In less than two years, the generative AI chatbot ChatGPT has become a household name alongside products like the iPhone, Windows and Google Search.
OpenAI is the company behind the chatbot, which made its public debut in November 2022, spurring a wave of AI-powered creativity that quickly mesmerized us, in text, images and videos. Other companies — tech titans and fellow startups — saw the effusive public reaction and jumped on the bandwagon with their own tools.
Meta’s Llama model debuted in February 2023. Google’s Gemini chatbot, initially called Bard, came out in May 2023. Anthropic’s Claude and Adobe’s Firefly followed shortly thereafter. Llama distinguishes itself as an open-source model, which Meta CEO Mark Zuckerberg said will ultimately make it more competitive, while Firefly is focused on image generation and editing. Otherwise these tools more or less have the same goals: to help us brainstorm, write, learn and plan with reasoning, vision analysis, code generation and multilingual processing capabilities.Â
Though the initial novelty of AI-generated content may have worn off a bit, tech companies continue to push boundaries with increasingly powerful models, new ways to interact with chatbots, and additional functionality such as search. Just this week, for instance, OpenAI unveiled its new o1 model, which it says is “designed to spend more time thinking” before responding to a prompt.
Earlier this year, SEO strategy site Backlinko found that ChatGPT had nearly 70% market share among subscription-based AI tools. And as the original gen AI pioneer, OpenAI may hold further advantage in this burgeoning market as the first to capture our imaginations and show us what chatbots can really do at home and at work.
If you’re trying to get a handle on OpenAI, keep scrolling for a look at everything you need to know.
What is OpenAI?
Founded in 2015, OpenAI is still considered a startup, despite its prominence in the field.
Its mission at the start wasn’t to put AI tools in the hands of consumers. Instead, it was to “ensure that artificial general intelligence benefits all of humanity.” Artificial general intelligence, or AGI, is a more advanced form of AI that rivals human intelligence and can outperform us at many tasks. It remains an open question how close we are to actually achieving AGI.
OpenAI began as a nonprofit, but in 2019 it split into what it calls a hybrid for-profit and nonprofit organization, to raise more capital in order to acquire the necessary computing resources to develop AGI. The company is based in San Francisco and its CEO is Sam Altman.
OpenAI is the for-profit arm. It’s released, in its “GPT” family, large language models, or LLMs, which are AI systems trained on huge data sets to understand and generate human language. It’s also released multimodal models in the GPT family. Those are deep-learning models that process additional content types, like video, audio and images.
GPT-4o and GPT-4o Mini are the latest GPT models, along with the newly announced 01 model. These AI models facilitate our interactions with the ChatGPT chatbot.
OpenAI has also developed text-to-image models in the Dall-E family and a text-to-video model called Sora that is expected to be released later this year.
How does ChatGPT work?
The gen AI technology underpinning OpenAI’s ChatGPT allows the bot to generate responses to user prompts. Those prompts can include text or verbal requestsin plain English for nearly anything, as long as the query falls within OpenAI’s safety standards. You can request help planning a Labor Day barbecue, for instance, or ask the bot to tell you about the significance of the Louisiana Purchase or explain what causes the aurora borealis. The prolific chatbot can also write poetry and code, and it has passed the CPA exam and the bar exam (though some people are skeptical about its bar results).
How does OpenAI train its models?
ChatGPT was originally based solely on a large language model. That’s the AI system trained on large data sets to understand and generate human language. The latest GPT model, GPT-4o, is multimodal, which means it understands images, audio and video as well.
OpenAI says its LLMs use information that’s publicly available on the internet; info the company licenses from third parties; and data from OpenAI’s users and human trainers. However, the training data is current only up to a certain date. In the case of GPT-4o and GPT-4o Mini, that date is October 2023.Â
Further, OpenAI says it filters out data it doesn’t want its models to learn, like hate speech, adult content and spam. The information fed into the LLM is called training data, and OpenAI, like other generative AI makers, hasn’t shared exactly what information is in its training data. Â Â
What was OpenAI’s first GPT model and when did it come out?
OpenAI introduced the concept of generative pretrained transformers, or GPTs, in a 2018 research paper. The name refers to the model’s ability to generate text, as well as its use of an AI technology called a transformer, which is a type of deep learning that can translate text and speech nearly in real time. (Deep learning is a branch of machine learning that uses neural networks — models that make decisions like the human brain.)
The first transformer-based language model was 2018’s OpenAI-GPT, or GPT-1. That was followed by GPT-2 and GPT-3, in 2019 and 2020, respectively.
These models were long available to developers, but it was the release of GPT-3.5 and the ChatGPT interface in 2022 that made it possible for virtually anyone to use generative AI, sparking the transformative era we’re in now.
CNET reviewer Imad Khan noted earlier this year that ChatGPT 3.5 will “get the job done for most people,” providing “serviceable” answers even though “it’s always best to do a bit of fact-checking.” Â Â
He found that ChatGPT 4 is smarter and generates more-thoughtful answers that can synthesize complex information. “ChatGPT 4 really impresses when you need more-specialized answers to specific questions (like college-level philosophy questions),” Khan wrote.
What are the newest GPT models?
The latest widely available GPT models are GPT-4o and GPT-4o Mini, released in May and July, respectively. In September, OpenAI unveiled the preview of a model called o1.Â
GPT-4o provides responses that are more up to date than those of its predecessors, and it can understand — and generate — larger chunks of text. (The “o” stands for “omni” as it can accept any combination of text, audio, image and video inputs and generate any combination of text, audio and image outputs.)
