An AI could be said to understand language if it used language as one of a selection of tools to operate on itself, a peer or other being, or its environment. The idea of "meaning = co-ocurrance" overlooks things like need, cause, and effect that appear when language is used as a tool to operate on its environment.<p>Most of what I read about ML and AI is about creating these monolithic models that treat networks and clusters of neurons as a single entity, but that would be like treating a species of individuals with lifecycles as a single entity. The comment in the article about how GPT models are like a shadow compared to a 3D world suggests the bottleneck to evolving them is really us, as we're trying to make just one that emmulates many of us, instead of letting one loose on the internet to divide and proliferate to evolve millions where the best few will be exponentially better. Right now we're building expert systems that are individual specimens without an ecosystem.<p>There isn't yet a botnet of GPT nodes compromising machines and harvesting compute for training and evolving through participating in forums, but then again how would I know? (There's nothing worse than failing a modern catchpa and having a flash of existential dread at the stark possibility I may have indeed been a robot all along. Now I do them at random just to be sure.)
This is a neat idea, but I think it's missing a large and important area for generalization, and that's the process of seeking and exploring exceptions or counter-examples (see my other comments for examples).<p>Language defines things through subtraction, inversion, comparison, and contrast as much as construction and straightforward language.<p>Engineering and computer science rely too heavily on induction, but deduction and other non-linear processes are largely missing from these kinds of analyses/approaches. And until they are accounted for I don't think we'll reach any kind of true approach to generalization.
Nice post! I work on NLP and I think a lot of ideas in this post resonate with what I find exciting about working on the intersection of language + the real world: large text datasets as sources of abundant prior knowledge about the world, structure of language ~ structure of concepts that matter to humans, etc.<p>I feel like the bottleneck is getting access to paired (language, other modality) data though (if your other modality isn't images). i.e. "bolt on generalization" is an intuitively appealing concept, but then it reduces to the hard problem of "how do I learn to ground language to e.g. my robot action space?" I haven't seen a robotics + language paper that actually grapples with the grounding problem / tries to think about how to scale the data collection process for language-conditioned robotics beyond annotating your own dataset as a proof-of-concept. Unlike language modeling / CLIP-type pretraining, it seems (fundamentally?) more difficult to find natural sources of supervision of (language, action). I'd be curious about your thoughts on this!<p>> When it comes to combining natural language with robots, the obvious take is to use it as an input-output modality for human-robot interaction. The robot would understand human language inputs and potentially converse with the human. But if you accept that “generalization is language”, then language models have a far bigger role to play than just being the “UX layer for robots”.<p>You should check out Jacob Andreas's work, if you haven't seen it already - esp. his stuff on learning from latent language (<a href="https://arxiv.org/abs/1711.00482" rel="nofollow">https://arxiv.org/abs/1711.00482</a>).
I like the design of your website!<p>What do you mean when you say words are disentangled, standalone concepts? I see words as being very much related to each other.<p>I assume I may be misinterpreting what you mean by "disentangled, standalone concepts”.<p>Barbara Tversky's research seems to contradict linguistic relativism. I definitely don’t think language is the foundation of cognition.
What are your thoughts on the externalism of Putnam and Kripke, i.e. that meanings aren't just defined by use, but that they are also determined by objects themselves? It feels like that puts a crimp in meaning = co-occurance, but maybe not?
This seems very similar to the research program led by the late Patrick Henry Winton: <a href="https://groups.csail.mit.edu/genesis/index.html" rel="nofollow">https://groups.csail.mit.edu/genesis/index.html</a><p>Besides, I wish that causality had been mentioned more than once in passing. Due to the existence of the ladder of causality, many important queries cannot be answered by mere observation, or even by intervention; such queries require counterfactual reasoning, and structural causal models generalize because they describe something that is very invariant in the world.