Better language models and their implications

OpenAIWe’ve trained a large-scale unsupervised language model which generates coherent paragraphs of text, achieves state-of-the-art performance on many language modeling benchmarks, and performs rudimentary reading comprehension, machine translation, question answering, and summarization—all without task-specific training.Our model, called GPT‑2 (a successor to GPT), was trained simply to predict the next word in 40GB of Internet text. Due to our concerns about malicious applications of the technology, we are not releasing the trained model. As an experiment in responsible disclosure, we are instead releasing a much smaller model(opens in a new window) for researchers to experiment with, as well as a technical paper(opens in a new window).GPT‑2 is a large transformer(opens in a new window)-based language model with 1.5 billion parameters, trained on a datasetA of 8 million web pages. GPT‑2 is trained with a simple objective: predict the next word, given all of the previous words within some text. The diversity of the dataset causes this simple goal to contain naturally occurring demonstrations of many tasks across diverse domains. GPT‑2 is a direct scale-up of GPT, with more than 10X the parameters and trained on more than 10X the amount of data.GPT‑2 displays a broad set of capabilities, including the ability to generate conditional synthetic text samples of unprecedented quality, where we prime the model with an input and have it generate a lengthy continuation. In addition, GPT‑2 outperforms other language models trained on specific domains (like Wikipedia, news, or books) without needing to use these domain-specific training datasets. On language tasks like question answering, reading comprehension, summarization, and translation, GPT‑2 begins to learn these tasks from the raw text, using no task-specific training data. While scores on these downstream tasks are far from state-of-the-art, they suggest that the tasks can benefit from unsupervised techniques, given sufficient (unlabeled) data and compute.SamplesGPT‑2 generates synthetic text samples in response to the model being primed with an arbitrary input. The model is chameleon-like—it adapts to the style and content of the conditioning text. This allows the user to generate realistic and coherent continuations about a topic of their choosing, as seen by the following select samples.BAs the above samples show, our model is capable of generating samples from a variety of prompts that feel close to human quality and show coherence over a page or more of text. Nevertheless, we have observed various failure modes, such as repetitive text, world modeling failures (e.g., the model sometimes writes about fires happening under water), and unnatural topic switching. Exploring these types of weaknesses of language models is an active(opens in a new window) area(opens in a new window) of research(opens in a new window) in the natural language processing community.Overall, we find that it takes a few tries to get a good sample, with the number of tries depending on how familiar the model is with the context. When prompted with topics that are highly represented in the data (Brexit, Miley Cyrus, Lord of the Rings, and so on), it seems to be capable of generating reasonable samples about 50% of the time. The opposite is also true: on highly technical or esoteric types of content, the model can perform poorly. Fine-tuning offers the potential for even more detailed control over generated samples—for example, we can fine-tune GPT‑2 on the Amazon Reviews dataset and use this to let us write reviews conditioned on things like star rating and category.These samples have substantial policy implications: large language models are becoming increasingly easy to steer towards scalable, customized, coherent text generation, which in turn could be used in a number of beneficial as well as malicious ways. We’ll discuss these implications below in more detail, and outline a publication experiment we are taking in light of such considerations.Zero-shotGPT‑2 achieves state-of-the-art scores on a variety of domain-specific language modeling tasks. Our model is not trained on any of the data specific to any of these tasks and is only evaluated on them as a final test; this is known as the “zero-shot” setting. GPT‑2 outperforms models trained on domain-specific datasets (e.g. Wikipedia, news, books) when evaluated on those same datasets. The following table shows all our state-of-the-art zero-shot results.