# MIT Technology Review reports OpenAI’s iGPT experiments with image generation [Canonical reading](https://escapemeridian.com/readings/mit-technology-review-reports-openai-s-igpt-experiments-with-c2c1cc9d) MIT Technology Review reported on July 16, 2020, that OpenAI was adapting GPT-2 to generate images. The resulting model, iGPT, was trained by turning images into sequences of pixels and asking the model to predict missing parts. The article says iGPT could predict the second half of an image from its first half, producing completions that appeared sensible to a human. It describes the work as a new path for unsupervised learning, which trains on unlabeled data and can reduce the labor of labeling examples. The article also raises a risk: it says iGPT could generate plausible images without training only on a narrow category such as faces, unlike the GANs it discusses. It notes that training remained computationally expensive, while warning that this barrier might not last. These results concern image completions, not demonstrated real-world reliability or the future video generation discussed in a quoted forecast by OpenAI policy director Jack Clark. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source MIT Technology Review · Article Original source date: Jul 16, 2020 (2020-07-16T13:59:09Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-10-29T00:48:25.311Z. First observed: 2026-10-28T04:13:54.981Z. Retrieved: 2026-10-28T04:15:49.745Z. - [OpenAI’s fiction-spewing AI is learning to generate images](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Limits Text as retrieved on October 28, 2026; originally published July 16, 2020. The page is marked as edited after publication and includes an update; passages may reflect later changes to the page. The article’s prediction about future video generation is not evidence that it occurred. The article describes image completions but does not provide measures of their performance across unfamiliar images or the computation required to generate them. ## Open question How well would iGPT’s image completions hold up across unfamiliar images, and how much computation would generating them require? ## Source-stated MIT Technology Review reports that OpenAI trained iGPT on images converted into pixel sequences, and that it could predict an image’s second half from its first. > they unfurled the images into a single sequence of pixels. They found that the new model, named iGPT, was still able to grasp the two-dimensional structures of the visual world. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Source-stated The article presents unsupervised learning as a way to train on unlabeled data and reduce manual labeling work. > it allows an AI system to learn about the world without a human filter, and significantly reduces the manual labor of labeling data. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Source-stated The article says iGPT could make plausible images without being trained only on a narrow category, while noting that training was computationally expensive. > While training the model is still computationally expensive, offering a natural barrier to its access, that may not be the case for long. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Escape Meridian interpretation Escape Meridian reads the article’s account as leaving open whether iGPT’s image completions generalize beyond the examples described. > Given the sequence of pixels for the first half of an image, it could predict the second half in ways that a human would deem sensible. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## OpenAI OpenAI develops AI models and products and publishes research, release announcements and evaluations. - [Explore this context](/entities/openai) - [Reference](https://openai.com/news/) ## Jack Clark Jack Clark writes Import AI, a newsletter covering AI research, development and its implications. - [Explore this context](/entities/jack-clark) - [Reference](https://jack-clark.net/) ## Editorial provenance Version 1. Publication ID: 4e42e1c2-65cc-fa8a-5d03-6c604840d775. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T00:48:03.003Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: ba0f52f3c701b7b8ca9d457a2b45d21b33bb3d8644ceca8fdf0d45d5a3d12487. Source text hash: f03222b85d32b9f4e43f39abfbfb1b0d593959caa587a73cd6a7be9833e0081b. ## Revision and source-check history ### Version 1 — 2026-10-29T00:48:25.311Z MIT Technology Review reported on July 16, 2020, that OpenAI was adapting GPT-2 to generate images. The resulting model, iGPT, was trained by turning images into sequences of pixels and asking the model to predict missing parts. The article says iGPT could predict the second half of an image from its first half, producing completions that appeared sensible to a human. It describes the work as a new path for unsupervised learning, which trains on unlabeled data and can reduce the labor of labeling examples. The article also raises a risk: it says iGPT could generate plausible images without training only on a narrow category such as faces, unlike the GANs it discusses. It notes that training remained computationally expensive, while warning that this barrier might not last. These results concern image completions, not demonstrated real-world reliability or the future video generation discussed in a quoted forecast by OpenAI policy director Jack Clark. ## Original source MIT Technology Review · Article Original source date: Jul 16, 2020 (2020-07-16T13:59:09Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-10-29T00:48:25.311Z. First