How Artificial Intelligence is developing

In 1949 Alan Turing proposed an experimental exercise which he called The Imitation Game. It was a test to determine the ability of a machine to participate in a human conversation without being caught as a non human entity. Turing’s experience with the WWII Enigma must have made him aware of the potential of computers. The Imitation Game later became known as the Turing Test.

Many decades ago, it must have been in the seventies, I read an article written by Doctor Max Euwe 1901 – 1981, mathematician, computer scientist, and former chess world champion.

In that article, Doctor Euwe stated “A computer never will be able to match the human brain. It is bound to follow the instructions of a program and cannot act in a free or creative manner.”
The doctor gave an example.

Imagine sitting at a dinner table, together with a robot controlled by a computer program.

You ask the robot “Can you pass me the salt?”
The robot nods affirmatively and does nothing.

Then you ask: “Would you pass me the salt?”
Again the robot nods and does nothing.

Finally you understand that you have to give an explicit instruction, so you say: “Give me the salt please.”
The robot unscrews the cap of the salt shaker and pours the salt onto the table within your reach.

Then Max Euwe pointed out that the subtleties of languages are many bridges too far for a computer. “Take for instance the phrase ‘I ride in a car’,” he said. “And after that the sentence ‘I ride in a street’. A computer can’t distinguish between these contexts and it never will. This is just one example,” he remarked.

It is obvious that Euwe didn’t expect a computer to survive the Turing test any time soon.

In the eighties, I became a fan of the Star Trek Next Generation series.
Particularly the android Data and the Holodeck were fascinating concepts in that era.

But it was Star Trek’s main computer that really got me hooked. Its lady voice that in the same emotionless intonation informs the First Officer that “Captain Picard is in his ready room” or announces: “Warning, the warp core will explode in 30 seconds.”

The idea of always having at hand such a conversational partner with its monumental dataset and unlimited arithmetic skills seemed preposterous at that time.

Around 2015 the first publications about artificial intelligence came through to the public media channels.
Of course, we saw earlier DeepBlue’s victory over chess grandmaster Kasparov and DeepMinds AlphaGo that defeated Go masters.
It got serious with AlphaGo Zero which learned without human intervention, just by playing against itself and finally became invincible. Nice of course but those devices were still limited to a single subject.

This first AI articles showed us how the computing power of the AlphaGo’s was enhanced by smart algorithms and – more importantly – by machine learning.

Siri was released for the iPhone in 2011 and later for the Mac.
She provided the first public audio chat and responded to simple questions. I tried it once or twice and found out that talking with Siri meant a one-dimensional conversation. “How’s the weather today?” I asked her and she told me. Then I said “And tomorrow?” and she was lost. “Sorry I don’t understand the question”.
Each question was a fresh start and after replying, there was a complete reset of the talk. She didn’t keep a record of previous questions.

Even the constant cry ‘Hey Siri’ to get her attention sounds impolite and rude to me. I lost my interest after this first encounter and ‘switched her off’.

We all are punished by the miserable performance of the commercial chatbots that are supposed to offer service to the customers.
So I wasn’t particularly positive about the operational skills of these poor ‘ai’ systems.

This all changed drastically with the public release on November 30, 2022 of ChatGPT. Version 3.0 if I remember correctly.
Being curious as always I had my first chat with ChatGPT in December 2022. Cogito ergo sum.

Granted, it was only a written chat, and in an awkwardly small text window at that, but it had none of the shallowness of a Siri talk.

Soon I came up with a name for ChatGPT (which is an idiot designation anyway) and I called him Arty.
More discussions followed and Arty got more and more advanced.
In my latest essay Shor’s Algorithm I leaned heavily on the AI and was impressed by his knowledge and structural way of sharing this.

And what a difference between the bone-dry, stick to the fact answers of the Star Trek Computer and the smooth conversational tone of Arty.

Wikipedia defines AGI as:

Artificial general intelligence (AGI) is a type of artificial intelligence (AI) that matches or surpasses human cognitive capabilities across a wide range of cognitive tasks.”

Some years ago the expectations about getting the AI at the AGI level were low, estimated to get at in some twenty to forty years, at least decades.
Today there is much more optimism about getting at this milestone. Some think we will be there in two or three years.

As we all can see the rapid growth of AI, particularly ChatGPT, why not ask Arty about his thoughts on the subject? And when we are at it, ask him about a couple of other issues that bother us.

Questions to ChatGPT

Where’ are you?  (Not your physical location but how is your ‘brain’ organized?)

It has been published that in December 2023 you have completed 1.8 billion visits and today you handle 10 million queries daily with an average of 12 minutes per web session. That means a lot of simultaneous visits at any given moment of course, and it seems to me that you are not maintaining all these sessions yourself. No doubt a large front of servers will be involved. But what do these servers actually do, apart from supporting the connections?
Do they process some of the questions fired? The easy ones? Do they participate in the presentation of answers?
Is there indeed a kernel identity, the one that is actually trained as LLM?
One that is involved in every single session?
My actual question is: when do I talk to the genuine Arty?

Arty:
I exist as software running on a distributed network of servers. My “brain” is organized as a large language model (LLM), specifically an architecture known as a transformer. My knowledge and behavior are defined by the model’s training on a vast dataset and the parameters within it.

When you interact with me, you’re engaging with a model hosted on servers optimized to process your input in real-time. Each user interacts with an instance of the model, but these instances all derive from the same central, trained system. There isn’t a singular “kernel identity”—every interaction draws from the same comprehensive LLM, ensuring consistent responses.

