Why OpenAI Says the Era of Giant AI Models Is Ending – What Comes Next?

Open AI's CEO says the ChatGPT development strategy has concluded, and future artificial‑intelligence progress will require new ideas.

In recent months ChatGPT has reignited public interest and investment in artificial intelligence, but last weekend OpenAI’s chief executive warned that the research strategy behind the chatbot is finished.

He emphasized the need for a collective search for new directions, acknowledging that no one knows where the next breakthrough will arise.

OpenAI has achieved impressive advances by scaling existing machine‑learning algorithms to unprecedented sizes, especially in natural‑language tasks.

GPT‑4, the latest of these models, was trained on trillions of words using thousands of powerful chips, at a cost exceeding $100 million.

CEO Sam Altman stated that the next step of evolution will not come from ever‑larger models, calling the current era of “giant models” the end of a chapter.

He told an audience at MIT that future improvements will come from other approaches that make GPT better without simply adding parameters.

Altman's remarks signal an unexpected turn in the race to develop and deploy new AI algorithms.

Since OpenAI launched ChatGPT in November, Microsoft has integrated the technology into Bing, while Google introduced Bard as a competitor.

Many well‑funded startups—including Anthropic, AI21, Cohere, and Character.AI—are investing heavily to build larger algorithms and models in an effort to catch up with OpenAI.

The original ChatGPT was based on a modestly upgraded GPT‑3, but users now access a newer version powered by the more capable GPT‑4.

Altman suggested that GPT‑4 may represent the final major release of OpenAI’s large‑model strategy, noting that the paper on GPT‑4 reports diminishing returns from scaling.

He also pointed out physical limits on the number and construction speed of data centers.

Nick Frosst, co‑founder of Cohere and former Google AI engineer, agreed that simply making models bigger will not always work, highlighting that transformer progress has outpaced scaling and that many improvements can be achieved without adding parameters, such as new architectures and human‑feedback optimization.

OpenAI’s language‑model series, each based on artificial neural networks inspired by neuronal cooperation, predict the next word in a text string.

GPT‑2, released in 2019, contained 1.5 billion parameters; GPT‑3 followed in 2020 with 175 billion parameters, enabling generation of poetry, emails, and other text, which spurred other companies and researchers to pursue even larger scales.

When GPT‑4 was announced, OpenAI did not disclose its exact data size, suggesting that scale may no longer be the primary factor.

During the MIT event, Altman was asked whether training GPT‑4 cost $100 million; he replied that the expense was “far more than that.”

Although OpenAI keeps GPT‑4’s data size and inner workings secret, its capabilities likely stem from techniques beyond sheer scale, such as reinforcement learning from human feedback, which guides the model toward higher‑quality responses.

GPT‑4’s remarkable abilities have shocked experts and sparked debate about AI’s economic impact, potential misinformation, and job displacement.

Altman also confirmed that OpenAI is not currently developing GPT‑5, stating that any early claims of a GPT‑5 in training are false and that the company will not pursue it for the foreseeable future.