GPT-4o Mini is a small language model that provides AI horsepower and speed without the high cost or computing resources required for an LLM. (Microsoft’s Phi-3 Mini, which is built to run on phones and PCs, is another example. So is Google’s Gemini 1.5 Flash.)
It has a context window of 128,000 tokens, which is a measurement of how much it can remember in a single conversation. GPT-4o has the same context window, while a prior model, GPT-3.5 Turbo, has a context window of 16,000 tokens.
The OpenAI 01-preview model, the company says, can better handle complex tasks and can solve harder math, science and coding problems than previous models. “We trained these models to spend more time thinking through problems before they respond, much like a person would,” OpenAI said in a blog post. “Through training, they learn to refine their thinking process, try different strategies, and recognize their mistakes.Â
What does OpenAI do besides language models and chatbots?
OpenAI has a family of text-to-image models called Dall-E. The latest, Dall-E 3, was released in October 2023.
In his review, CNET’s Stephen Shankland called Dall-E 3 “a marvel” among image generators that does well with both realistic and surreal images and encourages you to get creative.
OpenAI’s text-to-video model, Sora, offers the same premise for realistic videos, but it’s still being tested for potential harms and risks — like creating misleading content, as well as extreme violence, sexual content, hateful imagery, celebrity likenesses and the intellectual property of others — so it isn’t yet available to the public.
OpenAI also offers APIs for developers who want to build new applications based on OpenAI technology or custom AI apps called GPTs, which you can create and share in OpenAI’s app store. There are millions of GPTs available, including ones for fitness, haikus and books.
What is SearchGPT?
OpenAI is testing a search engine prototype, called SearchGPT. It’s currently available to only a small group of testers.
Instead of you having to ask questions and comb through links to find an answer, as in traditional search, SearchGPT generates answers to questions, with links to the online sources where it found the information. It’s a comparable experience to Google’s AI Overviews or the search functionality from startups like Perplexity.ai.
OpenAI eventually plans to integrate search functionality into ChatGPT.
Who is on OpenAI’s leadership team?
OpenAI was founded by a group of research engineers and scientists, as well as CEO Altman, entrepreneur Elon Musk, machine learning expert Ilya Sutskever and president and chairman Greg Brockman. Mira Murati later joined as CTO and Sarah Friar has come on board as CFO.
Sutskever, who was the chief scientist at OpenAI until June, disagreed with Altman over how rapidly AI should develop amid concerns it could eventually harm humanity without the right constraints. The same month he left OpenAI, Sutskever founded an AI company called Safe Superintelligence Inc., or SSI. According to the website, its singular goal is safe superintelligence, or AGI.
In November 2023, Altman was briefly ousted as CEO by the board of directors for, according to one director, withholding and misrepresenting information and lying. He was reinstated five days later, following what was reportedly a weekend of internal conflict and pressure from investors. OpenAI has said Altman didn’t do anything that justified his removal as CEO.
What happened between OpenAI and Elon Musk?
When OpenAI was thinking about switching to a for-profit model in 2017, Musk, according to an OpenAI blog post, wanted the startup to merge with his electric-car company, Tesla, or to give him majority equity, board control and the CEO title. Asked by CNN for a response to the blog post at the time, lawyers for Musk declined to comment. Musk departed in February 2018 with the intent to build his own AGI competitor, OpenAI’s post said.
Musk filed a lawsuit against OpenAI, accusing the startup of abandoning its nonprofit mission, but he later dropped it, and then he refiled it, earlier this month, alleging fraud and breach of contract. In response, OpenAI referred to its blog post about Musk’s initial lawsuit. The post said the company intends to move to dismiss all of Musk’s claims.
Who has invested in OpenAI?
Founders Altman, Brockman and Musk, along with VCs Reid Hoffman and Peter Thiel, investor Jessica Livingston, Amazon’s cloud computing arm Amazon Web Services, IT company Infosys and nonprofit YC Research (now OpenResearch) committed an initial $1 billion.
Since then, Microsoft has invested $13 billion, according to reports. While details of investment are confidential, a Microsoft spokesperson said, “it is important to note that Microsoft does not own any portion of OpenAI and is simply entitled to a share of profit distributions.” Other backers include investment firm Tiger Global and VC firms Sequoia Capital, Founders Fund, Thrive Capital, Andreessen Horowitz and K2 Global.
Microsoft’s first investment, in 2019, helped fund supercomputing technology. Later investments have supported OpenAI research. In addition, Microsoft’s Copilot AI assistant uses GPT-4o to answer queries and generate content with greater accuracy, as well as to open apps and edit photos.
What is OpenAI’s relationship with Apple?
At its annual developers conference in June, Apple announced a partnership with OpenAI. The iPhone maker plans to integrate ChatGPT into its iOS smartphone operating system; its tablet operating system, iPadOS; and its computer operating system, MacOS. It also plans to offer ChatGPT as an option to users querying its Siri voice assistant.
What’s going on with OpenAI and publishers?
The New York Times is among the publications that have sued OpenAI (and Microsoft) over unauthorized use of their content to train AI models.
The New York Daily News, The Chicago Tribune, The Orlando Sentinel, The Sun Sentinel of Florida, The San Jose Mercury News, The Denver Post, The Orange County Register and The St. Paul Pioneer Press have also sued OpenAI over the use of their content to train chatbots.
OpenAI has sought to have parts of the lawsuits dismissed, saying that chatbots aren’t replacements for news articles. The AI startup has also begun to sign deals with a number of media companies to license news stories, including the Associated Press, Axel Springer, News Corp and Vox Media.
Correction, Sept. 6: This article earlier misstated aspects of Microsoft’s investment in OpenAI. Microsoft does not own any portion of OpenAI.