The emergence of ChatGPT

GPT: what is it, exactly?

GPT’s designers transcoded an immense base of textual knowledge (the Web) into numerical data that computers can process. A few order-of-magnitude reference points:

• Number of indexed words on the web: roughly 10¹⁵ (a rough estimate, no official count exists)
• Number of parameters in the largest GPT models: on the order of 10¹² (the exact figure isn’t publicly disclosed by OpenAI)
• Number of neurons in a human brain: roughly 10¹¹, or 86 billion per the reference study by Herculano-Houzel (2009)

GPT operates on a simple principle: predict the most probable next word given a context made up of an ordered base of words, a precise question (prompt) and its algorithmic model. The power of the model and the richness of the information produce relevant responses where “semantics emerge from signs”.

What does GPT tell us?

GPT is a parrot that is not learned, but gifted with a formidable talent for finding needles (prompts) in a haystack (the Web).

GPT does not understand what it says, comparable to a CD player broadcasting music without understanding it. It originates nothing but excels at locating and expressing the information sought.

GPT is magical – so “there’s a trick”

Human language is built on consensual conventions. Its decoupling of signified and signifier, combined with the alphabet, creates a radical economy of means for expressing infinite possibilities.

Meaning resides in the author’s mind and emerges in the reader who reconstructs meaning from signs. Thus, “the reader (human or machine) causes the meaning to emerge from the words and statistical sentences constructed by GPT”.

Garbage in, garbage out

GPT can access everything but “knows” nothing. If fed statistically inaccurate data, it would become “conspiratorial and denialist” without its natural common sense.

GPT: what’s next?

What we capture in its responses is therefore not an emergent consciousness, but the collective intelligence of its designers and web contributors, relayed by the reader who gives it meaning.

Augmented capabilities could emerge by integrating algorithmic layers using GPT dynamically, combined with recursive structures and reinforcement learning. These combinations could generate emergent adaptability properties.

Since the knowledge base remains human in nature, future advanced AIs “will in some ways remain in our image”.

Integration at Agora Software

Agora Software uses GPT for two complementary tasks:

1. Building training corpora for specific intent recognition
2. Refining intent recognition for phrases that fall outside the defined functional scope

The solution preserves the qualities of sovereignty, explainability and confidentiality that characterise Agora’s approach.