March 2023
GPT-4, Claude v1, and Gemini
OpenAI launches the multimodal GPT-4; Anthropic releases Claude v1; Google launches Bard (later Gemini). The AI race at the frontier level has begun.
Three giants, one year
March 2023 marked the moment when the AI race moved into high gear simultaneously at multiple labs. OpenAI launched GPT-4 — a multimodal model that could process both text and images, performed at human level on numerous professional exams (bar exam, SAT, GRE), and showed significantly improved reasoning over GPT-3.5. Anthropic launched Claude v1 — its first publicly available model, built with Constitutional AI and a focus on safety and reliability. Google launched Bard, initially based on LaMDA and later PaLM 2, as its answer to ChatGPT.
GPT-4's capabilities
GPT-4 was a qualitative leap over its predecessor. It scored in the top 10% on the bar exam, 88th percentile on the LSAT, and 99th percentile on the SAT. It could analyze and describe images. It was significantly more consistent, less likely to hallucinate, and better at following instructions. The technical report published by OpenAI deliberately omitted details about model architecture, training data, and compute — a choice that sparked a debate about openness in AI research.
Claude and Constitutional AI
Anthropic, founded in 2021 by former OpenAI researchers including Dario Amodei and Ilya's sister Daniela Amodei, took a different approach to safety. Constitutional AI — in which the model is trained to evaluate its own responses against a set of principles — was the methodological foundation of Claude. The approach aimed to produce a model that was helpful, harmless, and honest without relying solely on human feedback for every response.
Sources
- OpenAI (2023). GPT-4 Technical Report. arXiv:2303.08774.
- Wikipedia — GPT-4
- Wikipedia — Claude