The OpenAI-Anthropic Story
Anthropic Was Not Catching Up. OpenAI Was.
This piece is the fifth in a series on Anthropic, OpenAI, and AI governance:
Dario Amodei: The Self-Appointed Ethics Czar for Planet Earth — The Pentagon standoff and the gap between architectural choices and moral authority
Time for Open Source Large Language Models — Why the AI cartel’s safety arguments are market protection dressed as principle
They Were Going to Save Us From This. Then They Became This. — How OpenAI’s idealistic governance structure was the vulnerability Dario walked out through
What Is Dario’s Lawsuit All About? — Why the Pentagon lawsuit is narrative management, not legal strategy
Act I — What Was Happening Inside OpenAI
By 2020, the knowledge of how to build large language models at commercial scale existed in exactly one place: inside OpenAI. Not because the research was secret by design — the original mission was open research — but because OpenAI had quietly stopped being open. GPT-3 was published as a paper but the model wasn’t released. The weights weren’t shared. The training details stayed internal. “OpenAI” had become a misnomer, and the outside world had no way to engage with what was actually happening.
Nobody else had the resources to find out independently. Training GPT-3 cost somewhere between $4 and $12 million in compute alone — a figure that ruled out every university lab, most research institutions, and all but a handful of well-capitalized companies. The knowledge existed in papers but the ability to act on it required a compute budget that essentially nobody outside of Google, Microsoft, and OpenAI could access. The field was not just small. It was captive.
Inside that captive world, two distinct research philosophies had been developing under the same roof. Ilya Sutskever — co-inventor of AlexNet and one of the architects of the GPT series — represented one approach: RLHF, Reinforcement Learning from Human Feedback. Train a model to satisfy human raters. Also opaque: the model learns to please without being able to explain why any given decision was made.
Dario Amodei and his team were developing the alternative. Constitutional AI: give the model an explicit set of principles and train it to reason against them in a self-critique loop. No armies of human raters required — the feedback is synthetic. The result is auditable by design. A compliance officer can examine the constitution. A regulator can trace a decision back to a principle.
This was not simply a safety disagreement — everyone in the field claims to care about safety. It was a technical argument about methodology. Dario’s team’s position was that Constitutional AI produces more predictable, more auditable, more consistent behavior than RLHF. A model trained against explicit principles behaves in ways you can reason about in advance. A model trained to satisfy human raters learns statistical patterns that are harder to predict in edge cases and impossible to fully explain after the fact. The enterprise case for Constitutional AI was not moral — it was engineering: if you can tell a compliance officer exactly what principles the model reasons against, and trace a decision back to a specific rule, you have a product that regulated industries can actually deploy with confidence.
Dario saw the trajectory clearly. He was VP of Research. He had complete visibility into every experiment, every training run, every capability threshold. He understood something that almost nobody in the media or investment community understood: the application code for a large language model is startlingly lean. A working GPT-2 fits in roughly 500 lines of Python. This is possible because AI development builds on vast open source libraries — PyTorch alone, the framework that dominates the field, represents millions of lines of highly optimized infrastructure that every lab uses equally. The proprietary layer on top is thin.
The intelligence is not in the software — it’s in knowing the process needed to make the model. His team carried in their heads the knowledge of what to write and the far more valuable intuitions about what data to use and which experiments to run. Reconstruction would take months, not years. This was not a leap of faith. It was a calculated extraction with a known timeline.
And the outside world still had no idea any of this was happening.
Act II — The Founding
In late 2020, Dario left. He took his sister Daniela, who ran Safety and Policy. He took a carefully selected team — not random disgruntled employees but the specific people required to reconstruct what OpenAI had built, for themselves, somewhere else.
Kaplan & McCandlish — Physicists who wrote the Scaling Laws: the mathematical rules governing how compute buys intelligence.
Tom Brown — Lead author of the GPT-3 paper. The engineer who could build what the theorists conceived.
Chris Olah — The world’s foremost mechanistic interpretability researcher. The person who could look inside a model and explain why it reasons as it does.
Jack Clark — Policy and narrative: positioning the work to attract capital and credibility.
Daniela Amodei — Operations and recruiting: building the company around the research.
Theory, engineering, interpretability, narrative, operations. Every component required to build and commercialize a frontier model. No redundancy. No gaps. And crucially — a team that had already solved the hardest problem in team-building. They knew how to work together. They shared a technical vision and philosophy, a common vocabulary, and years of accumulated experimental intuition. The “forming-storming-norming” phase that kills many startups was already behind them.
OpenAI’s governance structure — a nonprofit controlling a capped for-profit, with limited non-competes and a mission-driven culture that gave researchers enormous latitude — had no effective remedy. Nothing legally prevented a senior VP from walking out with his entire team. And so he did.
The “safety” framing was the story told publicly. The more complete explanation: Constitutional AI was a genuine technical belief, a cleaner corporate structure offered a clearer path to liquidity, and the Microsoft entanglement made OpenAI an increasingly uncomfortable home for researchers who wanted to control their own direction. “Safety” attracted mission-aligned early capital, created regulatory positioning, and repackaged a product argument — our alignment approach is auditable, theirs is not — as a values argument.
