GPT-4 Launch
OpenAI releases GPT-4, demonstrating powerful multimodal and reasoning capabilities, sparking global discussion about AGI.
Overview
In March 2023, OpenAI released GPT-4, a large multimodal model capable of accepting both image and text inputs and producing text output. GPT-4 demonstrated near-human performance on multiple benchmarks, including the bar exam and medical exams. Its powerful reasoning, coding, and multimodal understanding capabilities once again sparked widespread discussion about the possibility of artificial general intelligence (AGI). The advent of GPT-4 marked that large language models have entered a more mature and practical stage.
Study Notes
Background, mechanics, impact, and key takeaways in one read
Background & Context
GPT-4 was released by OpenAI on March 14, 2023 in the "GPT-4 Technical Report." It is the first large multimodal model in the GPT line, accepting image and text inputs and emitting text. The motivation was to push complex reasoning, instruction following, and safety well beyond GPT-3.5, and to validate "predictable scaling"——using models trained with 1/1000th the compute to accurately forecast the large model's behavior, lowering training risk. On a simulated bar exam GPT-4 scored around the top 10% of test takers (GPT-3.5 was near the bottom 10%), and it led by wide margins on multilingual and multi-subject benchmarks, marking a key capability threshold.
Key Points
- 1It is the first multimodal GPT, taking image and text inputs while still outputting text.
- 2On a simulated bar exam it reached the top 10% (GPT-3.5 was near the bottom 10%), far more reliable on hard tasks.
- 3Six months of adversarial testing plus ChatGPT lessons improved factuality and safety through iterative alignment.
- 4It introduced "predictable scaling": extrapolating large-model behavior from tiny models to make training safer and controllable.
- 5It led by wide margins on MMLU across 57 subjects and beat GPT-3.5 in 24 of 26 languages.
- 6The report deliberately withholds architecture, parameter count, and training details for competitive and safety reasons.
How It Works
GPT-4 remains a Transformer pre-trained to predict the next token, then aligned with RLHF, but its public novelty is engineering and methodology rather than disclosed structure. OpenAI and Azure co-designed a supercomputer from scratch and rewrote the whole deep-learning stack, making training unusually stable——the first large model whose performance could be predicted ahead of time. "Predictable scaling" means running experiments on models with no more than 1/1000th of GPT-4's compute, fitting how loss changes with scale, and extrapolating to the full run to foresee capability and risk. For multimodality, images and text are combined into one input and handled by the same model, though image input was initially limited to a partner.
Impact & Legacy
GPT-4 pushed LLM capability past the "human-level professional exam" threshold and set the paradigms of multimodality and predictable scaling, shaping later model design. It enabled trials in high-stakes domains like medicine, law, and coding, while raising harder risks of bias, misinformation, privacy, and cyber-security. OpenAI open-sourced the Evals framework so the community could surface model gaps. Its refusal to disclose training details also sparked the "frontier-model transparency" debate. Capability leaps and safety risks together made AI governance more urgent. The model also showed that scaling alone does not remove hallucinations, keeping reliability a core open problem.
Key Takeaways
- Multimodality (image plus text) is becoming the default shape of large models, not a bolt-on.
- Predictable scaling turns frontier training from art into engineerable risk management.
- Reliability on hard tasks reveals generation gaps better than single benchmarks.
- Capability and safety must be designed together; transparency is central to governance.
Glossary
A model that handles multiple input modalities (such as image, text, audio) in one unified model; GPT-4 takes image+text and outputs text.
A methodology that extrapolates a large model's behavior and risk from tiny-model experiments, making training controllable and foreseeable.
An adversarial evaluation that probes a model's weaknesses with tricky inputs before release, to find and fix risks.
Related Links
Related Events
ChatGPT Launch
OpenAI releases ChatGPT, which takes the world by storm with its amazing conversational abilities, starting the Generative AI boom.
GPT-3: Scale is All You Need
OpenAI releases GPT-3 with 175 billion parameters, demonstrating emergent capabilities of large-scale language models.
Transformer: The Attention Revolution
The paper "Attention Is All You Need" is published, introducing the Transformer architecture and revolutionizing NLP.