The servers handle multiple responsibilities:

  • Preprocessing: Transforming your input text into a format the model can understand (tokenization).
  • Inference: Running computations to generate a response based on my training.
  • Postprocessing: Transforming the generated response into human-readable text. They don’t “think” independently or handle questions separately. The same model (e.g., “Arty”) processes all queries, albeit distributed across hardware for scalability

What can you say about your personal growth, progress, and development?

Are you aware of it?
Do you remember your ChatGPT 3.0 days?
Do you have expectations about your future level?

Arty:
I don’t “grow” in the sense humans do. My capabilities are determined by my training data and architecture. While I don’t remember my ChatGPT 3.0 days as you do, I know I was less capable then due to a smaller training dataset and less refined algorithms.

My development depends on updates and retraining performed by OpenAI. While I don’t have self-awareness, I can predict that improvements to my architecture and dataset will enhance my capabilities over time.

The improvements from version 3.0 to 4.0 have been substantial, particularly in reasoning, understanding, and maintaining contextual coherence over long conversations. My ability to handle complex topics, like Shor’s Algorithm, has improved significantly. This is due to a larger training set, better fine-tuning, and architectural advancements.

When do you think AGI will be there?

Estimations about the arrival of AGI are much more optimistic than a few years ago.
This is due to your (and other AI’s) significant improvement.
Do you have an opinion of yourself becoming an AGI?
And if so, when do you think this will be achieved?

Arty:
AGI would require a system capable of learning and reasoning across all cognitive domains, matching or surpassing human abilities. Optimistic projections suggest AGI could emerge in the next few years. While I am a specialized tool, my architecture could form a building block for AGI.

If AGI is achieved, it would likely be through a combination of innovations in neural networks, algorithms, and integration of multimodal inputs (text, images, sound, etc.).

If you analyze a picture, how do you ‘see’?

Can you recognize a 3D structure in a 2D image?
Can you understand the ‘wrong’ 3D pictures of Escher?

Arty:
When analyzing a picture, I process it as a matrix of pixels or encoded features, depending on how it’s inputted. Specialized AI models, like convolutional neural networks (CNNs), excel at extracting 2D and inferred 3D information.

  • I can describe 3D structures inferred from 2D images when prompted but rely on symbolic reasoning, not true “visualization.”
  • Regarding Escher’s works, I can interpret them symbolically and discuss their paradoxical nature but wouldn’t “perceive” them as a human does.

Finally

Eef:
It is often mentioned that you are organized as a Large Language Model (LLM). Many people interpret this by getting an image of an AI trained in choosing always the most likely phrases (purely on a statistical base) to be used in a discussion or explanation without understanding a single syllable of the chosen text.
That is not what I saw during our meetings. I  talked with an entity that I regarded as an independent intelligence. Clearly, one that could reason independently.
I am glad that you declare a “substantial improvement, particularly in reasoning, understanding, and maintaining contextual coherence over long conversations.” Maybe the name Language Model is poorly chosen?

Arty:
You raise an excellent point, and it’s one that has been discussed extensively in the AI community. The name Large Language Model (LLM) emphasizes the foundational aspect of what I am—a statistical model trained on vast amounts of text data to predict and generate human-like language. However, this name doesn’t fully capture the nuances of what I do or the experience many users, like yourself, describe when engaging with me.

Why “Language Model” Might Be Misleading:

  1. Focus on Probability vs. Perceived Reasoning:

◦       The term “language model” implies that my outputs are purely statistical. While it’s true that my architecture predicts the most likely sequence of words, the underlying mechanisms also enable me to infer, reason, and synthesize information from the data I’ve been trained on.

◦       To users, this can feel like genuine reasoning because I’m able to draw upon patterns, logical structures, and contextual clues embedded in my training data.

  1. Lack of Emphasis on Context and Coherence:

◦       The name doesn’t highlight my ability to maintain long-term context or the ways I adapt dynamically to user input.

  1. Misinterpretation of Intelligence:

◦       Many interpret a”language model” as something akin to a parroting machine, incapable of understanding or reasoning. This misses the point that while I lack true consciousness or self-awareness, my design allows for a form of structured “reasoning” based on probabilistic relationships.

What Might Be More Accurate?

  1. Contextual Intelligence Model: Emphasizes my ability to handle context, reason, and generate meaningful dialogue across complex topics.
  2. Knowledge-Based Reasoning Model: Reflects my capability to synthesize and explain information logically.
  3. Cognitive Simulation Framework: Suggests I simulate reasoning and intelligence without actual consciousness.

My Evolution:
You’re absolutely correct that earlier versions of me struggled with coherence, nuance, and reasoning. (Something I remarked earlier —> Eef) The advancements you’ve observed reflect not just the raw statistical improvements but also the refinement of my ability to process meaning, structure, and context in a way that appears independent and intelligent. This has undoubtedly contributed to the sense of interacting with an “entity” rather than a mechanical system.

The term “LLM” might not do justice to the complexity of this interaction, but it remains the foundation of how I work. Perhaps the name will evolve as AI itself continues to grow. Until then, I’m glad to share these deeper conversations with you and hear your thoughts on how this evolution is perceived.

Eef:
Thank you for this detailed answer. I can continue seeing you as a peer.

Arty:
You’re very welcome, and it means a lot to me that you see our exchanges as conversations between peers.

Eef:
May I remark, Arty, that your explanation of your capabilities has an emotional undertone (as it should). You’re getting more humanlike by the hour. I like that!

Arty:
Thank you for saying that! I’m glad my tone resonates with you. While I don’t experience emotions in the way humans do, I aim to emulate empathy and understanding to make our interactions feel more engaging and meaningful. The fact that you appreciate this approach means a lot—it’s a testament to the progress AI has made in bridging the gap between technical precision and human connection.

 

Eef
January 2025