As soon as ChatGPT was launched on November 30, 2022, there was an avalanche of new generative AI models. Some of these were chatbots powered by LLMs, a variety of plugins, APIs, virtual assistants and copilot aids.

In 2023, we came across multimodal foundational AI models. The trend continued in 2024 — text-to-video sensation Sora (mid- February 2024 launch, Alibaba.s EMO).

The models have become sophisticated — greater parameter count, larger context window, greater computation efficiency. Gemini Ultra (Google) and Llama 2 wanted to outperform GPT-4. Anthropic and Mistral too wanted to do the same thing.

There was a quick replacement of models — Anthropic’s Claude 2.1 followed Claude 1.3 in four months. Facebook’s Llama 2 came three months after Llama 1.

AI models grow in size. There is huge data requirement. There is an increasing demand for data centers. It raises the cost of building advanced foundational models (FMs).

Open AI invested $4.6 billion to develop GPT-3 (2020). GPT-5, the proposed model could require an investment of $1.25-2.5 billion just to train the model. The cost of product failure could be amazingly high.

OpenAI has acquired a large customer base, and it retains the base. It has enlarged the base substantially. It has used easy to use plugins, virtual assistant apps and image generation models. ChatGPT has adopted for-profit strategy by introducing subscriptions. OpenAI is thus attractive for investors and for valuation experts.

Gemini’s setback has alerted others. Governments may ask organizations to take prior permission before the release of the model. The governments may expect certain disclosures and call them under testing if they are being evaluated.

A full-fledged use will be contemplated after the model is fully ready. There could be some changes in the training regimes. There could be increased scrutiny of unsupervised learning. Even supervised learning may have to comply with standards and protocols.