Announced in 2016, Gym is an open-source Python library created to facilitate the development of reinforcement knowing algorithms. It aimed to standardize how environments are specified in AI research study, making published research more quickly reproducible [24] [144] while users with a basic user interface for connecting with these environments. In 2022, new advancements of Gym have actually been transferred to the library Gymnasium. [145] [146]
Gym Retro

Released in 2018, Gym Retro is a platform for support knowing (RL) research on video games [147] using RL algorithms and research study generalization. Prior RL research study focused mainly on optimizing representatives to solve single tasks. Gym Retro gives the capability to generalize in between games with comparable concepts however different looks.

RoboSumo

Released in 2017, RoboSumo is a virtual world where humanoid metalearning robotic agents initially do not have knowledge of how to even stroll, but are offered the goals of learning to move and to press the opposing agent out of the ring. [148] Through this adversarial learning procedure, the representatives find out how to adjust to altering conditions. When a representative is then eliminated 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 learned how to balance in a generalized way. [148] [149] OpenAI's Igor Mordatch argued that competitors between agents might create an intelligence "arms race" that could increase an agent's capability to work even outside the context of the competitors. [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 team of 5, the first public presentation happened at The International 2017, the yearly premiere championship tournament for the game, where Dendi, a professional Ukrainian gamer, lost against a bot in a live individually matchup. [150] [151] After the match, CTO Greg Brockman explained that the bot had actually learned by playing against itself for 2 weeks of actual time, which the learning software was an action in the direction of developing software application that can handle complex tasks like a surgeon. [152] [153] The system uses a kind of reinforcement knowing, as the bots learn in time by playing against themselves hundreds of times a day for months, and are rewarded for actions such as eliminating an enemy and taking map objectives. [154] [155] [156]
By June 2018, the ability of the bots expanded to play together as a full team of 5, and they were able to beat groups of amateur and semi-professional gamers. [157] [154] [158] [159] At The International 2018, OpenAI Five played in 2 exhibit matches against professional gamers, however wound up losing both video games. [160] [161] [162] In April 2019, OpenAI Five beat OG, the ruling 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 video games in a four-day open online competition, winning 99.4% of those video games. [165]
OpenAI 5's systems in Dota 2's bot gamer reveals the challenges of AI systems in multiplayer online battle arena (MOBA) video games and how OpenAI Five has shown the use of deep support knowing (DRL) representatives to attain superhuman skills in Dota 2 matches. [166]
Dactyl

Developed in 2018, Dactyl utilizes machine discovering to train a Shadow Hand, a human-like robotic hand, to manipulate physical objects. [167] It discovers totally in simulation using the exact same RL algorithms and training code as OpenAI Five. OpenAI dealt with the item orientation problem by utilizing domain randomization, a simulation approach which exposes the student to a variety of experiences instead of attempting to fit to reality. The set-up for Dactyl, aside from having motion tracking electronic cameras, also has RGB cams to enable the robot to control an arbitrary item by seeing it. In 2018, OpenAI showed that the system was able to manipulate a cube and an octagonal prism. [168]
In 2019, OpenAI demonstrated that Dactyl could resolve a Rubik's Cube. The robot had the ability to fix the puzzle 60% of the time. Objects like the Rubik's Cube present intricate physics that is harder to design. OpenAI did this by enhancing the toughness of Dactyl to perturbations by utilizing Automatic Domain Randomization (ADR), a simulation method of generating progressively harder 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 new AI designs established by OpenAI" to let designers get in touch with it for "any English language AI job". [170] [171]
Text generation

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

The initial paper on generative pre-training of a transformer-based language design was written by Alec Radford and his coworkers, and published in preprint on OpenAI's site on June 11, 2018. [173] It demonstrated how a generative model of language might obtain world knowledge and procedure long-range dependences by pre-training on a diverse corpus with long stretches of contiguous text.

