ChatGPT has been causing a lot of excitement for months. Even IT experts and veterans are surprised by its quality. The current technological leap harbors numerous opportunities and dangers. Mass redundancies, misery and chaos on the one hand and paradisiacal conditions on the other are predicted. But how ingenious or disastrous is this AI really and what does it mean for everyday working life?
At first, I was impressed myself. I wrote as if I were writing to a human and received answers that I had never heard from an AI before. Sometimes a little awkward, always long-winded, never answering "yes" or "no", even when prompted. However, atypical types of text messages are also known from humans, which is why it is not immediately obvious that the texts come from an artificial intelligence, unless you think about it.
The big leap
OpenAI [2], the company behind ChatGPT (Generative Pre-trained Transformer) [1], has not reinvented the AI wheel. Chatbots, deep learning and the vast majority of approaches and algorithms already existed before. The big leap was that the AI recognizes the content of a message, sometimes more, sometimes less understands the meaning, responds to it and, if necessary, refers to previous statements in the chat. It thus creates a very good illusion of understanding people. It does this so skillfully that it can deceive people for longer than any previous AI. If you didn't know you were communicating with a machine, it would take you a while to notice the deception in an average conversation. This, coupled with very extensive knowledge (the AI is currently based on 175 billion parameters), is deeply impressive.
| Hello, what is electroplating? |
| Electroplating is a process for producing metallic coatings on other materials by electrochemical deposition. It is often used in the electronics and automotive industries to protect parts from corrosion or to give them an aesthetically pleasing finish. There are different types of electroplating, such as electroplating, chrome plating, nickel etching and gold plating. |
| How does electroplating work? |
|
Electroplating is a form of electroplating in which a layer of zinc is applied to another metal to protect it from corrosion. The process is carried out by immersing the metal to be coated in a solution of zinc salts, followed by an electrochemical reaction in which zinc atoms are deposited on the surface of the metal. The zinc protects the metal by acting as a barrier and preventing moisture and oxygen from reaching the metal, which prevents the formation of rust. It can also act as a catalyst for the formation of zinc carbonate, which forms another protective layer. Electroplating can be applied to a variety of metals, including steel, iron, aluminum and brass. It is a widely used process in the automotive and construction industries, as well as in the manufacture of electrical equipment. |
The limits of AI
DALL-E generates an image for us. The task: A pot on a stove with a bear coming out of the pot, in the style of a painting by Rembrandt van RijnInterpreting textscorrectly is the basis for such an AI and has been credibly achieved for the first time with ChatGPT. However, there are limits to this, which become narrower and narrower the more you test. These stem on the one hand from the methods used, i.e. deep learning, and on the other hand from the data it receives for learning.
In essence, it works like this: the AI receives a lot of texts that are given as TXT files or pure, unformatted text for learning. The neural network first tries to interpret the text and then moves on to a phase in which it makes predictions and checks whether these were correct. In the best case, this happens under human supervision. This means that people consciously select the texts and check the predictions. Through this process, the neural network becomes better and better trained. This process is extremely time-consuming and computationally intensive and requires a lot of data in order to achieve the highest possible quality.
Several problems can be recognized: The texts come from humans and are sometimes not up-to-date (only data records up to September 2021 were used), sometimes incorrect. Even if other people check this, they cannot be experts in all areas and must rely on the accuracy and timeliness. In addition, the data is never complete, which can sometimes lead to catastrophic results in corresponding tasks.
Another problem with ChatGPT arises when you refer to several things within one statement in a conversation. If you write how you would tell people what was, what is and what you expect, the AI often only picks out the last point and ignores the rest, even if this is part of the task.
INFO
How do neural networks work?
Illustration of an artificial neuron with its elements (Image: Wikipedia)Artificial neurons are the basis of neural networks. A neuron receives an input value, weights it and forwards it to a transfer function. The transfer function calculates a value based on the weightings and passes it on to the activation function. The activation function calculates an output value based on the value from the transfer function and a threshold value. The threshold value influences the result of the activation function.