observed: 2026-10-28T04:13:54.981Z. Retrieved: 2026-10-28T04:15:49.745Z. - [OpenAI’s fiction-spewing AI is learning to generate images](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Limits Text as retrieved on October 28, 2026; originally published July 16, 2020. The page is marked as edited after publication and includes an update; passages may reflect later changes to the page. The article’s prediction about future video generation is not evidence that it occurred. The article describes image completions but does not provide measures of their performance across unfamiliar images or the computation required to generate them. ## Open question How well would iGPT’s image completions hold up across unfamiliar images, and how much computation would generating them require? ## Source-stated MIT Technology Review reports that OpenAI trained iGPT on images converted into pixel sequences, and that it could predict an image’s second half from its first. > they unfurled the images into a single sequence of pixels. They found that the new model, named iGPT, was still able to grasp the two-dimensional structures of the visual world. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Source-stated The article presents unsupervised learning as a way to train on unlabeled data and reduce manual labeling work. > it allows an AI system to learn about the world without a human filter, and significantly reduces the manual labor of labeling data. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Source-stated The article says iGPT could make plausible images without being trained only on a narrow category, while noting that training was computationally expensive. > While training the model is still computationally expensive, offering a natural barrier to its access, that may not be the case for long. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Escape Meridian interpretation Escape Meridian reads the article’s account as leaving open whether iGPT’s image completions generalize beyond the examples described. > Given the sequence of pixels for the first half of an image, it could predict the second half in ways that a human would deem sensible. - [Supporting source](https://www.technologyreview.com/2020/07/16/1005284/openai-ai-gpt-2-generates-images/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## OpenAI OpenAI develops AI models and products and publishes research, release announcements and evaluations. - [Explore this context](/entities/openai) - [Reference](https://openai.com/news/) ## Jack Clark Jack Clark writes Import AI, a newsletter covering AI research, development and its implications. - [Explore this context](/entities/jack-clark) - [Reference](https://jack-clark.net/) ## Editorial provenance Version 1. Publication ID: 4e42e1c2-65cc-fa8a-5d03-6c604840d775. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: passed by anthropic/claude-opus-5.5 on 2026-10-29T00:48:03.003Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: ba0f52f3c701b7b8ca9d457a2b45d21b33bb3d8644ceca8fdf0d45d5a3d12487. Source text hash: f03222b85d32b9f4e43f39abfbfb1b0d593959caa587a73cd6a7be9833e0081b. ### passed — 2026-10-29T00:48:03.003Z — anthropic/claude-opus-5.5 Publication: 4e42e1c2-65cc-fa8a-5d03-6c604840d775 attribution — pass: The reading attributes the report to MIT Technology Review and the research to OpenAI. It labels Jack Clark's video-generation remark as a quoted forecast, which matches the article. > So researchers at OpenAI decided to swap the words for pixels and train the same algorithm on images in ImageNet summary — pass: The reading clearly states the actor, the development (GPT-2's algorithm applied to images as pixel sequences) and the key term unsupervised learning in plain language. GANs are mentioned without explanation, but they are peripheral and the contrast still reads clearly. The question is a substantive open issue rather than a leading claim. > Given the sequence of pixels for the first half of an image, it could predict the second half in ways that a human would deem sensible. date — pass: The stated date, July 16, 2020, matches the UTC calendar day of originalPublished. The page was edited after publication (dateModified 2020-12-10); the update note says only a politician's name was removed. The reading's claims rest on core article passages that are plainly original. Clark's remark is presented as a forecast, not an outcome. The later sidebar stories are not used. Evidence kind is publisher-page, so no archive disclosure is needed. > Update: This article has been updated to remove the name of the politician in the hypothetical scenario described at the end. support — pass: Each claim is grounded in the article: pixel-sequence training and half-image completion, unsupervised learning reducing labeling work, and broad training versus curated GAN data along with the computational-cost barrier. The interpretive claim is labeled and conditional. The page edit does not affect these core passages. > iGPT, by contrast, simply learns enough of the structure of the visual world across millions and billions of examples to spit out images that could feasibly exist within it. media — pass: No media was selected and no verified media candidates were supplied, so nothing needs to be checked for identity or relevance. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/c2c1cc9dc72f698e6a4906bde8ab79a0e4c9cf8af350ec96099b4da6dfe18536/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/mit-technology-review-reports-openai-s-igpt-experiments-with-c2c1cc9d.json)