Announced in 2016, Gym is an open-source Python library developed to help with the development of support learning algorithms. It aimed to standardize how environments are defined in AI research study, making released research more quickly reproducible [24] [144] while supplying users with a simple user interface for engaging with these environments. In 2022, 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 learning (RL) research study on video games [147] utilizing RL algorithms and study generalization. Prior RL research focused mainly on enhancing representatives to resolve single jobs. Gym Retro offers the capability to generalize between games with comparable ideas however different appearances.

RoboSumo

Released in 2017, RoboSumo is a virtual world where humanoid metalearning robot representatives initially do not have understanding of how to even stroll, but are given the objectives of learning to move and to push the opposing representative out of the ring. [148] Through this adversarial learning process, the agents learn how to adapt to changing conditions. When a representative is then eliminated from this virtual environment and placed in a new virtual environment with high winds, the representative braces to remain upright, suggesting it had discovered how to balance in a generalized method. [148] [149] OpenAI's Igor Mordatch argued that competitors in between representatives could produce an intelligence "arms race" that might increase an agent'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 computer game Dota 2, forum.altaycoins.com that find out to play against human gamers at a high ability level totally through experimental algorithms. Before ending up being a group of 5, the first public presentation occurred at The International 2017, the annual best championship tournament for the 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 discovered by playing against itself for 2 weeks of actual time, and that the knowing software was a step in the instructions of creating software application that can manage complicated jobs like a cosmetic surgeon. [152] [153] The system utilizes a type of support knowing, as the bots discover with time by playing against themselves numerous times a day for months, and are rewarded for actions such as eliminating an enemy and taking map goals. [154] [155] [156]
By June 2018, the ability of the bots broadened to play together as a complete group of 5, and they had the ability to defeat groups of amateur and semi-professional gamers. [157] [154] [158] [159] At The International 2018, OpenAI Five played in two exhibit matches against professional players, however ended 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' last public look came later on that month, where they played in 42,729 overall games in a four-day open online competitors, winning 99.4% of those games. [165]
OpenAI 5's mechanisms in Dota 2's bot gamer shows the obstacles of AI systems in multiplayer online battle arena (MOBA) games and how OpenAI Five has actually demonstrated the use of deep reinforcement knowing (DRL) agents to attain superhuman skills in Dota 2 matches. [166]
Dactyl

Developed in 2018, Dactyl utilizes machine learning to train a Shadow Hand, a human-like robotic hand, to manipulate physical things. [167] It learns completely in simulation utilizing the exact same RL algorithms and training code as OpenAI Five. OpenAI took on the object orientation problem by utilizing domain randomization, a simulation technique which exposes the learner to a range of experiences instead of attempting to fit to reality. The set-up for Dactyl, aside from having movement tracking electronic cameras, likewise has RGB video cameras to permit the robotic to control an arbitrary object by seeing it. In 2018, OpenAI showed that the system was able to control a cube and an octagonal prism. [168]
In 2019, OpenAI showed that Dactyl could solve a Rubik's Cube. The robotic was able to resolve the puzzle 60% of the time. Objects like the Rubik's Cube introduce intricate physics that is harder to design. OpenAI did this by enhancing the robustness of Dactyl to perturbations by using Automatic Domain Randomization (ADR), a simulation method of generating progressively more difficult environments. ADR differs from manual domain randomization by not needing a human to define randomization varieties. [169]
API

In June 2020, OpenAI revealed a multi-purpose API which it said was "for accessing brand-new AI models established by OpenAI" to let designers call on it for "any English language AI task". [170] [171]
Text generation

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

The original paper on generative pre-training of a transformer-based language model was written by Alec Radford and his coworkers, and published in preprint on OpenAI's website on June 11, 2018. [173] It demonstrated how a generative model 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 an unsupervised transformer language design and the successor engel-und-waisen.de to OpenAI's original GPT design ("GPT-1"). GPT-2 was revealed in February 2019, with only minimal demonstrative versions initially launched to the public. The full version of GPT-2 was not immediately released due to issue about prospective misuse, including applications for writing fake news. [174] Some experts revealed uncertainty that GPT-2 positioned a substantial threat.