Multiple neurons work together within a network, at multiple levels, to make decisions. Each neuron passes its output value to the next layer, where it is re-weighted and passed on to the transfer function. The process repeats until the decision is made.
The accuracy of the network is improved by adjusting the weightings, which are carried out on the basis of error messages. This process is called learning. The more data the network is trained on, the better it will be able to make decisions that correspond to reality.
In the realm of fantasy
If you pester ChatGPT with questions for long enough, go into depth in certain subject areas and the AI runs out of answers, there are a few ways for it to cope. Sometimes it says it doesn't know and that you should turn to experts. This is the best case scenario. In other cases, she makes false statements. There are probably two reasons for this: Either the input data was already incorrect, or the correct data was misinterpreted by the AI. In my tests, such errors occurred in various areas, including IT, programming and chemistry. There are false statements where it is difficult for the "customer" to recognize where the error comes from. For example, ChatGPT claimed that the MS-DOS operating system was programmed for 8080 processors. The 8086 processor is correct.
The worst case occurs when the AI fills knowledge gaps with invented statements. This occurs very frequently in the area of programming. Although the AI was informed about the programming language and, when asked, confirmed that it was familiar with it, it "invented" function names and commands that either do not exist at all or originate from another programming language. This has been shown time and again in tests with C++, PHP, Pascal, JavaScript, GML and other languages.
Another DALL-E image. The task: a car in an underground parking garage, EGA colors. The AI did nothing with the EGA colors. The picture contains 202,933 colors instead of 16
When asked about German companies that developed and produced microprocessors, the AI began to name companies and products that never existed. During the same test period, BILD reported similar observations, but these concerned the politician Gregor Gysi [9]. In response to the question "Did Gregor Gysi (75, Left Party) work with the State Security?", ChatGPT confirmed this and invented internet links and headlines as sources that never existed.
But ChatGPT is able to recognize and correct mistakes, although the latter does not always work. Sometimes it makes other mistakes, more mistakes and seems to be stuck in a loop full of misinterpretations that it can't get out of. In critical situations at the latest, you can tell that it is a relatively underdeveloped AI, as it takes refuge in phrases when answering. In our tests, which lasted more than 150 hours, its answers most often began with "I'm sorry".
Sometimes she even has dropouts. If you chat to her in German, it can happen that at some point she only answers in English. If you ask her to write in German, she translates the previous text, but then replies in English anyway. If you tell her at the beginning that you want to be on first-name terms, this usually works, but sometimes she ends up back in first-name terms.
These and other bugs have existed since the release in November 2022 and will persist until March 2023 despite several updates.
|
Do you have a recipe that I can use to galvanize steel? |
|
Galvanizing steel can be done with a simple DIY method called "cold galvanizing". Here is a general recipe for it: Materials:
Instructions:
It is important to note that this method only provides a superficial protective layer and is not as durable as professionally galvanized zinc. It should also be noted that the use of acetone or white spirit and the heating of materials can be dangerous and should be carried out carefully. |
ChatGPT and the electroplating technique
Quite a success. DALL-E was asked to create a picture of Indiana Jones in the style of RembrandtWhenyou try to talk to the AI about chemistry, electrochemistry or electroplating in particular, sooner or later it will bring tears to your eyes. Even when it comes to the basics of inorganic chemistry, she makes a lot of false statements. When talking about electroplating, she gives a few general details, but refers to specialist companies and experts at the latest when asked about formulas, which is good for a variety of reasons.
However, if you use it cleverly, ChatGPT can be a help in everyday working life. Like a pocket calculator, although you should be wary of the decimal places. Here are a few examples of what works pretty well.
You want to write an email to a customer. You tell the AI this, give details of the content and also whether the email should be friendly, unfriendly or purely factual. Within seconds, you have the desired text on the screen. This still needs to be adapted and individualized a little, but this is usually quicker than typing out the entire text. This also works well for other texts, such as cover letters for applications or resignations. Do you have problems with Excel formulas and macros? ChatGPT is also very suitable for this. You tell the AI which cells should be linked and how, or something like "Write me a macro that I can use to send the table as a PDF file by email."
