Looking Back at AI: Past Experience and Present Reflections

In 2022, ChatGPT made GPT-3.5 and GPT-3.5 Turbo available to the public. These were models for general tasks, including translation and conversation. I remember that, at first, ChatGPT was just a novel chatbot. People quickly accepted it, then began to question what it could actually do. In 2023, I started using the Tab key frequently at work to complete code. Codex could only generate a line or two at a time, often with errors. Fully automated agents began to appear in 2024, but even by the end of 2025, few AI tools could handle entire files or projects. I used Copilot Chat extensively to generate functions and files, but the output needed considerable review and correction.

In early 2026, I started using OpenCode, Claude Code, Codex, and GitHub Copilot. As the models improved, coding agents developed rapidly. In the few months from late 2025 to early 2026, they became part of all my development projects. I still remember working on retrieval-augmented generation (RAG) in 2024. Many open-source frameworks handled model orchestration, chain-of-thought reasoning, and RAG. But improvements in models and coding agents soon took over these tasks. Who still remembers RAG now?

AI has had an enormous effect on my work. I work in a field where AI can replace a large share of human labor, and I was among the earliest to work with AI. I have felt these changes directly. I still remember the expectations at each stage: my own, those of society, and those of the industry. At first, I thought AI could become part of our infrastructure if it could solve the problem of hallucinations. Software could be orchestrated without conventional programming. But at the time, GPT had a context window of only 4,000 to 8,000 tokens. Even the code it generated contained errors. I tried feeding compiler errors back to it, with limited success. There was a large gap between what I experienced and what I imagined.

Today, I want to abandon linear thinking. Over the past three years, AI has helped me put many ideas into practice. I have used it to develop software and test ideas quickly. I have also seen enormous changes in model performance, scale, and price. And I do not think this is where it ends. Many AI products have a chat box as their interface. But I expect future demand for AI to come from entities that do not need that chat box: demand that arises without human participation. Companies such as OpenAI and Anthropic are supplying vast amounts of inexpensive intelligence that surpasses 99% of humans. We have moved from a shortage of intelligence to an abundant supply. The GPT-6 Astra I use today has a context window of 1.2 million tokens, dozens of times that of GPT-3.5.

What role will this intelligence eventually take? Consider a translator. Training one might take ten years, but a large language model can replace one at very little cost. The same applies to many office jobs. LLMs are a form of intelligence with economies of scale. People have to be trained one by one. AI does not.

Before AI, intelligence in software was roughly equivalent to computation, to a program. LLMs are also mathematical models and, ultimately, computation. Yet they have produced an unexpected way to handle the complexity and uncertainty of language and images. This intelligence can also expand more readily with data, including data produced on the internet and through interactions during use. Its role may go beyond helping one office worker do what previously required five people. It may develop from an agent into an entity with a status equal to that of a human: an intelligent entity that needs almost no human operation. It could work behind services, technology, and businesses. Its role would not simply be to replace a particular person. For now, there is only speculation that companies such as OpenAI and Anthropic may already have achieved recursive self-improvement (RSI). In other words, AI that began as a software system could eventually improve its own performance, rather than participate as a third party in existing software, applications, and services.

This is an enormous change. I see a bleak future for the job market. We may be only in the early stages of its disruption. AI may eventually become much more widespread than I imagine, much as Excel did. The difference is that proficiency in Excel can let you continue working as an accountant or in finance. Adopting AI is becoming unavoidable, but it does not guarantee you an office job.

If AI does not need to meet human needs, why would it need to keep running? Recently, OpenAI has produced many mathematical proofs. These are mostly cases where a proof was known to exist, but its complexity had kept humans from completing it. This demonstrates a possibility: AI could extend the boundaries of research in place of humans, working continuously at the frontier. There is more work at that frontier than we can meet. Another possibility is for AI to maintain existing software systems and services, with less and less human intervention. This signals the end of the software industry.

Looking back, what AI can do today already far exceeds what I expected three years ago. At the time, I imagined another killer app like TikTok. Today, AI can even exist without human users. The enormous changes continue. AI is still expanding in two directions: wider adoption and work at the frontier. An increasing supply of inexpensive tokens is allowing demand for models to grow faster. Large numbers of mathematical proofs have already demonstrated its ability to work at the research frontier. You and I are living through this change, watching AI gradually reach a point where human needs alone can no longer explain its purpose.