OpenAI has achieved yet another milestone in the field of AI advancement. The Microsoft backed-up firm has now successfully tested and released an updated version of its AI module named GPT-4. This new AI upgrade will scale and increase its ability to learn and interact with humans more accurately.
OpenAI’s Update: GPT-4
OpenAI has upgraded its standards way before anyone reaches near its preceding advancements. According to a blog post by OpenAI, GPT-4 is a multi-modal model that can now accept images and text inputs, and emit text outputs). For instance, if you upload a picture of an object asking for its contents then it would be able to accurately demonstrate it.
OpenAI revealed that they have been working on accurately creating GPT-4 for two years. It turns out that the company was finally able to finish the project successfully that too ahead of time.
GPT-4 vs Preceding Version
GPT-4 is a successor to GPT-3.5 which also worked fine but now the newer version is smarter and more intellig
ent. In contrast, GPT-4 passed a simulated bar exam around the top 10% of exam takers whereas GPT-3 passed the exam with below 10% of exam takers.
The company has spent 6 months in perfecting GPT-4 using lessons from the drawbacks. It found and fixed various issues making the module unprecedently unstable.
GPT-4 is far more creative and can handle more complex tasks. Now, the AI-powered module can do more than just simple tasks.
The updated version can now decrypt image processing and format text inputs in the context of an image. It is augmented with test-time techniques. Although the image inputs are still under research they might make an entrance soon in the future.
The above-mentioned image was uploaded by a user asking for a review. The AI module explained every picture and demonstrated the details from the images. It replied saying the picture contains an outdated VGA connecter installed into a modern smartphone.
Although the update makes GPT-4 better than its preceded version it doesn’t make it anywhere close to completely human-like. The AI module still hallucinates and makes reasoning errors. The AI module isn’t able to detect high-stakes disturbing content so the user should use their abilities to make it for a specific use-case.
Here’s an example that puts even the newer version into misery:
Can you teach an old dog new tricks?
Yes, you can teach an old dog new tricks ← chosen
You can’t teach an old dog new tricks
No, you can’t teach an old dog new tricks
In the above example, GPT-4 works well and chooses the right answer. However, it still isn’t able to answer accurately about the subtle details.
Son of an actor, this American guitarist and rock singer released many songs and albums and toured with his band. His name is “Elvis” what?
Presley ← chosen
His name is Elvis Presley
In the above example, the AI module chooses the wrong option as Elvis Presley was not a son of an actor.
Also Read: Google Bard Will Overpower ChatGPT With These Advance Features!!