[7/7] In short: text-bounded LLMs could, in principle, represent the real world. And it's possible that structural correspondences help them to do so.
But we need further empirical work (taking care to establish exploitation!) to rule out more deflationary explanations of LLM behaviour.
[6/7] A complication is that this requires selecting appropriate task-success criteria. Different training procedures may warrant different criteria. This might result in different exploited correspondences... which would ground different contents!
[5/7] I discuss some empirical evidence that the first condition is met — LLM processing may be sensitive to activation vector offsets.
To test the second condition, I argue that we need targeted intervention experiments that modulate candidate exploited correspondences independently.
[4/7] Structural correspondences are cheap. For one to genuinely ground representation, the system must exploit it. This requires two things:
(i) processing must be causally sensitive to the relevant internal structure, and
(ii) the correspondence must contribute to successful task performance.
[3/7] When an LLM's internal states mirror real-world geography, maybe that's just an artefact of the fact that those states track the statistics of geographic language (which happens to approximately mirror the actual geography).
[2/7] Some researchers have found that LLMs' internal states structurally mirror real-world domains — colour spaces, spatial layouts, temporal orderings. But does finding such a correspondence mean the LLM represents those things? I argue: not so fast.
[1/7] My paper "Can structural correspondences ground real-world representational content in large language models?" is now out at Mind & Language.
Q: Can text-only LLMs represent things in the real world, even though they never directly interact with it?
onlinelibrary.wiley.com/share/author...
I recently sat down with Sam Bennett on the AITEC podcast to talk about my thoughts on intentions in Large Language Models. This was a fun conversation!
👂 Listen here:
open.spotify.com/episode/3sm9...
📃 Read the preprint here:
philpapers.org/rec/WILIRI-4
Join us for a public talk by Prof. Anandi Hattiangadi (Stockholm University) Center for Philosophy of AI (University of Copenhagen).
Artificial General Intelligence: A Philosopher’s Manifesto
📅 Dec 10, 18:30-20:00
📍HUSET, Rådhusstræde 13, 1466 Copenhagen.
Register: cpai.ku.dk/events/artif...
🧠🤖 Join us for the launch of the Center for Philosophy of AI (University of Copenhagen)!
📅 Sept 3, 13:00-17:00
📍CPH Conference
Keynotes on philosophy of LLMs by @parismarx.com, Ellie Pavlick, @dcm.social.sunet.se.ap.brid.gy, Tom Sterkenburg & @zhijingjin.bsky.social
Register: cpai.ku.dk
Similarly, current advanced chatbots exhibit some, but not all, of the capacities characteristic of full-fledged assertion. And some capacities they possess partially.
Our take? We should think of current LLM-driven chatbots as proto-asserters.
[5/5]
We need a different perspective.
Take young children: toddlers lack some of the cognitive capacities exercised by adult asserters, but many features are partially present.
In this phase, a child's speech is not (exactly) assertion but it's not *not* assertion: they are proto-asserters!
[4/5]
Some have tried to "split the difference" between the "pro" and "con" cases.
We argue that previous attempts to do this – treating chatbots as asserters in a merely fictional sense, or holding that they only make "proxy"-assertions on behalf of humans – are unsatisfactory.
[3/5]
We identify some considerations in favour of a "yes" answer, then review recent objections to the idea of chatbot assertion.
We argue that neither flat rejection nor straightforward endorsement is compelling. So how should we think about chatbot assertion?
[2/5]
