sublab

Announced in 2016, hb9lc.org Gym is an open-source Python library designed to facilitate the development of reinforcement knowing 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 basic user interface for communicating with these environments. In 2022, new advancements 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 computer game [147] utilizing RL algorithms and research study generalization. Prior RL research focused mainly on enhancing agents to resolve 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 robot agents initially lack knowledge of how to even walk, but are given the objectives of learning to move and hb9lc.org to push the opposing representative out of the ring. [148] Through this adversarial knowing procedure, the representatives discover how to adjust to altering conditions. When a representative is then eliminated from this virtual environment and put in a new virtual environment with high winds, the agent braces to remain upright, suggesting it had found out how to stabilize in a generalized method. [148] [149] OpenAI's Igor Mordatch argued that competitors between agents might create an intelligence "arms race" that could increase a representative's capability to function 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 find out to play against human gamers at a high skill level totally through trial-and-error algorithms. Before becoming a group of 5, the first public presentation occurred at The International 2017, the yearly premiere champion competition for the game, where Dendi, an expert Ukrainian player, lost against a bot in a live individually matchup. [150] [151] After the match, CTO Greg Brockman explained that the bot had found out by playing against itself for two weeks of genuine time, and that the learning software was a step in the instructions of creating software that can handle intricate tasks like a cosmetic surgeon. [152] [153] The system utilizes a type of reinforcement learning, as the bots discover with time by playing against themselves numerous times a day for months, and are rewarded for actions such as killing an opponent and taking map goals. [154] [155] [156]
By June 2018, the ability of the bots broadened to play together as a complete team 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 exhibition matches against expert players, 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' last 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 games. [165]
OpenAI 5's systems in Dota 2's bot gamer shows the challenges of AI systems in multiplayer online battle arena (MOBA) games and how OpenAI Five has shown making use of deep support learning (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 robot hand, to control physical objects. [167] It learns completely in simulation using the exact same RL algorithms and training code as OpenAI Five. OpenAI dealt with the object orientation problem by utilizing domain randomization, a simulation technique which exposes the student to a variety of experiences rather than attempting to fit to reality. The set-up for Dactyl, aside from having motion tracking cameras, likewise has RGB cameras to allow the robotic to manipulate an arbitrary item by seeing it. In 2018, OpenAI revealed that the system was able to control a cube and pipewiki.org 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 complicated physics that is harder to model. OpenAI did this by enhancing the effectiveness of Dactyl to perturbations by utilizing Automatic Domain Randomization (ADR), a simulation approach of producing progressively more challenging environments. ADR varies 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 models established by OpenAI" to let developers get in touch with it for "any English language AI task". [170] [171]
Text generation

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

The original paper on generative pre-training of a transformer-based language design was written 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 knowledge and process long-range dependencies by pre-training on a diverse corpus with long stretches of contiguous text.

GPT-2

Generative Pre-trained Transformer 2 ("GPT-2") is an unsupervised transformer language design and the successor to OpenAI's initial GPT model ("GPT-1"). GPT-2 was revealed in February 2019, with only limited demonstrative variations at first launched to the public. The full variation of GPT-2 was not immediately released due to issue about possible misuse, consisting of applications for writing phony news. [174] Some experts expressed uncertainty that GPT-2 posed a significant threat.

In response 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, 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 difficult to filter". [176] In November 2019, OpenAI launched the total variation of the GPT-2 language design. [177] Several sites host interactive demonstrations of different instances of GPT-2 and other transformer designs. [178] [179] [180]
GPT-2's authors argue without supervision language designs to be general-purpose students, illustrated by GPT-2 attaining cutting edge accuracy and perplexity on 7 of 8 zero-shot jobs (i.e. the design was not further trained on any task-specific input-output examples).

The corpus it was trained on, called WebText, engel-und-waisen.de contains somewhat 40 gigabytes of text from URLs shared in with at least 3 upvotes. It prevents certain concerns encoding vocabulary with word tokens by utilizing byte pair encoding. This allows 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 design and the follower to GPT-2. [182] [183] [184] OpenAI stated that the full variation of GPT-3 contained 175 billion parameters, [184] two 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 criteria were likewise trained). [186]
OpenAI stated that GPT-3 prospered at certain "meta-learning" tasks and could generalize the function of a single input-output pair. The GPT-3 release paper gave examples of translation and cross-linguistic transfer knowing in between English and Romanian, and between English and German. [184]
GPT-3 considerably improved benchmark results over GPT-2. OpenAI cautioned that such scaling-up of language designs could be approaching or encountering the fundamental capability constraints of predictive language designs. [187] Pre-training GPT-3 needed several thousand petaflop/s-days [b] of calculate, gratisafhalen.be 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 right away launched to the general public for issues 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 licensed specifically to Microsoft. [190] [191]
Codex

