Technical Optimizations and Breakthroughs of GPT‑4: Multimodal Capabilities, Alignment Strategies, and Predictable Scaling

Less than five months after the release of ChatGPT, OpenAI launched GPT‑4, a large multimodal model that can process both images and text and generate textual responses.

The piece compiles two highly up‑voted Zhihu answers that discuss GPT‑4’s technical optimizations and breakthroughs, aiming to provide readers with insights for understanding and applying the model.

What is GPT‑4? GPT‑4 is a massive multimodal transformer that accepts image‑text inputs and outputs text. It claims human‑level performance on many professional and academic benchmarks, ranking in the top 10% on simulated exams, while its predecessor GPT‑3.5 often falls in the bottom 10%.

The model required about six months of alignment iterations to improve factuality and controllability, and its training infrastructure was rebuilt on Azure, enabling more stable and predictable training.

Capabilities Compared with GPT‑3.5, GPT‑4 shows dramatically better understanding of complex instructions, excels at challenging tasks such as Olympiad‑level problems, and demonstrates strong multilingual performance on MMLU. Its multimodal vision‑language ability allows it to solve physics problems presented entirely as images by performing OCR, visual reasoning, and step‑by‑step logical deduction.

Limitations GPT‑4 still hallucinates, though its truthfulness (e.g., on TruthfulQA) has improved. The base model is only modestly better than GPT‑3.5, but reinforcement learning from human feedback (RLHF) yields a large performance jump, and the model now avoids generic filler responses.

Technical Report Highlights The official 98‑page technical report (https://cdn.openai.com/papers/gpt-4.pdf) describes GPT‑4 as a transformer‑style model pre‑trained on publicly available and licensed data, then fine‑tuned with RLHF. Two major contributions are emphasized:

Alignment Process : includes red‑blue adversarial training, RLHF, and a novel rule‑based reward model (RBRM) where human annotators define safety rules that the model evaluates on its own outputs.

Predictable Scaling : uses a variant of OpenAI’s 2020 scaling law to fit loss versus compute, allowing performance prediction with as little as 0.1% of the full compute budget, and even extrapolates to evaluation metrics for budgeting purposes.

The report does not disclose many multimodal implementation details, but the author speculates that GPT‑4’s vision encoder follows the same approach as Microsoft’s KOSMOS‑1 (ViT/CLIP embeddings combined with text tokens).

Parameter count is not officially released; based on scaling curves the author estimates GPT‑4 likely exceeds GPT‑3 by over 100×, suggesting a scale of tens of trillions of parameters.

Overall, the article provides a concise overview of GPT‑4’s strengths, weaknesses, and the research directions of alignment and scaling that underpin its development.