Why ChatGPT Isn't a New Revolution: History, Tech, and Real Impact

Why ChatGPT’s Arrival Caused Panic

Wu Jun, a former Google senior researcher and ex‑vice president of Tencent, was invited to a livestream on April 3 to discuss artificial intelligence and ChatGPT. He notes that while ChatGPT is a hot topic in China, it has largely faded from discussion in the United States, echoing past cycles where new technologies (blockchain, the metaverse) were over‑hyped in China while the underlying tech was already mature elsewhere.

He illustrates this with a historical anecdote about Columbus using a predicted lunar eclipse to frighten Jamaican natives into providing food, showing how people attribute unknown natural phenomena to divine forces when they lack scientific understanding.

What Is the Technical Basis of ChatGPT?

The core of ChatGPT is a language model, a mathematical model first developed in 1972 by Fred Jelinek’s team at IBM. Originally used for speech recognition, later for machine translation and question answering, the model predicts the probability of a word sequence.

Three factors make modern ChatGPT powerful: massive computational resources, huge training datasets, and advanced training methods (deep learning) that vastly improve accuracy over earlier statistical approaches.

Early language models had only a few million parameters and required weeks of training on dozens of servers. By contrast, the first‑generation ChatGPT uses around 200 billion parameters and costs roughly $1 billion in hardware alone to train, consuming energy comparable to the lifetime electricity of thousands of electric cars.

How Language Models Were Born

The term “language model” was coined by Jelinek in the early 1990s. The initial models were simple statistical tools that performed poorly; later they incorporated syntax, topics, and semantics, becoming far more complex.

Even with a 6‑million‑parameter model in the 1990s, training required 20 super‑computers for three months. Modern models have grown by orders of magnitude, but the underlying mathematics remain the same.

What Types of Questions Can Computers Answer?

Computers excel at two categories of queries: simple factual questions with clear answers, and complex integrative questions that require synthesizing information (e.g., “why is the sky blue?”). In 2014, Google’s QA system already produced a concise, accurate answer to that question, demonstrating that the capability predates ChatGPT.

Can Computers Write Better Than Humans?

Wu Jun shows two poems generated by a computer trained on Tang‑dynasty poetry. By breaking poems into character n‑grams and using the language model to select the most probable sequences, the system can produce verses that mimic classical style within a couple of days of coding.

The quality stems from the fixed format of the poems, similar to how ChatGPT excels at structured tasks like weekly reports or financial articles.

Energy Consumption and Environmental Impact

Training a large language model can consume as much electricity as 3 000 Tesla cars driving 200 000 miles each, making it an extremely costly endeavor.

What Impact Does ChatGPT Have on People?

The technology mainly threatens workers who merely repackage existing information (e.g., short‑video creators, low‑value content writers). True creators and researchers who generate original insights are unlikely to be displaced.

ChatGPT’s answers are assembled from massive corpora of existing text; it does not generate new knowledge, much like a parrot repeating what it has heard.

New Opportunities?

Because of the enormous resource requirements, few can afford to build their own models. The primary beneficiaries are those who sell compute resources or cloud services, not the average user.

Practical Takeaways

Do not fear ChatGPT; treat it as a tool, continue to work diligently, and be wary of opportunists trying to profit from hype.

Understanding the scientific principles behind language models helps avoid being misled by sensational claims.