The most consequential sentence in the GPT-4 report isn't about the bar exam — it's that OpenAI predicted GPT-4's performance from models trained with less than 1/1,000th of its compute. Reliable scaling forecasting turns frontier training from a gamble into engineering, and that infrastructure claim matters as much as any benchmark. Just as telling is what the report omits: architecture, model size, dataset, and training method are all withheld.
Key Capabilities
- Multimodal input. GPT-4 accepts interleaved text and images and reasons over both, though it produces only text output.
- Exam-level performance. Human-level or better across a broad battery of professional and academic tests (bar exam ~top 10%, GRE, AP suites) — a shift from NLP benchmarks toward human credential exams as the yardstick.
- Predictable scaling. Loss and some capabilities were extrapolated accurately from far smaller runs, validating scaling-law-driven development.
- Alignment via post-training. RLHF-style post-training improved factuality and adherence to intended behavior over the raw pretrained model.
Great Fit If
Read it for the scaling-prediction methodology and the safety-and-evaluation framing that set the template for later frontier-model reports. Look elsewhere if you want reproducibility or architecture details — this is a capabilities-and-safety report, not a recipe — or independent benchmark numbers, since the results are OpenAI's own.