In reaction to GPT-2, the Allen Institute for Artificial Intelligence reacted with a tool to identify "neural phony news". [175] Other scientists, such as Jeremy Howard, alerted of "the technology to absolutely 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 launched the complete version of the GPT-2 language model. [177] Several websites host interactive presentations of different circumstances of GPT-2 and other transformer designs. [178] [179] [180]
GPT-2's authors argue without supervision language designs to be general-purpose learners, shown by GPT-2 attaining modern accuracy and perplexity on 7 of 8 zero-shot tasks (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 with at least 3 upvotes. It prevents certain problems encoding vocabulary with word tokens by using byte pair encoding. This allows 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 without supervision transformer language model and the follower to GPT-2. [182] [183] [184] OpenAI mentioned that the complete variation of GPT-3 contained 175 billion specifications, [184] 2 orders of magnitude bigger than the 1.5 billion [185] in the full variation of GPT-2 (although GPT-3 models with as few as 125 million specifications were likewise trained). [186]
OpenAI stated that GPT-3 was successful at certain "meta-learning" tasks and wiki.dulovic.tech might generalize the function 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 between English and German. [184]
GPT-3 significantly improved benchmark outcomes over GPT-2. OpenAI warned that such scaling-up of language designs might be approaching or coming across the fundamental ability constraints of predictive language designs. [187] Pre-training GPT-3 required numerous thousand petaflop/s-days [b] of calculate, compared to tens of petaflop/s-days for the complete GPT-2 model. [184] Like its predecessor, [174] the GPT-3 trained model was not immediately launched to the public for concerns of possible abuse, although OpenAI prepared to enable gain access to through a paid cloud API after a two-month complimentary personal beta that started 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 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 personal beta. [194] According to OpenAI, the model can produce working code in over a lots shows languages, the majority of efficiently in Python. [192]
Several issues with glitches, style flaws and security vulnerabilities were cited. [195] [196]
GitHub Copilot has been accused of producing copyrighted code, with no author attribution or license. [197]
OpenAI revealed that they would terminate support for Codex API on March 23, 2023. [198]
GPT-4

On March 14, 2023, OpenAI announced 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 test 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 likewise read, evaluate or generate as much as 25,000 words of text, and write code in all significant programs languages. [200]
Observers reported that the iteration of ChatGPT utilizing GPT-4 was an improvement on the previous GPT-3.5-based model, with the caveat that GPT-4 retained some of the problems with earlier revisions. [201] GPT-4 is also capable of taking images as input on ChatGPT. [202] OpenAI has actually decreased to reveal numerous technical details and stats about GPT-4, such as the precise size of the design. [203]
GPT-4o

On May 13, 2024, OpenAI announced and released GPT-4o, which can process and produce text, images and trademarketclassifieds.com audio. [204] GPT-4o attained state-of-the-art 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) criteria compared to 86.5% by GPT-4. [207]
On July 18, 2024, OpenAI launched GPT-4o mini, a smaller variation 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 $15 respectively for GPT-4o. OpenAI anticipates it to be particularly beneficial for business, startups and developers looking for to automate services with AI agents. [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 consider their reactions, leading to higher accuracy. These models are especially efficient in science, coding, and thinking tasks, and were made available to ChatGPT Plus and Staff member. [209] [210] In December 2024, o1-preview was changed by o1. [211]
o3

On December 20, 2024, oeclub.org OpenAI unveiled o3, the follower of the o1 thinking design. OpenAI likewise unveiled o3-mini, a lighter and much faster variation of OpenAI o3. As of December 21, 2024, this design is not available for public usage. 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 designs. [214] The model is called o3 instead of o2 to avoid confusion with telecoms companies O2. [215]
Deep research study

