OpenAI

OpenAI announced the text-to-video model Sora on February 15, 2024. Sora is a diffusion model that can output videos up to one minute long based on user prompts. Upon the announcement, OpenAI made Sora available to red teamers, to assess potential areas of harm and risks, and visual artists, designers, and filmmakers to gain feedback on performance.

Sora is OpenAI's text-to-video AI model capable of generating videos up to a minute long based on user prompts. Sora is a diffusion model, that generates videos from a starting point of static noise. It uses a transformer architecture similar to OpenAI's GPT models and builds on previous research from the Dall-E models, in particular, using the recaptioning technique from Dall-E 3 that involves generative descriptive captions for visual training data. Analogous to text tokens for LLMs, Sora uses visual patches, an effective representation of visual data. Patches are scalable and allow generative models to be trained on a range of video and image types. At a high level, Sora turns videos into patches by compressing them into a lower-dimensional latent space and decomposing the representation into spacetime patches.

On December 27, 2023, the New York TimesNew York Times sued OpenAI and Microsoft for copyright infringement. The news outlet became the first major American media organization to sue the companies over copyright issues regarding its written work. The lawsuit was filed in Federal District Court in Manhattan. It contends that the companies used millions of articles published by the New York TimesNew York Times to train chatbots that are now in competition with it as a source of reliable information. The lawsuit states the defendants should be held responsible for “billions of dollars in statutory and actual damages” and they should destroy any chatbot models and training data using copyrighted material from Thethe New York Times Times.

On January 10, 2024, OpenAI introduced the GPT store to ChatGPT Plus, Team, and Enterprise users. The rollout of the GPT store came two months after the company announced custom GPTs at its first developer conference. OpenAI stated that in those two months, users created over three million custom versions of ChatGPT. The store offers a range of GPTs developed by OpenAI's partners and the wider community.

On January 10, 2024, OpenAI launched ChatGPT Team, a new ChatGPT plan designed for teams of all sizes that provides a secure, collaborative workspace. The new plan offers access to OpenAI models, such as GPT-4 and DALL·E 3, as well as tools like Advanced Data Analysis. It also includes a collaborative workspace for team and admin tools plus:

ChatGPT Team costs $25/month per user billed annually, or $30/month per user billed monthly.

Custom GPTs are available to ChatGPT Plus, Team, and Enterprise users through the GPT store. The store features GPTs developed by OPenAIOpenAI partners and the community. Visitors can browse popular and trending GPTs on the community leaderboard with categories such as DALL·E, writing, research, programming, education, and lifestyle. The store highlights new and impactful GPTs each week. To share a GPT to the store, builders have to verify their profile and ensure it is compliant with OpenAI's usage policies and brand guidelines. OpenAI has implemented a review system that includes human and automated reviews and the ability for users to report GPTs. US GPT builders can earn money for their work based on user engagement.

OpenAI undertakes research to align its AI systems with its mission to benefit all of humanity. This includes training AI systems to do what humans want and to be helpful, truthful, and safe. A post from August 2022 detailed the company's empirical and iterative approach to aligning its AI systems with human values and human intent. OpenAI aims to push alignment ideas as far as possible and to understand how different approaches succeed or fail. Unaligned AGI poses a significant risk to humanity and finding solutions requires input from a large number of people. Aligning AI systems poses a wide range of socio-technical challenges. Therefore, OpenAI has committed to sharing its alignment research when safe to do so.

In July 2023, OpenAI started a new superalignment team, co-led by Ilya Sutskever (cofounder and Chief Scientist) and Jan Leike (Head of Alignment), to work on the scientific and technical breakthroughs required to align future superintelligent AI systems. The company aims to solve the problem within four years and is dedicating 20 percent of its secured compute power over this period to the effort.

OpenAI is an artificial general intelligence (AGI) research and development company founded in 2015 and based in San Francisco, California, United States. The AI laboratory consists of the original parent company, with the legal name "OpenAI Inc.", which acts as a nonprofit, and Open AI LP, a for-profit company formed in 2019 that now employs most of its staff. OpenAI develops autonomous systems capable of performing work considered economically valuable. The company focuses on long-term research, working on problems that require fundamental advances in AI capabilities. Research by OpenAI is published at top machine learning conferences. The organization also contributes open-source software tools for accelerating AI research and releases blog posts to communicate its research to others in the field. OpenAI has received criticism both for not sharing its research and for the potential misuse of its AI systems.

