Cezary Gesikowski
2 min readApr 30, 2023

--

Thanks for the fascinating questions, Lannie. Here is my attempt at an answer.

OpenAI employed advanced techniques like 'adversarial fine-tuning' for scaling up GPT-4, which was not used in GPT-3. This resulted in a more human-like text generation model. However, OpenAI has not shared the training code or detailed architecture for GPT-4: https://www.newscientist.com/article/2364375-gpt-4-openai-says-its-ai-has-human-level-performance-on-tests/. This paper: https://arxiv.org/abs/2211.01910 has sparked heated debates on the human-like capabilities of LLMs (major geek alert here): https://news.ycombinator.com/item?id=35506472

Some consider emergent behavior in AI as 'AI magic black box behavior,' recently endorsed by Sundar Pichai (though it sounds a bit like 'snake oil' to me in context of 'trying-to-catch-up-Bard'): https://www.searchenginejournal.com/googles-sundar-pichai-envisions-a-future-with-bard-at-your-side/484757/. However, GPT-4's capabilities, including potential reasoning or logic components, are determined by design choices made by OpenAI—some were disclosed at the 2021 OpenAI Summit, while others remain undisclosed. The new architecture enables GPT-4 to outperform previous language models in speed, performance, and workload throughput, improving natural language understanding without necessarily adding specific "reasoning engines." But Sam Altman wants us to think ChatGPT-4 is magical unicorn too ;-)... (New ChatGPT Business subscription for professionals is coming soon!)

Emergent behaviour of LLMs is a fascinating topic with origins in large biological systems. I might write more about this in the future, so stay tuned!

BTW, just for fun, I poked ChatGPT with your questions, and here are the results:

"GPT-4 does not have a specific reasoning or logic engine built into it. The apparent reasoning and logic capabilities are emergent behaviors resulting from the large-scale language model's architecture and training process. These behaviors arise from the model's ability to understand and generate human-like text based on patterns it has learned from vast amounts of data."

"While the exact structural details of GPT-4 have not been fully disclosed by OpenAI, it is more than just a larger GPT-3.5. GPT-4 incorporates advanced techniques such as adversarial fine-tuning and other improvements that contribute to its enhanced performance, natural language understanding, and capabilities. These advancements make GPT-4 a more sophisticated and powerful language model compared to its predecessors."

--

--

Cezary Gesikowski
Cezary Gesikowski

Written by Cezary Gesikowski

Human+Artificial Intelligence | Photography+Algography | UX+Design+Systems Thinking | Art+Technology | Philosophy+Literature | Theoria+Poiesis+Praxis

Responses (1)