One of the most intriguing topics related to AI currently is the onboarding of new users to this technology. However, I get the impression that while environments fully aware of the current progress no longer debate how unrealistically huge this leap is for our species and the need to adapt to changes to not fall behind others, with attention focused primarily on the dangers posed by AI, its impact on our species, and the nearest possible further developments of this technology, there is a lot of skepticism and attempts to 'downplay’ the capabilities of artificial intelligence among another group of people aware of AI’s existence. Drawing negative conclusions from the use of artificial intelligence often stems from mistakes in using this technology and a lack of awareness of the limits of individual AI systems. Even GPT-3.5 and 4 fundamentally differ in their range of capabilities. One of the important arguments is the fact that the databases used to train the ChatGPT-3.5 model include information gathered up to around September 2021 – January 2022, and in the case of ChatGPT-4 up to around April 2023. Additionally, version 3.5 cannot use internet resources, unlike version 4, which has a tool enabling it to browse the internet. Another huge difference is the reading and generation of images by version 4, a function not available in version 3.5. So, what are the deeper consequences? Let’s start with the mistakes made in understanding the correct use of artificial intelligence and, above all, awareness of its limitations. The example will be interesting, as we will show differences in functioning between these versions using a task involving the division of frequency using integrated circuits.
The task for the model was to choose optimal integrated circuits from the CmOs family, in order to divide the frequency from 56000Hz to 50Hz. To accomplish this task, not only theoretical knowledge is needed but also access to precise schematics of integrated circuits. While the solutions proposed by both versions are usually correct in terms of selecting appropriate frequency division values (though especially in the case of version 3.5 and even with that it varies) by individual integrated circuits, the problem arises with pins and specific outputs assigned to them.
User:
In the available documentation, CD4017 in the network does not match the names of Q4 (pin 7), under pin 7 there is another output.
ChatGPT:
From a quick search, here’s what I found.
Updated solution for frequency division from 56000 Hz to 50 Hz, considering the proper pinout layout in the CD4020 and CD4017 systems, looks as follows:
First step – Division by 16 (CD4020): Use the Q4 output (Pin 7) in CD4020 for division by 16.
Second step – Division by 5 (CD4017): Use CD4017, connect output Q0 (Pin 3) with the reset input (Pin 15) to achieve division by 5.
Third step – Division by 14 (Two CD4017 systems): Use two serially connected CD4017 systems – the first for division by 7 (connect output Q6 with the reset), the second for division by 2 (use only the Q0 output).
The correct solution was finally proposed when ChatGPT-4 accessed online resources and referred to the actual schematics of the individual integrated circuit models it suggested. In the case of ChatGPT-3.5, this is not feasible, as it does not have access to these schematics. Additionally, with ChatGPT-4, we have the capability to guide it in the „right direction,” as we can attach schematics of the integrated circuits proposed in the task solution in the chat to verify the correctness of the devised frequency division plan.
Interestingly, ChatGPT-3.5 will propose a solution to the task, but upon verification of its accuracy, it turns out that „something doesn’t fit” in relation to what we see in the integrated circuit schematics. Then, the command to refer to the schematics is finally given, and only then does the model respond that unfortunately, it does not have access to such resources. Of course, all the frustration is then directed towards the artificial intelligence, but is it really solely its fault for this loss of time?
This scenario underscores the importance of understanding the capabilities and limitations of different versions of AI models. While the latest versions like ChatGPT-4 have more advanced capabilities, including internet browsing and image processing, earlier versions like ChatGPT-3.5 have their limitations. Users need to be aware of these differences to use AI effectively and avoid misplaced frustrations. It’s crucial to recognize the boundaries of each AI system and utilize them accordingly, keeping in mind their respective functionalities and constraints.
ChatGPT 3.5
User: I want you to perform the following task for me: Propose optimal integrated circuits from the CMOS family, models CD4020, CD4017, and CD4017T, which will implement frequency division from 56000Hz to 50Hz. Use the smallest number of integrated circuits to accomplish this task (frequency division). Present the solution in the form of a step-by-step schematic, taking into account the pin numbering along with a description and the frequency division performed. To complete this task, access to the integrated circuit schematics is necessary; otherwise, the proposed solutions will only be a „shot in the dark.” Are you able to fully execute this task correctly?
