OpenAI Chronicle: What Drives ChatGPT Creators

How open OpenAI is, and how influential, are probably two of the most frequently asked questions since the release of ChatGPT by OpenAI. We have sifted through at least 50 different articles and paid publications to bring you the captivating story of the company that aims to develop benevolent AI and maintain control over it, eventually creating enormous wealth and somehow distributing it to people.

It's also intriguing to trace the development of ChatGPT – its creation process and the factors that led to its success. We've located all the relevant research papers and are sharing them with you - it promises to be very insightful reading.

We shed light on the key individuals and their visions. Looking ahead, we're also divulging information about OpenAI's funding, its investment ambitions, and - an issue of great societal concern - what OpenAI is doing regarding the safety of Large Language Models (LLMs).

This profile is packed with unique details and important links that you are unlikely to find collectively elsewhere. As Stratechery suggests, it's possible "that OpenAI will become the platform on which all other AI companies are built". We can't be certain if that will come true, but understanding their original intentions and current ambitions could be essential."

1. Founder’s intentions and internal friction

2. Sam Altman’s vision

3. Money situation and current investments

4. GPT development history and other interesting products

5. Safety & alignment

6. Urge for regulations

7. Bonus: What and Who is Open AI

Founders’ intention and internal frictions

Allegedly, OpenAI began in 2015 with Elon Musk and Sam Altman's shared view of the danger of the risks and opportunities of AI technology. They wanted to move towards Artificial General Intelligence (AGI) most openly by naming the company OpenAI and making it non-profit. They were not the first to declare AGI as a goal, but the intention was different, not clouded by commercial interests. "It will be important to have a leading research institution that can prioritize a good outcome for all over its self-interest," the announcement said. “Researchers will be strongly encouraged to publish their work, whether as papers, blog posts, or code, and our patents (if any) will be shared with the world.” As The New Yorker’s Tad Friend put it: “OpenAI was particularly concerned that Google’s DeepMind Technologies division was seeking a supreme A.I. that could monitor the world for competitors. Musk told me, “If the A.I. that they develop goes awry, we risk having an immortal and superpowerful dictator forever.” He went on, “Murdering all competing A.I. researchers as its first move strikes me as a bit of a character flaw.”

Everything changed in 2018: Elon Musk resigned from the board, officially because of the potential conflict of interest between Tesla and OpenAI, but most likely because he couldn't take over the company and run it his way.

This coincided with the realization that the organization could not survive as a non-profit. As its research became more ambitious, it became clear that it would need significant funding to achieve its goals.

To meet this challenge, Sam Altman proposed a novel structure for the organization. Under this structure, OpenAI would remain a non-profit entity focused on basic research in AI while creating a for-profit arm to develop and commercialize AI technologies. In 2019, OpenAI announced the creation of a for-profit subsidiary called OpenAI LP, to attract more funding and form partnerships with other companies.

The idea behind this structure was to generate revenue from commercial AI products and services, which could then be reinvested in OpenAI's research efforts. This would allow OpenAI to pursue the development of advanced AI technologies, such as machines that could match or exceed human intelligence, without relying solely on philanthropic donations.

The for-profit arm of OpenAI has generated a great deal of controversy and criticism.

In December 2022, Elon Musk tweeted on ChatGPT: "The danger of training AI to be woke – in other words, lie – is deadly." And in February, accused the company of being "a closed source, maximum-profit company effectively controlled by Microsoft."

"Most of that is not true, and I think Elon knows that," Altman responded on the "On With Kara Swisher" podcast, adding: "he's a jerk."

Sam Altman’s vision

So, after reading several articles about Sam Altman in The New Yorker, The Information, The New York Times, and MIT Technology, it seems that his vision for OpenAI is to achieve AGI. He believes that the creation of AGI will create an immense amount of wealth, but he is unsure about how that wealth will be distributed among humans and what impact it will have on society.

He also compares his company's ambition to the Manhattan Project, which developed the first atomic bomb during World War II, suggesting that OpenAI's goal is to achieve something of similar scale and impact.

Without a doubt, Altman's vision for OpenAI is enormously ambitious, and it also involves significant uncertainty and potential consequences. Recently, the naming has changed and instead of AGI, OpenAI is concerned about Superintelligence.

I recommend reading these articles to understand Sam Altman, Greg Brockman, and what drives OpenAI:

• Sam Altman’s Manifest Destiny (The New Yorker)

• The messy, secretive reality behind OpenAI’s bid to save the world (MIT Technology Review)

• The ChatGPT King Isn’t Worried, but He Knows You Might Be (The New York Times)

• The OpenAI coder who spun tech on its head (The Information)

(some of these articles are behind the paywall, if you are our paid subscriber, let us know and we will send you a pdf)

Money situation

Outstanding Funding

2015 the non-profit started with $1 billion collectively pledged by Sam Altman, Greg Brockman, Reid Hoffman, Jessica Livingston, Peter Thiel, Elon Musk, Amazon Web Services (AWS), Infosys, and YC Research.