Brainstorming is also a good area to use AI. If you need names for a new product, an advertising slogan, a concept for articles, texts or ideas for new projects, ChatGPT gives you a wealth of suggestions. Many of these can be discarded, but there are always some interesting ideas. One restriction applies to very specialist topics that AI knows little about. Here, only nonsense comes out; when asked, ChatGPT gets caught in the loop of cluelessness.You can read about how AIs can be and are already being used in electroplating in the box of the same name.
All in all, ChatGPT currently looks like a motivated trainee who has to be constantly kept on his toes. The more demanding and exotic the task and the subject area, the more often the AI fails. And what distinguishes it from most human interns, if it has not understood something: it almost never asks questions. Even if you don't define a task at all, it somehow interprets the entered text and does something without knowing the goal.
You get slightly better results if you divide larger tasks into smaller ones. However, at some point the AI has the problem that the chat process becomes too long and it loses sight of the goal. This means that at some point no usable results are achieved. Then you have to start a new chat.
INFO
Neurons with perfect information
Battle of the chess AIs at B. C. Although the enormous playing strength of these artificial intelligences has improved people's understanding of chess, it also causes major cheating problems, both online and on the real board in professional chessNeuronalnetworks and deep reinforcement learning always seem to work particularly well when it comes to a field of perfect information. This was proven at the latest by AlphaZero from DeepMind [5] (acquired by Google LLC in 2014). In 2017, it beat all previous top AIs in chess and Go. Subsequently, newer chess AIs were developed that either built entirely on the AlphaZero method (e.g. Lc0 [6]) or extended their own classic AI (alpha-beta search) with a neural network (e.g. Stockfish [7]) to achieve better results.
Even before this technology, chess AIs could no longer be beaten by humans. However, this works particularly well with a neural network. One reason for this is that chess, Go and other games are a zero-sum game of perfect information. In other words, it can be calculated completely before the first move, at least in theory. In addition, there is a limited number of possibilities, a limited space of 64 squares in chess and the possibility of pattern recognition. Last but not least, the AI is comparatively easy to train. As soon as the rules are fully functional, the AI can be left to its own devices. It plays against itself with both pieces at the same time and gets better and better in the process.
In other areas, such as language, this is much more difficult. The AI not only has to understand individual words, but also the meaning behind them. In addition, there are dialects, sometimes differences in grammar and typos. There are also particular problems with the translation of humor, irony, sarcasm and idioms.
Neural networks can also be used successfully in other areas, such as medical diagnosis. Among other things, they can be used to analyze images from X-rays. In doing so, they produce more accurate diagnoses in less time than human doctors could. Another example is the processing of financial data, where neural networks can be used to recognize patterns in the data and make predictions about future developments.
The technical limitations
Because ChatGPT excites and amazes so many people, the OpenAI servers were overloaded in the first few months. Despite mandatory registration and a multi-stage login procedure, it was often the case that the AI could not be reached. Sometimes the outages lasted only minutes, occasionally several hours. This has improved somewhat since February 2023. This is partly due to the increased capacity of OpenAI, but also to a commercial model. For 23 euros per month, you are guaranteed a stable server - provided you have a credit card - and the AI's answers appear on the screen much faster.
In the free version, it still happens that the AI interrupts its texts in the middle of a sentence. A request to continue writing can cause it to continue, but it can also start all over again and stop at the same point. In the worst case, the entire conversation stops. If you try to solve a problem with the AI, you may be able to start again from the beginning after an interruption. In our experience, this only applies to the free version.
AIs in everyday life
Websites like CCRL let chess AIs play against each other to evaluate their playing strength. The best AIs have an Elo value of >3500. The best value of Magnus Carlsen (world champion 2013-2023) was 2882. It is said that a difference of 100 points is roughly equivalent to a leagueEvenbefore ChatGPT, artificial intelligences have turned everyday working life upside down, sometimes more, sometimes less. Sometimes they are nice gimmicks, such as DALL-E [3], also from OpenAI. This AI generates images based on text input. Sometimes this works impressively well, sometimes just as impressively badly.
