When GPT-1 was introduced, the idea was quite mocked by many experts in the field.

True: contemporaneous (2018–2019) commentary or later credible accounts show that many prominent AI or NLP experts dismissed or ridiculed the GPT-1 approach (generative pretraining or scaling of unsupervised language models) when it was introduced. Mostly true: there was notable skepticism or dismissal from a number of experts, but it was milder than 'mocked' or came from fewer people than 'a lot'. Mixed: the reception was divided, with comparable amounts of praise and skepticism. Mostly false: there was only isolated skepticism, and the main expert reception was neutral or positive; the claim overstates a small amount of criticism. False: the evidence shows GPT-1 was broadly well received by experts, with no meaningful mockery.

Set before the research began.

1. E1 GPT-1 — Igniting the Pre-training Revolution with Decoder-only Transformer awesome-papers project (papernotes) · 17 September 2026

   “A 12-page report widely underestimated at the time, it was the first to prove that”

2. E2 GPT-1 — Igniting the Pre-training Revolution with Decoder-only Transformer awesome-papers project (papernotes) · 17 September 2026

   “"Pretraining is useless" (2017 community consensus): GPT-1 directly refuted — 100 epochs of pretraining gives 5–10 point jumps on small downstream tasks”

3. E3 Paper Close Read — GPT-1: Generative Pretraining (Radford et al., 2018) | AI Science II: Large Language Models Socratopia · 1 January 2026

   “In 2018, the prevailing view was that left-to-right language models were too "limited" for understanding tasks — they could only see half the context at any given position.”

4. E4 GPT-1 Deep Dive: How a 'Rambling' 117M-Parameter Model Changed the World in 2018 zhichai.net · 16 May 2026

   “Few realized that this rambling language model would ultimately change the world forever.”

5. E5 GPT-1 — Igniting the Pre-training Revolution with Decoder-only Transformer awesome-papers project (papernotes) · 17 September 2026

   “A 12-page report widely underestimated at the time, it was the first to prove that”

6. E6 GPT-1 — Igniting the Pre-training Revolution with Decoder-only Transformer awesome-papers project (papernotes) · 17 September 2026

   “"Pretraining is useless" (2017 community consensus): GPT-1 directly refuted — 100 epochs of pretraining gives 5–10 point jumps on small downstream tasks”

7. E7 Paper Close Read — GPT-1: Generative Pretraining (Radford et al., 2018) | AI Science II: Large Language Models Socratopia · 1 January 2026

   “In 2018, the prevailing view was that left-to-right language models were too "limited" for understanding tasks — they could only see half the context at any given position.”

8. E8 Paper Close Read — GPT-1: Generative Pretraining (Radford et al., 2018) | AI Science II: Large Language Models Socratopia · 1 January 2026

   “It lost the 2018 benchmarks by a fair margin, yet its architecture eventually became the backbone of every frontier model.”

9. E9 llm-evolution/02-the-transformer-revolution/02-gpt-1.md at main · maraja/llm-evolution GitHub (maraja/llm-evolution repository) · 24 February 2026

   “GPT-1 was seen as a competent but less powerful approach because it couldn't use bidirectional context.”

10. E10 llm-evolution/02-the-transformer-revolution/02-gpt-1.md at main · maraja/llm-evolution GitHub (maraja/llm-evolution repository) · 24 February 2026

    “The decoder-only choice was initially seen as a limitation compared to BERT's bidirectional approach, but it proved critical for generation capabilities and for the scaling properties that emerged in larger models.”

11. E11 GPT-1 | AI Wiki AI Wiki (aiwiki.ai) · 18 March 2026

    “The model was modest by later standards, but it served as the proof-of-concept that launched a paradigm shift in NLP research and laid the foundation for the much larger models that followed.”

12. E12 The First LLM thundergolfer.com (personal blog) · 23 March 2025

    “GPT-1 (Generative Pre-Train One) is widely accepted as an LLM.”

The sources are all later educational or aggregator write-ups. They say GPT-1 was "widely underestimated" (E1), that the prevailing 2018 view saw left-to-right language models as too limited (E3), and that it was seen as "competent but less powerful" than BERT (E9, E10). E2 describes a 2017 view that pretraining was useless, which is about pretraining in general rather than GPT-1. None of these sources names or quotes experts reacting to GPT-1, and none describes mockery or ridicule. These weak secondary accounts point to underestimation and technical doubt, but they don't show how widespread or how sharp the expert reaction was, so they cannot settle whether the claim was true, overstated or false.