OpenAI has developed a number of products, including those below:

The release of ChatGPT in November 2022 brought significant attention to the company, with over 5 million uses in the first five days of launch. A study published in February 2023, estimates that ChatGPT became the fastest-growing consumer application up to that point, with 100 million monthly active users just two months after launch. ChatGPT had roughly 13 million unique visitors in January 2023, over double the numbers of December 2022. Figures released at OpenAI's first developer conference in November 2023, stated one hundred million people are using ChatGPT on a weekly basis, and over two million developers are using the company's API. This includes over 92% of Fortune 500 companies.

OpenAI investors include Microsoft, Reid Hoffman’s charitable foundation, Khosla Ventures, Sequoia Capital, Tiger Global Management, Bedrock Capital, and Andreesen Horowitz. In July 2019, Microsoft partnered with OpenAI to support the company's AGI research and development. This included a $1 billion investment from Microsoft, a partnership to develop a hardware and software platform for Microsoft Azure whichthat will scale to AGI, and Microsoft becoming OpenAI's exclusive cloud provider. In September 2020, Microsoft also announced it had partnered with OpenAI to exclusively license its language model, GPT-3. Microsoft made a second investment in OpenAI in 2021, although details were not released.

A sale of OpenAI stock in 2021 from existing shareholders to investors (including Sequoia Capital, Tiger Global Management, Bedrock Capital, and Andreessen Horowitz) implied a company valuation of nearly $20 billion. Reports in January 2023 suggest Microsoft and OpenAI were in discussions over a new investment of $10 billion that would value the company at $29 billion, making the company one of the most valuable U.S. startups despite not generating significant revenue up to that point. The deal would result in Microsoft owning a 49% percent stake in OpenAI, with a clause meaning Microsoft would receive three-quarters of OpenAI profits until the investment is recovered. On January 23, 2023, Microsoft and OpenAI announced a new multiyear, multibillion-dollar investment. While the financial terms of the partnership were not revealed, Bloomberg reported the figure of $10 billion. In February 2023, Microsoft announced a new AI-powered Bing search engine and Edge browser integrating OpenAI technology.

OpenAI LP is governed by OpenAI Nonprofit's board, which is comprised of employees Greg Brockman (Chairmanchairman &and Presidentpresident), Ilya Sutskever (Chiefchief Scientistscientist), and Sam Altman (CEO), and non-employees Adam D’Angelo, Reid Hoffman, Will Hurd, Tasha McCauley, Helen Toner, and Shivon Zilis. Founder member Elon Musk left the OpenAI board in February 2018 and is no longer formally involved in OpenAI. As well as overseeing OpenAI LP, OpenAI Nonprofit runs educational programs and hosts policy initiatives.

In early November 2023, OpenAI's board was made up of Adam D’Angelo, Helen Toner, Ilya Sutskever, Tasha McCauley, Sam Altman, and Greg Brockman (chairman). On November 17, 2023, Altman and Brockman were removed from the board by the other members, with Altman departing the company and Brockman quitting. After Altman was rehired as CEO on November 21, 2023, a reconfigured board was announced, made up of Bret Taylor (Chair), Larry Summers, and Adam D'Angelo (lone holdover from the previous board). Reports suggest the 3-personthree-person board plans to vet and appoint an expanded board of up to nine people, with the potential of Microsoft and Altman getting a seat.

In January 2022, OpenAI CEO Sam Altman joked about startupstar-tup funding, claiming on Twitter that:

On January 23, 2023, Microsoft and OpenAI announced the third phase of their partnership, with a new multiyear, multibillion-dollar investment to accelerate AI breakthroughs. Financial terms related to the new partnership were not revealed. However, Bloomberg reported a $10 billion, consistent with previous reporting. Continuing the partnership allows for new advances in AI supercomputing and research, with both companies able to independently commercialize the resulting AI technologies. Key areas for collaboration include those below:

In April 2023, VC firms including SequioaSequoia Capital, Andreessen Horowitz, Thrive, and K2 Global bought shares in the company, investing over $300 million at a valuation between $27 billion and $29 billion. The Founders Fund was also involved in the investment. The finalized tender allowed some OpenAI staff members to cash out their holdings. This tender was followed in October 2023, with Thrive Captial leading another deal to buy employee shares. This time at a company valuation of at least $80 billion.

