OpenAI quietly dropped a safety-focused retrospective on GPT-4 this week, and the headline numbers are worth a second look. The company says GPT-4 is 82% less likely to respond to requests for disallowed content and 40% more likely to produce factual responses than GPT-3.5 on internal evaluations. Those aren’t marginal tweaks — they’re the kind of improvements that actually shift how aggressively a model gets deployed in production environments where compliance teams have veto power.
The post traces the lineage from GPT through GPT-2 and GPT-3, framing GPT-4 as the latest milestone in scaling deep learning. But the more interesting thread is the safety pipeline. OpenAI incorporated ChatGPT user feedback directly into training, worked with over 50 experts across AI safety and security for early red-teaming, and — perhaps most notably — used GPT-4 itself to help generate training data for fine-tuning and iterate on classifiers. That’s the self-referential loop that’s become standard practice since: models helping to align models.
OpenAI is also candid about what hasn’t been fixed. Social biases, hallucinations, and adversarial prompts remain open problems. The company frames transparency and user education as part of the mitigation strategy, which reads as an acknowledgment that the technical safeguards alone won’t carry the load. GPT-4’s original training ran on Microsoft Azure AI supercomputers, a partnership that’s only deepened since with Azure’s AI-optimized infrastructure now serving the model globally.
GPT-4 was available on ChatGPT Plus and via API at launch, which feels almost quaint given how quickly OpenAI has shipped GPT-4o, o1, and successors since. But the safety architecture described here — human feedback loops, expert red-teaming, model-assisted classifier iteration — became the template for everything that followed. The fact that OpenAI is revisiting this two years later suggests the alignment playbook hasn’t changed as much as the model names have.
