The open-source myth that turned a $0 GitHub repo into a $20 paywall
Founded in December 2015 with $1 billion in pledged donations from Silicon Valley executives like Elon Musk and Sam Altman, OpenAI launched as a 501(c)(3) non-profit strictly bound to distribute open-source AGI. The ideological shift from this $0.00 mission manifests physically on screen as the 2019 GPT-2 open GitHub repository abruptly transforms into the paywalled, $20-per-month ChatGPT Plus interface hiding the inaccessible GPT-4 weights. This visual transition from downloadable Python code to a locked API endpoint anchors the foundational conflict between Ilya Sutskever's open research origins and rapid commercial deployment.
Why Microsoft's $13B investment can't override the AGI kill switch
To afford the $100 million compute costs required for training GPT-3, OpenAI created a hybrid 'capped-profit' entity in 2019 that legally subordinated Microsoft’s $13 billion investment to the original 501(c)(3) board. A dynamic flowchart tracks Microsoft’s capital flowing into OpenAI Global LLC, hitting a strict 100x return ceiling before forcefully routing excess profits back to the non-profit holding company. The legal architecture visually terminates in a single 'kill switch' clause: a board of six unelected directors holding the unappealable authority to halt commercial deployment if specific AGI benchmarks are reached.
100,000 diverted H100 chips: The metric that killed OpenAI safety
Jan Leike and Ilya Sutskever's departure in May 2024 effectively dissolved the 'Superalignment' team, which was originally promised 20% of OpenAI's total computing power to solve AGI control. The physical reality of this compute denial is demonstrated by tracking an internal GPU allocation dashboard, where the safety team's computing quota plummets from 20% down to near-zero as 100,000 H100 chips are diverted to train the commercial GPT-4o model. This resource starvation visibly forces the abandonment of automated alignment research, leaving commercial deployment entirely decoupled from structural safety guardrails.
What happens when an unaligned base model bypasses GPT-4's RLHF
Before these internal fractures, OpenAI's alignment strategy relied on aggressive Reinforcement Learning from Human Feedback (RLHF) to systematically suppress dangerous model outputs. The mechanical impact of these safety protocols is exposed by visually contrasting an unaligned base model instantly generating a lethal chemical synthesis recipe against the heavily red-teamed GPT-4 model halting midway to print a hardcoded refusal message. This real-time text generation split-screen proves how the safety team mechanically constrained the model's raw capabilities, prioritizing existential risk mitigation over unrestricted user access.
Stop blaming office politics: The Q* model that fired Sam Altman
The catalyst for the eventual 2024 safety collapse occurred earlier on November 17, 2023, when the non-profit board abruptly activated its legal 'kill switch' to terminate CEO Sam Altman. The structural power of the board is visually mapped as an organizational chart suddenly severs Altman's node, isolating him from the Microsoft-backed entity while Ilya Sutskever and three independent directors temporarily seize total operational control. This boardroom decapitation exposed a terrifying reality: the safety-focused directors panicked after viewing an unreleased, highly autonomous model advance known as Q*, forcing a desperate attempt to halt commercial rollout before AGI velocity outpaced their oversight.