ML//model//GPT

GPT is a decoder-only model family, with no encoder and no cross-attention. If it translates, it learned translation as a pattern during pre-training, without a dedicated encoder. It is trained with causal language modeling, predicting the next token from the previous tokens only, and it does not use a [CLS] token: it simply takes the last token, already optimized to carry the meaning of the full sequence through causal masking.

GPT is a decoder-only model family, with no encoder and no cross-attention. If it translates, it learned translation as a pattern during pre-training, without a dedicated encoder. It is trained with causal language modeling, predicting the next token from the previous tokens only, and it does not use a [CLS] token: it simply takes the last token, already optimized to carry the meaning of the full sequence through causal masking.

The generations differ mainly in how they were trained: each one added a layer to the training pipeline.

GPT-1 (2018) used pre-training plus SFT, with SFT for specific tasks like classification rather than for chatting.

GPT-2 used pure pre-training: OpenAI wanted to see whether it learned from volume alone, with no public fine-tuning.

GPT-3 used pure pre-training at massive scale. It predicted text beautifully but did not follow instructions.

InstructGPT and GPT-3.5 were the inflection point: they took GPT-3 and added SFT and RLHF, the birth of the assistant, and the proof that human preference training turns a raw language model into a useful one.

GPT-4 and GPT-4o use pre-training, SFT, RLHF and RLAIF, the same pipeline scaled, with AIs helping to score and correct.