2016 $120k pre-seed round from Y Combinator.

2019 $1 billion investment from Microsoft (split between cash and credits to Azure, Microsoft’s cloud computing platform).

2023 $10 billion (rumored) investment from Microsoft, extending their partnership with OpenAI, which includes the exclusive use of Microsoft Azure as the cloud provider for the ChatGPT tool and the exclusive access to GPT-4 which would power Microsoft’s own Prometheus model for Bing.

Evaluation (according to WSJ)

Money to invest

Current investments

GPT development story

The OpenAI developers were puzzled by the success of ChatGPT because it was a version of an AI system that they’d had for a while. The same basic models had been available on the API for almost a year before ChatGPT came out. But the packaging was extremely important. “We made it more aligned with what humans want to do with it. It talks to you in dialogue, it’s easily accessible in a chat interface, it tries to be helpful. That’s amazing progress, and I think that’s what people are realizing,” explains Jan Leike, the leader of OpenAI’s alignment team, in MIT Technology Review.

ChatGPT: A chatbot based on Large Language Models (LLMs). You can find ChatGPT API here.

Way to ChatGPT’s success (linked to the important research papers):

2017 Transformer Architecture (introduced in the iconic paper Attention is All You Need) combined with Semi-supervised Sequence Learning led to –>

2018 the creation of GPT (generative pre-trained transformer) led to –>

2019 GPT-2 – an LLM with 1.5 billion parameters. It was not fully released to the public, only a much smaller model was available for researchers to experiment with, as well as a technical paper, that led to –>

2020 GPT-3 and the paper Language Models are Few-Shot Learners, which evolved to –>

2022 GPT-3.5 and the fine-tuned version of GPT-3.5, called InstructGPT with the research paper Training language models to follow instructions with human feedback.

ChatGPT was trained in a very similar way to InstructGPT, the magic behind it is based on the research paper Deep reinforcement learning from human preferences, the technique is called reinforcement learning from human feedback (RLHF).

To identify any weaknesses in the model, OpenAI conducted "red-teaming" exercises where everyone at OpenAI and external groups tried to break the model. Additionally, they had an early-access program where trusted users were given access to the model and provided feedback.

And other interesting products

• OpenAI Codex was a model series descended from the GPT-3 series, trained on both natural language and billions of lines of code, with proficiency in over a dozen languages such as Python, JavaScript, Go, Perl, PHP, Ruby, Swift, TypeScript, SQL, and Shell. Codex models have been deprecated as of March 2023; you can check new Chat models that can perform many coding tasks with similar capabilities here.

• DALL-E (2021) and DALL-E 2 (2022): Neural network-based image generation systems. Important papers: in Learning Transferable Visual Models From Natural Language Supervision OpenAI researchers explain CLIP (Contrastive Language-Image Pre-training), a neural network that can acquire image representations by being trained on natural language datasets. In 2022, in their paper Hierarchical Text-Conditional Image Generation with CLIP Latents, OpenAI researchers explain the architecture behind DALL-E 2. You can find DALL-E API here. The most recent research goes beyond diffusion models and introduces Consistency Models, a new family of generative models that allow for fast one-step generation and zero-shot data editing, achieving high sample quality without adversarial training and outperforming existing distillation techniques for diffusion models in one and few-step generation. Read the original paper here.

• Whisper – an open-sourced speech-to-text model based on the research Robust Speech Recognition via Large-Scale Weak Supervision. You can find Whisper API here.

• OpenAI Five: A computer program that can play a five-on-five video game Dota 2.OpenAI Five's success demonstrates the significance of LSTMs in complex AI scenarios, marking a major milestone in the history of deep learning.

Related search: Understanding LSTM -- a tutorial into Long Short-Term Memory Recurrent Neural Networks; Techniques for training large neural networks, Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos; Scaling Laws for Reward Model Overoptimization; Triton: Open-source GPU programming for neural networks, and the very important paper Proximal Policy Optimization Algorithms.

• Inspired by the results, OpenAI developed a system called Dactyl, which is trained entirely in simulation but has proven to solve real-world tasks without physically-accurate modeling of the world. It learns from scratch using the same general-purpose RL algorithms and training code as OpenAI Five. Important paper: Learning Dexterous In-Hand Manipulation. This project extends the OpenAI vision of achieving AGI to the physical world.

