What is the full form of GPT in ChatGPT, recently seen in the news?

Question

UnlockIAS · Accepted Answer

Correct — D, Generative Pre-Trained Transformer. ChatGPT is OpenAI's conversational front end to its GPT family of large language models, and each of the three words in the expansion names a separate design decision, which is why the full form is worth knowing rather than merely recognising. Generative means the model produces new text instead of only classifying or scoring text that already exists: it is trained to predict the next token — roughly a word-piece — given everything before it, and writing an answer is simply that prediction run over and over. Pre-Trained names the two-stage recipe OpenAI set out in its 2018 paper 'Improving Language Understanding by Generative Pre-Training' (Radford, Narasimhan, Salimans and Sutskever): an expensive unsupervised first stage in which the model reads an enormous unlabelled corpus and absorbs general language structure, then a far cheaper second stage adapting it to a task — for ChatGPT, supervised fine-tuning followed by reinforcement learning from human feedback (RLHF), which is what teaches it to follow an instruction and to decline an unsafe request. Transformer is the neural-network architecture introduced in the Google paper 'Attention Is All You Need' (Vaswani and co-authors, 2017); its self-attention mechanism lets every token be weighed directly against every other token in parallel, replacing the step-by-step recurrence of earlier models and making training at internet scale practical on GPUs. The scale-up was rapid — GPT-1 had 117 million parameters in 2018, GPT-2 1.5 billion in 2019 and GPT-3 175 billion in 2020. ChatGPT itself was released as a free research preview on 30 November 2022 running on GPT-3.5, GPT-4 followed on 14 March 2023, and the product was estimated to have crossed 100 million monthly users by January 2023 — which is exactly why the acronym was 'recently seen in the news' when this paper was set in September 2023.

Why the other options are wrong — (a) A genuine GPT, and the most tempting wrong answer for anyone with a biology background: glutamic-pyruvic transaminase is the liver enzyme now more usually called alanine aminotransferase (ALT), and Indian pathology reports still print it as SGPT in a liver function test, where a raised value points to hepatocellular injury such as viral hepatitis or drug damage. It is decades older than the AI sense but has nothing to do with language models, and the stem's qualifier 'in ChatGPT' rules it out.  (b) Also a real GPT, and the closest trap inside computing itself. The GUID Partition Table is the disk-partitioning layout defined by the UEFI specification, brought in to replace the legacy Master Boot Record scheme, which could address only about 2 TB of disk and hold four primary partitions; GPT uses 64-bit addressing and stores a backup copy of the table at the end of the disk. It is a storage format, not a model architecture, and nothing about it was in the news in 2023.  (c) A third real expansion, this one from clinical neuropsychology: the Grooved Pegboard Test times how quickly a subject can rotate keyed pegs and drop them into randomly oriented slots, and is used to measure fine motor speed and eye-hand coordination, often to detect impairment on one side of the body. Its presence here shows the examiner's design — every option is a true GPT, so elimination of nonsense will not work and only the correct sense of the acronym does.

Concept: A large language model (LLM) is a neural network trained on a very large body of text to predict the next token given everything before it; ChatGPT is such a model wrapped in a chat interface. Its architecture — the transformer — works by self-attention: for every token the network computes how strongly it should be influenced by every other token in the input, so context is captured directly rather than passed hand to hand along a chain, and the whole sequence can be processed in parallel. That parallelism is why transformers scaled where the earlier recurrent networks (RNNs and LSTMs) stalled. Training runs in two phases, a broad unsupervised pre-training phase and a small alignment phase using human-ranked responses. Models built this way are called foundation models, because one pre-trained base is adapted to many downstream tasks, and generative AI, because the output is new content rather than a label. The same architecture now underlies image, speech and protein-structure models, not text alone.

State PCS 2023 Q6 — Science & Technology PYQ with Answer

State PCS 2023 Q6 — Science & Technology PYQ with Answer