GPT-2

Generative Pre-trained Transformer 2 ("GPT-2") is a without supervision transformer language model and the follower to OpenAI's initial GPT model ("GPT-1"). GPT-2 was announced in February 2019, with only minimal demonstrative variations initially launched to the public. The full version of GPT-2 was not immediately launched due to concern about possible abuse, consisting of applications for writing phony news. [174] Some professionals expressed uncertainty that GPT-2 presented a considerable hazard.

In response to GPT-2, the Allen Institute for Artificial Intelligence reacted with a tool to find "neural fake news". [175] Other scientists, such as Jeremy Howard, forum.altaycoins.com warned 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 impossible to filter". [176] In November 2019, OpenAI released the total version of the GPT-2 language design. [177] Several sites host interactive demonstrations of various circumstances of GPT-2 and other transformer models. [178] [179] [180]
GPT-2's authors argue not being watched language models to be general-purpose learners, illustrated by GPT-2 attaining modern precision and perplexity on 7 of 8 zero-shot jobs (i.e. the model was not further trained on any task-specific input-output examples).

The corpus it was trained on, called WebText, contains somewhat 40 gigabytes of text from URLs shared in Reddit submissions with at least 3 upvotes. It avoids certain issues encoding vocabulary with word tokens by utilizing byte pair encoding. This allows representing any string of characters by encoding both specific 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 full version of GPT-3 contained 175 billion criteria, [184] 2 orders of magnitude larger than the 1.5 billion [185] in the full version of GPT-2 (although GPT-3 designs with as couple of as 125 million parameters were also trained). [186]
OpenAI specified that GPT-3 succeeded at certain "meta-learning" tasks and could generalize the purpose of a single input-output pair. The GPT-3 release paper offered examples of translation and cross-linguistic transfer learning between English and Romanian, and between English and German. [184]
GPT-3 dramatically improved benchmark outcomes over GPT-2. OpenAI cautioned that such scaling-up of language models could be approaching or encountering the essential 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 design. [184] Like its predecessor, [174] the GPT-3 trained model was not immediately launched to the public for issues of possible abuse, although OpenAI planned to permit gain access to through a paid cloud API after a two-month complimentary private beta that began in June 2020. [170] [189]
On September 23, 2020, GPT-3 was certified exclusively to Microsoft. [190] [191]
Codex

Announced in mid-2021, Codex is a descendant of GPT-3 that has actually furthermore 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 private beta. [194] According to OpenAI, the model can create working code in over a dozen programs languages, the majority of efficiently in Python. [192]
Several issues with glitches, style flaws and security vulnerabilities were pointed out. [195] [196]
GitHub Copilot has been accused of producing copyrighted code, with no author attribution or license. [197]
OpenAI revealed that they would discontinue 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 revealed that the updated innovation passed a simulated law school bar exam with a score around the top 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 compose code in all major programming languages. [200]
Observers reported that the model of ChatGPT using GPT-4 was an enhancement on the previous GPT-3.5-based model, with the caution that GPT-4 retained a few of the issues with earlier modifications. [201] GPT-4 is likewise capable of taking images as input on ChatGPT. [202] OpenAI has decreased to reveal numerous technical details and stats about GPT-4, such as the exact size of the model. [203]
GPT-4o

On May 13, 2024, OpenAI announced and launched GPT-4o, which can process and produce text, images and audio. [204] GPT-4o attained advanced outcomes in voice, multilingual, and vision criteria, setting brand-new records in audio speech acknowledgment 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 replacing 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 particularly useful for business, 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 developed to take more time to consider their reactions, resulting in higher accuracy. These designs are particularly efficient in science, coding, and thinking jobs, and were made available to ChatGPT Plus and Staff member. [209] [210] In December 2024, o1-preview was replaced by o1. [211]
o3

