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

Released in 2018, Gym Retro is a platform for support knowing (RL) research study on computer game [147] using RL algorithms and research study generalization. Prior RL research study focused mainly on enhancing representatives to fix single jobs. Gym Retro provides the capability to generalize between video games with similar ideas but different appearances.

RoboSumo

Released in 2017, RoboSumo is a virtual world where humanoid metalearning robot agents initially do not have understanding of how to even stroll, but are provided the goals of learning to move and to push the opposing representative out of the ring. [148] Through this adversarial knowing process, the agents find out how to adjust to altering conditions. When a representative is then gotten rid of from this virtual environment and placed in a new virtual environment with high winds, the representative braces to remain upright, recommending it had actually discovered how to stabilize in a generalized method. [148] [149] OpenAI's Igor Mordatch argued that competitors in between representatives might produce an intelligence "arms race" that might increase a representative's capability to work even outside the context of the competition. [148]
OpenAI 5

OpenAI Five is a team of 5 OpenAI-curated bots utilized in the competitive five-on-five video game Dota 2, that find out to play against human players at a high skill level totally through experimental algorithms. Before becoming a team of 5, the very first public demonstration occurred at The International 2017, the annual best championship competition for the video game, where Dendi, an expert Ukrainian player, lost against a bot in a live individually match. [150] [151] After the match, CTO Greg Brockman explained that the bot had actually learned by playing against itself for 2 weeks of real time, which the learning software was a step in the instructions of producing software application that can handle intricate jobs like a cosmetic surgeon. [152] [153] The system utilizes a type of support learning, as the bots find out gradually by playing against themselves hundreds of times a day for months, and are rewarded for actions such as eliminating an opponent and taking map objectives. [154] [155] [156]
By June 2018, the ability of the bots broadened to play together as a full group of 5, and they had the ability to beat teams of amateur and semi-professional gamers. [157] [154] [158] [159] At The International 2018, OpenAI Five played in 2 exhibit matches against expert players, however ended up losing both video games. [160] [161] [162] In April 2019, OpenAI Five defeated OG, the reigning world champions of the video game at the time, 2:0 in a live exhibition match in San Francisco. [163] [164] The bots' final public look came later that month, where they played in 42,729 total games in a four-day open online competition, winning 99.4% of those video games. [165]
OpenAI 5's mechanisms in Dota 2's bot gamer shows the difficulties of AI systems in multiplayer online battle arena (MOBA) video games and how OpenAI Five has demonstrated using deep reinforcement knowing (DRL) agents to attain superhuman proficiency in Dota 2 matches. [166]
Dactyl

Developed in 2018, Dactyl utilizes device discovering to train a Shadow Hand, a human-like robotic hand, to control physical items. [167] It finds out completely in simulation using the exact same RL algorithms and training code as OpenAI Five. OpenAI tackled the things orientation issue by utilizing domain randomization, a simulation approach which exposes the student to a range of experiences instead of trying to fit to truth. The set-up for Dactyl, aside from having movement tracking cameras, also has RGB cams to enable the robot to control an approximate object by seeing it. In 2018, OpenAI showed that the system had the ability to control a cube and an octagonal prism. [168]
In 2019, OpenAI demonstrated that Dactyl might solve a Rubik's Cube. The robot had the ability to resolve the puzzle 60% of the time. Objects like the Rubik's Cube present complex physics that is harder to model. OpenAI did this by improving the effectiveness of Dactyl to perturbations by using Automatic Domain Randomization (ADR), a simulation technique of creating gradually more tough environments. ADR differs from manual domain randomization by not requiring a human to define randomization varieties. [169]
API

In June 2020, OpenAI announced a multi-purpose API which it said was "for accessing brand-new AI designs developed by OpenAI" to let developers get in touch with it for "any English language AI task". [170] [171]
Text generation

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

The initial paper on generative pre-training of a transformer-based language design was written by Alec Radford and his associates, and published in preprint on OpenAI's site on June 11, 2018. [173] It showed how a generative model of language might obtain world understanding and procedure long-range dependences 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 follower to OpenAI's original GPT model ("GPT-1"). GPT-2 was announced in February 2019, with only minimal demonstrative variations initially released to the general public. The complete version of GPT-2 was not instantly released due to issue about possible abuse, including applications for composing fake news. [174] Some professionals revealed uncertainty that GPT-2 positioned a significant threat.

