Mike, list I think the second paper is showing us how induction can expand the perimeters of explanation within an already existent hypothesis [ understanding the term ‘hypothesis’ to mean a particular Form of existence]. For example - an expansion that "led to the discovery of 75 previously unknown connections between constants”
As to abduction, which to my understanding, is the creation [ self organized or other organized] of an entirely novel Form - that’s a different action in my view. That is, if we understand an hypothesis to be a FORM, ie, a stable pattern of organization of matter and/or concepts….then, how readily does such an event take place? I can think of several intellectual abductive actions - ie - the discovery that various diseases were not caused by evil humours or..but due to the presence of microorganisms - bacterial cells. This was enabled by technological advances [ microscopes etc]. Discovery of DNA. Develoopment of agriculture. Constitutional democracy as a mode of governance - which is dependent on population size]... What about, in the natural world, the emergence of an entirely new Form [ habit]? It does happen - via a reactive adaptation to environmental pressures, isolation of the host population…and I’d say it’s more a process of adaptation of an existent FORM [ elephants without tusks] than emergence of an entirely new FORM, for such would be destabilizing in the ecological network. .. and in the less complex realms of chemistry - would be too destabilizing [ a new atom? A new molecule} Man-made abductive changes in the natural world? Covid, for example, was not a natural development, but, due to the gain-of-function development in a human laboratory,to change the infrastructure of the virus. Edwina > On Dec 19, 2024, at 2:00 AM, Mike Bergman <[email protected]> wrote: > > Hi All, > > When I initially responded to this thread I knew it had the prospect for > lengthy to-and-fro, so I silently told myself I would say my piece and then > hold my tongue. > > But I can't resist on two levels. The first is to understand abduction in > Peircean terms. The second is to comment critically on the actual > capabilities of current AI systems (and to make sure we distinguish LLMs and > chatbots from other more dedicated AI systems, which are broader in toto). > Cosmically, two new papers came out in the past day that speak to both of > these topics. Their appearance convinced me I should release my tongue. > > I find Peirce's concept of abduction (which I prefer to retroduction, another > argument) to be more subtle and precise and embracing than standard uses. My > understanding of Peirce's concept of abduction has two complementary aspects. > The first, based on the newly encountered, surprising fact, is to span the > possible problem and solution 'space' to begin formulating possible > explanations or hypotheses. The second, given the inexhaustible scope of the > possible solution space, is to winnow possible testable approaches to the > open question based on pragmatic prospects of likely rewards at acceptable > investigation costs. It is a beautiful, yet difficult to operationalize, > design. I think this aligns with Edwina's basic understanding, as well. > > In the general literature, I find the general use of 'hypothesis' and > 'abductive reasoning' to not conform to these subtleties. LLMs are inductive > in their approaches, that conjure up hypothesis largely based on priors, but > one can also prompt with new hypotheses and work interactively to explore new > territory. That is the research assistant-mode I mentioned previously. > > So, the first paper from today [1] really gets at this general topic. I think > it pretty well speaks to the lack of innovative, 'new' hypotheses arising > from current LLMs. The hypotheses offered by LLMs are generally priors. Also, > please note there are AI capabilities from outfits like DeepMind and their > protein folding that work on different bases than chatbots. There are some > new hypothesis-generating capabilities in AI, but these approaches (to my > knowledge) are not yet incorporated into LLMs. > > The second paper that came out today [2] begins to show a glimmer of > approaches that might get closer to a Peircean understanding of abduction. > This paper [2] embraces the two main concepts: first, a way to create a > representation of the general problem space (a lattice in this instance); and > then, second, a way to use an ROI-like approach to scan the possible solution > space. This paper is very mathematically and constant (also using precision) > inclined one, but the general idea is something that could be adopted to > imprecise text and LLMs perhaps. Or, possibly, a Peircean abduction would > require even other approaches . . . . > > My own conclusion is that using LLMs as human-in-the-loop research assistants > is the current, most productive way to use AI for Peircean-related questions > and investigations. They do not innovate new solutions without the > interpretant (human) prompting the right questions and premises, but may > surface previously unforeseen connections that the human might use to > formulate new hypotheses. One good thing: ChatGPT or whatever never gets > tired! > > I will again try to hold my tongue. > > Best, Mike > > [1] https://arxiv.org/pdf/2412.13645 > [2] http://arxiv.org/abs/2412.12361 > > > On 12/18/2024 3:19 PM, Edwina Taborsky wrote: >> >> Gary R, Mike, Tuezuen, Frederik, >> >> I remain unconvinced that AI has any capacity for abduction. As I outlined >> in a previous post, I consider that induction [ Secondness] and Deduction ] >> Thirdness, are its strengths. Abduction, [Firstness] is, I feel outside of >> its range. I outlined my reasoning and won’t repeat it .. >> >> I think that the massive and rapid inductive capacities of AI enable it to >> expand the range of a hypothesis, which could indeed, suggest the inclusion >> of observed facts within a hypothesis. Abduction, to my understanding, is >> without a hypothesis and is focused on choosing one.. >> >> As Peirce points out [7.218, 1902] abduction operates by resemblance [ which >> I would define as Firstness] while induction operates by contiguity [ >> Secondness]. >> >> Or “the invention, selection and entertainment of the hypothesis’ HP 2.895. >> 1901 >> >> Therefore, I continue to think that AI is extremely useful in inductive >> exploration - with the capacity to organize and critique data within >> existent hypotheses [ and thus, expand the function of these hypotheses]. >> >> But - since it is operative in Firstness, ‘guessing’, I don’t see it as >> having the capacity for abduction, ie, the creation of totally novel >> hypotheses or Forms or Habits.. An example - the development of a totally >> new beak in a finch ; the emergence of a new species. [genetic mutation >> etc]. >> >> Edwina >> >> >> >>
_ _ _ _ _ _ _ _ _ _ ARISBE: THE PEIRCE GATEWAY is now at https://cspeirce.com and, just as well, at https://www.cspeirce.com . It'll take a while to repair / update all the links! ► PEIRCE-L subscribers: Click on "Reply List" or "Reply All" to REPLY ON PEIRCE-L to this message. PEIRCE-L posts should go to [email protected] . ► To UNSUBSCRIBE, send a message NOT to PEIRCE-L but to [email protected] with UNSUBSCRIBE PEIRCE-L in the SUBJECT LINE of the message and nothing in the body. More at https://list.iupui.edu/sympa/help/user-signoff.html . ► PEIRCE-L is owned by THE PEIRCE GROUP; moderated by Gary Richmond; and co-managed by him and Ben Udell.