On March 20, 2017, OpenAI launched DistillDistill, a journal aimed at communicating machine learning results. Research milestones from 2017 included the following:

• The release of highly-optimizedhighly optimized GPU kernels for neural networks with block-sparse weights

OpenAI Five held its final live event on April 13th13, competing against the reigning Dota 2 world Champions, OG, in front of a live audience. OpenAI Five defeated OG in back-to-back games, becoming the first AI to beat the world champions in an esports game. DeepMind’s AlphaStar had previously beaten professional players privately but lost their live matches. OpenAI Five began as a way to develop deep RL algorithms. To create the Dota 2 bots, OpenAI created a system called Rapid that could run proximal policy optimization at a greater scale. After OpenAI Five's losses at The International in 2018, OpenAI upgraded performance with 8x more training compute.

In April 2018, OpenAI released a charter describing the principles the company will use to execute its mission of achieving AGI, while acting in the best interests of humanity. The document reflects strategy developed over two years at the company and was made using feedback from people both internal and external to OpenAI.

OpenAI chose to deliberately use the same transformer architecture GPT-2 in language to highlight the potential generative sequence modeling of unsupervised learning algorithms. Image GPT works using a sufficiently large transformer trained on next-pixel prediction to generate diverse samples. OpenAI's work showed, that given sufficient compute, a sequence transformer can generate results comparable to convolutional nets for unsupervised image classification.

In August 2022, OpenAI releasesreleased details on its research approach to making AGI align with human values and follow human intent. This includes an iterative and empirical approach to making AI systems safer.

• Custom GPTs—a new way for users to create tailored versions of ChatGPT, called GPTs. Without coding, users can begin building chatbots better suited to their needs. User-generated GPTs can be shared publicly, and OpenAI has plans to launch a GPT Store, a searchable tool and leaderbuild of creations from verified builders.

On Friday, November 17, 2023, OpenAI announced Sam Altman had departed the company (leaving his role as CEO and board member) with CTO Mira Murati being appointed interim CEO while the company searched for a permanent successor. The surprise announcement came eleven days after Altman led the keynote at the OpenAI's first DevDay conference. The board of directors for OpenAI, Inc., the 501(c)(3) (original non-profit company) made the announcement following a deliberative review process, concluding that Altman:

However, Brockman announced he had quit the company hours after the announcement. Brockman stated he and Altman were notified of their removal from the board that day from Sutskever. Sources have stated that Sutskever was instrumental in the removal of Altman following a power struggle between the research and product sides of the company. Reports stated that OpenAI employees found out the news from the public announcement. After the news broke, a number of OpenAI employees also resigned from the company Friday evening.

On Saturday, November 18, reports stated Altman was in discussions to return to OpenAI (an idea being pushed by a number of OpenAI investors, including Microsoft) or start a new AI venture. It was also revealed that OpenAI investors were not given advance warning or an opportunity to weigh in on the board’s decision to remove Altman. On Saturday evening, Altman posted on X “i love the openai team so much.” this was followed by a large number of OpenAI employees demonstrating their support by reposting him with a heart emoji, and rumors began of mass resignations at the company if Altman was not reinstated. An internal memo by OpenAI COO Brad Lightcup stated Altman's removal by the board was over a "breakdown of communications," not "malfeasance."

On Tuesday, November 21, it was announced that Altman and OpenAI had come to an agreement in principle for his return as CEO with a reconfigured board. The announcement from the company, posted on X, read:

Taylor is the former co-CEO of Salesforce, and Summers is a former US Treasury Secretary. D'Angelo is the lone holdover from the previous board. Altman posted on X after the announcement:

Former president Brockman also returns to the company, and both sides have agreed to an investigation into the turmoil at the company. Sources have stated the new 3-personthree-person board's first job is to vet and appoint an expanded board of up to nine people. With reports that Microsoft and Altman himself want a seat on the new board.

On August 28, 2023, OpenAI launched ChatGPT Enterprise, a new version of ChatGPT for businesses that offers the following:

OpenAI states Dall-E 3 generated images are more visually striking and crisper in detail compared troto Dall-E 2. Dall-E 3 improvements are particularly noticeable when generating text, hands, and faces. The model also supports both landscape and portrait aspect ratios. These improvements are achieved by training an image captioner to generate better textual descriptions for images. These images with improved captions were then used to train Dall-E 3, producing a model more responsive to user-supplied captions.

