i3AI

The Capability–Sentience Gap

Amol Kelkar (kelkar.amol@gmail.com)

Discussions about “conscious AI” often go in circles because two different questions get mixed into one word. Separating them clears up most of the confusion.

Two axes

Almost all of the recent progress in AI is on the first axis. We think more of the same will not get us to the second axis, because feelings require a homeostatic stake.

Evidence from interpretability

Anthropic’s A Global Workspace in Language Models (2026) is a good example. It finds what it calls the J-space (named after the Jacobian based method used to find it), a small set of internal patterns, less than a tenth of the model’s internal activity, that behave like consciously accessible thoughts. The model can report what is in it. It can bring things into it on request (e.g. when asked to concentrate on citrus fruits while copying an unrelated sentence, “orange” and “fruits” show up in its J-space). It uses the J-space for multi-step reasoning, and the J-space is strongly connected to the rest of the network.

The paper also shows the model privately noticing when it is being tested, a kind of early “point of view”. This is real evidence on the functional axis, and a functional self in current models is a real possibility.

On the second axis, the paper makes no claim. In its words, the experiments “don’t show Claude can have experiences, or feel things in the way humans do”, and it adds that it is unclear whether any experiment could settle the question. It separates access consciousness, which it argues the J-space supports, from phenomenal consciousness, which it does not claim.

This is the same distinction we make. The space between the two axes is where i3AI works.

How could a functional self emerge?

Let’s look at the mechanism, because the same mechanism also shows the limit.

A deep autoregressive Transformer has a kind of recurrence: each generated token is fed back in as input, so the model reasons over a growing trace of text. But this loop goes through a token bottleneck, i.e. at every step a rich internal distribution gets collapsed into a single token, and a lot of information is lost. This is external recurrence through generated text, and not a persistent internal state carried forward by the “self”.

This gives the model an alien form of access consciousness. Its whole sensorium is the prompt, and almost nothing carries over from one pass to the next except the text itself, so each pass is a fresh “observation” with a single token as output. This is how standard Transformer models work. Newer experimental architectures, e.g. recurrent and state-space models, have started to carry internal state from one step to the next. Within that pass, the model can access and report what is in its prompt. Biological access consciousness works the other way around. Recurrent internal state dominates and relatively little new input is used at each moment, i.e. the system is highly stateful, and it is bound to a body in a tight interoceptive, proprioceptive and exteroceptive loop.

Newer depth-recurrent and looped transformer architectures pass continuous hidden state through reused layers without emitting tokens. This internal recurrence is a better substrate for a lasting self model. Several of these approaches describe the loop as settling into an attractor, which is the same kind of dynamics we have been focusing on for a long time (see Pattern Machine). The gains so far are limited and the internals are not well understood yet, but the direction is towards the kind of internal recurrence a functional self would need.

But none of this gives the model a stake.

Why do feelings need a homeostatic stake?

According to HTCC, feeling (valence) is the constitutive viability relation of a system that maintains itself. It is not a module that can be added on top of intelligence. If a system has to stay within certain bounds to keep existing, then some states are actually bad for it. That “bad-for-it” is real, and it comes before any felt quality. A system with no viability boundary, i.e. nothing at stake in its own persistence, has nothing for a feeling to be about.

This gives us a 2 ingredient test for machine sentience:

  1. An intrinsic stake: a real viability boundary that the system can move towards or away from. Today’s LLMs don’t have one. Their goals come from a prompt, and nothing is at risk for them.
  2. A self model that is its own implementation: a self that is re-instantiated by the same dynamics it reports on, so its states are felt and not just observed.

Current models may be getting the second ingredient, especially as internal recurrence improves. They don’t have the first at all. This leads to a clear prediction (still a prediction, since there is no agreed test for sentience):

Scaled Transformers can become capable, and maybe functionally self-aware, without becoming sentient.

Adding parameters won’t close this gap, because the missing ingredient is a stake and not more capability. HTCC explains why, Pattern Machine is the architecture that builds homeostasis into the substrate, and the testable claims are on Predictions.