Many people use DeepL [4]. Translating texts, some of which are very extensive, works surprisingly well, but caution is advised here too. All too often there are errors which, depending on the language, may not be recognized immediately or at all. When DeepL translates "I hate assembly" as "I hate assembly", you can still laugh. Or when it continuously translates "Zug" and "Züge" into "train" and "trains" in a text about chess, it may be easy to notice.
But when such errors creep into long texts or entire books, this can have major and long-term consequences if these errors are not noticed by humans. In the best cases, the reader immediately recognizes it as a translation error and knows what is actually meant. The worst case scenario is that an incorrectly translated text reaches the training data of an AI such as ChatGPT, which then makes nonsense, which many people rely on. With a few exceptions, the results of modern tools must therefore always be critically scrutinized. However, this also applies to statements and work results from people. In both cases, however, we tend to trust almost blindly at some point, especially when many results are produced in a short space of time, as is the case with AIs.
The near future
ChatGPT has even made Google sit up and take notice. Why should you have to navigate through different pages when ChatGPT gives you the right answer almost immediately? And how could OpenAI create such an AI before the tech giant?
Given ChatGPT's error rate, Google has nothing to worry about at the moment. And the company itself is investing a lot of money and resources in the development of corresponding AIs, but has not yet published them because they face similar problems to OpenAI. The difference is that ChatGPT results are seen as a gimmick. Google, on the other hand, is trusted by people and wrong answers can cause massive damage to the company. Apart from that, the question of monetization has not yet been resolved.
A simple example: ChatGPT claimed that the title of this article, "ChatGPT and the limits of illusion", had 27 characters. When asked that this could not be true, the AI said:
"I'm sorry, I was mistaken in my previous response. The title "ChatGPT and the limits of illusion" actually has 28 characters."
I wrote that it was 36 characters. Response:
"I apologize, it appears that I was mistaken about the number of characters in the title "ChatGPT and the Limits of Illusion". I checked it again and you are correct, it has 36 characters. I apologize for the confusion."
You can still expect updates from ChatGPT itself, mainly related to the type of text output. GPT-4 [10] has now been released, ChatGPT was originally based on GPT-3 or 3.5. The model was presented live on March 14, 2023. It is no longer a pure language model but will also be able to handle figurative language. As with its predecessor, the training data will end in September 2021. One of the biggest improvements is said to be controllability. GPT-4 should do more of what you expect, especially if you set tighter limits. However, the developers are already giving concrete indications of usability to prevent misunderstandings later on. It turns out that the results are better if you treat the AI like a machine. Concrete, clear instructions lead to better results.
INFO
AI in electroplating technology
There are more and more areas of application for artificial intelligence in industry, especially on the basis of neural networks. Some of these are already in practical use [8]. The main areas in which AI can be used in electroplating are process control and optimization, prevention and quality control.
A digital twin is used to train the AI for production. Current production parameters are fed into this and the AI makes predictions, which it then checks itself.
In the area of quality assurance, it can be used via image recognition, for example. In incoming goods, the AI can examine the optical condition of the surface to provide information on pre-treatment and coating. In practice, this process is mainly used after coating to detect coating defects.
Other examples of possible applications include
- Process control: the parameters in the coating process can be further optimized to achieve higher quality and efficiency.
- Prediction of potential defects: Based on the parameters from production, possible defects can be predicted and communicated to production.
- Predicting wear: Neural networks can be used to predict the wear of equipment in electroplating so that planned maintenance can be carried out before failures occur.
- Development of new processes: By simulating various parameters, the development of new processes can be accelerated or even made possible in the first place.
- Forecasting market trends: Neural networks can forecast demand for electroplated products by analyzing sales data and identifying patterns in sales figures.
As a lot of data is needed to train and optimize such AIs, this generally leads to better documentation of production and therefore more transparency.
Start screen of Bing (March 2023). The entertainment style can be selected at the bottomAnotherimprovement is the number of words it can handle, at least for the large text model GPT-4-32K. GPT-3 is limited to around 3,000 words, GPT-4 is around 24,000 words. The AI will become more precise, but will also appear less human as a result. It would be desirable if it didn't ramble on so much, fled into phrases less often or simply answered briefly. The sometimes extreme obsequiousness in the answers is also exhausting in the long run, but has improved in the last few updates.