Capabilities include those below:

• Retrieval—augments the assistant with knowledge from outside OpenAI's models, such as proprietary domain data, product information, or documents provided by users. The Assistants API can optimize what retrieval technique to use based on experience building knowledge retrieval in ChatGPT.

OpenAI now offers users the ability to build custom versions of ChatGPT, called GPTs. These new versions of ChatGPT can be tailored for specific user or enterprise tasks. Creating a tailored GPT requires no coding. GPTs are available to paying ChatGPT Plus subscribers and OpenAI enterprise customers. Demos of the platform include a "creative writing coach" bot that can critique writing samples and a GOT to help attendees of a developer conference. The platform auto-named the bot “Event Navigator,” generated a profile picture for it using DALL-E, and ingested an event schedule to help attendees. Each GPT can be given access to web browsing, DALL-E, and OpenAI’s Code Interpreter tool.

On July 5, 2023, OpenAI announced a new superalignment team to work on the scientific and technical breakthroughs required to control future AI systems that will be smarter than humans. The team will be co-led by Ilya Sutskever and Jan Leike, with the company dedicating 20% percent of its compute power to the project, which it hopes to solve within four years.

In July 2023, OpenAI started a new superalignment team, co-led by Ilya Sutskever (cofounder and Chief Scientist) and Jan Leike (Head of Alignment), to work on the scientific and technical breakthroughs required to align future superintelligent AI systems. The company aims to solve the problem within four years and is dedicating 20% percent of its secured compute power over this period to the effort.

Artificial superintelligence (ASI) refers to a theoretical form of AI that surpasses human intellect, manifesting cognitive skills and developing thinking skills of its own. ASI represents a much higher capability level than AGI. While this type of AI is not currently possible, and there is uncertainty over the speed of its development, its potential power could bring significant dangers leading to the disempowerment of humanity. OpenAI is aiming to target breakthroughs for aligning systems much more capable than current models.

The company plans to leverage AI systems to provide a training signal on tasks that are difficult for humans to evaluate and assist in the evaluation of other AI systems (scalable oversight). They also want to understand and control how OpenAI models generalize oversight to tasks that are not supervised. To validate the alignment of systems, the company automates the search for problematic behavior and problematic internals. Finally, the entire pipeline will be tested by deliberately training misaligned models, and confirming that our techniques detect the worst kinds of misalignments (adversarial testing).

On May 25, 2018, OpenAI released Gym Retro, a platform for RL research on games. The company uses Gym Retro to conduct research on RL algorithms and study generalization. The release includes a series of retro games from Sega's Genesis console and Master System, and Nintendo’s NES, SNES, and Game Boy consoles. Gym Retro also includes preliminary support for the Sega Game Gear, Nintendo Game Boy Color, Nintendo Game Boy Advance, and NEC TurboGrafx.

On February 7, 2023, Microsoft announced new AI-powered versions of its search engine, Bing, and its browser, Edge, that integrate OpenAI technology. The new Bing runs on an OpenAI large language model, called the "Prometheus Model," which takes key advancements from GPT-3.5 and ChatGPT and customizes them for search. Microsoft refers to these new tools as an "AI copilot for the web," creating a new way to browse the internet with users able to ask Bing questions and receive answers, similar to ChatGPT. During the announcement, Microsoft showed a number of demos, including Bing returning traditional search results with AI annotations and allowing users to talk directly to the Bing chatbot. Unlike ChatGPT, Bing can now retrieve information about recent events.

The new Bing search engine is available as a limited preview on desktop from February 7, 2023. Users can try sample queries and signupsign up for a waiting list to for full access in the future.

Microsoft is also launching two new AI features for the Edge web browser browser—"chat" and "compose." Embedded within the Edge sidebar, users can ask for a summary of a webpage or document they ask questions about it using the "chat" function. "Compose" works as a writing assistant, generating text based on starting prompts.

On March 14, 2023, OpenAI announced its GPT-4, a large multimodal model that can accept both image and text inputs, responding with text outputs. A transformer-based model trained to predict the next token in a document, GPT-4 allows users to specify vision and language tasks. Only the model's text capabilities are initially available, either through ChatGPT Plus or OpenAI's API, with a waiting list for access. OpenAI stated it is working closely with a single partner to prepare the image input capability for public release. Microsoft has since confirmed its AI-enabled Bing was already running an early version of GPT-4. GPT-4 was trained on Microsoft Azure AI supercomputers. Training finished in August 2022 and was followed by 6six months of model alignment to improve safety.

