Claude's 2023 Launch: Closed Alpha & Constitutional AI

A closed alpha, not a public release

Anthropic's March 14, 2023 announcement offered Claude only to a short list of partners: Notion, Quora, DuckDuckGo, Juni Learning, Robin AI and AssemblyAI, through a closed alpha. Two sizes shipped: Claude, described as Anthropic's high-performance model, and Claude Instant, a faster and cheaper option. Neither had a public sign-up page or a chat interface a reader could try that day. [1]

What Constitutional AI actually replaces

Constitutional AI, described in a paper Anthropic published in December 2022, trains a model in two stages. First the model critiques and rewrites its own answers against a written list of principles and is fine-tuned on the results. Then a preference model is trained not on human judgments of which of two answers is better, but on a second model's judgments made against those same principles, and reinforcement learning optimizes against that preference model. [2]

Atlas interpretation: The swap against RLHF is specifically in the preference stage. Human labelers rating pairs of harmful-versus-not responses are replaced by a model applying the constitution's rules, so the principles document is where human judgment enters the pipeline instead of a labeling workforce rating individual outputs. [2]

The document came two months after the model

Anthropic did not publish the constitution's actual text on launch day. That came May 9, 2023, in a post that opened: "Since launching Claude, our AI assistant trained with Constitutional AI, we've heard more questions about Constitutional AI." The March 14 announcement itself never uses the phrase Constitutional AI or names a training method. [3][1]

Atlas interpretation: So the summary's shorthand, that Claude arrived trained with a written constitution, compresses two dates into one. The method behind it predates the launch by three months in Anthropic's own research; the actual written principles postdate it by nearly two months. On March 14 a reader had Anthropic's word that Claude was built to be safer, not a document to check that claim against. [3][2]

Two labs, one afternoon, unequal reach

OpenAI shipped GPT-4 the same day, at 10:06 a.m. Pacific, live immediately to paying ChatGPT Plus subscribers, with a developer API waitlist behind it. [4]

Atlas interpretation: The shared date obscures a real gap in reach that day. One model launched to anyone willing to pay a monthly subscription; the other to six named companies with no public access point. A rivalry grew over the following years, but on March 14 only one of the two models was available to the general public. [1][4]