"Why don't scientists trust atoms? Because they make up everything". This 'dad joke' was cracked by GPT4 in a demo by its creator OpenAI during its launch. You might find it funny, you might not, but the back-story of the joke will certainly impress you. Greg Brockman, a co-founder of OpenAI, showed GPT4 a picture of a handwritten note on what kind of website he wanted, and GPT4 actually built a functioning website on the fly.
The appalling dad jokes on the site were purely incidental.
This was perhaps the singular demo of the highly anticipated next version of GPT3.5 - made famous as ChatGPT - when OpenAI introduced it to an unsuspecting world. It also announced the launch of ChatGPT Plus, which runs on GPT4 and costs $20 a month. GPT4 is both an evolutionary and a revolutionary change from ChatGPT.
Three revolutionary changes are:
1. Making the blind 'see': GPT4 has the impressive ability to describe images in detail and be a 'visual volunteer'. The Be My Eyes demo revealed how GPT4 could describe any visual, say a painting like Starry Night by Van Gogh, in great visual detail. This multimodal feature is perhaps the killer app of GPT4 and could potentially work across images, music, and video. This feature has not been released to the public yet, as OpenAI thinks there is a huge potential of misuse in its current form and is trying to build some ethical guardrails around it.
2. GPT4 is a polyglot: It uses Microsoft Azure Translator as a service with the ability to translate and converse in multiple languages. GPT4 was tested on its reasoning abilities across 26 languages (which included Bengali) and its reasoning was superior to that of GPT3.5/ChatGPT in 24.
3. Building websites from language: Described earlier is this jaw-dropping ability of GPT4 to build a rudimentary website out of English language prompts and images, something which could democratise web content building in a very dramatic way.
In addition, there were four very impressive evolutionary step-ups over GPT3.5/ChatGPT:
1. More accurate: GPT4 promises to be more precise than ChatGPT. Though it still has flaws (It happily continues to 'hallucinate' through conversations), the magnitude seems less. This could be the larger training data sets which power GPT4, though OpenAI has refused to divulge how big these were. We know that ChatGPT was designed to be plausible rather than truthful, to "make stuff up as it goes along". GPT4 has a similar design principle but has been tweaked such that it has a higher element of 'truth' in it.
2. GPT4 has reasoning capabilities: OpenAI showed off a roster of standardised tests across law, education, medicine, etc. that GPT4 passed with flying colours. This included the SAT and the Bar exam. In fact, it passed a simulated Bar exam with a score in the top 10 per cent of test takers, while GPT 3.5 had languished in the bottom 10 per cent.
3. GPT4 has business use cases from the word go. Khan Academy demo-ed a tutor for its young students, which was available 24x7 and almost 'human' in its conversational and guidance skills. BeMyEyes built a GPT4-powered 'Virtual Volunteer' for blind and low-vision users, as we saw above. Morgan Stanley is using GPT4 for wealth management. Harvey, a UK startup, has signed a deal with PWC and Bain for internal knowledge management. As GPT4 APIs are open, expect a mushrooming of such examples across industries.
4. GPT4 is possibly safer: OpenAI claimed that "GPT4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on our internal evaluations. Hopefully, the guardrails around it are stronger and higher, and it will have more responsibility and explainability built in. This is very important since the dark side of GPT can be really, really dark. Sam Altman, the founder CEO of OpenAI, has gone on record recently saying they are "a little scared" of their own creation, and that the last thing the world wants today is a new Frankenstein.
But is it worth paying $20 a month for ChatGPT Plus built on GPT4? You bet it is.
(Jaspreet Bindra is a technology expert, author of 'The Tech Whisperer', and is currently pursuing his Masters in AI and Ethics from Cambridge University)
Disclaimer: These are the personal opinions of the author.