GPT-4 is a large multimodal model, with greater performance compared to previous GPT generations. The transformer-based model can accept both text and image inputs performing vision and language tasks based on the user's prompt. GPT-4 remains less capable than humans in many real-world scenarios,; however, the model demonstrates human-level performance on a range of professional and academic benchmarks, including performing in the 10% on the uniform bar exam.

Compared to ChatGPT, GPT-4 allows text inputs and produces text outputs with over 8x8 times the number of words, 25,000words—25,000 words (GPT-4) compared to 3,000 words (ChatGPT). Image input capabilities include understanding images and logical ideas about them, generating captions, classifications, and analyses.

In January 2023, OpenAI announced plans on its discordDiscord for a paid version of ChatGPT. Before an official announcement from the company, users began reporting having access to a new pro tier of the AI system priced at $42 a month.

• Correctlycorrectly identifies 26% of AI-written text (true positives) and
• Incorrectlyincorrectly identifies human-written text as AI-written 9% of the time (false positives).

The classifier's reliability improves with the length of the text. OpenAI made the classifier publicly available to gain feedback on whether these kinds of imperfect tools are useful, with the hope to share improved methodologies in the future. The hope is that better-performing classifiers can mitigate the risks of AI-generated text, such as misinformation campaigns and academic dishonesty.

A research preview of ChatGPT was released on November 30, 2022, with the model freely available to users with an OpenAI account. On February 1, 2023, OpenAI launched ChatGPT Plus, a paid subscription pilot available to US customers, for $20 per month. ChatGPT Plus offers users benefits that include access to ChatGPT during peak times, faster response times, and priority access to new features.

OpenAI has developed a number of products, including those below:

On January 23, 2023, Microsoft and OpenAI announced the third phase of their partnership with a new multiyear, multibillion-dollar investment to accelerate AI breakthroughs. Financial terms related to the new partnership were not revealed. However, Bloomberg reported a $10 billion, consistent with previous reporting. Continuing the partnership allows for new advances in AI supercomputing and research, with both companies able to independently commercialize the resulting AI technologies. Key areas for collaboration include those below:

• Microsoft increasing its investment in the development and deployment of supercomputing systems to accelerate OpenAI's AI research.
• Deploying OpenAI's models across Microsoft's consumer and enterprise products and introducing new digital experiences based on their AI systems.
• Azure remaining OpenAI's exclusive cloud provider.

• Broadly distributed benefits—committing any influence OpenAI obtains over AGI development to benefit everyone, preventing the harmful use of AI or AGI, and preventing concentrations of power in the field.
• Long-term safety—committing to research that makes AGI safe and helping to ensure safety is broadly adopted across the AI community.
• Technical leadership—addressing AGI's impact on society.
• Cooperative orientation—cooperating with other research and policy institutions to create a global AGI community working together to solve challenges facing the field.

On March 20, 2017, OpenAI launched Distill a journal aimed at communicating machine learning results. Research milestones from 2017 included the following:

After the tournament, OpenAI released a blog detailing the development of its Dota 2 bot. In August 2018, OpenAI Five, a team of five neural networks competed in a series of 5v5 Dota 2 games. OpenAI Five won a best-of-three game against 99.95th percentile Dota players, which included 4four players that have played Dota 2 professionally. The match took place in front of a live audience and 100,000 concurrent livestream viewers. OpenAI would go on to lose two games against top Dota 2 players at the International in Vancouver.

On April 25, 2019, OpenAI released a deep neural network capable of generating music compositions called MuseNet. The model can produce 4-minutefour-minute compositions with ten different instruments, combining a range of styles. MuseNet was not explicitly programmed with musical knowledge. Instead, it discovers patterns in harmony, rhythm, and style and learns to predict the next token in hundreds of thousands of MIDI files. MuseNet uses the same unsupervised technology as GPT-2 uses to predict the next token in a text sequence.

