HomeArtificial IntelligenceOpenAI Unveils Next Gen AI Model GPT-4 That Accepts Image, Text Inputs

OpenAI Unveils Next Gen AI Model GPT-4 That Accepts Image, Text Inputs

chatgpt 1 Thanks to an unwitting slip-up, we’ve known that the next version of GPT was coming this week, and sure enough it has arrived. OpenAI has pulled the covered off its new large multimodal model GPT-4 which accepts image and text inputs.

We know what’s on your mind—Can GPT-4 generate AI videos? That’s what everyone was led to believe a few days ago. The answer is no, unfortunately.

All we’re being told is that compared to GPT-3.5, the new AI model is more reliable, creative and capable of handling complex instructions. So it’s a better version of the old GPT, and does not introduce any revolutionary new tools.

Or let’s just say we’ve gotten so familiar with the wondrous nature of ChatGPT and generative AI over the past few months, ChatGPT 4 does not look like an exciting new product.

“We’ve created GPT-4, the latest milestone in OpenAI’s effort in scaling up deep learning,” the company said on Tuesday.

“We’ve spent 6 months iteratively aligning GPT-4 using lessons from our adversarial testing program as well as ChatGPT, resulting in our best-ever results on factuality, steerability, and refusing to go outside of guardrails,” it added.

The AI tool is now capable of taking in text and image prompts and replying in text, which is pretty amazing. See the image below, for example:

ChatGPT 4 Output
Image credit: OpenAI

Many machine learning benchmarks are in English. OpenAI translated the MMLU benchmark—a suite of 14,000 multiple-choice problems covering 57 subjects—into over two dozen languages using Azure Translate.

In 24 out of 26 languages tested, GPT-4 beat the English-language performance of GPT-3.5 and other large language models including Chinchilla and Google’s PaLM.
OpenAI has taken care to point out that this next gen version of ChatGPT is still liable to make up facts, a phenomenon in AI dubbed “hallucinating”, which is why it’s not a good idea to use any kind of AI to write news without thorough editing.
It can commit reasoning errors too. Folks are being cautioned not to use it for sensitive tasks such as performance reviews. GPT-4 can also offer harmful advice (so probably don’t use it for your blog on mental well being or general health tips) and write buggy code.
The new GPT incorporates an additional safety reward signal during Reinforcement Learning from Human Feedback (RLHF) to reduce harmful outputs. Its tendency to respond to prompts for disallowed content has been lowered by 82% compared to GPT-3.5. It also handles sensitive requests in a better manner.

Those who’re paying the $20 per month fee for ChatGPT Plus will get access to GPT-4 with a usage cap. Developers can sign up for a spot on the GPT-4 API’s wait-list.