Announced in mid-2021, Codex is a descendant of GPT-3 that has 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 released in private beta. [194] According to OpenAI, the model can develop working code in over a lots programs languages, many effectively in Python. [192]
Several concerns with problems, design defects and security vulnerabilities were cited. [195] [196]
GitHub Copilot has actually been implicated of discharging copyrighted code, with no author attribution or license. [197]
OpenAI revealed that they would cease 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), efficient in accepting text or image inputs. [199] They announced that the upgraded technology 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, evaluate or produce up to 25,000 words of text, and write code in all major programs languages. [200]
Observers reported that the version of ChatGPT utilizing GPT-4 was an enhancement on the previous GPT-3.5-based model, with the caution that GPT-4 retained some of the issues with earlier modifications. [201] GPT-4 is also capable of taking images as input on ChatGPT. [202] OpenAI has decreased to reveal different technical details and statistics about GPT-4, such as the precise size of the model. [203]
GPT-4o

On May 13, 2024, OpenAI revealed and launched GPT-4o, which can process and generate text, images and audio. [204] GPT-4o attained advanced outcomes 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) criteria compared to 86.5% by GPT-4. [207]
On July 18, 2024, OpenAI launched GPT-4o mini, a smaller sized 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 expects it to be particularly helpful for enterprises, startups and designers seeking to automate services with AI representatives. [208]
o1

On September 12, 2024, OpenAI launched the o1-preview and o1-mini models, which have been designed to take more time to consider their responses, resulting in greater accuracy. These designs are especially efficient 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 revealed o3, the follower of the o1 thinking design. OpenAI also revealed o3-mini, a lighter and faster version 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 scientists had the opportunity to obtain early access to these designs. [214] The model is called o3 instead of o2 to prevent confusion with telecommunications providers O2. [215]
Deep research

Deep research study is an agent developed by OpenAI, revealed on February 2, 2025. It leverages the abilities of OpenAI's o3 design to perform extensive web browsing, data 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) benchmark. [120]
Image classification

CLIP

Revealed in 2021, CLIP (Contrastive Language-Image Pre-training) is a design that is trained to evaluate 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 model that develops images from textual descriptions. [218] DALL-E uses a 12-billion-parameter version of GPT-3 to analyze natural language inputs (such as "a green leather handbag shaped like a pentagon" or "an isometric view of a sad capybara") and produce matching images. It can produce pictures of sensible items ("a stained-glass window with a picture of a blue strawberry") in addition to objects 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 announced DALL-E 2, an upgraded version of the design with more realistic outcomes. [219] In December 2022, OpenAI released on GitHub software application for Point-E, a brand-new primary system for transforming 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 prompt engineering and render complicated details like hands and text. [221] It was released to the general 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 short detailed prompts [223] as well as extend existing videos forwards or backwards in time. [224] It can create videos with resolution as much as 1920x1080 or 1080x1920. The optimum length of generated videos is unidentified.

Sora's advancement team called it after the Japanese word for "sky", to signify its "limitless creative capacity". [223] Sora's technology is an adaptation of the innovation behind the DALL · E 3 text-to-image model. [225] OpenAI trained the system using publicly-available videos in addition to copyrighted videos certified 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, stating that it could generate videos up to one minute long. It likewise shared a technical report highlighting the approaches utilized to train the model, and the model's capabilities. [225] It acknowledged a few of its drawbacks, consisting of struggles simulating intricate physics. [226] Will Douglas Heaven of the MIT Technology Review called the presentation videos "outstanding", however noted that they need to have been cherry-picked and might not represent Sora's normal output. [225]
Despite uncertainty from some academic leaders following Sora's public demonstration, notable entertainment-industry figures have actually revealed significant interest in the innovation's capacity. In an interview, actor/filmmaker Tyler Perry expressed his astonishment at the technology's capability to create sensible video from text descriptions, citing its potential to reinvent storytelling and content creation. He said that his enjoyment about Sora's possibilities was so strong that he had actually chosen to pause prepare for broadening 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 disgaeawiki.info is likewise a multi-task design that can carry out multilingual speech acknowledgment as well as speech translation and language recognition. [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 tunes with 10 instruments in 15 designs. According to The Verge, a tune created by MuseNet tends to start fairly however then fall into mayhem the longer it plays. [230] [231] In popular culture, preliminary applications of this tool were used as early as 2020 for the web psychological thriller Ben Drowned to produce 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 genre, artist, and gratisafhalen.be a bit of lyrics and outputs tune samples. OpenAI mentioned the songs "show regional musical coherence [and] follow standard chord patterns" however acknowledged that the tunes do not have "familiar larger musical structures such as choruses that repeat" and that "there is a considerable space" in between Jukebox and human-generated music. The Verge specified "It's highly outstanding, even if the outcomes seem like mushy versions of songs that might feel familiar", while Business Insider specified "remarkably, a few of the resulting tunes are memorable and sound genuine". [234] [235] [236]
Interface

Debate Game

In 2018, OpenAI launched the Debate Game, which teaches makers to dispute toy issues in front of a human judge. The purpose is to research study whether such a technique may 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 considerable layer and nerve cell of 8 neural network models which are often studied in interpretability. [240] Microscope was produced to analyze the functions that form inside these neural networks quickly. The models included are AlexNet, VGG-19, various variations 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 user interface that permits users to ask concerns in natural language. The system then reacts with an answer within seconds.