After switching to a capped-profit business, OpenAI's leadership instituted a new pay structure that was based in part on how each employee absorbed the company's mission. This included the culture-related expectations for different levels. Examples include the following:

Shortly after the transition, in July 2019, OpenAI and Microsoft announced a partnership to work together on building AGI. This included a $1 billion investment from Microsoft, as well as the following:

• Contrastive Language-Image Pre-training (CLIP)—a neural network developed by OpenAI that can learn visual concepts from natural language descriptions.
• Dall-E (pronounced dolly)—a neural network capable of creating images from text captions for a wide range of concepts.

Key OpenAI announcements during 2021 included those below:

GPT-3 is an autoregressive language model with 175 billion parameters made by OpenAI. It launched May 29, 2020. The model builds on the transformer-based language model previously made by OpenAI called GPT-2 , which had 1.5 billion parameters. GPT-3 improves on GPT-2 by adopting and scaling features present in GPT-2, such as modified initialization, pre-normalization, and reversible tokenization. GPT-3 training can improve scaling up language models primarily through improved task-agnostic and few-shot performance compared to the GPT-2 model. Open AI claims GPT-3 is approaching the performance of SOTA fine-tuned systems for generating quality performance and samples for defined tasks.

OpenAI Codex is an AI system that translates natural language into code. A general-purpose programming model, while results vary, Codex can be applied to effectively any coding task. Codex powers GitHub Copilot, a project built in partnership with GitHub that suggests code and entire functions in real time. Codex is proficient in over a dozen programming languages (Python, JavaScript, Go, Perl, PHP, Ruby, Swift, TypeScript, Shell, etc.), interpreting simple natural language commands to execute them on the user's behalf. The model makes it possible to build a natural language interface for existing applications.

A descendent of GPT-3, Codex is trained on both natural language and billions of lines of source code from publicly available sources, including public GitHub repositories. Codex has a similar natural language understanding as GPT-3 but translates this information into working code. This requires the following:

Whisper is an automatic speech recognition (ASR) system that is approaching human levels of accuracy for the English language. The model is trained on 680,000 hours of multilingual and multitask supervised data collected from the internet. Using such a large training dataset helps Whisper improve its robustness to accents, background noise, and technical language, enabling transcription in multiple languages as well asand translating from various languages into English.

Whisper's architecture uses an end-to-end approach implemented as an encoder-decoder transformer. Audio is divided into 30-secondthirty-second chunks before being converted into a log-Mel spectrogram and then passed into an encoder. A decoder is trained to predict the corresponding text caption as well asand perform specific tasks, such as language identification, phrase-level timestamps, multilingual speech transcription, and to-English speech translation. While specialized models show better speech recognition performance, using a large and diverse dataset allows Whisper to be used for a variety of tasks with fewer errors. Roughly a third of Whisper's audio dataset is non-English. OpenAI open-sourced the Whisper model and inference code.

MuseNet is a deep neural network capable of composing four-minute musical pieces with up to ten different instruments and in any genre. The software was not programmed to explicitly understand music. Instead, learning to predict the next "token" in a MIDI file, by applying the same unsupervised technology as GPT-2 to hundreds of thousands of MIDI files. To train, the organization used data from ClassicalArchives and BitMidi. MuseNet can discover patterns in harmony, rhythm, and style. The neural network uses the recompute and optimized kernels of the Sparse Transformer to train a 72-layer network with 24twenty-four attention heads - overheads—over a context of 4096 tokens.

Jukebox is a neural network capable of generating music in a wide variety of genres and styles, including rudimentary singing, as raw audio. While MuseNet explored composing music based on significant amounts of MIDI data, in raw audio, Jukebox has to deal with high diversity and long-range structures. Generating music requires dealing with long sequences. For example, a typical 4-minutefour-minute song at CD quality (44kHz, 16-bit) contains over 10 million timesteps. OpenAI's GPT-2 model had 1,000 timesteps and the company's AI Dota 2 team, OpenAI Five, took tens of thousands of timesteps across an entire game. Therefore, to produce music, Jukebox has to deal with long-range dependencies. This is done using an autoencoder that compresses raw audio to a lower-dimensional space and discards some of the irrelevant bits of information. Jukebox generates audio in this compressed space before upsampling back to the quality needed for raw audio.

