Announced in 2016, Gym is an open-source Python library designed to help with the advancement of support learning algorithms. It aimed to standardize how environments are specified in AI research, making released research more quickly reproducible [24] [144] while providing users with a simple interface for engaging with these environments. In 2022, new developments of Gym have been transferred to the library Gymnasium. [145] [146]
Gym Retro

Released in 2018, Gym Retro is a platform for support learning (RL) research on computer game [147] using RL algorithms and research study generalization. Prior RL research study focused mainly on optimizing agents to fix single tasks. Gym Retro provides the ability to generalize between video games with similar principles but different appearances.

RoboSumo

Released in 2017, RoboSumo is a virtual world where humanoid metalearning robot representatives at first do not have knowledge of how to even stroll, but are given the objectives of finding out to move and to press the opposing agent out of the ring. [148] Through this adversarial knowing process, the agents learn how to adapt to altering conditions. When a representative is then gotten rid of from this virtual environment and put in a brand-new virtual environment with high winds, the agent braces to remain upright, recommending it had actually discovered how to stabilize in a generalized way. [148] [149] OpenAI's Igor Mordatch argued that competitors in between representatives might develop an intelligence "arms race" that could increase a representative's capability to operate even outside the context of the competition. [148]
OpenAI 5

OpenAI Five is a group of 5 OpenAI-curated bots utilized in the competitive five-on-five video game Dota 2, that discover to play against human players at a high skill level entirely through trial-and-error algorithms. Before ending up being a group of 5, the very first public demonstration happened at The International 2017, the yearly premiere champion competition for the video game, where Dendi, an expert Ukrainian gamer, lost against a bot in a live one-on-one matchup. [150] [151] After the match, CTO Greg Brockman explained that the bot had actually found out by playing against itself for 2 weeks of real time, which the learning software was a step in the direction of creating software application that can handle intricate tasks like a surgeon. [152] [153] The system utilizes a form of reinforcement learning, as the bots learn over time by playing against themselves hundreds of times a day for months, and are rewarded for actions such as eliminating an opponent and taking map goals. [154] [155] [156]
By June 2018, the capability of the bots broadened to play together as a complete group of 5, and they were able to beat groups of amateur and semi-professional players. [157] [154] [158] [159] At The International 2018, OpenAI Five played in 2 exhibit matches against expert players, however wound up losing both video games. [160] [161] [162] In April 2019, OpenAI Five beat OG, the reigning world champions of the game at the time, 2:0 in a live exhibit match in San Francisco. [163] [164] The bots' final public look came later on that month, where they played in 42,729 overall games in a four-day open online competition, winning 99.4% of those games. [165]
OpenAI 5's mechanisms in Dota 2's bot player reveals the obstacles of AI systems in multiplayer online battle arena (MOBA) video games and how OpenAI Five has actually shown the use of deep support knowing (DRL) agents to attain superhuman proficiency in Dota 2 matches. [166]
Dactyl

Developed in 2018, Dactyl utilizes machine learning to train a Shadow Hand, a human-like robotic hand, to control physical objects. [167] It learns completely in simulation using the very same RL algorithms and training code as OpenAI Five. OpenAI dealt with the item orientation issue by utilizing domain randomization, a simulation method which exposes the student to a variety of experiences rather than attempting to fit to reality. The set-up for Dactyl, aside from having movement tracking cams, also has RGB cameras to allow the robotic to control an approximate item by seeing it. In 2018, OpenAI showed that the system was able to control a cube and an octagonal prism. [168]
In 2019, OpenAI demonstrated that Dactyl might solve a Rubik's Cube. The robotic had the ability to fix the puzzle 60% of the time. Objects like the Rubik's Cube introduce complicated physics that is harder to model. OpenAI did this by enhancing the toughness of Dactyl to perturbations by utilizing Automatic Domain Randomization (ADR), a simulation approach of generating progressively harder environments. ADR varies from manual domain randomization by not requiring a human to define randomization ranges. [169]
API

