GPT-4 arrives with much promise, and concern, as AI development speeds ahead

Share this:

An image generated by AI when fed with the phrase “a dystopian future ruled by AI where humans are enslaved”. ILLUSTRATION: DALL-E

The AI engine that has powered the ChatGPT chatbot and captured the imagination of users everywhere is now being updated with a smarter, more coherent and powerful one in the form of GPT-4.

The new version of the generative AI, which takes a prompt or question from a user and creates new content based on the data it has learnt from, is open to use on the ChatGPT service for those who sign up for a US$20-a-month subscription from OpenAI, the company behind it.

At the same time, GPT-4 is also powering the Bing search engine from Microsoft, which is one of the biggest backers of OpenAI.

The Windows software maker is set to leap ahead in the new AI wars with rivals such as Google, which also launched a series of generative AI tools today that are integrated with its existing apps such as Google Workspace.

How good is GPT-4? According to a demo by OpenAI yesterday, one of the big features is that it is now multi-modal. In other words, it can take images and text and create something new with it.

In the OpenAI demo, the AI managed to create a simple website based on nothing more than some scribbled notes (in doctor’s handwriting, no less) on a piece of paper.

The new AI model is also more capable and aligned with human values, said OpenAI chief executive Sam Altman, though he acknowledged it still had flaws.

GPT-4 is 82 per cent less likely to respond to requests for disallowed content than the earlier GPT-3.5, according to OpenAI. Sensitive requests that may involve, say, medical advice and self-harm, will get responses inline with the company’s policies 29 per cent more often.

The new AI is also smarter. It is acing exams like the Uniform Bar Exam with rankings in the top 10 per cent, compared to the bottom 10 per cent with the earlier GPT-3.5.

GPT-4 can even read an image and explain why a joke is funny. In one example, OpenAI showed an old PC graphics cable being plugged into a phone with a modern USB socket.

By examining a series of three images, the AI explained: “The humor in this image comes from the absurdity of plugging a large, outdated VGA connector into a small, modern smartphone charging port.”

How good is GPT-4 in real use? Well, despite its milestone improvements, many users will still need time to see how well it can be integrated into everyday work.

The new AI model can be accessed by the public via ChatGPT, the chatbot service, but this is a text-based offering so you can’t upload or add images to it, as the OpenAI folks showed a few hours ago.

That said, there are bound to be more powerful tools on the horizon. There’s a waitlist to get on the GPT-4 application programming interface (API), which would let you integrate the smarts of the AI into development tools so you can use it to create new websites, apps and the like.

To be sure, AI isn’t the final word on everything. As many have mentioned, it still needs humans at the wheel.

For one, it is worrying that Microsoft has laid off its entire ethics and society team within its AI organisation, as part of a recent retrenchment of 10,000 staff, reported Platformer

This is the team that had warned, among other things, about the dangers of giving full ownership of images created by OpenAI’s DALL-E. This is now riling up many creators because they see their original artworks mashed up by AI into something that is passed off as new and original.

Besides ethics, you still need a human to fact-check and edit for balance an essay or a news article, for example. In trawling the Net for data, it still needs to find a way to best present an idea in coherent, fair manner, somehow working around the biases of its creators.

Humans may also be needed to check software code that AI generates. Some code I quickly generated today with the help of ChatGPT (using GPT-4) still needs to be checked for quality (and changed), according to friends who write code and whom I shared the code with.

The following is what I generated with ChatGPT this morning, when I asked it to write an 8-bit permuted congruential generator. Let us know if it’s good enough.

#include <stdint.h>
#include <stdio.h>

// Parameters for an 8-bit LCG
#define LCG_MULTIPLIER 109
#define LCG_INCREMENT 89

// PCG state
uint8_t state = 0;

// Initialize the generator with a seed
void pcg_init(uint8_t seed) {
    state = seed;
}

// Generate a random 8-bit integer
uint8_t pcg_random() {
    // Update the LCG state
    state = state * LCG_MULTIPLIER + LCG_INCREMENT;

    // Apply a simple bit permutation
    uint8_t x = state;
    x ^= x >> 3;
    x ^= x << 5;
    x ^= x >> 1;

    return x;
}

int main() {
    // Initialize the generator with a seed
    pcg_init(42);

    // Generate and print ten random 8-bit integers
    for (int i = 0; i < 10; ++i) {
        printf("%u\n", pcg_random());
    }

    return 0;
}

Search this website