FACT-CHECK GUARANTEE (References): [Official documentation: 1 source] [Peer-reviewed research papers: 1 source] [Academic website: 1 source]
| By Richard Ahern — Last year, ChatGPT set the world on fire as one of the most advanced AI chatbots in existence, but now Elon Musk’s OpenAI has raised the bar yet again.
Even if you live under a rock, you probably experienced some of the excitement around Open AI’s chatbot, ChatGPT, which was released in November 2022.
While tech companies frequently tout their new products as the “next big thing,” Open AI’s group of GPT large language models turned heads everywhere.
On the surface, it was a text-based messenger service with a computer talking back on the other end. It didn’t speak audibly or produce any visual feedback — it just read and spat out lines of text.
So why did people fall in love with it?
Because it made life easier, it got the job done and did it well. But, of course, it depends on what you use it for; it won’t do the laundry or cook for you — but it will give you some decent recipe ideas!
However, for writers and coders is where it shines, ask it to write a computer program in any language, and it does a pretty impressive job.
Its uniqueness lies in the way that you can give it very simplistic or unclear instructions, and it will often fill in the blanks and make the correct assumptions.
For writers, they could copy and paste a chunk of text and ask to summarize it in one paragraph — no problem. You can use it as a basic spelling and grammar checker, but that’s wasting its talents. Not only will it correct mistakes and improve clarity, just like any high-end AI writing assistant, but you can also ask it to rewrite your entire piece or write the whole thing from scratch (should you be lazy).
Lest we forget…
It’s been a dismal nightmare for teachers and examiners as it opened a new can of worms in the battle against cheating. But, of course, it doesn’t help that OpenAI has tested the GPTs by giving them standard school exams, and as you’ll see below, with remarkable results.
To truly understand its power, you must experiment for yourself, but on the whole, the output quality is impressive, mainly because it can produce extended and detailed swaths of content, not just a sentence or two.
But that was just GPT-3.5…
Yesterday, news broke that GPT-4 is ready, and it is a whole new monster.
Firstly, it can reportedly process image content as well as text, something the tech community was begging for. Safety appears to be a focal point for GPT-4, with it “82% less likely to respond to requests for disallowed content.”
In a nutshell, it’s bigger …
The GPTs are called large language models — they are fed giant sets of data about a language and use probabilities to predict the sequence of words. By examining billions of parameters about the structure of a language, the program will look at a word or set of words, calculate the probabilities of what words follow, and then pick the highest likelihood.
For example, take the sentence “I ran up the…” — then take the following words, “dog,” “ball,” “stairs,” or “hill.”
Intuitively, we know that “dog” and “ball” make no sense, but “stairs” and “hill” are both viable choices. However, a deep learning program doesn’t have human intuition; it will look at a large amount of text and calculate the probabilities of each word following the sentence “I ran up the…”.
Let’s say “dog” and “ball” occur less than 0.001% of times after that sentence and say “stairs” has a 20% likelihood of following those words, but “hill” scores a 21% probability. So, the machine will pick “hill” and output: “I ran up the hill.”
Could it be wrong? Of course, but it has a higher probability of being correct, and the more data it has, the more accurate it will be.
It’s not quite that simple; once the model has the data, it is tested and fine-tuned by human reviewers for accuracy and to minimize “hallucination,” the tendency to produce nonsensical garbage — picking the wrong words!
GPT-4 is the largest model yet, by many orders of magnitude, although the exact number of parameters hasn’t been disclosed. Previously, GPT-3 was over 100 times larger than GPT-2, with 175 billion parameters to GPT -2’s 1.5 billion. We can assume a similar increase with GPT-4. In addition, we know the program has undergone intense fine-tuning using reinforcement learning from human feedback. This involves asking humans to rate the chatbot’s responses, and these scores are fed back to “teach it” to produce better outputs.
Open-AI has remained secretive about GPT-4, citing “both the competitive landscape and the safety implications.” Hence, the exact model size, hardware, and training methods are all unknown.
They have said this:
“GPT-4 can solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem solving abilities.” It is 82% less likely than GPT-3.5 to respond to requests for banned content and 60% less likely to make stuff up.
Here’s the scary part:
GPT-4 did significantly better than most human test takers and GPT-3.5 on school exams. For example, in the Uniform Bar Exam (law), it scored in the top 90%, compared to GPT-3.5, which scored in the pitiful 10th percentile. In AP statistics, AP psychology, AP biology, and AP art history (A-level equivalents in the UK), GPT-4 scored between the 80th and 100th centiles — in other words, sometimes beating everyone!
It is not all good:
Interestingly, it did the poorest (8th to 22nd centile) in English literature and composition and could have been more impressive in calculus (43rd to 59th centile).
On Twitter, some people demonstrated how GPT-4 turned a scribbled outline of a website on a napkin into a fully functional online application.
Overall, OpenAI emphasized improved accuracy and safety as the critical improvements of GPT-4. It is far less likely to respond to users asking for instructions to create a bomb, for example. It’s also capable of handling much longer content than its predecessor, processing 25,000 words compared to roughly 1,500 words.
GPT-4 has been touted as more “creative” than before — according to OpenAI, “It can generate, edit, and iterate with users on creative and technical writing tasks, such as composing songs, writing screenplays…”
Finally, perhaps biggest of all, it has “vision,” being able to analyze and classify the content of images.
AI has arrived, and whether you find its evolution thrilling or frightening, there’s no denying it’s here to stay. While some may worry about being replaced, those who embrace its potential will wield it as the most powerful tool available.
We need YOUR help! We bring you the uncensored news for FREE, but we can only do this thanks to the support of loyal readers just like YOU! If you believe in free speech and enjoy real news, please consider supporting our mission by becoming a patron or by making a one-off donation here. 20% of ALL funds are donated to veterans!
This article is only possible thanks to our sponsors and patrons!
Join the discussion!