Why ChatGPT Isn't a New Revolution: Insights from AI Pioneer Wu Jun

Why ChatGPT attracted public attention

ChatGPT became a hot topic in China while discussion in the United States has largely subsided. The surge of interest is largely driven by hype, similar to previous cycles around blockchain and the metaverse, which created opportunities for opportunists to profit.

Technical foundation of ChatGPT

ChatGPT is built on a language model , a statistical mathematical model for estimating the probability of a word sequence. The concept dates back to 1972 when Fred Jelinek and his IBM team introduced the first language‑model framework for speech recognition, later extended to machine translation and question answering.

Three factors make modern ChatGPT powerful:

Massive computational resources (thousands of GPUs, high‑performance clusters).

Enormous training corpora (hundreds of billions of tokens from the public web).

Advanced training algorithms (deep‑learning architectures, transformer models, large‑scale optimization techniques).

Evolution of model size and training cost

Early language models (circa 1990s) contained about 6 × 10⁶ parameters and required dozens of servers running for weeks to train. By contrast, the first publicly released ChatGPT version uses roughly 2 × 10¹¹ parameters. Training such a model costs on the order of US$1 billion in hardware and electricity, comparable to the energy consumption of thousands of electric cars or the lifetime energy of ~3 000 Tesla vehicles.

Types of questions a language model can answer

Questions can be categorized as:

Simple factual queries – e.g., “Why is the sky blue?” – which require retrieving known facts.

Complex procedural or explanatory queries – e.g., “How to bake a cake?” – which require synthesizing multiple steps.

Example answer to “Why is the sky blue?” generated by a Google QA system in 2014:

AI‑generated poetry and the underlying mechanism

ChatGPT can produce well‑structured text such as Tang‑style poems because the task has a fixed format. The model tokenizes input text into short word groups (e.g., bi‑grams or tri‑grams), computes the conditional probability of each possible next token, and selects the highest‑probability continuation. The process repeats until the desired length is reached.

Two poems generated in a two‑day experiment are shown below:

Impact on different types of work

ChatGPT does not replace high‑value creative or research activities. Its primary economic impact is on low‑value, format‑driven content creation (e.g., repetitive reports, simple news summaries). Professionals who generate original insights remain largely unaffected.

Economic opportunities

The main beneficiaries are cloud service providers and hardware manufacturers that supply the massive compute infrastructure required for training and inference. Individual developers or small organizations face prohibitive costs.

Energy and cost considerations

Training a large language model consumes energy comparable to the lifetime electricity of several thousand electric vehicles. Estimates suggest the training energy is equivalent to powering ~3 000 Tesla cars for 200 000 km each.

Key takeaways

Do not over‑interpret the hype; the underlying technology is a well‑studied statistical model enhanced by modern compute.

Focus on tasks that require genuine creativity or deep domain expertise, as these are less likely to be automated.

Be aware that the primary economic gains accrue to providers of compute resources rather than to end‑users of the model.