Following our third AI Forum on “Chat GPT In Action,” we’re pleased to share “Towards the AI Transition,” an article by Šimun Strukan, the featured speaker at COTRUGLI Business School.
The Pandora Moment
As we stand at the threshold of a technological revolution, questions, fears, and bold predictions surround the AI transition. Some envision dystopian futures, mass unemployment, and even the end of humanity. While not entirely baseless, these fears can obscure the technology’s transformative potential.
Much like Pandora’s myth, where hope remained in the box after evils were unleashed, ChatGPT’s launch symbolized the opening of new possibilities. At its core, hope persists—offering us a chance to shape AI responsibly.
Math, Not Magic
Every human-made technology begins with mathematics. The same applies to AI. Neural networks, the foundation of today’s most powerful AI tools, were first theorized in 1943 by Warren McCulloch and Walter Pitts. Their work sparked an 80-year journey from theory to application.
In the 1950s and 60s, researchers at Stanford developed the first functional neural networks, ADALINE and MADALINE, which eliminated echo on telephone lines. Since then, the main hurdle to progress wasn’t ideas, but computational power.
ChatGPT-4, for example, has approximately 100 trillion parameters. Running it requires dozens of GPUs, each costing about $15,000, with daily maintenance costs potentially reaching millions. OpenAI keeps these numbers close, but the infrastructure demands are clear.
How ChatGPT Works
To understand ChatGPT, one must grasp basic AI terminology:
- Linear Algebra: Essential for neural network operations.
- Vectors: Numeric representations of words.
- Embeddings: Mapping words into high-dimensional space.
- Tokens: The smallest data units, such as words or punctuation.
- Training vs. Inference: Training adjusts model parameters; inference generates responses.
- Parameters: Weights the model learns to predict outcomes.
Despite its sophistication, ChatGPT is stateless. Each query is treated independently. Context is passed along with each prompt, but the model retains no memory. It operates as a statistical machine, calculating the probability of the next word in a sentence, one word at a time.
Critically, ChatGPT has no consciousness or emotions. It doesn’t “think” before responding but instead calculates outcomes based on training data—a kind of statistical intuition.
Storing Knowledge: An Analogy
Think of training ChatGPT like learning multiplication. You remember the results (e.g., 7×8=56), but not the exact moment you learned them. Similarly, ChatGPT retains what it learned during training but not the data itself. This distinction is crucial for understanding privacy and data use.
Real Dangers vs. Science Fiction
While headlines often hype superintelligence, the real, near-term dangers lie in cybersecurity. AI weapons are becoming powerful tools, not for kinetic warfare but for automated cyberattacks. Tools like STUXNET in 2007 showed the impact of such tactics.
Automated phishing tools now operate in the dark web, capable of targeting millions using AI-generated communications. These systems require no human oversight—only a cryptocurrency wallet and malicious intent.
One chilling possibility: combining generative AI with harmful content such as child exploitation. These are the threats that need urgent attention, more so than distant AI overlords.
Proceeding with Purpose
Despite valid concerns, AI still holds immense promise. We must not allow exaggerated fears to distract from real risks—or from AI’s potential to solve pressing global problems.
History shows that widely discussed fears often receive enough scrutiny to avoid catastrophe (e.g., nuclear war). Yet less visible threats can slip by unnoticed. The AI transition demands vigilance, clarity, and thoughtful governance—not panic.
Like Pandora, we must keep hold of hope while facing what AI unleashes. The future depends not just on what technology becomes, but on how wisely we choose to use it.
