Key takeaways
- GPT is the model; ChatGPT is the chat product built on it, with memory, tools and safety layers around it.
- Generative means it produces text, pre-trained means it first learned from a huge corpus, transformer is the neural network design it uses.
- GPT actions let a custom GPT call an external API you describe with an OpenAPI schema.
- GPT-2 hidden states can feed a classifier, but purpose-built sentence-embedding models usually work better for the same effort.
- AI data centres use water mainly for cooling; how much depends on location, season and cooling design.
GPT in ChatGPT is the large language model underneath the chat app. GPT stands for Generative Pre-trained Transformer: a neural network that generates text, was first pre-trained on a huge amount of text, and uses the transformer architecture. ChatGPT is the product built around those models, adding a chat interface, memory, tools such as search and code execution, and safety systems. This page explains each word in GPT, which GPT powers Copilot, what GPT actions are, whether GPT-2 embeddings work for classification, how GPT uses water, and what GPT can do with video.
What is GPT in ChatGPT?
GPT is the family of large language models made by OpenAI, and ChatGPT is the chat application that runs on them. When you type a message, ChatGPT sends it, with your conversation and instructions, to a GPT model, which generates the reply.
The distinction matters when you build things. ChatGPT has features the bare model does not: conversation history, memory across chats, web search, file uploads, image generation and a code tool. Developers who use the OpenAI API get the model and must add those pieces themselves. ChatGPT also chooses which model handles a request in some modes, routing simple questions to fast models and hard ones to reasoning models. So "ChatGPT said X" really means a GPT model, plus ChatGPT's tools and settings, produced X.
What does GPT stand for?
GPT stands for Generative Pre-trained Transformer. OpenAI introduced the name with its first GPT model in 2018, and every later generation, from GPT-2 and GPT-3 to the current models, keeps it.
The 2018 paper showed that a transformer pre-trained on a large body of unlabelled text, then fine-tuned for specific tasks, beat models designed for each task separately. GPT-2 in 2019 showed that a larger model could do many tasks without task-specific fine-tuning. GPT-3 in 2020 scaled this further and made prompting, simply describing the task in text, a practical way to use a model. ChatGPT, launched in late 2022, added instruction tuning and human feedback so the model follows requests conversationally.
What does GPT stand for in the context of AI?
Each word describes a design choice. Generative: the model produces new text one token at a time. Pre-trained: it first learns general language patterns from a massive dataset before any specific use. Transformer: the neural network architecture it is built on.
In more detail:
- Generative: it predicts the next token given everything before it, and repeating that prediction produces sentences, code and answers.
- Pre-trained: the expensive learning happens once, on a large corpus; later steps such as instruction tuning and fine-tuning adapt it cheaply.
- Transformer: an architecture from 2017 built on attention, which lets every token weigh every earlier token, so the model can use context across long passages.
GPT models are decoder-only transformers: each token attends only to earlier tokens, which suits generation.
Which GPT does Copilot use?
Microsoft Copilot runs mainly on OpenAI's GPT models, and Microsoft has moved it to newer GPT generations as they launch, including the GPT-5 family. It chooses the model for each request automatically and also uses Microsoft's own models for some features.
Microsoft 365 Copilot similarly uses GPT models hosted in Microsoft's cloud, and Microsoft has added models from other providers in some features. GitHub Copilot is different: it has a model picker with options from OpenAI, Anthropic and Google, and new models are announced on GitHub's changelog. Microsoft rarely promises a fixed model for a feature, because it updates the service continuously. For most users the model name matters less than the grounding: whether Copilot is answering from the web, your files or your code.
What are GPT actions?
GPT actions let a custom GPT in ChatGPT call an external API. You describe the API with an OpenAPI schema, and the GPT decides when to call it, fills in the parameters from the conversation, and uses the response in its answer.
For example, a college events GPT could have an action get_events(date) that calls your events API, so a student can ask "what is happening on campus on Friday?" and get live data instead of a guess. Actions support authentication with an API key or OAuth, so a GPT can act for a signed-in user. They are the no-code end of tool use. For richer apps with their own interface inside ChatGPT, OpenAI's newer Apps SDK, built on the Model Context Protocol, is the more capable route.
