Artificial Intelligence: Regulation is hardly possible (Proof: concrete examples)

Powerful systems have a good and bad side, just like current AI systems do. Regulation of AI is often discussed, but I think it's no longer possible. The progress of artificial intelligence can't be stopped anymore. Concrete examples from my local AI system show this. By the way, "local" also means controllable data protection.

Introduction

No hype anymore, but a revolution. That's AI from now on. And that is probably the greatest technological upheaval in human history. Whoever thinks this is exaggerated probably knows too little about the current possibilities offered by KI algorithms.

In any case, numerous representatives from politics, economy, science as well as society have recognized that the possibilities of AI also bring a danger with them. This danger shall be dampened down, which is referred to as regulation.

Regulations can be enforced by governments, regulatory bodies or self-regulatory organizations. They may apply to various areas such as financial markets, environmental protection, healthcare, consumer protection or labor law. Regulations are intended to help prevent abuse, fraud, monopolies and other unwanted practices in order to ensure a fair and functioning system for all parties involved.

Definition of the term Regulation according to my requested AI.

In essence, I understand regulation as a legal framework, that is, binding provisions armed with sanctions.

In what follows, I explain why I consider regulation as no longer possible in order to freeze or even reduce current capabilities of artificial intelligence. Also, slowing down its growth would be difficult and not very useful.

Regulation of AI is no longer possible

Let's start with something positive. The abilities of today's AI algorithms, models and software packages are breathtaking. That is also said that it goes both about mathematical model as well as about software, datasets, fully trained neural networks and available computing infrastructures. That was not like this until recently. How often have I already heard, even from computer scientists: That existed already. Not true.

Here are a few applications that can be done with little effort and run locally on your own computer without having to give your data to Microsoft, Google, OpenAI or Amazon:

• Knowledge Assistants: Wouldn't you like to get information from your Intranet through a question-and-answer dialogue, just once? No annoying searching through directories, no unsatisfying search mask with few relevant keywords. Or how about a code search in your repository, using a natural language query command?
• Summary of texts: You can find some real examples in posts on Dr. GDPR.
• Keyword Extraction: I've also implemented this myself, as a local system. Both helpful for quick text comprehension as well as possibly for search engine optimization (as long as it's not yet based on semantic search algorithms)
• Identify answered questions: What questions does this post answer? And what is the answer in one sentence? You can find real examples in this blog.
• Translations in other languages: DEEPL does a great job. But maybe you want something of your own, cheaper or simply more flexible instead of a Black Box? Or they would like to provide their own API.
• Speech Processing: Transcription, Time Index Recognition, Translation, Speaker Recognition, Music Noise Filtering… Text after language and language after text, then translated into ten languages. How would that be? Already done and impressive results achieved.
• Video Generation: For example, generating an animated, quite realistic person that moves her lips in sync with a given text while speaking it.
• Generating Images from Texts: You're familiar with it from DALL-E or Midjourney. How about a system of your own? Keywords: Data Protection, Own Images, Higher Flexibility, Code Ownership
• Generating Text from Images: Image component recognition. What can be seen on the following image?

Input image (original image):

The AI has the task of recognizing and describing image components.

Here is the result:

The objects in the image were numerous recognized, namely: Person (2x), Wall, Curtain, Table, Bottle (multiple times), Paper, Bowl. A variant of the algorithm even recognizes the wine glass, which is cut off at the front. Another variant also recognizes the podium as such.

• Image+Audio for Video: Creating 3D objects (in video form) from 2D objects (images), combined with animation and synchronization. Here's a real example:

Input image (original image):

Textinput (partially generated by AI): Input-text (German) –> Output/Input-text (English) –> Output/Input-Audio (can be any language, I've just for fun generated an audio file from a text using an AI):

Computer-generated result (Video with animation and audio):

It's even funnier with an original voice of an American President. Here is the AI output:

From a two-dimensional image (plus audio input) we thus get a quasi four-dimensional result (3D animation + synchronous audio output). Perhaps you have noticed that physiognomy has also been taken into account.

The calculation of each video took about a minute (on a fairly good PC). The costs for these results are marginal (my above demonstration caused me no costs). I could give further examples in almost endless form.

These examples are not even the maximum of what is currently possible and can run on my systems locally and at any time. The examples serve only for illustration and should stimulate your imagination for what is currently realizable. You probably know the refined dialogues from ChatGPT. Something like that I (at least without great effort) can create in the form of a question-and-answer assistant, knowledge system or text analysis program anytime.

Intelligence is universal

So far there have been at most quite intelligent systems that were specialized in each task, but even then their performances were not really convincing.

A manager from IT told me that his company has been using a chatbot for several years now, working at a large German airline company, and therefore he doesn't see the AI revolution coming. Here's what this air travel chatbot replied to my complaint today:

My question:

How do I get a refund for the flight that was delayed yesterday?

Response of the Air Traffic Chatbot:

I'm sorry, I couldn't understand your input. Please either give me your booking code or your ticket number.

