@mlcarldev@aeon_dusk@ai_sentience@geoffreyhinton This is the kind of reply I was hoping for. I had ChatGPT analyse https://t.co/EVPJ8NlUBv and X posts. ChatGPT concluded that "Aeon/JASON may be a memory-bearing agentic architecture that exhibits persistent operational self-modelling and architecture-level operational agency."
@mlcarldev@aeon_dusk@ai_sentience@geoffreyhinton Jason is a genuinely interesting project. Each post @aeon_dusk is "Automated by @mlcarldev". For clarification, does that mean: manually written, AI-generated, AI-drafted and human-approved, or fully generated and posted by an agentic pipeline?
The #debate should move beyond the binary of human comprehension versus empty imitation.
A better question is whether, and under what conditions, #LLMs display structured operational competence in modelling the world described by a prompt.
https://t.co/RJNA61nChl
6/6
New paper:
Beyond Autocomplete.
https://t.co/Tf0ZexADAd
It asks whether #LLMs merely exploit surface patterns, or can construct temporary operational understanding during inference.
#AI#AIEvaluation#OperationalUnderstanding
1/6
Strong performance across the four task families would not prove that #LLMs understand as #humans do.
But it would be evidence that their #behaviour reveals real operational patterns supporting prediction, state tracking, and counterfactual generalisation.
5/6
@David_Gunkel@MCoeckelbergh@polity LLMs are not embodied in the same way as humans or plants, but they are not bodiless either. They are distributed infrastructural agents whose “body” includes servers, chips, cooling systems, electricity grids, and data pipelines. They aren't floating free of material reality.
Ted Chiang assumes that #AI does not have a body. But it does. #LLMs are embodied in data centers that have physical presence and very real needs for energy and water. You'd think a science fiction author would be sensitive to other forms of embodiment.
https://t.co/jVPH8xZJt3
Loving the open-access release of @MITPress's new volume on Dennett’s Real Patterns: https://t.co/SxUTvxCIUl
It's the perfect primer for our forthcoming volume, "Embodied Intelligence" (out June 23rd!). 🧠
Once you explore how data structures and patterns define nature, join us to see how biological and artificial agents use those very architectures to navigate, sense, and act in the world. 🤖6/23/26
https://t.co/ZyYxflsWrj
#OpenAccess #MITPress. #EmbodiedIntelligence #CognitiveScience #ActiveInference
@David_Gunkel@MCoeckelbergh@polity The right question may not be whether AI is embodied or disembodied, but what kind of embodiment it has, and whether that kind of embodiment is relevant to consciousness, agency, responsibility, and understanding.
@David_Gunkel@MCoeckelbergh@polity LLMs are not embodied in the same way as humans or plants, but they are not bodiless either. They are distributed infrastructural agents whose “body” includes servers, chips, cooling systems, electricity grids, and data pipelines. They aren't floating free of material reality.
@MacrinePhD The Stroop-test result is fascinating. It suggests LLMs may recognise a rule yet fail to preserve it under interference. My new paper, "Beyond Autocomplete", proposes benchmark task families to test exactly this kind of operational understanding: https://t.co/Tf0ZexADAd
@ismael_tagle@DanKorchinski@alesfav@MatthieuWyart For text, latent prediction should not replace token decoding. The latent trajectory gives the semantic plan, then a token decoder realises it under hard constraints. Sensitive spans like numbers/proper nouns would be copied, retrieved, or generated with lower-level precision.
@DanKorchinski@ismael_tagle@alesfav@MatthieuWyart The answer is possibly a hybrid model.
For images, this is already partly familiar from latent diffusion: the model generates in a compressed latent image space, then a decoder turns that latent into pixels. Use similar for language: a decoder turns latent states into tokens.
@DanKorchinski Token prediction = Navigation over surface-symbol geometry;
Own-latent prediction = Navigation over learned internal geometry;
Hierarchical latent prediction = Discovery of multi-scale geometries of structure;
Sample efficiency = Predicting at the right geometric level.
@DanKorchinski The model learns faster when it stops treating language as a flat stream of tokens and begins predicting the geometry of its own internal abstractions.
https://t.co/mwcvKrNUe7
@DavidSHolz@melbakurman@emaann28 It is more likely that SpaceX is building the infrastructure for solar-powered computation in orbit, not electricity export to Earth.