# Why we built Anthropic

## Executive summary

Anthropic was founded out of a deep commitment to AI safety and alignment, driven by the realization that scaling models (the 'big blob of compute') leads to unpredictable, powerful, and potentially dangerous emergent abilities. The company operates as a public benefit corporation, prioritizing safety and societal good over pure profit. The discussion covers the development of advanced models like Claude 3 Opus and Mythos, which demonstrate high capability but also reveal significant cybersecurity risks, necessitating new deployment methods like the Glasswing project.

## Key takeaways

- The Imperative of Safety: The founders believed that AI, like pharmaceuticals or unlocking the secrets of the atom, holds immense value but carries enormous risks, necessitating a dedicated organization focused on safety and alignment.
- Public Benefit Structure: Anthropic incorporated as a public benefit corporation (January 26, 2021) to ensure that the company's mission was not solely governed by economic incentives, allowing them to prioritize the mission over maximum shareholder value.
- Frontier Capability and Risk: The development of highly capable models, such as Mythos, demonstrated concrete, specific ways that the technology could be misused by malicious actors, forcing a change in operational strategy.

## Technical details

- Model Evolution and Capabilities: Early models were characterized as 'incompetent' but charming. A model code-named Junior, later called Claude, was released in mid-2022 and publicly in March 2023, demonstrating improved reasoning (passing the 'helicopter test').
- Advanced AI Systems: The company developed Claude Code, a tool designed to allow users to ask the AI to perform tasks on their computer and solve problems. The release of Claude 3 Opus and Sonnet 3.5 marked a significant increase in model utility for programming.
- Vulnerability Detection (Mythos): Mythos is a class of model trained to find vulnerabilities and exploits in a variety of software applications, which was found to be highly effective by cybersecurity experts.
- Safety Deployment (Glasswing): To mitigate the risk posed by Mythos, Anthropic implemented the Glasswing project, which provided limited access to the model to a smaller set of companies while safeguards were being developed for broader availability.

## Practical implications

- AI is being used in biology to discover new gene-editing mechanisms, with the potential for rapid advancement in medical treatments.
- The technology is being utilized to assist in cybersecurity, allowing defenders to use Claude to patch vulnerabilities identified by the models.
- The company's focus on alignment aims to deliver AI that helps users start businesses or pursue hobbies, demonstrating a focus on positive societal impact.

## Topics

AI Safety, Large Language Models (LLMs), Corporate Governance, Cybersecurity

Source: https://www.youtube.com/watch?v=sof3vW4mHTs
