Architecture and deliberate opacity

GPT-4 is multimodal, accepting both image and text inputs and producing text outputs. OpenAI disclosed almost nothing about the architecture, parameter count, training data composition, or hardware. The report is explicit about this decision, citing competitive and safety considerations. The opacity generated significant debate in the research community about publication norms and scientific reproducibility, since the technical report functions more as a capability disclosure than a scientific paper in the traditional sense.

Human exam benchmarks

GPT-4 is evaluated extensively on standardized human exams. It scores around the 90th percentile on the Uniform Bar Exam, around the 93rd percentile on SAT Evidence-Based Reading and Writing, around the 80th percentile on GRE Quantitative, and passes the USMLE Step exams. These benchmarks matter because they provide a concrete and interpretable comparison to human population performance. Unlike NLP benchmarks, which can be gamed by training data contamination, professional exams have fixed scoring rubrics and broad familiarity, making performance claims harder to dispute.

RLHF alignment pipeline

GPT-4 uses reinforcement learning from human feedback, the same pipeline used for InstructGPT and GPT-3.5. Human raters compare pairs of model outputs and their preferences are used to train a reward model. The reward model then provides a training signal through PPO, with a KL penalty from the supervised fine-tuning checkpoint to prevent the model from drifting too far from coherent language generation. This pipeline is responsible for the consistency improvements over GPT-3.5: fewer refusals on benign requests, better instruction following across long conversations, and meaningfully reduced hallucination rates.

Predictive scaling

One technically notable contribution is predictive scaling. OpenAI used smaller proxy models to forecast GPT-4’s final training loss before committing the full compute budget. The relationship between proxy loss and final loss was consistent enough to make reliable predictions, which significantly reduces the risk of large training runs.

Results and impact

GPT-4 became the foundation of ChatGPT Plus and the early versions of Bing Chat. It drove a sharp acceleration in commercial AI adoption and raised concrete questions about benchmark saturation. Its staged multimodal capability and RLHF alignment pipeline became reference points for subsequent closed and open models. The decision to withhold architectural details sparked ongoing debate about whether the field had moved from open science to proprietary competition.