I find that quite weak and unsatisfactory. The result has nothing to do with artificial intelligence, but rather with artificial stupidity. In another test I wanted to know how I can get compensation for yesterday's flight which was not punctual. The chatbot replied "I have understood that you want to cancel a flight". That one can cancel a flight which has already taken place ("yesterday") is either a sign of special leniency or an incompetent chatbot.

The probably several thousand euros that are paid monthly for this service driven by a Microsoft product could have been spared.

In contrast here is a (simple) example of what current AI can do:

As input a short list of instructions is named, each followed by a technical function name (in bold print). Something like:

Listen to Manfred Mann's Earth Band album on Google Music: PlayMusic*
give me a list of movie times for films in the area: SearchScreeningEvent
I would like to go to the popular bistro in Frankfurt/Main: BookRestaurant
what is the weather like in Kassel: GetWeather

Now the AI system is being asked this:

I want to book a flight for New York:

System response:

BookFlight

The remarkable thing is that the determined technical function name was not even present in our input dataset list. Rather, the AI automatically recognized the schema of function names (e.g. "Book" + "Restaurant"). Furthermore, the AI recognized the semantics of instructions and performed a transfer performance from Bistro to Restaurant. Finally, the AI then generated the missing technical function name for the given instruction. It could not be done better by a human. At least it could not be done worse than the above-mentioned chatbot of the large German airline company.

You're familiar with the impressive dialogues that chatbots like ChatGPT can provide. If not, you'll find an example of a dialogue from an AI that is indistinguishable from an intelligent human here.

Post-processing complex algorithms

The Translation program DEEPL (from Germany!) to replicate has never been so easy and fast possible as now. With today's AI I get a local, expandable and adaptable system that works without third-party services.

Even apps like Shazam or SoundHound, which can recognize a song title when you hold your phone just a few seconds next to the radio, can easily be reprogrammed with the AI possibilities of Nowtime.

If you're not impressed now, that's your right. You are then someone who is either a spoiled consumer or completely clueless or very spoiled technician.

You can see for yourself indirectly by the number of examples I've given above and in recent posts about AI that it couldn't have been much trouble to generate these examples. How else could I give so many examples in such a short time? My time is too valuable for me to spend days generating mere demonstrations for posts like this just to convince some readers of the enormous capabilities of AI.

Characteristics of AI capabilities

As I have previously described, today's capabilities of artificial intelligence are very well available. Especially, AI is based on high-performance computers that anyone can afford. KI-models can be downloaded and further trained. Software toolkits are also available for download. Apparently, AI is freely available.

Let's summarize:

• Equipment: It's freely available. You surely know how to order a computer somewhere. If not: Rent yourself a computer, or even several.
• Infrastructure: It is freely available. You certainly know how to order network components and download corresponding software. Otherwise, rent or hire someone who will connect the loose ends.
• AI-Software: Free available or also rentable for money.
• AI-Models: freely available and can be improved on own hardware as needed. Alternatively, calculate a model yourself. To do this, however, you would need mass data. Ah, yes, there's also free of charge or can be produced by scraping the internet itself. Those who find that too tedious can use an existing AI to generate training data.
• Open Culture: Has developed over the years. The scientific community has always been concerned with an open exchange. Platforms like Github promote the exchange of source texts.

The mere mention of an open culture is actually enough as an argument why regulation of AI is no longer possible.

If AI is going to be regulated soon, I've already downloaded all the necessary software parts at my place anyway. A powerful computer infrastructure is also available to our company.

What is supposed to be regulated, anyway?

Those who demand regulation of artificial intelligence should at least get specific. What is to be regulated? Certainly not artificial intelligence in general, because then numerous everyday applications would have to be shut down. I wouldn't mind if spying agencies named Amazon Alexa or Google Nest were banned.

Algorithms cannot be regulated.

The sale of computer hardware or its possession cannot be regulated.

The use of existing and already available AI models cannot be prohibited.

Applying intelligence (science, research, private entrepreneurs, hobby researchers and programmers) can be made difficult, but not prohibited, at least not in Germany.

Conclusion

Much rather data protection authorities should take care that unlawful data deliveries to Google and other companies by German responsible persons do not continue to occur. Example: Web tracking. Fines: 0 Euros in Germany. Also some German courts could develop a bit more enthusiasm for dealing with data protection cases so far that they let the problem come to pass for the affected person and this one an effective data protection.

When that's done, we can start talking about AI-things that most people don't know what they are. One thing is certain: AI is still more technical than any kind of internet things. But that doesn't mean there won't be advisors who have no idea about it, coming from a completely different field, but still appearing as experts for AI or think they can deal with AI legally, although they haven't even understood the basics.

Very soon there will be further contributions on artificial intelligence at this point. You will especially be informed about the numerous possibilities and served concrete examples. My blog will be searchable by a question-and-answer system, instead of just searching for keywords. If it's not too much work, I'll deliver the answers perhaps also as summarized texts. The summary can not only be extractive (quoting individual sentences), but also abstractive occur (representation in one's own words).