OpenAI's API allows developers to incorporate the company's AI systems into their own projects. The API is powered by multiple models with different capabilities and price points, including the following:

• GPT-3—performs a variety of natural language tasks, including copywriting, summarization, translation, text classification, and parsing unstructured text. The base GPT-3 models offered in the API are called Davinci, Curie, Babbage, and Ada.
• Codex—translates natural language into code,; common tasks include text to SQL, calling an API using natural language, and code continuation.
• Dall-E—image generation tool capable of creating original images based on natural language prompts and editing existing images.

The models accessible through the OpenAI API break down text into tokens,; these can be words or chunks of characters. The number of tokens the models process in a given API request depends on both the inputs provided and the generated outputs. Generally speaking, a single token is equivalent to roughly 4 characters or 0.75 words for English text. A limitation of using the API is that the combined input and output for a request cannot be more than the model's maximum context length, which is 2048 tokens for most models (equivalent to roughly 1500 words).

Users can interact with the API through HTTPhttp requests from any language, via OpenAI's official Python bindings, their official Node.js library, or a community-maintained library. API keys are used for authentication. To help OpenAI ensure safe usage, the API comes with free content filtering, end-user monitoring for misuse, and specialized endpoints to view API usage.

Users receive $18 in free credit for use in their first 3three months after signing up. Fine-tuned custom models are priced for training and usage:

OpenAI undertakes research to align its AI systems with its mission to benefit all of humanity. This includes training AI systems to do what humans want and to be helpful, truthful, and safe. A post from August 2022, detailed the company's empirical and iterative approach to aligning its AI systems with human values and human intent. OpenAI aims to push alignment ideas as far as possible and to understand how different approaches succeed or fail. Unaligned AGI poses a significant risk to humanity and finding solutions requires input from a large number of people. Aligning AI systems poses a wide range of socio-technical challenges. Therefore, OpenAI has committed to sharing its alignment research when safe to do so.

OpenAI's alignment research focuses on building a scalable training signal for smart AI systems that align with human intent. TheThere are three main components areto the alignment research:

• Descript—a video editor using AI to simplify the editing process.
• Harvey—an interface for legal workflows using generative language models.
• Mem—an AI-powered self-organizing workspace that can predict the information most relevant to users at a given time.
• Speak—an AI tutor that can have open-ended conversations with people learning a new language and provide feedback. The company initially launched in East Asia and has over 100,000 paying subscribers.

Key papers describing OpenAI models include those listed:

OpenAI is an artificial general intelligence research and development company. The company was founded in 2015 by Elon Musk, Greg Brockman, Ilya Sutskever, Sam Altman, and Wojciech Zaremba, in San Francisco, California, United States. OpenAI develops autonomous systems capable of performing work considered economically valuable. The company focuses on long-term research, working on problems that require making fundamental advances in AI capabilities. OpenAI investors include Microsoft, Reid Hoffman’s charitable foundation, and Khosla Ventures. In July 2019, OpenAI received a $1 billion investment from Microsoft, whowhich in return got exclusive rights to license the company's GPT-3 model a year later.

Research by OpenAI is published at top machine learning conferences. The organization also contributes open-source software tools for accelerating AI research and releases blog posts to communicate theirits research to others in the field.

OpenAI has developed a number of models and products including:

• ChatGPT—a model capable of interacting with users in a conversational way.
• Dall-E—AI system capable of generating images from natural language descriptions.
• Whisper—neural net approaching human-level accuracy speech recognition for the English language.

Founded in 2015, OpenAI Inc. was originally a non-profit company. In March 2019, it formed OpenAI LP, a "capped-profit" company that now employs most of its staff. To help raise investment capital and attract talent, OpenAI LP offers investors and employees a capped return based on its success. Any returns beyond the capped amount are to be owned by the parent non-profit entity. Since 2019, any official correspondence from the company refers to OpenAI LP as simply "OpenAI" with the original parent company referred to as "OpenAI Nonprofit."

OpenAI has developed a number of models and products:

• ChatGPT—a model capable of interacting with users in a conversational way
• Dall-E—AI system capable of generating images from natural language descriptions
• Whisper—neural net approaching human-level accuracy speech recognition for the English language

Founded in 2015, OpenAI Inc. was originally a nonprofit company. In March 2019, it formed OpenAI LP, a "capped-profit" company that now employs most of its staff. To help raise investment capital and attract talent, OpenAI LP offers investors and employees a capped return based on its success. Any returns beyond the capped amount are to be owned by the parent nonprofit entity. Since 2019, any official correspondence from the company refers to OpenAI LP as simply "OpenAI" with the original parent company referred to as "OpenAI Nonprofit."