How do GPT actions work?
When you build a custom GPT, you paste an OpenAPI schema describing your endpoints, set authentication, and write instructions about when to use them. During a chat, the model decides to call an endpoint, ChatGPT makes the HTTP request, and the model reads the JSON response.
Also asked as: how do gpt actions work youtube
Steps to build one:
- Create a GPT in the GPT editor and write clear instructions.
- Under Actions, add your OpenAPI schema with clear
operationIds and descriptions for each endpoint. - Configure authentication: none, API key or OAuth.
- Add a privacy policy URL if you plan to share it.
- Test in the preview pane and inspect each call and response.
Most failures come from vague endpoint descriptions; the model chooses tools from the text you write, so describe each one as you would to a new teammate.
Can we use GPT-2 sentence embeddings for classification tasks?
Yes. You can take GPT-2's hidden states, pool them into one vector per sentence, and train a simple classifier such as logistic regression on top. It works, but sentence-embedding models trained for this purpose usually give better results.
Also asked as: can we use gpt-2 sentence embedding for classification tasks
GPT-2 is a decoder-only model, so each token only sees the tokens before it. Mean pooling over all tokens or using the last token's state are the usual choices:
from transformers import GPT2Tokenizer, GPT2Model
import torch
tok = GPT2Tokenizer.from_pretrained("gpt2"); tok.pad_token = tok.eos_token
model = GPT2Model.from_pretrained("gpt2")
batch = tok(["great product", "terrible support"], return_tensors="pt", padding=True)
with torch.no_grad():
h = model(**batch).last_hidden_state
mask = batch["attention_mask"].unsqueeze(-1)
emb = (h * mask).sum(1) / mask.sum(1) # one 768-d vector per sentence
Compare against a Sentence-BERT style model on the same data before choosing.
How does GPT use water?
GPT models do not use water directly; the data centres that train and run them do, mainly for cooling servers and indirectly through the water used to generate their electricity. The amount depends heavily on location, season and cooling technology.
A 2023 study from researchers at the University of California estimated that training GPT-3 in Microsoft's US data centres directly evaporated hundreds of thousands of litres of fresh water, and that running models adds more over time. Such figures are estimates with wide ranges, because companies do not publish per-model water use. Data centres in hot, dry regions use more water for evaporative cooling; others use air cooling or recycled water and use more electricity instead. Major cloud providers publish annual sustainability reports with their overall water use.
Can GPT watch videos?
Not in the way people do. GPT models in ChatGPT can analyse images, so they can examine frames taken from a video, and transcripts from its audio, but they do not reliably watch a full-length video from start to finish.
The practical approach is to extract a transcript with a speech-to-text tool and a set of key frames, for example one every few seconds or at scene changes, and give those to the model with a precise question. In ChatGPT's mobile app, voice mode can use the live camera, which helps for quick questions about what is in front of you. Some other models, such as Google's Gemini, accept video files directly through their API and are a better fit when you need to analyse long footage.
Can GPT make videos?
GPT models themselves generate text and, in ChatGPT, images. Video generation comes from a separate OpenAI model, Sora, available through its own app and website for eligible users.
You can still use GPT for most of video production: scripts, storyboards, shot lists, captions, titles and thumbnail ideas. With its code tool, ChatGPT can sometimes assemble a simple slideshow video from images by writing Python, but that is a workaround. For generated footage, write a detailed prompt in Sora describing the subject, setting, camera movement and style, generate several takes, and edit the best clips in a normal video editor. Label AI-generated footage when you publish it.
Sources
- Radford et al., Improving Language Understanding by Generative Pre-Training (2018)cdn.openai.com
- Radford et al., Language Models are Unsupervised Multitask Learners (2019)cdn.openai.com
- Vaswani et al., Attention Is All You Need (2017)arxiv.org
- Reimers & Gurevych, Sentence-BERT (2019)arxiv.org
- Li et al., Making AI Less Thirsty: Uncovering and Addressing the Secret Water Footprint of AI Models (2023)arxiv.org
GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).
Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.