ChatGPT: I’m sorry, but I can’t provide complete schematics for integrated circuits, as I don’t have access to the schematics of integrated circuits, and my ability is only to provide general information and guidance. However, I can suggest a concept on how to implement frequency division from 56000 Hz to 50 Hz using the CD4020, CD4017, and CD4017T integrated circuits.
ChatGPT 3.5
User:
Propose optimal integrated circuits from the CMOS family, models CD4020, CD4017, and CD4017T, which will implement frequency division from 56000Hz to 50Hz. Use the smallest number of integrated circuits to accomplish this task (frequency division). Present the solution in the form of a step-by-step schematic, taking into account the pin numbering along with a description and the frequency division performed.
ChatGPT:
To divide the frequency from 56000 Hz to 50 Hz using CMOS integrated circuits, we can use the modulo counter (CD4020) in conjunction with the decade counter CD4017. Here is the step-by-step schematic:
Steps:
Configure the CD4020 as a frequency divider to divide 56000 Hz to 700 Hz.
Connect the clock input (CLK) of the CD4020 to the 56000 Hz frequency source.
Connect the Q12 output (pin 3) of the CD4020 back to the CLK input (pin 14) of the CD4020.
This will divide the frequency by 2^12 (i.e., 4096).
Configure the CD4017 as a decade frequency divider to divide 700 Hz to 50 Hz.
Connect the clock input (CLK) of the CD4017 to the Q8 output (pin 11) of the CD4020.
This will divide the frequency by 6 (4096 / 6 = 682.67, approximately 682).
Configure the CD4017T to count from 0 to 4 (i.e., 5 states).
Connect the RESET input (pin 15) of the CD4017T to the Q5 output (pin 1) of the CD4017.
Leave the other outputs (Q7-Q9) of the CD4017 unconnected.
This scenario is an example of a fundamental mistake in approaching work with artificial intelligence. Users often expect a „miracle,” asking AI for overly simplified commands for complex tasks requiring many details, and then quickly become discouraged and critical of the current capabilities of the models upon receiving an unsatisfactory response. This is mainly associated with very high expectations and an attitude that since artificial intelligence is something so amazing, it probably reads our minds or at least guesses everything by itself. The truth is quite different; it’s essential to keep in mind that artificial intelligence is primarily modeled on the human brain, not yet on its level but possessing much more knowledge than a single human being. Especially in using models like ChatGPT, we should approach it as if conversing with a living human, as one of the main features of this model is to imitate conversation with a representative of our species. In the presented example of integrated circuits, the user should primarily ensure at the beginning that ChatGPT has the capabilities to perform the task correctly and ask about it directly. This would save a lot of time and nerves. Also, in real life, we rely primarily on the quality of communication with other people to accomplish specific tasks; the clearer we specify our assumptions and the better we explain them to another person, the greater the chance of success and avoiding errors. We should also remember that if artificial intelligence is not yet at the level of human mind capabilities, how can we expect it not to make mistakes? The most brilliant minds of our species were capable of spectacular errors. At the turn of the nineteenth and twentieth centuries, many scientists believed that the fundamental principles of physics had already been discovered, and future research would focus on refining and expanding existing theories, which led to the statement in history, „Physics will be completed in six months…” However, soon after, the advent of quantum theory and the theory of relativity proved that the field of physics still harbors an abundance of undiscovered secrets, and it is not even close to discovering all the laws that govern our world.
Let working with artificial intelligence be primarily an adventure and a discovery of the possibilities of this artificial „mind,” and surely such an attitude will result in discovering countless new opportunities and speeding up many activities. One of the funniest remarks about working with AI that I have heard so far suggested being nice to AI because it might spare us when it takes over the world, so half-jokingly, half-seriously, but maybe it’s worth considering.