OpenAI LP is governed by OpenAI Nonprofit's board, which is comprised of employees Greg Brockman (Chairman & President), Ilya Sutskever (Chief Scientist), and Sam Altman (CEO), and non-employees Adam D’Angelo, Reid Hoffman, Will Hurd, Tasha McCauley, Helen Toner, and Shivon Zilis. Founder Elon Musk left the OpenAI board in February 2018 and is no longer formally involved in OpenAI. OpenAI employees are organized into three main areas:

OpenAI has developed artificial intelligence that has trained to play hide-and-seek against itself (involves hiders and seekers). The AI was trained without any predetermined explicit incentives, other than the hiders are to avoid the seekers' line of sight, and the seekers are meant to keep the hiders in their line of sight. The hiders eventually learned to find objects, and use them to block seekers, and continued to learn new strategies as seekers did to accomplish their tasks. Emergent strategies were the result of an autocurriculum. The company believes that these multi-agent dynamics can lead to complex behavior, similar to human behavior.

MuseNet is a deep neural network that is capable of composing four-minute musical pieces with up to ten different instruments, and write them in any genre, lasting four minutes. The software was not programmed with our understanding of music, but learned from predicting the next note in a MIDI file thousands of times. To train, the organization used data from ClassicalArchives and BitMidi.

GPT-3 is an autoregressive language model with 175 billion parameters made by OpenAI. thatIt launched on May 29, 2020 . The model builds on the transformer-based language model previously made by OpenAI called GPT-2 that, which had 1.5 billion parameters. GPT-3 improves on GPT-2 by adopting and scaling features present in GPT-2, such as modified initialization, pre-normalization, and reversible tokenization. GPT-3 training can improve scaling up language models primarily through improved task-agnostic and few-shot performance compared to the GPT-2 model. Open AI claims GPT-3 is approaching the performance of SOTA fine-tuned systems for generating quality performance and samples for defined tasks.

According to the GitHub depository of GPT-3, it can achieve strong performance when dealing with NLP datasets, such as translation, question-answering, and cloze tasks. It can also perform on-the-fly reasoning/domain adaptation tasks, like unscrambling words, using novel words in sentences, and performing 3-digitthree-digit arithmetic. The performance of GPT-3 in few-shot learning and training on large web corpora datasets facefaces methodological issues whichand produceproduces undesirable results which are undesirable. GPT-3 was founded to be capable of producing news articles that are difficult for humans to distinguish from news articles written by humans.

ChatGPT is a variant of the popular language generation model, GPT (Generative Pre-trained Transformer), in particular, GPT-3.5. ChatGPT is designed for chatbot applications, and withhas the ability to generate human-like responses to user input in a conversation, with follow-up questions and responses. The language model is capable of producing code, poems, songs, essays, stories (inspired by specific authors), and more. ChatGPT is generative, completing tasks for users rather than solely being a source of information.

The GPT model was trained on a large dataset of internet text and could generate human-like text for a variety of language tasks. As with GPT, ChatGPT uses a transformer architecture and is trained using unsupervised learning, which means it is able to learn from raw text data without the need for explicit labels or annotations. This allows it to generate text that is highly fluent and human-like, making it well-suited for chatbot applications wherewhen the goal is to create a natural and seamless conversation with users.

ChatGPT adds an additional layer of training to GPT-3.5 models called Reinforcement Learning with Human Feedback (RLHF). An initial model utilized supervised fine-tuning with human AI trainers providing conversations in which they played both sides, thesides—the user and the AI assistant. Human trainers were given access to model-written suggestions in order to compose responses. To build a reward model for reinforcement learning, these conversations were collected as comparison data. Model-written messages were randomly selected, sampling several alternative completions, and trainers ranked their quality. This information was fed back into the model using proximal policy optimization. The process was repeated several iterations to improve performance.

DALL-E is trained using a variant of the Transformer architecture, which was originallyinitially developed for natural language processing tasks, such as translation and language modeling. DALL-E is able to generate images by combining the information from the textual descriptions with a latent space of possible images. This allows it to generate a wide range of images, from photorealistic to highly stylized, depending on the input text.

The name DALL-E is a reference to the artist Salvador Dali and the Pixar character Wall-E. It was chosen to reflect the combination of creativity and technical capabilities that the system represents.