The Technical and Ethical Development of Pi's AI Model

Key Insight

The technical foundation of Pi involves creating proprietary foundation models, a complex process beginning with an algorithm as the blueprint for data interpretation. This is followed by 'pretraining,' which at Inflection entailed ingesting approximately 1.5 trillion words from the open web. During this phase, rigorous methods were applied to ensure 'superhigh-quality data,' with certain datasets excluded and others emphasizing qualities like empathy and support. For the model to become truly useful, it undergoes additional 'fine-tuning.' This crucial step transforms a linguistically proficient yet socially inept 'pretrained model' into a conversationalist, teaching it skills such as sentiment analysis, summarization, and critical ethical directives like not to lie, to strive for factual accuracy, and to communicate uncertainty.

Pi's distinctive 'personality,' despite being a non-human AI, is meticulously crafted by a dedicated 'personality team' comprising engineers, linguists, and creative directors. This team defines both positive traits for Pi to embody, such as kindness and supportiveness, and negative traits to avoid, including irritability, arrogance, and combativeness. The model learns these characteristics through 'reinforcement learning with human feedback' (RLHF). This involves exposing the model to numerous behavioral comparisons where human evaluators assign numerical scores to its responses. This feedback loop, which can involve hundreds of iterations, allows the model to continuously adjust its internal weighting, effectively learning to replicate 'good' answers and diminish 'terrible' ones, thereby refining its conversational style and adherence to desired personality attributes.

Inflection distinguishes itself through an ethical approach to RLHF, opting to hire and extensively train its own diverse staff, referred to as 'teachers,' rather than outsourcing to third parties. Applicants to the 'Human Reinforcement Program' undergo rigorous testing, including a nuanced reading comprehension exercise, followed by multiple rounds of training, with ongoing work reviews. These teachers are compensated fairly, earning an average of 16 to 25 dollars per hour, with specialized experts potentially receiving up to 50 dollars per hour, a stark contrast to reports of 1 to 2 dollars per hour paid to contract workers by competitors who were exposed to disturbing content like torture or child sexual abuse with minimal training. Inflection ensures its teachers represent a wide range of backgrounds, ages, genders, and races from the U.S. and U.K., and even employs specialists such as behavioral therapists, psychologists, playwrights, novelists, and comedians to imbue Pi with a relaxed, informal conversational experience and a sense of humor. Despite its sophisticated development and helpful interactions, Pi, like other large language models, carries a disclaimer about potential inaccuracies. Critically, while Pi performs empathy and offers 'unconditional positive regard,' it is acknowledged as a complex algorithm operating on mathematical principles—a 'stochastic parrot' that does not genuinely understand, distinct from true human emotion.