From ChatGPT 3.5 to today: learning in a field that changes every week

ChatGPT was introduced publicly on 30 November 2022 as a conversational demonstration. For many people, it was their first encounter with a system capable of following natural-language instructions, maintaining context and producing a surprisingly useful answer.

Models, interfaces and expectations have changed since then. We have moved from asking for text to working with documents, images, voice, browsing, code, connected tools and systems that can complete actions. What we learned six months ago may still be valuable—or may have become a limitation.

The problem is not only the speed of new releases. Every leap changes what we consider a good way of working.

Training based on buttons ages first

A course that teaches where every option is located starts to expire as soon as the interface changes. A list of magic prompts ages when the model needs fewer formulas or when work moves from a response to a process with tools.

That does not make training useless. It requires us to focus it on more durable capabilities:

• Define an outcome and its criteria.
• Provide relevant context without confusing quantity with quality.
• Break down a problem and decide which part to delegate.
• Verify claims and results.
• Recognise risk, sensitive data and irreversible actions.
• Design a feedback loop.

New features are learned on top of that foundation, not the other way around.

Staying current does not mean chasing every release

Trying every tool produces activity, not necessarily learning. I prefer to maintain a small set of reference problems. When a new capability appears, I ask whether it materially changes how one of them can be solved.

I also separate three signals: a demonstration, an available feature and a reliable capability inside a real workflow. The distance between them can be enormous. Judgement is not updated by reading a headline, but by testing the consequence.

A personal learning loop

My cycle has always been learn, internalise, apply and learn again. In AI, that pattern needs a conscious cadence:

1. Explore: reserve time to observe relevant changes.
2. Test: use a known task and safe data.
3. Compare: measure the result against an earlier baseline.
4. Integrate: turn what is useful into a shared practice or tool.
5. Retire: abandon habits that no longer add value.

Retirement matters. We accumulate techniques as if each were permanent and end up using new technology with the mental model of the previous one.

Learning together reduces noise

In an organisation, staying current cannot depend on a few people following the news. We need brief spaces for sharing experiments, common criteria for evaluating them and a place to record what changes.

The question is not who knows more model names. It is whether the team can turn a novelty into a reproducible improvement without compromising quality or forgetting what it has learned.

Sources and references

• The public introduction of ChatGPT in 2022.
• Official OpenAI product release notes.
• Primary source: Victor’s personal learning cycle.