ChatGPT reveals that while its story begins around 2015, its current capabilities are the result of years of research, development, and most significantly learning from vast amounts of data, per the below.
· GPT-1 – June 2018 (117 million parameters)
· GPT-2 – February 2019 (1.5 billion parameters)
· GPT-3 – June 2020 (175 billion parameters)
· GPT-3.5 – November 2022 (further refinements on GPT-3)
· GPT-4 – March 2023 (multimodal, improved reasoning)
· GPT-4 Turbo – November 2023 (faster, more cost-efficient variant)
Turbo, the last version, is the prime engine processing all queries since its release, both paid and unpaid. This vast amount of data includes the near totality of human savoir-faire, professional and scientific knowledge bases in all fields, to the point that it can pass strict professional exams and write theses at the doctorate level.
The question is: with this humongous amount of data, and their extensive language-based reasoning capabilities, why have we not seen any scientific breakthrough by these LLM’s in nearly 15 years of fending on their part altogether? Does that say something about our model of science (scientific method), and the value and validity of what we know in science, in particular the fundamental premises in all disciplines? Is this a verdict on the quality of what we know in terms of our scientific principles? In light of this null result, can we expect what we know to tell us something in the least amount about or toward the resolution of what we don’t know? If there is a hard breaking between our knowns and the unknowns, can the LLM’s help at all leapfrog the barrier? Given ceiling being currently hit in their learning capacity, would more time make any difference?
https://qgmindarchitects.com/downloads/ai-threats-to-the-human-knowledge-base
While the humongous amounts of data absorbed by LLMs endow them with a breadth of general knowledge (less reliably, particular facts) unequaled by any human, those data don't provide many examples of explicit, sophisticated reasoning, or purposeful concept formation. Another shortcoming is accurate spatial visualization of objects and events. Methods of guiding rational, purposeful inference and creative concept formation will have to be developed by different means.
The so-called training provides them with a sense of association between elements of the data sets, and an ability to synthesize the same, which you could all call reasoning. Association and synthetic training have also enabled them for pattern recognition. Yes, they do fail in spatial visualization. Yet they are able to make close graphic representations to what is requested from them by prompters. While it may be true that other methods can help spur their analytic and composition skills, one must acknowledge that passing strict medical exams is not an insignificant achievement.
The question remains why in almost 2 decades, no scientific breakthrough had come along with them. And what does that mean?
- Badges
- Science topic
- Similar topics
- Engineering
- Materials Engineering
- Powders