My take on the AI bubble

Timeline

• End of 2022 – Release of ChatGPT and subsequent adoption by 100 million users in two months (LLMs).
• From GPT-4, pivot toward more “efficient” models (using the same amount of training data but achieving better results).
• 2024 – Training data recognized as a bottleneck (“training data is the fossil fuel of AI” – Ilya Sutskever).
• Continued scaling of model size and compute, along with new developments in training techniques.
• Late 2024 / Early 2025 – Adoption of “thinking models” (DeepSeek R1, OpenAI o1).
• Problem: Makes prompts significantly more expensive, which is a considerable issue for prompt-heavy applications like so-called “vibe coding apps.”
• GPT-5 performs below expectations.
• Primary AI providers (xAI, OpenAI, Anthropic) start pivoting to the secondary AI market.
• Secondary market companies start getting massively overinflated.

But:

• It seems people are recognizing that the “early adopters” (developers trying to apply AI in areas they know nothing about) maybe shouldn’t be the ones driving adoption.
• Instead, it should be people with deep industry knowledge trying to adopt AI.
• This will take longer but will undoubtedly bring revenue to the primary providers.
• So it’s basically a waiting game until different sectors figure out how exactly AI can help them grow revenue.
• What didn’t happen: AI didn’t suddenly turn the world upside down and revolutionize every economic sector.
• Instead, ChatGPT rapidly took on the role of the world’s internet search engine, information gatherer, essay writer, and coding assistant—but nowhere else did it gain rapid adoption.