In June 2020, OpenAI announced a multi-purpose API which it said was "for accessing new AI models developed by OpenAI" to let developers contact it for "any English language AI job". [170] [171]
Text generation

The business has actually popularized generative pretrained transformers (GPT). [172]
OpenAI's original GPT design ("GPT-1")

The original paper on generative pre-training of a transformer-based language design was composed by Alec Radford and his colleagues, and published in preprint on OpenAI's website on June 11, 2018. [173] It demonstrated how a generative design of language might obtain world understanding and procedure long-range dependencies by pre-training on a varied corpus with long stretches of adjoining text.

GPT-2

Generative Pre-trained Transformer 2 ("GPT-2") is a without supervision transformer language design and the successor to OpenAI's original GPT model ("GPT-1"). GPT-2 was revealed in February 2019, with just restricted demonstrative variations at first launched to the public. The complete version of GPT-2 was not right away released due to concern about prospective misuse, including applications for writing phony news. [174] Some specialists expressed uncertainty that GPT-2 postured a significant threat.

In reaction to GPT-2, the Allen Institute for Artificial Intelligence responded with a tool to identify "neural phony news". [175] Other researchers, such as Jeremy Howard, cautioned of "the innovation to completely fill Twitter, email, and the web up with reasonable-sounding, context-appropriate prose, which would hush all other speech and be impossible to filter". [176] In November 2019, OpenAI launched the total version of the GPT-2 language design. [177] Several websites host interactive presentations of different circumstances of GPT-2 and other transformer designs. [178] [179] [180]
GPT-2's authors argue unsupervised language models to be general-purpose students, illustrated by GPT-2 attaining state-of-the-art precision and perplexity on 7 of 8 zero-shot jobs (i.e. the design was not more trained on any task-specific input-output examples).

The corpus it was trained on, called WebText, contains a little 40 gigabytes of text from URLs shared in Reddit submissions with at least 3 upvotes. It prevents certain problems encoding vocabulary with word tokens by using byte pair . This permits representing any string of characters by encoding both individual characters and multiple-character tokens. [181]
GPT-3

First explained in May 2020, Generative Pre-trained [a] Transformer 3 (GPT-3) is an unsupervised transformer language model and the follower to GPT-2. [182] [183] [184] OpenAI specified that the full variation of GPT-3 contained 175 billion specifications, [184] two orders of magnitude bigger than the 1.5 billion [185] in the complete variation of GPT-2 (although GPT-3 designs with as few as 125 million parameters were likewise trained). [186]
OpenAI specified that GPT-3 prospered at certain "meta-learning" tasks and might generalize the purpose of a single input-output pair. The GPT-3 release paper provided examples of translation and cross-linguistic transfer knowing in between English and Romanian, and in between English and German. [184]
GPT-3 considerably improved benchmark outcomes over GPT-2. OpenAI warned that such scaling-up of language models could be approaching or encountering the essential ability constraints of predictive language designs. [187] Pre-training GPT-3 needed numerous thousand petaflop/s-days [b] of compute, compared to 10s of petaflop/s-days for the full GPT-2 design. [184] Like its predecessor, [174] the GPT-3 trained design was not immediately released to the general public for issues of possible abuse, although OpenAI planned to allow gain access to through a paid cloud API after a two-month free personal beta that started in June 2020. [170] [189]
On September 23, 2020, GPT-3 was certified specifically to Microsoft. [190] [191]
Codex

Announced in mid-2021, Codex is a descendant of GPT-3 that has actually additionally been trained on code from 54 million GitHub repositories, [192] [193] and is the AI powering the code autocompletion tool GitHub Copilot. [193] In August 2021, an API was launched in private beta. [194] According to OpenAI, the model can produce working code in over a lots programs languages, most successfully in Python. [192]
Several concerns with problems, style defects and security vulnerabilities were mentioned. [195] [196]
GitHub Copilot has actually been implicated of emitting copyrighted code, without any author attribution or license. [197]
OpenAI revealed that they would discontinue support for Codex API on March 23, 2023. [198]
GPT-4