Deep research is an agent established by OpenAI, revealed on February 2, 2025. It leverages the capabilities of OpenAI's o3 model to carry out substantial web surfing, information analysis, and synthesis, providing detailed reports within a timeframe of 5 to 30 minutes. [216] With searching and Python tools allowed, it reached an accuracy of 26.6 percent on HLE (Humanity's Last Exam) standard. [120]
Image classification

CLIP

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

DALL-E

Revealed in 2021, DALL-E is a Transformer design that produces 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 purse shaped like a pentagon" or "an isometric view of an unfortunate capybara") and generate corresponding images. It can create pictures of reasonable items ("a stained-glass window with a picture of a blue strawberry") as well as objects 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 revealed DALL-E 2, an upgraded variation of the model with more reasonable results. [219] In December 2022, OpenAI released on GitHub software application for Point-E, a new primary 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 much better able to create images from complicated descriptions without manual timely engineering and render intricate details like hands and text. [221] It was launched to the public as a ChatGPT Plus function in October. [222]
Text-to-video

Sora

Sora is a text-to-video design that can produce videos based upon brief detailed prompts [223] as well as extend existing videos forwards or backwards in time. [224] It can create videos with resolution up to 1920x1080 or 1080x1920. The optimum length of generated videos is unidentified.

Sora's development team called it after the Japanese word for "sky", to symbolize its "limitless innovative capacity". [223] Sora's innovation is an adjustment of the technology behind the DALL · E 3 text-to-image design. [225] OpenAI trained the system using publicly-available videos along with copyrighted videos accredited for that function, but did not reveal the number or the specific sources of the videos. [223]
OpenAI showed some Sora-created high-definition videos to the general public on February 15, 2024, stating that it might produce videos up to one minute long. It also shared a technical report highlighting the techniques utilized to train the model, and the design's capabilities. [225] It acknowledged a few of its shortcomings, including battles imitating complex physics. [226] Will Douglas Heaven of the MIT Technology Review called the presentation videos "excellent", but noted that they need to have been cherry-picked and might not represent Sora's common output. [225]
Despite uncertainty from some scholastic leaders following Sora's public demo, noteworthy entertainment-industry figures have actually revealed significant interest in the technology's capacity. In an interview, actor/filmmaker Tyler Perry expressed his awe at the innovation's ability to produce realistic video from text descriptions, citing its possible to transform storytelling and content production. He said that his enjoyment about Sora's possibilities was so strong that he had chosen to stop briefly plans for broadening his Atlanta-based film studio. [227]
Speech-to-text

Whisper

Released in 2022, Whisper is a general-purpose speech acknowledgment design. [228] It is trained on a large dataset of diverse audio and is likewise a multi-task model that can carry out 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 predict subsequent musical notes in MIDI music files. It can produce songs with 10 instruments in 15 designs. According to The Verge, a song created by MuseNet tends to start fairly but then fall under chaos the longer it plays. [230] [231] In pop culture, preliminary applications of this tool were used as early as 2020 for the internet psychological 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 specified the songs "reveal local musical coherence [and] follow traditional chord patterns" however acknowledged that the tunes do not have "familiar larger musical structures such as choruses that duplicate" and that "there is a substantial space" in between Jukebox and human-generated music. The Verge mentioned "It's technologically outstanding, even if the outcomes seem like mushy variations of songs that may feel familiar", while Business Insider stated "surprisingly, a few of the resulting songs are appealing and sound legitimate". [234] [235] [236]
User user interfaces

Debate Game

In 2018, OpenAI introduced the Debate Game, which teaches machines to dispute toy issues in front of a human judge. The function is to research study whether such a technique might 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 significant layer and neuron of 8 neural network models which are often studied in interpretability. [240] Microscope was developed to evaluate the functions that form inside these neural networks quickly. The designs consisted of are AlexNet, fishtanklive.wiki VGG-19, various versions of Inception, and different 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 interface that permits users to ask questions in natural language. The system then reacts with an answer within seconds.