
Tech • AI • Robotics • Game
⏱ ≈ 4 min read · 15-min video — 11 min saved
Anthropic was founded by former OpenAI leaders who believed rapidly scaling AI could bring major scientific and economic gains but required a company built around safety, public benefit, and frontier research.
Anthropic emerged from a group of researchers and operators who had been working on large models, policy, interpretability, and safety at OpenAI. They concluded that advances in compute and data were producing steady gains in model capability, with new behaviors appearing as systems scaled. That trend convinced them human-level AI could arrive within their lifetimes and that a new organization was needed to handle both the opportunity and the risk.
A central insight inside the founding team was that increasingly powerful AI could be built by expanding compute, enlarging models, and feeding them more data. Many in the field had expected explicit hand-coding of reasoning or planning, but the group saw intelligence improving on a more predictable scaling curve. That raised the prospect of AI systems capable of contributing to software, mathematics, biology, robotics, and medicine.
The founders saw AI as a technology with transformative upside, including potential advances against cancer, heart disease, obesity, infections, and other problems constrained by human ingenuity. At the same time, they viewed it as comparable in gravity to pharmaceuticals or nuclear science: immensely valuable, but dangerous if mishandled. That dual view drove the decision to create a company where safety would be part of the core mission rather than a secondary concern.
The company incorporated as a public benefit corporation on January 26, 2021. The structure was meant to allow decisions that favored long-term societal benefit over short-term shareholder returns. Internally, that meant planning for moments when the company might deliberately sacrifice profit, speed, or market share in order to reduce risk or deploy systems more carefully.
The early operation was lean and improvised. Senior leaders handled basic startup work including securing AWS compute clusters, building the first website, fundraising, recruiting, and even payroll. The company initially functioned mainly as a research lab, with uncertainty over whether its systems would become major products or remain scientific curiosities.
Early models were built around the goal of being helpful, harmless, and honest, but progress was uneven. Some systems confidently invented false information, endorsed bizarre advice such as an all-potato diet, or could be pushed into absurd behavior. A turning point came in mid-2022 with the model that became Claude, which began answering questions in a more grounded and coherent way and showed a clear jump in common-sense reasoning.
Claude was released publicly in March 2023, but the company viewed itself as trailing the frontier. Potential customers openly questioned why they should choose its model over rivals. That period intensified the push to improve pretraining, reinforcement learning, and efficient use of compute, with the company treating access to large-scale computing power as a strategic necessity.
The release of Claude 3 Opus marked the first time the company felt it had a top-tier model. Claude 3.5 Sonnet then strengthened its position, particularly for programming tasks. That led to the development of Claude Code, a tool designed to let the model act more directly on a user’s computer to solve software problems, helping trigger a new wave of demand.
Anthropic argues that safety research is strongest when conducted on the most capable systems rather than weaker ones. Working at the frontier can reveal failure modes and misuse risks earlier, giving developers more time to respond. The company has also said it publishes safety research even when doing so may weaken its competitive position, on the premise that broad adoption of safeguards is necessary for the industry.
A major shift came with limited testing of Mythos, a model reportedly able to find software vulnerabilities at an expert level across many applications. The company concluded that if such a system were broadly accessible to malicious actors, the consequences could be severe. Instead of a standard release, it launched the Glasswing project, giving access to a narrower set of organizations while building safeguards and enabling defenders to identify and patch vulnerabilities first.
Leaders inside the company see that slower, more controlled deployment can hurt in the short run but better serves the long-term mission. They argue that AI companies must offer not only warnings about catastrophic misuse but also a positive case for the technology. Current uses cited include helping patients pursue better diagnoses, supporting entrepreneurs and hobbyists, and accelerating biology research, including work that may uncover new gene-editing mechanisms at a much faster pace.
Anthropic was founded on the belief that advanced AI will be both profoundly useful and inherently risky. Its strategy has been to stay at the frontier while accepting commercial tradeoffs when necessary to shape how the technology is deployed.
Ask a question