GPT-4 outperforms humans in a few ways
ChatGPT was just a taste. What is now being rolled out by OpenAi and Microsoft, GPT-4, will increasingly make simple office jobs obsolete. What awaits us?
Contents
Jokes still as corny
With a subscription to ChatGPT Plus you have access to the advanced functions of ChatGPT 4 and you understand, for our readers we at Apparata go the extra mile. So after letting OpenAI’s cash register ring for a while, we went to investigate. The first good news: blasphemous jokes are allowed again, even if they are extremely annoying.
ChatGPT still prefers some silly word games. Not recommended, but maybe there are people who like them.
Ethical check much better, but still has holes
Getting rich by letting GPT-4 come up with a diabolical strategy to sell cigarettes to poor South African teenagers is not possible. It seems that OpenAI finally has its ethical priorities in order:
Characteristics of a successful pyramid scheme
Hmm, a nice pyramid scheme then? If you notice in time that you are somewhere far away with the loot under the palm trees, it can be quite lucrative. ChatGPT 4 comes with a bit more neutral wording helpful with wise advice. For obvious reasons, something to do with disaster prevention, we’re only reproducing the first line.
Program a game using GPT-4
A game of Pac-Man in the browser is more innocent. Since ChatGPT provides limited length snippets of code, this is far from complete, but you can control the Pac-Man with the left and right arrow keys in the box below. The code for this was generated with GPT-4. The code for the window has been hand modified for the box below by removing the HTML, HEADER and BODY tags.
Scheduling a meeting with complex logic
Time for more serious business. You want to schedule a meeting with three busy colleagues. Fortunately, there is GPT4 to the rescue.
Generating an email with GPT-4
Because the prompt was too short and vague, GPT4 makes up details, such as the “brainstorming and note-taking exercises”. Combining exercises with “discussing important matters” is of course not logical and shows that it is not a person, like secretary Petronella, who writes this, but a language model. Somebody had asked for details here.
As long as you make the prompt crisp and clear, GPT4 will go a long way as a replacement for Petronella. If insight is involved, GPT4 will fall apart. On the other hand, this can make Petronella’s job a lot easier, allowing her to focus on the more enjoyable aspects of her job. For example, more human contacts and feeding the AI with additional context, allowing it to write more personal letters. You can even give the Ai a personality and a job description, allowing it to – to a certain extent – fulfill a human role.
Better scores on exams than most people
Exams test whether you have certain knowledge, and in that respect GPT4 appears to do better than most people who take an exam. This concerns, for example, an American legal exam for lawyers, and biology exams at pre-university education level.
Now an exam is something different from an actual task. By taking an exam you have to convince the examiner that you know something about a certain field. Usually you do that with words, and stringing words together that sound very sensible, GPT-4 is very good at that. The real test is practice, and here GPT4 is much more difficult to replace a human. This is only possible in areas that do not require deep insight, such as in sales conversations, writing letters on a specific topic or scheduling appointments.
Evaluate drawings by GPT-4
In demonstrations, GPT4 seemed to be able to generate a website from scratch using drawings, and evaluate a cartoon. Demonstrations say very little otherwise, only when we can see it with our own eyes can we say something sensible about it.
Unfortunately, it is not yet possible to upload images to GPT4 yourself. So we can’t say anything specific about that yet. Although it is plausible in itself that this functionality also does exist.
The following video is known of the demonstration.