On December 20, 2024, OpenAI revealed o3, the follower of the o1 thinking model. OpenAI also unveiled o3-mini, a lighter and quicker variation of OpenAI o3. Since 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 scientists had the opportunity to obtain early access to these designs. [214] The design is called o3 instead of o2 to prevent confusion with telecoms providers O2. [215]
Deep research

Deep research is a representative established by OpenAI, unveiled on February 2, 2025. It leverages the abilities of OpenAI's o3 design to perform comprehensive web browsing, data analysis, and synthesis, providing detailed reports within a timeframe of 5 to 30 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 design that is trained to evaluate the semantic resemblance in between text and images. It can notably be utilized for image classification. [217]
Text-to-image

DALL-E

Revealed in 2021, DALL-E is a Transformer model that creates images from textual descriptions. [218] DALL-E utilizes a 12-billion-parameter version 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 develop images of realistic items ("a stained-glass window with a picture of a blue strawberry") in addition to items that do not exist in reality ("a cube with the texture of a porcupine"). As of March 2021, no API or code is available.

DALL-E 2

In April 2022, OpenAI announced DALL-E 2, an updated version of the design with more practical outcomes. [219] In December 2022, OpenAI released on GitHub software application for Point-E, a new fundamental system for transforming a text description into a 3-dimensional design. [220]
DALL-E 3

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

Sora

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

Sora's development team named it after the Japanese word for "sky", to symbolize its "endless creative potential". [223] Sora's technology is an adaptation of the technology behind the DALL · E 3 text-to-image model. [225] OpenAI trained the system using publicly-available videos as well as copyrighted videos accredited for that purpose, but did not expose 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, mentioning that it might produce videos as much as one minute long. It likewise shared a technical report highlighting the methods used to train the model, and the design's capabilities. [225] It acknowledged a few of its imperfections, consisting of battles mimicing intricate physics. [226] Will Douglas Heaven of the MIT Technology Review called the presentation videos "impressive", but noted that they must have been cherry-picked and may not represent Sora's normal output. [225]
Despite uncertainty from some academic leaders following Sora's public demo, significant entertainment-industry figures have actually revealed substantial interest in the technology's capacity. In an interview, actor/filmmaker Tyler Perry expressed his astonishment at the innovation's ability to produce sensible video from text descriptions, mentioning its prospective to transform storytelling and material creation. He said that his excitement about Sora's possibilities was so strong that he had chosen to pause strategies for broadening his Atlanta-based film 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 diverse audio and is likewise a multi-task model that can perform multilingual speech acknowledgment in addition to speech translation and language identification. [229]
Music generation

MuseNet

Released in 2019, MuseNet is a deep neural net trained to anticipate subsequent musical notes in MIDI music files. It can generate songs with 10 instruments in 15 styles. According to The Verge, a tune generated by MuseNet tends to start fairly however then fall under turmoil the longer it plays. [230] [231] In popular culture, preliminary applications of this tool were used as early as 2020 for the internet mental thriller Ben Drowned to develop music for the titular character. [232] [233]
Jukebox

Released in 2020, Jukebox is an open-sourced algorithm to generate music with vocals. After training on 1.2 million samples, the system accepts a category, artist, and a snippet of lyrics and outputs tune samples. OpenAI stated the songs "show local musical coherence [and] follow conventional chord patterns" but acknowledged that the songs do not have "familiar larger musical structures such as choruses that repeat" which "there is a considerable gap" between Jukebox and human-generated music. The Verge stated "It's technologically impressive, even if the outcomes sound like mushy versions of songs that might feel familiar", while Business Insider mentioned "surprisingly, a few of the resulting tunes are memorable and sound legitimate". [234] [235] [236]
Interface

Debate Game

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

Released in 2020, Microscope [239] is a collection of visualizations of every substantial layer and neuron of 8 neural network designs which are often studied in interpretability. [240] Microscope was produced to examine the functions that form inside these neural networks quickly. 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 constructed on top of GPT-3 that provides a conversational interface that allows users to ask questions in natural language. The system then responds with a response within seconds.