The sophisticated early investors — Eric Schmidt, Amazon, Google — were not buying a moral position. Schmidt said publicly he invested in the person, not the concept. Amazon put in $4 billion because Constitutional AI produces the auditable, consistent behavior their enterprise customers need. Google invested because they understood the technical methodology was genuinely differentiated and wanted access to it. They were buying a product architecture and a proven team. The investors who mattered read the footnotes.
The safety narrative made it legible to a broader audience.
“Safety” was a well-chosen word for that audience. The hallucination problem was real and visible: models that confidently made things up (hallucinated) looked unsafe to every enterprise buyer, every legal team, every regulator who touched them. Constitutional AI’s auditability spoke directly to that.
There is a technical dimension to this that conventional analysis misses entirely. In traditional software, the code is the moat — millions of lines of accumulated architecture that takes as long to rebuild as to build. That is why non-competes and trade secret law developed the way they did. Machine learning broke that model. The application code for a large language model is startlingly lean — a few hundred lines. Constitutional AI compounded this further: by removing the human rater bottleneck, it meant reconstruction required primarily compute and time, not infrastructure. The uncertainty was not “can we rebuild this?” — it was only “how much compute do we need?”
It is entirely plausible that the team arrived at Anthropic and had working prototypes running within weeks. They took nothing — no code, no weights, no data. They didn’t need to. Everything that mattered lived in their heads: the experimental intuitions built over years of running the same training loops, the architectural decisions they had already made and unmade, the failure modes they had already ruled out. Writing the code was a short order task for people who had already written it. The hard part — knowing what to write and what to feed it — was already solved.
The outside world still didn’t know any of this was happening. ChatGPT wouldn’t launch for another two years.
Act III — What Remained
The team that remained at OpenAI inherited the factory but lost the architects — and lost the technical rudder entirely.
Ilya Sutskever represented the competing methodology — RLHF — and was never part of Dario’s camp. Dario took everyone he needed. He did not take Ilya. After Dario left, what remained at the top was no coherent technical vision. Mira Murati, a capable product and operations executive, absorbed the research portfolio. The direction got filled by momentum and market pressure — not by anyone with the standing to set it deliberately.
Ilya tried to fight from inside. He voted to fire Altman in November 2023, reversed himself days later when 770 employees threatened to walk, was sidelined, and left in May 2024. By September 2024, Murati and the VP of Research had also gone. By late 2024 the original research leadership was gone entirely.
Ilya’s successor as Chief Scientist is Jakub Pachocki — a serious researcher who led GPT-4 and the o1 reasoning models. He is not a lightweight. Whether he and Chief Research Officer Mark Chen can generate the kind of integrated technical vision that Dario’s group had is an open question. OpenAI may be building that team. But it is a question mark, not an answer.
Anthropic’s seven co-founders are all still there — more than four years later. The real retention question is the second wave: researchers who joined in 2022 and 2023, fully vested or approaching it, who know exactly how the sausage is made. The co-founders have golden handcuffs. The second wave does not.
The Takeaway
The story most people tell is that Anthropic emerged from OpenAI’s shadow, played catch-up for years, and gradually proved itself as a serious competitor. That story is wrong.
Anthropic was not catching up to OpenAI. OpenAI was catching up to Anthropic. The knowledge, the team, the methodology, and the head start all belonged to the people who left — not to the organization they left behind. OpenAI spent the next three years reconstructing what had walked out the door in 2020, while Anthropic spent those same years executing on a plan that was already mapped before the company was formally announced.
That head start was built on a specific and unrepeatable set of conditions: captive knowledge inside one organization, a governance structure with no defection remedy, application code lean enough to reconstruct in months, and a methodology that removed the human rater bottleneck.
Dario did not leave because of safety concerns. He left because he had a complete team, a technical conviction, a cleaner corporate vehicle, and the precise understanding that reconstruction would take months not years. The “safety” narrative made it sound like a sacrifice. The technical reality made it look like a business plan.
By 2025, OpenAI had published a Model Spec and introduced Deliberative Alignment — both close peers to Constitutional AI in structure and intent. The field converged on Dario’s methodology without announcing it. In hindsight, the technical bet was the right one. The RLHF approach that Dario walked away from has been quietly supplemented or replaced across the industry by variations of what he built. Dario won the argument. He just won it in a way that made the methodology everyone’s rather than anyone’s.
That is the story of how the AI industry’s defining rivalry actually began. Not a moral crusade. Not a safety emergency. A technical schism, a governance gap, and a calculated extraction — executed before the world knew the race had started.
One final observation. In 2020, the lean codebase made reconstruction fast for a team that already knew what to write. In 2026, agentic coding tools have advanced to the point where anyone who understands what to build can direct an AI to build it — even a million-line codebase, if you know what to ask. The execution barrier is gone. What remains is the knowledge of what to ask: the experimental intuition, the architectural judgment, the understanding of which training decisions matter and which don’t. That knowledge still lives in heads, not in repositories. It always did.
Even the models themselves now understand a great deal of what to do at a high level.
The moat may now just be a puddle to step over.