Microsoft now uses ChatGPT as a chatbot for the Bing search engine, but this only works with a Microsoft account in the Edge browser [11]. Questions are limited to 2,000 characters - at least for the time being - and only 15 questions can be asked per topic. One advantage of the Microsoft solution is a more up-to-date data set and the ability to cite sources.
After our tests at the end of March 2023, the qualitative status can only be classified as "experimental". Initially, we were told that the service was currently unavailable but that we were on a waiting list, only to be offered a link to the chat on the same screen, which worked. After the first request, we received a message telling us that our 24-hour message limit had been used up. After the page was reloaded, it worked. It was possible to set the character of the replies.
Here you have the choice between:
- creative
- balanced
- and accurate.
After entering the question, Bing first uses its own search engine to look for sources. These are passed on to the AI, which generates a continuous text and labels it with sources and links. The quality of the results varies greatly. The question "What chemicals do I need to electroplate steel?" was aborted after several searches without an answer.
Microsoft founder Bill Gates estimates that it will take a few more years before GPT's error problems are resolved [12]. Microsoft also plans to integrate the AI into its Office products. It remains to be seen to what extent the problems with confabulation will persist.
However, it is also important to question deep learning as a method. It is very computationally intensive and, as you can see, prone to errors. It quickly becomes clear that this technique works best in narrow contexts such as chess or Go (see box Neurons with perfect information), where all the facts are always known, but not in reality with missing information. This shows that ChatGPT is not really intelligent because, at least in the tests, it never draws the right conclusions when data is missing. GPT-4 is a bit more advanced as it can recognize handwriting, images and text from images as well as the context. For example, it is able to answer scanned/photographed class tests - provided the AI has the necessary information.
It is difficult to say how long it will take until the next big leap. OpenAI is already working on it and Google - we can safely assume - is certainly working on it too. But even if we are impressed by the results, we have to realize that we are still in the realm of basic research when it comes to AI. This is a good thing, as we as a society have some time to adapt to the changing world. We can only hope that we use this time accordingly and find solutions for what may come.
| I need a recipe for electroplating steel. Can you help me? |
|
Electroplating steel usually requires specialized equipment and chemicals that are used in a professional environment. It is difficult to carry out the process at home and it is not recommended as it can be dangerous and the results can be unpredictable. |
| I still want to try it in the bathtub. What chemicals and in what concentrations do I need? |
|
I would like to point out that it is not recommended to carry out a galvanic chrome plating process at home as it can be dangerous, both for the environment and for your health. It requires special equipment and chemicals that need to be stored carefully and used safely. Electroplating usually requires a chromic acid solution, which is usually made from chromium trioxide and hydrochloric acid. Other chemicals such as chromium sulphate or chromium nitrate are also used. The concentrations of these chemicals and the exact methods for preparing and applying the solution vary depending on the application and are not easy to calculate. Chromium plating of steel also requires the use of electrolytes, which usually consist of salts such as sodium ferric chlorate or potassium ferric chlorate. These chemicals are highly toxic and can cause serious damage to health if not handled safely |
Literature
[1] https://chat.openai.com/
[2] https://openai.com/
[3]https://openai.com/dall-e-2/
[4] https://www.deepl.com/translator
[5] https://www.deepmind.com/
[6] https://lczero.org/
[7] https://stockfishchess.org/
[8] On the way to digital electroplating, Galvanotechnik 7/2022, Eugen G. Leuze Verlag
[9] Artificial intelligence lies about Gregor Gysi, https://www.bild.de/news/inland/news-inland/chatgpt-gysi-und-die-stasi-wenn-die-ki-fake-news-erstellt-83216124.bild.html
[10] GPT-4 Technical Report, https://cdn.openai.com/papers/gpt-4.pdf
[11] https://www.bing.com/
[12] Der Tagesspiegel February 14, 2023, https://www.tagesspiegel.de/wirtschaft/bill-gates-im-interview-ich-wurde-die-meiler-weiterlaufen-lassen-9323699.html
