GPT

GPT stands for Generative Pre-trained Transformer. This is the neural network architecture that OpenAI has been using since 2018. ChatGPT, GPT-4o, GPT-4.5 are all specific models based on it. When people say “ChatGPT” in conversation, they usually mean the interface, not the architecture itself.

In my practice, GPT-4o is the first tool that I try for any new task: writing a brief, analyzing a competitor, generating headline options for an A/B test. He copes well with the Russian language, understands the context of long dialogues and knows how to work with images - this is important when analyzing screenshots from advertising accounts.

Limitations of the GPT architecture that I notice in my work: the model confidently hallucinates numbers and dates if it is not stopped by a specific request. I usually ask the model to indicate the source for any number - this reduces hallucinations by a factor of three. Knowledge cutoff is also critical: GPT does not know events after its training date, which is important when working with current news or recent algorithmic updates.