In response to GPT-2, the Allen Institute for Artificial Intelligence responded with a tool to detect "neural phony news". [175] Other researchers, such as Jeremy Howard, alerted of "the innovation to completely fill Twitter, email, and the web up with reasonable-sounding, context-appropriate prose, which would muffle all other speech and be difficult to filter". [176] In November 2019, OpenAI released the total variation of the GPT-2 language model. [177] Several websites host interactive presentations of different instances of GPT-2 and other transformer designs. [178] [179] [180]
GPT-2's authors argue not being watched language designs to be general-purpose students, shown by GPT-2 attaining advanced accuracy and perplexity on 7 of 8 zero-shot tasks (i.e. the model 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 avoids certain concerns encoding vocabulary with word tokens by utilizing byte pair encoding. This permits representing any string of characters by encoding both private characters and multiple-character tokens. [181]
GPT-3

First explained in May 2020, Generative Pre-trained [a] Transformer 3 (GPT-3) is a not being watched transformer language model and the successor to GPT-2. [182] [183] [184] OpenAI mentioned that the complete variation of GPT-3 contained 175 billion parameters, [184] two orders of magnitude bigger than the 1.5 billion [185] in the complete variation of GPT-2 (although GPT-3 models with as few as 125 million criteria were likewise trained). [186]
OpenAI mentioned that GPT-3 was successful at certain "meta-learning" jobs and could generalize the function of a single input-output pair. The GPT-3 release paper gave examples of translation and cross-linguistic transfer learning in between English and Romanian, and between English and German. [184]
GPT-3 drastically improved benchmark results over GPT-2. OpenAI cautioned that such scaling-up of language models could be approaching or coming across the basic capability constraints of predictive language designs. [187] Pre-training GPT-3 needed several thousand petaflop/s-days [b] of calculate, compared to 10s of petaflop/s-days for the complete GPT-2 model. [184] Like its predecessor, [174] the GPT-3 trained model was not instantly released to the general public for concerns of possible abuse, although OpenAI prepared to permit 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 in addition 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 released in personal beta. [194] According to OpenAI, the model can produce working code in over a lots programming languages, most successfully in Python. [192]
Several concerns with problems, style flaws and security vulnerabilities were cited. [195] [196]
GitHub Copilot has actually been accused of giving off copyrighted code, with no author attribution or license. [197]
OpenAI announced that they would stop assistance 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 announced that the upgraded technology passed a simulated law school bar exam with a score 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, analyze or produce approximately 25,000 words of text, and write code in all significant programming languages. [200]
Observers reported that the model of ChatGPT utilizing GPT-4 was an improvement on the previous GPT-3.5-based iteration, with the caution 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 declined to reveal numerous technical details and statistics about GPT-4, such as the exact size of the model. [203]
GPT-4o

On May 13, 2024, OpenAI revealed and launched GPT-4o, which can process and create text, images and audio. [204] GPT-4o attained state-of-the-art lead to voice, multilingual, and vision standards, setting brand-new records in audio speech acknowledgment and translation. [205] [206] It scored 88.7% on the Massive Multitask Language Understanding (MMLU) standard compared to 86.5% by GPT-4. [207]
On July 18, 2024, OpenAI launched GPT-4o mini, a smaller version of GPT-4o changing GPT-3.5 Turbo on the ChatGPT interface. Its API costs $0.15 per million input tokens and $0.60 per million output tokens, compared to $5 and 135.181.29.174 $15 respectively for GPT-4o. OpenAI anticipates it to be especially useful for enterprises, start-ups and developers 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 been created to take more time to think of their actions, resulting in greater precision. These designs are especially reliable in science, coding, and reasoning jobs, and were made available to ChatGPT Plus and Team members. [209] [210] In December 2024, o1-preview was replaced by o1. [211]
o3

On December 20, 2024, OpenAI unveiled o3, the follower of the o1 thinking model. OpenAI also unveiled o3-mini, a lighter and much faster version of OpenAI o3. Since December 21, 2024, this model is not available for public use. According to OpenAI, they are testing o3 and o3-mini. [212] [213] Until January 10, 2025, security and security researchers had the opportunity to obtain early access to these models. [214] The design is called o3 instead of o2 to prevent confusion with telecoms companies O2. [215]
Deep research

Deep research is a representative developed by OpenAI, unveiled on February 2, 2025. It leverages the abilities of OpenAI's o3 model to perform extensive web browsing, information analysis, and synthesis, delivering detailed reports within a timeframe of 5 to thirty minutes. [216] With searching and Python tools enabled, it reached an accuracy of 26.6 percent on HLE (Humanity's Last Exam) benchmark. [120]
Image classification