On March 14, 2023, OpenAI revealed the release of Generative Pre-trained Transformer 4 (GPT-4), capable of accepting text or image inputs. [199] They revealed that the updated technology passed a simulated law school bar exam with a rating around the leading 10% of test takers. (By contrast, GPT-3.5 scored around the bottom 10%.) They said that GPT-4 could also read, examine or produce as much as 25,000 words of text, and write code in all significant shows languages. [200]
Observers reported that the version of ChatGPT utilizing GPT-4 was an improvement on the previous GPT-3.5-based iteration, with the caveat that GPT-4 retained a few of the issues with earlier revisions. [201] GPT-4 is likewise efficient in taking images as input on ChatGPT. [202] OpenAI has actually declined to expose various technical details and statistics about GPT-4, such as the exact size of the design. [203]
GPT-4o

On May 13, 2024, OpenAI announced and launched GPT-4o, which can process and create text, images and audio. [204] GPT-4o attained cutting edge results in voice, multilingual, and vision benchmarks, setting brand-new records in audio speech recognition and translation. [205] [206] It scored 88.7% on the Massive Multitask Language Understanding (MMLU) benchmark compared to 86.5% by GPT-4. [207]
On July 18, 2024, OpenAI released GPT-4o mini, a smaller sized version of GPT-4o changing GPT-3.5 Turbo on the ChatGPT user interface. Its API costs $0.15 per million input tokens and $0.60 per million output tokens, compared to $5 and $15 respectively for GPT-4o. OpenAI anticipates it to be especially beneficial for enterprises, startups and designers looking for to automate services with AI representatives. [208]
o1

On September 12, 2024, OpenAI launched the o1-preview and o1-mini models, which have actually been developed to take more time to think of their reactions, resulting in greater accuracy. These designs are especially effective in science, coding, and reasoning tasks, and were made available to ChatGPT Plus and Employee. [209] [210] In December 2024, o1-preview was changed by o1. [211]
o3

On December 20, 2024, OpenAI unveiled o3, the follower of the o1 thinking model. OpenAI likewise unveiled o3-mini, a lighter and much faster variation of OpenAI o3. As of December 21, 2024, this model is not available for public use. According to OpenAI, they are checking o3 and o3-mini. [212] [213] Until January 10, 2025, security and security researchers had the chance to obtain early access to these models. [214] The design is called o3 rather than o2 to prevent confusion with telecoms services service provider O2. [215]
Deep research study

Deep research study is an agent developed by OpenAI, unveiled on February 2, 2025. It leverages the capabilities of OpenAI's o3 design to perform substantial web browsing, data analysis, and synthesis, providing detailed reports within a timeframe of 5 to thirty minutes. [216] With browsing and Python tools made it possible for, it reached an accuracy of 26.6 percent on HLE (Humanity's Last Exam) benchmark. [120]
Image category

CLIP

Revealed in 2021, CLIP (Contrastive Language-Image Pre-training) is a model that is trained to analyze the semantic resemblance between text and images. It can significantly be utilized for image classification. [217]
Text-to-image

DALL-E

Revealed in 2021, DALL-E is a Transformer design that creates images from textual descriptions. [218] DALL-E utilizes a 12-billion-parameter version of GPT-3 to interpret natural language inputs (such as "a green leather bag shaped like a pentagon" or "an isometric view of a sad capybara") and create matching images. It can develop images of sensible items ("a stained-glass window with an image of a blue strawberry") in addition to objects that do not exist in truth ("a cube with the texture of a porcupine"). Since March 2021, no API or code is available.