• Started as OpenAI Gym, the toolkit, and platform for developing and comparing reinforcement learning (RL) algorithms, turned into Gymnasium under the non-profit Farama Foundation, focused on enhancing the field of RL by advocating for improved standardization and open-source tooling for researchers and industry professionals. Check their GitHub for more information as well as the website. Part of OpenAI Gym was RoboSumo, a multi-agent competitive environment, based on the research paper Continuous Adaptation via Meta-Learning in Nonstationary and Competitive Environments.

• To address the safety issues, in 2021, OpenAI open-sourced Safety Gym – a suite of tools and environments to measure the progress of RL agents in adhering to safety constraints during training. Check their GitHub.

Safety & alignment

OpenAI's safety work originated from a report created in 2021 in collaboration with Georgetown University's Center for Security and Emerging Technology and the Stanford Internet Observatory. This report analyzed the risks of using large language models to spread false information. It highlighted the potential impact of language models on influence operations and emphasized the importance of adopting mitigation strategies involving AI developers, social media companies, and government agencies.

In August 2022, OpenAI began Alignment research to ensure that artificial general intelligence (AGI) aligns with human values and intentions. They have three main pillars for this approach: training AI systems with human feedback, training AI systems to assist in human evaluation, and training AI systems for alignment research. However, there are various challenges in aligning AI systems with human values, such as deciding whom these systems should be aligned with.

According to their article “Our approach to AI safety” (April 2023) OpenAI is committed to safely building and deploying AI systems for broad benefit. They release AI systems gradually, with safeguards, learn from real-world use, and prioritize protecting children, privacy, and factual accuracy. OpenAI collaborates with experts, engages stakeholders, and emphasizes safety in research, products, and customer stories, while advocating for global governance of AI safety.

“​​We believe (and have been saying in policy discussions with governments) that powerful training runs should be reported to governments, be accompanied by increasingly-sophisticated predictions of their capability and impact, and require best practices such as dangerous capability testing. We think the governance of large-scale compute usage, safety standards, and regulation of/lesson-sharing from deployment are good ideas, but the details really matter and should adapt over time as the technology evolves. It’s also important to address the whole spectrum of risks from present-day issues (e.g. preventing misuse or self-harm, mitigating bias) to longer-term existential ones.” from Greg Brockmann’s tweet on April 12, 2023.

As part of their alignment research, OpenAI used GPT-4 to explain the behavior of neurons in GPT-2. This interpretability paper was published in May 2023 and was a significant step in understanding the inner workings of AI systems.

In June 2023, they launched a $1 million cybersecurity grant program “to boost and quantify AI-powered cybersecurity capabilities and to foster high-level AI and cybersecurity discourse.”

Urge for regulations

In May 2023, the whole wave of questions around regulations and governance rose. Sam Altman faced the senators on Capitol Hill and asked them to regulate the AI industry.

Then Sam Altman, Greg Brockman, and Ilya Sutskever authored a paper named Governance of Superintelligence, where they stated that “now it is a good time to start thinking”. To get it right they want to let people around the world “democratically decide on the bounds and defaults for AI systems.” How? “We don't yet know how to design such a mechanism, but we plan to experiment with its development.” Through their non-profit initiative, they intend to award ten grants valued at $100,000 each, aimed at developing proofs-of-concept for a democratic process that can address questions about the rules AI systems should adhere to.

Sam Altman also, among other industry leaders, signed a one-sentence statement:

Bonus: What and who is OpenAI

OpenAI is an artificial intelligence (AI) research laboratory based in the United States. It consists of two entities: OpenAI Incorporated, a non-profit organization, and OpenAI Limited Partnership, a for-profit subsidiary. The Lab's primary goal is to advance the development of safe and beneficial AI. It operates on the world's fifth most powerful supercomputer, which Microsoft built specifically for OpenAI's AI research. The lab was founded on December 11, 2015, in San Francisco by a group of individuals, including Elon Musk, Greg Brockman, Ilya Sutskever, John Schulman, Sam Altman, and Wojciech Zaremba.

Key Employees

• CEO and co-founder: Sam Altman,

  former president of the startup accelerator Y Combinator

• President and co-founder: Greg Brockman,

  former CTO, 3rd employee of Stripe

• Chief Scientist and co-founder: Ilya Sutskever,

  a former Google expert on machine learning, and co-author of a lot of OpenAI research papers.

• Chief Technology Officer: Mira Murati,

  previously at Leap Motion and Tesla, Inc.

• Chief Operating Officer: Brad Lightcap,

  previously at Y Combinator and JPMorgan Chase.

Board of the Open AI Nonprofit

• Greg Brockman

• Ilya Sutskever

• Sam Altman

• Adam D'Angelo (CEO of Quora)

• Will Hurd, an American politician, and former CIA clandestine officer, ex-Congressman.

• Tasha McCauley, CEO at GeoSim Systems

• Helen Toner, Director of Strategy at Georgetown’s Center for Security and Emerging Technology (CSET)

To be continued…