow API users to significantly customize their user experience to a certain extent.

8. Limitations of GPT-4

OpenAI frankly explained the current limitations of GPT-4. Its answers are still impossible to be 100% reliable. Hallucinations and reasoning errors are still common, especially when users ask questions that are too small to be included in the database. if. GPT-4’s database is only updated until September 2021, and GPT-4 does not know what happened after that, and naturally it cannot help you predict tomorrow’s stock price (it feels like this is the most popular question).

9. GPT with more controllable risks

Compared to GPT-3.5, the GPT-4 model was 82% less likely to respond to requests with disallowed content, while responding to sensitive requests such as medical advice and self-harm was 29% more often in-policy.

OpenAI has hired more than 50 experts in the field of artificial intelligence security to conduct adversarial tests on GPT-4 to ensure the safety of users. Let us praise this effort. After all, no one wants GPT-4 to give advice to criminals , or negatively induce minors with insufficient judgment ability.

10. Predictable Scaling

A big focus of the GPT-4 project is to build deep learning stacks that scale predictably.

The main reason is that for very large training runs like GPT-4, extensive model-specific tuning is not feasible. The infrastructure and optimizations developed by OpenAI have very predictable behavior across multiple scales.

To verify this scalability, OpenAI accurately predicted the final loss of GPT-4 on OpenAI’s internal codebase (not part of the training set) by inferring from a model trained with the same method but using 10,000 times less computation. At present, the core value of this matter is to reduce the cost of training large models in the future.