(+) means a higher score is better for this domain. (–) means a lower score is better.DatasetMetricOur resultPrevious recordHumanWinograd Schema Challengeaccuracy (+)70.70%63.7%92%+LAMBADAaccuracy (+)63.24%59.23%95%+LAMBADAperplexity (–)8.699~1–2Children’s Book Test Common Nouns (validation accuracy)accuracy (+)93.30%85.7%96%Children’s Book Test Named Entities (validation accuracy)accuracy (+)89.05%82.3%92%Penn Tree Bankperplexity (–)35.7646.54unknownWikiText-2perplexity (–)18.3439.14unknownenwik8bits per character (–)0.930.99unknowntext8bits per character (–)0.981.08unknownWikiText-103perplexity (–)17.4818.3unknownGPT‑2 achieves state-of-the-art on Winograd Schema, LAMBADA, and other language modeling tasks.On other language tasks like question answering, reading comprehension, summarization, and translation, we are able to get surprising results without any fine-tuning of our models, simply by prompting the trained model in the right way (see below for examples of how we do this), though we do still fall short of state-of-the-art for specialized systems.We hypothesize that since these tasks are a subset of general language modeling, we can expect performance to increase further with more compute and data. Others have published similar hypotheses(opens in a new window). We also expect fine-tuning to help performance on downstream tasks, though we have yet to do thorough experiments.Policy implicationsLarge, general language models could have significant societal impacts, and also have many near-term applications. We can anticipate how systems like GPT‑2 could be used to create:AI writing assistantsMore capable dialogue agentsUnsupervised translation between languagesBetter speech recognition systemsWe can also imagine the application of these models for malicious purposes, including the following (or other applications we can’t yet anticipate):Generate misleading news articlesImpersonate others onlineAutomate the production of abusive or faked content to post on social mediaAutomate the production of spam/phishing contentThese findings, combined with earlier results on synthetic imagery, audio, and video, imply that technologies are reducing the cost of generating fake content and waging disinformation campaigns. The public at large will need to become more skeptical of text they find online, just as the “deep fakes(opens in a new window)” phenomenon calls for more skepticism about images.CToday, malicious actors—some of which are political in nature—have already begun to target the shared online commons, using things like(opens in a new window) “robotic tools, fake accounts and dedicated teams to troll individuals with hateful commentary or smears that make them afraid to speak, or difficult to be heard or believed.” We should consider how research into the generation of synthetic images, videos, audio, and text may further combine to unlock new as-yet-unanticipated capabilities for these actors, and should seek to create better technical and non-technical countermeasures. Furthermore, the underlying technical innovations inherent to these systems are core to fundamental artificial intelligence research, so it is not possible to control research in these domains without slowing down the progress of AI as a whole.Release strategyDue to concerns about large language models being used to generate deceptive, biased, or abusive language at scale, we are only releasing a much smaller version of GPT‑2 along with sampling code(opens in a new window). We are not releasing the dataset, training code, or GPT‑2 model weights. Nearly a year ago we wrote in the OpenAI Charter: “we expect that safety and security concerns will reduce our traditional publishing in the future, while increasing the importance of sharing safety, policy, and standards research,” and we see this current work as potentially representing the early beginnings of such concerns, which we expect may grow over time. This decision, as well as our discussion of it, is an experiment: while we are not sure that it is the right decision today, we believe that the AI community will eventually need to tackle the issue of publication norms in a thoughtful way in certain research areas. Other disciplines such as biotechnology and cybersecurity have long had active debates about responsible publication in cases with clear misuse potential, and we hope that our experiment will serve as a case study for more nuanced discussions of model and code release decisions in the AI community.We are aware that some researchers have the technical capacity to reproduce and open source our results. We believe our release strategy limits the initial set of organizations who may choose to do this, and gives the AI community more time to have a discussion about the implications of such systems.We also think governments should consider expanding or commencing initiatives to more systematically monitor the societal impact and diffusion of AI technologies, and to measure the progression in the capabilities of such systems. If pursued, these efforts could yield a better evidence base for decisions by AI labs and governments regarding publication decisions and AI policy more broadly.We will further publicly discuss this strategy in six months. If you’d like to discuss large language models and their implications, please email us at: [email protected]. And if you’re excited about working on cutting-edge language models (and thinking through