DALL-E 2

In April 2022, OpenAI announced DALL-E 2, an upgraded version of the model with more practical results. [219] In December 2022, OpenAI released on GitHub software for Point-E, a brand-new simple system for converting a text description into a 3-dimensional model. [220]
DALL-E 3

In September 2023, OpenAI announced DALL-E 3, a more powerful design better able to create images from complicated descriptions without manual timely engineering and render complex details like hands and text. [221] It was released to the public as a ChatGPT Plus function in October. [222]
Text-to-video

Sora

Sora is a text-to-video model that can generate videos based on short detailed prompts [223] along with extend existing videos forwards or backwards in time. [224] It can produce videos with resolution up to 1920x1080 or 1080x1920. The maximal length of produced videos is unidentified.

Sora's advancement group named it after the Japanese word for "sky", to represent its "endless creative capacity". [223] Sora's technology is an adaptation of the technology behind the DALL · E 3 text-to-image design. [225] OpenAI trained the system utilizing publicly-available videos along with copyrighted videos accredited for that function, however did not reveal the number or the precise sources of the videos. [223]
OpenAI demonstrated some Sora-created high-definition videos to the general public on February 15, 2024, specifying that it might generate videos approximately one minute long. It also shared a technical report highlighting the techniques utilized to train the model, and the design's abilities. [225] It acknowledged some of its drawbacks, wakewiki.de including struggles replicating complex physics. [226] Will Douglas Heaven of the MIT Technology Review called the demonstration videos "outstanding", however noted that they must have been cherry-picked and might not represent Sora's typical output. [225]
Despite uncertainty from some scholastic leaders following Sora's public demonstration, significant entertainment-industry figures have revealed significant interest in the technology's potential. In an interview, actor/filmmaker Tyler Perry expressed his awe at the innovation's ability to generate reasonable video from text descriptions, citing its prospective to revolutionize storytelling and content creation. He said that his excitement about Sora's possibilities was so strong that he had actually decided to pause strategies for expanding his Atlanta-based movie studio. [227]
Speech-to-text

Whisper

Released in 2022, Whisper is a general-purpose speech recognition design. [228] It is trained on a large dataset of diverse audio and is also a multi-task model that can carry out multilingual speech recognition along with speech translation and language identification. [229]
Music generation

MuseNet

Released in 2019, MuseNet is a deep neural net trained to predict subsequent musical notes in MIDI music files. It can generate songs with 10 instruments in 15 designs. According to The Verge, a song created by MuseNet tends to begin fairly but then fall under chaos the longer it plays. [230] [231] In popular culture, initial applications of this tool were used as early as 2020 for the internet mental thriller Ben Drowned to create music for the titular character. [232] [233]
Jukebox

Released in 2020, Jukebox is an open-sourced algorithm to produce music with vocals. After training on 1.2 million samples, the system accepts a genre, artist, and a snippet of lyrics and outputs song samples. OpenAI mentioned the songs "reveal local musical coherence [and] follow standard chord patterns" however acknowledged that the songs lack "familiar bigger musical structures such as choruses that duplicate" which "there is a substantial space" in between Jukebox and human-generated music. The Verge stated "It's technically impressive, even if the results seem like mushy versions of songs that may feel familiar", while Business Insider specified "remarkably, a few of the resulting songs are memorable and sound genuine". [234] [235] [236]
Interface

Debate Game

In 2018, OpenAI launched the Debate Game, which teaches makers to dispute toy problems in front of a human judge. The function is to research whether such a technique may assist in auditing AI choices and in developing explainable AI. [237] [238]
Microscope

Released in 2020, Microscope [239] is a collection of visualizations of every considerable layer and nerve cell of eight neural network designs which are typically studied in interpretability. [240] Microscope was developed to analyze the functions that form inside these neural networks easily. The designs consisted of are AlexNet, VGG-19, various versions of Inception, and various variations of CLIP Resnet. [241]
ChatGPT

Launched in November 2022, ChatGPT is a synthetic intelligence tool built on top of GPT-3 that supplies a conversational user interface that enables users to ask concerns in natural language. The system then reacts with an answer within seconds.