Back in August 2020 I wrote here about the first machine-written text that sounds human. At the time, GPT-3 was a toy for developers, with a waiting list and programmer-oriented interfaces. This summer I tested DALL-E 2 and Midjourney and saw what these models can do with images. On 30 November 2022, OpenAI put the same family of technology into a chat window, for free, available to anyone. It is called ChatGPT, and for three weeks now I have been opening it almost every day.
I am not easily impressed by technology. I lived through the dot-com era, social media, the 2016–2017 chatbots that promised everything and understood nothing, crypto and the metaverse. But this time, for the first time in a long while, I had the feeling I was looking at something that changes how we work. Here are my impressions, with the enthusiasm and the caveats that come with them.
What it is, in short
ChatGPT is a language model from the GPT-3.5 series, fine-tuned specifically for dialogue. OpenAI launched it as a free “research preview”, a test version meant to gather feedback. The interesting part is how it was refined: real people rated the model’s answers, and the system learned from those ratings what a useful answer looks like (the technique is called reinforcement learning from human feedback). That is what sets it apart from “raw” GPT-3: you no longer need tricks to get the answer you want. You just write to it as you would to a person.
And people came. Sam Altman, OpenAI’s CEO, announced that ChatGPT passed one million users in about five days. Many of the social networks we use every day needed months or years to get there. The side effect: the servers are often “at capacity”, and that polite message asking you to come back later has already become an internet joke.
What I have used it for in these three weeks
I have treated it like a new assistant you are still testing: I threw everything at it to see where it shines and where it does not. In short:
- Email drafts: I give it the idea in three lines and get back a clean version that I then adjust.
- Article outlines: a skeleton of sections to start from, not the finished piece.
- Brainstorming: business names, marketing campaign ideas, messaging angles. Out of 30 suggestions a few are good, and that is already a lot.
- Explanations: a technical or financial concept explained simply, “as if to someone who has never heard of it”.
- Summaries of long texts.
- Small code snippets for simple tasks, a formula or a short script.
- Translations between Romanian and English, where it is surprisingly natural.
Nothing on this list is spectacular on its own. What is spectacular is that it does all of it, in the same window, in a few seconds.
What impressed me
First, the conversation. I can go back and say “shorter”, “less formal”, “now in English”, “keep only ideas two and four”. It remembers what we discussed earlier in the same chat and builds on it. That is the difference from a search engine: there you ask a question and get ten links; here you work on a text together with someone.
Second, the tone. It can be formal, friendly, ironic, write for a child or for an investor. For someone who has written marketing copy for years, I find that remarkable. And third, how quickly it jumps between topics: from an email to a business partner, to a recipe for cozonac (the Romanian Christmas sweet bread), to an explanation of interest rates.
When I wrote in 2017 about chatbots as the future of customer relationships, I said their problem was that they did not understand context. Here, for the first time, I feel that problem has started to be solved.
“Understands”? Let’s be honest here
I put “understands” in the title on purpose, so I could say this clearly: no, it does not understand the way we do. ChatGPT predicts, word by word, the most likely continuation of a text, based on a huge amount of text it read during training. The result looks like understanding, and is often as useful as if it were, but the mechanism is different.
The practical consequence matters: it can be wrong with perfect self-confidence. It makes up dates, quotes, book titles, code functions that do not exist. The industry calls this “hallucination”, and it is the biggest risk for anyone who uses it without checking. It gives you no sources, so you cannot directly verify where the information comes from. And its knowledge stops, broadly, in 2021: about anything that happened since, it knows little or nothing.
My rule so far: I use it for form, not for facts. Structure, tone, variations, ideas, yes. Numbers, dates and claims I sign with my own name, only after I have checked them elsewhere.
The data you feed it
Something few people think about: everything you paste into the chat window ends up on someone else’s servers. It is a product in a test phase, and conversations may be used to improve the model. I would not put contracts, customer data, internal figures or anything I would not email to a stranger in there. For founders and their teams, I think now is the time to set a few simple rules, before every employee invents their own.
The first reactions: Stack Overflow and schools
The signs that this is serious came quickly. In early December, Stack Overflow, the largest Q&A site for programmers, temporarily banned answers generated with ChatGPT. The reason: the answers look convincing but are wrong too often, and checking them eats up volunteers’ time. That is exactly the problem described above, at scale.
Meanwhile, teachers and schools are asking what a take-home essay even means when a machine can write a decent one in ten seconds. I do not think a ban is the answer. I think education will have to put more weight on thinking, on asking good questions and on verifying information, which are precisely the things the machine still does not do well.
What might change (cautiously)
I do not like making prophecies, especially after one month. But a few directions seem likely:
- Search: many people will prefer a direct answer to a list of links. For anyone who lives off organic traffic, that is a serious question.
- Content: there will be far more text, far cheaper. Value shifts to what cannot be generated: real experience, owned opinions, verified information.
- Customer support: repetitive questions can get good answers, provided the model is connected to accurate information about your company.
- Jobs: I do not think they disappear overnight. But I do think a person who knows how to work with an assistant like this will do the work of two or three people who do not.
There is also the part that interests me as an investor. In 2017 I wrote about what artificial intelligence can actually do in marketing. Back then you needed teams of specialists and lots of data. Now anyone has access to a very capable model through a chat window. That lowers the barrier for small startups, and for their competitors too.
What I would tell founders
Try it. Today, not in six months. Not to replace anyone, but to understand what it can and cannot do. Give it the boring tasks, check everything that matters and do not feed it sensitive data. And think, very concretely, about which part of your business is becoming cheaper for everyone, your competitors included.
I am three weeks in and still discovering something new almost every day. Have you tried it? Where did you find it genuinely useful, and where did it disappoint you?