.png)
December 1, 2024
The term "Ask the AI" has become ubiquitous in recent years. But what exactly happens when we query an AI? A former OpenAI researcher provides insights into the background and explains that the answers from AI systems like ChatGPT do not originate from a magical source of knowledge, but are based on the evaluation of human data.
Large language models, as used in AI assistants, undergo a two-stage training process. First, they learn from massive amounts of internet text and other data. In a second phase, called fine-tuning, they are trained with conversations between "human" and "assistant." Human data labelers define the assistant's responses. Thus, the AI learns to imitate human response patterns.
When we ask an AI for the "Top 10 sights in Amsterdam," it generates an answer based on how human data labelers have previously answered similar questions. For questions not included in the training data, the system creates statistically similar answers based on the learned human response behavior. So it's a kind of "averaged response" of the human data labelers.
For specialized topics, experts are employed as data labelers. For example, medical questions are answered by doctors and mathematical problems are worked on by mathematicians. However, the AI does not need an answer for every conceivable question. Sufficient examples are enough to learn how to simulate professional answers.
This does not mean, however, that the AI can provide expert knowledge on all questions. It may lack underlying knowledge or the ability to think logically. Nevertheless, its answers generally surpass those of an average internet user. AI systems can therefore be both very limited and very useful, depending on the application.
Especially with complex political questions, such as optimal governance, AI reaches its limits. The answers one would receive from an AI system would be comparable to the answers one would receive from a human layperson who had an hour to research.
While the AI can compile information from its training data, it cannot draw independent conclusions or grasp complex relationships. It merely reflects the knowledge and opinions of the people who created the training data.
The "personality" of an AI assistant is shaped by fine-tuning. This is where the AI learns how to act as a helpful assistant. It retains its basic knowledge but adapts its style to the fine-tuning data. When AI models respond to controversial topics with phrases like "it's a controversial issue," it's because human labelers were instructed to use such phrases to maintain neutrality.
Some AI researchers, including former OpenAI employees, are critical of Reinforcement Learning from Human Feedback (RLHF). This method, in which the AI learns through human feedback, is considered an interim solution because it lacks objective success criteria, unlike systems like DeepMind's AlphaGo.
Bibliographie: https://www.youtube.com/watch?v=ccHVRMrO7Uk https://www.reuters.com/technology/sam-altmans-ouster-openai-was-precipitated-by-letter-board-about-ai-breakthrough-2023-11-22/ https://www.youtube.com/watch?v=om5KAKSSpNg https://fortune.com/2023/05/03/openai-ex-safety-researcher-warns-ai-destroy-humanity/ https://decrypt.co/138310/openai-researcher-chance-ai-catastrophe https://www.reddit.com/r/OpenAI/comments/1d8h6bl/former_openai_researcher_agi_by_2027_is/ https://www.windowscentral.com/software-apps/former-openai-researcher-says-agi-could-be-achieved-by-2027 https://www.youtube.com/watch?v=WLJJsIy1x44 https://www.transparencycoalition.ai/news/former-openai-researcher-says-the-company-broke-copyright-law https://www.thedeepview.co/p/former-openai-researcher-says-the-world-is-not-ready-for-agi