Multimodal Models · LLM Reasoning
GPT-4 Technical Report Explained: Benchmarks, Not Blueprints
OpenAI's GPT-4 report is a measurement document, not a recipe. It hits human-level scores on professional and academic exams — bar exam ~top 10% — yet discloses no architecture, data, or compute.
Quick answer
GPT-4 is a multimodal model that accepts image and text inputs and returns text, and it reaches human-level performance on many professional and academic benchmarks — including a simulated bar exam scored around the top 10% of test takers. The catch worth quoting: the report tells you what GPT-4 scores, not how it was built. Model size, dataset, and training compute are deliberately withheld.
A report that withholds the method
Most technical reports exist to let others reproduce the work. This one inverts that contract. OpenAI states plainly that “given both the competitive landscape and the safety implications” it discloses no details about architecture, parameter count, training data, or hardware. What remains is a Transformer pretrained to predict the next token, followed by a post-training alignment stage — a one-sentence method for a 100-page document.
That is the genuinely new thing here, and it is not a capability. The 2020 GPT-3 paper was a buildable account; the GPT-4 report is the moment a frontier lab decided that “frontier” and “open recipe” no longer travel together. Everything substantive in the paper is evaluation, not construction.
Predictable scaling
The most technically interesting disclosure is that OpenAI built infrastructure and optimization methods that behave predictably across scales, then used small runs to forecast the large one. They accurately predicted some aspects of GPT-4’s performance from models trained with no more than 1/1,000th of its compute.
This matters more than any single benchmark. Training a frontier model is a single, enormously expensive bet, and you cannot afford to discover after the fact that it underperformed. Reliable extrapolation from cheap runs turns that bet into something closer to engineering — you know roughly what you are buying before you spend the compute. It is also the part of the report a competitor would most want and least receive: the claim is stated, the method is not.
Key results
GPT-4 shows human-level performance across a wide set of professional and academic exams, with the simulated Uniform Bar Exam at roughly the top 10% of test takers — versus GPT-3.5, which sat near the bottom 10% on the same exam. Post-training alignment improved measured factuality and adherence to desired behavior over the base model.
The multimodal claim is real but narrow: GPT-4 accepts interleaved image and text as input and produces text, demonstrated on charts, diagrams, and exam figures. It does not generate images. At launch, image input was not generally available, so for most users GPT-4 behaved as a stronger text model. The report is honest that GPT-4 remains “less capable than humans in many real-world scenarios” despite the exam numbers.
Why this reset the field
GPT-4 changed the vocabulary of model releases. After it, a launch was expected to ship benchmark breadth, a safety/system card, and documented deployment behavior — not just a loss curve. It also normalized a tension that still defines the field: the most influential systems can be exhaustively evaluated in public while their construction stays private. If you build on top of frontier APIs, that opacity is now a permanent condition you design around, not a temporary state.
Limits and open questions
The honest read: exam percentiles are the report’s strongest marketing and its weakest evidence. Standardized tests reward exactly what a next-token model trained on web text does well — recall and pattern-matching on problems with clean answers — and say little about reliability on open-ended, multi-step real work. The report concedes GPT-4 still hallucinates, has a fixed knowledge cutoff, and does not learn from experience.
Reproduction is impossible by design, so no external party can audit the training data for contamination, bias, or copyrighted material; the benchmark scores must be taken partly on trust. Who should skip this paper: anyone hoping to learn how to train such a model. Read it instead as a primary source on capability and as the document that defined what “closed frontier model” means.
GPT-4’s report tells you what the model scores, not how it was made — and that omission is the point. Full paper: https://arxiv.org/abs/2303.08774
FAQ
What is the GPT-4 Technical Report?
It is OpenAI’s March 2023 paper (arXiv 2303.08774) introducing GPT-4, a multimodal model that takes image and text inputs and outputs text. It documents benchmark performance and safety work but withholds architecture, dataset, and compute details.
Does the GPT-4 Technical Report reveal the model size or training data?
No. OpenAI explicitly declines to disclose parameter count, dataset, training compute, or hardware, citing competitive and safety concerns. The method is summarized only as a next-token Transformer followed by post-training alignment.
How well did GPT-4 do on the bar exam?
GPT-4 scored around the top 10% of test takers on a simulated Uniform Bar Exam, compared with GPT-3.5 near the bottom 10% on the same test.
Is GPT-4 actually multimodal?
Yes for input: GPT-4 accepts interleaved images and text and reasons over figures, charts, and diagrams. It outputs only text and does not generate images, and image input was not generally available at launch.
What is predictable scaling in the GPT-4 report?
It is OpenAI’s claim that they forecast aspects of GPT-4’s performance from models trained with up to 1,000x less compute, using infrastructure built to behave consistently across scales — making a frontier training run more predictable before committing the compute.