CLIP

Revealed in 2021, CLIP (Contrastive Language-Image Pre-training) is a model that is trained to evaluate the semantic similarity in between text and images. It can notably be used for image category. [217]
Text-to-image

DALL-E

Revealed in 2021, DALL-E is a Transformer model that develops images from textual descriptions. [218] DALL-E utilizes a 12-billion-parameter variation of GPT-3 to translate natural language inputs (such as "a green leather handbag formed like a pentagon" or "an isometric view of an unfortunate capybara") and produce corresponding images. It can create pictures of sensible things ("a stained-glass window with an image of a blue strawberry") along with things that do not exist in reality ("a cube with the texture of a porcupine"). Since March 2021, no API or code is available.

DALL-E 2

In April 2022, OpenAI revealed DALL-E 2, an updated variation of the model with more reasonable outcomes. [219] In December 2022, OpenAI released on GitHub software for Point-E, a new simple system for converting a text description into a 3-dimensional design. [220]
DALL-E 3

In September 2023, OpenAI revealed DALL-E 3, a more powerful model much better able to produce images from complicated descriptions without manual prompt engineering and render complicated details like hands and text. [221] It was launched to the public as a ChatGPT Plus feature in October. [222]
Text-to-video

Sora

Sora is a text-to-video model that can produce videos based upon brief 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 optimum length of created videos is unknown.

Sora's development group called it after the Japanese word for "sky", to represent its "limitless innovative capacity". [223] Sora's technology is an adaptation of the technology behind the DALL · E 3 text-to-image model. [225] OpenAI trained the system utilizing publicly-available videos in addition to copyrighted videos certified for that function, however did not reveal the number or the exact sources of the videos. [223]
OpenAI showed some Sora-created high-definition videos to the public on February 15, 2024, mentioning that it could generate videos up to one minute long. It likewise shared a technical report highlighting the approaches used to train the design, and the design's abilities. [225] It acknowledged a few of its imperfections, including struggles simulating complicated physics. [226] Will Douglas Heaven of the MIT Technology Review called the demonstration videos "remarkable", but noted that they need to have been cherry-picked and might not represent Sora's common output. [225]
Despite uncertainty from some academic leaders following Sora's public demonstration, notable entertainment-industry figures have actually revealed substantial interest in the innovation's potential. In an interview, actor/filmmaker Tyler Perry expressed his awe at the technology's capability to produce practical video from text descriptions, mentioning its prospective to transform storytelling and content creation. He said that his excitement about Sora's possibilities was so strong that he had chosen to stop briefly prepare for expanding his Atlanta-based motion picture studio. [227]
Speech-to-text

Whisper

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

MuseNet

Released in 2019, MuseNet is a deep neural net trained to forecast subsequent musical notes in MIDI music files. It can create songs with 10 instruments in 15 styles. According to The Verge, a tune created by MuseNet tends to begin fairly but then fall into mayhem the longer it plays. [230] [231] In popular culture, initial applications of this tool were utilized as early as 2020 for the web mental thriller Ben Drowned to produce 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 category, artist, and a bit of lyrics and outputs song samples. OpenAI stated the songs "reveal regional musical coherence [and] follow standard chord patterns" however acknowledged that the songs lack "familiar larger musical structures such as choruses that duplicate" and that "there is a significant space" in between Jukebox and human-generated music. The Verge mentioned "It's technically impressive, even if the results sound like mushy versions of songs that may feel familiar", while Business Insider specified "remarkably, some of the resulting tunes are memorable and sound legitimate". [234] [235] [236]
Interface

Debate Game

In 2018, OpenAI launched the Debate Game, which teaches makers to debate toy problems in front of a human judge. The purpose is to research whether such a method might help in auditing AI decisions and in developing explainable AI. [237] [238]
Microscope

Released in 2020, Microscope [239] is a collection of visualizations of every substantial layer and nerve cell of eight neural network models which are studied in interpretability. [240] Microscope was produced to analyze the features that form inside these neural networks quickly. The models included are AlexNet, VGG-19, different versions of Inception, and various variations of CLIP Resnet. [241]
ChatGPT

Launched in November 2022, ChatGPT is an expert system tool developed on top of GPT-3 that supplies a conversational user interface that allows users to ask questions in natural language. The system then responds with a response within seconds.