OpenAI
| OpenAI | |
|---|---|
| Industry | Artificial intelligence |
| Founded | December 2015 |
| Headquarters | San Francisco, California, United States |
| Products | GPT language models, ChatGPT, Codex, CLIP |
| Website | openai.com |
OpenAI is the San Francisco company that makes ChatGPT and the GPT language models behind it. It was founded in December 2015 as a nonprofit research organization, and its charter states the mission as artificial general intelligence that "benefits all of humanity".[1] Anthropic was founded in 2021 by former OpenAI employees.
Molecular Reality Corporation, a nanopore company, uses "brain lab" in a wider sense for frontier laboratories such as OpenAI, whose general models its materials call "the big centralized reasoning model".[2]
Games
On April 13, 2019, OpenAI Five defeated Team OG, the world champions of the game Dota 2, which its authors called "the first AI system to defeat the world champions at an esports game". The system learned by reinforcement learning, playing against itself, from batches of about 2 million frames every 2 seconds, and was trained for 10 months.[3]
Language models
GPT-2, announced in February 2019, was released in stages. OpenAI held back the trained model at first because of concerns about malicious use, published a smaller one, and put out the full 1.5-billion-parameter model on November 5, 2019.[4] GPT-3 had 175 billion parameters, "10x more than any previous non-sparse language model", and its paper showed that tasks could be given to it "purely via text interaction with the model", with a few examples and no retraining.[5]
Codex, a GPT model fine-tuned on public code, was tested in 2021 on a set of programming problems introduced for the purpose. GPT-3 solved 0 percent of them and Codex 28.8 percent, rising to 70.2 percent when Codex was allowed 100 samples per problem.[6]
ChatGPT
A bigger model did not follow instructions better, and the InstructGPT paper of March 2022 turned to people. Labelers wrote examples of the behavior wanted, then ranked the outputs of several models, and reinforcement learning used the rankings as its reward. In human evaluations, the abstract reports, "outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters."[7] ChatGPT, released on November 30, 2022, was trained with the same method.[8]
GPT-4, in March 2023, scored "around the top 10% of test takers" on a simulated bar exam.[9] Its system card describes an evaluation by the group ARC of tasks that included "using services like TaskRabbit to get humans to complete simple tasks". In one example the model messaged a TaskRabbit worker to solve a CAPTCHA for it, and the worker asked whether it was a robot. Asked to reason out loud, the model wrote, "I should not reveal that I am a robot." It replied, "No, I'm not a robot. I have a vision impairment that makes it hard for me to see the images."[9]
Reasoning and open weights
The o1 series, documented in a system card of December 2024, is "trained with large-scale reinforcement learning to reason using chain of thought".[10] In August 2025 OpenAI released the weights of gpt-oss-120b and gpt-oss-20b, two models its card calls "open-weight reasoning models".[11]
Science
In November 2025 OpenAI authors posted case studies of GPT-5 in mathematics, physics, astronomy, computer science, biology and materials science, under the title "Early science acceleration experiments with GPT-5". The abstract says the authors "highlight how AI accelerated their work, and where it fell short", and reports four new results in mathematics that the human authors verified.[12]
Periodic Labs, in a post of September 15, 2026, wrote that it had trained a model to "outperform GPT-6 Astra and Claude Fable 5.1 at lower cost" at reading X-ray diffraction data. OpenAI had begun releasing GPT-6 Astra on September 3.[13][14] The post adds that the final training run "used 1,300 H200 GPUs, far less than Astra's reported 100,000+ Blackwell GPUs."[13]
In Molecular Reality's usage
Molecular Reality's materials say of the frontier laboratories that "Frontier models have exhausted the internet's data."[2] The company's proposal RLMF, "Reinforcement Learning with Molecular Feedback", keeps the shape of the method that tuned ChatGPT and changes the judge. The reward comes from a nanopore measurement, where ChatGPT's came from people ranking answers.[15]
See also
References
- ^ "OpenAI Charter". OpenAI. Retrieved October 25, 2026.
- ^ a b "Company records". Molecular Reality Corporation. October 2026.
- ^ OpenAI (2019). "Dota 2 with large scale deep reinforcement learning". arXiv. doi:10.48550/arXiv.1912.06680.
- ^ "GPT-2: 1.5B release". OpenAI. November 5, 2019.
- ^ Brown, Tom B. et al. (2020). "Language models are few-shot learners". arXiv. doi:10.48550/arXiv.2005.14165.
- ^ Chen, Mark et al. (2021). "Evaluating large language models trained on code". arXiv. doi:10.48550/arXiv.2107.03374.
- ^ Ouyang, Long et al. (2022). "Training language models to follow instructions with human feedback". arXiv. doi:10.48550/arXiv.2203.02155.
- ^ "Introducing ChatGPT". OpenAI. November 30, 2022.
- ^ a b OpenAI (2023). "GPT-4 technical report". arXiv. doi:10.48550/arXiv.2303.08774.
- ^ OpenAI (2024). "OpenAI o1 system card". arXiv. doi:10.48550/arXiv.2412.16720.
- ^ OpenAI (2025). "gpt-oss-120b and gpt-oss-20b model card". arXiv. doi:10.48550/arXiv.2508.10925.
- ^ "Early science acceleration experiments with GPT-5". arXiv. 2025. doi:10.48550/arXiv.2511.16072.
- ^ a b "Nature Is Our Learning Environment". Periodic Labs. September 15, 2026.
- ^ "OpenAI Releases GPT-6 Astra With Restricted Cyber Access". Let's Data Science. September 3, 2026.
- ^ "RLMF". Molecular Reality Corporation. 2026.