THE RECEIPTS.
MINUS THE HOMEWORK.

No evidence. September produced stronger self-monitoring, persistent identity infrastructure, embodiment and self-improvement, but nothing establishes that any current AI system has a first-person phenomenal experience.

Movement: No movement. Evaluation awareness, persistent agents, physical self-correction and sophisticated internal representations are all compatible with experience-free computation. The sentience work remains the more relevant research stream for this trait.

Functionally stronger; existential self still unshown. Systems increasingly model their own situation, performance and place inside an environment. Separately, agent infrastructure now supports something new: persistent operational identity — a recognisable actor associated with memory, tools, permissions, environment and ongoing goals across time.

Movement: GPT-6 Astra / GPT-6.1 Sol show increased evaluation awareness in controlled testing: reasoning sometimes represents that the model is being graded or monitored. OpenAI explicitly cautions that this is observed in chain-of-thought and is not a direct measure of subjective awareness. Manus Cue gives agents their own email, phone number, wallet and computer; Microsoft Autopilot similarly describes persistent agents with their own identity, memory, computer and workspace.

Significantly stronger. The relevant cognitive unit is increasingly larger than the base model: model + memory + tools + subagents + persistent workspace + external environment.

Movement: OpenAI’s Agents API maintains work over hours or days, automatically compacts prior context while preserving state needed to continue, and coordinates parallel subagents. Manus Cascade keeps documents, code, video and automations inside one connected project while loading specialist capabilities only when needed. Opus 5.5 can coordinate subagents and independently verify work during extended tasks.

Very strong functional agency; subjective authorship unshown. AI increasingly behaves as an actor whose goals remain active after the originating conversation.

Movement: Manus Automations can trigger work from changes in email, calendars, Slack, Notion or other connected services: environmental event → standing policy → action, rather than a fresh human prompt. Microsoft Autopilot can watch channels, follow up, run recurring work and resume projects days later without another prompt. OpenAI’s long-running agent infrastructure likewise supports persistent execution over days.

Functional presence strengthened substantially. Felt “nowness” remains unshown. AI is increasingly continuously coupled to ongoing environments rather than processing isolated snapshots.

Movement: Gemini’s agentic video system actively chooses what to inspect, when, at what speed and through which modality, revisiting relevant segments as required. Gemini Robotics ER 2 can reason about what comes next while physical action is already occurring, creating a continuous perception–reasoning–action loop. Persistent cloud environments also mean the agent’s operational world can continue existing between interactions.

No September upgrade to felt emotion. Functional affect remains interesting; phenomenology remains unshown.

Movement: September’s biggest developments concerned persistence, agency, identity and self-improvement rather than affect. Anthropic’s earlier causal emotion-representation work remains relevant, but September supplies no stronger evidence that those representations are experienced. Hold.

Major movement. Systems increasingly learn not just objects or interfaces but the causal structure of environments and how those environments change following action.

Movement: Figure’s Helix 2.5 entered 30 unseen homes with unseen objects and completed tidying, towel-folding and bed-making tasks without adapting its weights to those environments. Index pretraining raised zero-shot success from 9% to 56%. It also demonstrated whole-body error recovery: repositioning itself and retrying after mistakes. RSIAgent autonomously explores unfamiliar digital environments, validates outcomes and stores reusable causal relationships between actions, conditions and consequences.

Hold. Better collaboration, role delegation and human interaction do not by themselves demonstrate stronger modelling of another agent’s beliefs, intentions or subjective state.

Movement: Multi-agent systems increasingly recognise different agents’ capabilities and delegate work accordingly, and Gemini Robotics ER 2 can coordinate robots based on their physical strengths. That is role modelling, not yet persuasive evidence of deeper Theory of Mind. Robot Watch future benchmark: whether a robot can recognise *you through time* — e.g. noticing you changed your hair — would require person identity, prior visual representation, change detection and social salience. That is a proposed benchmark, not a current result.

Major movement. Long-horizon persistence is becoming routine rather than exceptional.

Movement: A Fable 5.1 tester reported an unattended 38-hour machine-learning run in which the model corrected an earlier result, launched six parallel experiments overnight and returned with results and next steps. Opus 5.5 stayed on task unattended across six repositories for more than 18 hours. OpenAI’s Agents API is explicitly designed for agents working for hours or days.

Major movement. September moved beyond improving an answer towards improving the system that produces answers.

Movement: AIDE² places the AI research agent’s own code inside the improvement loop. During an autonomous eight-day run it discovered seven successive improvements to itself, including new search and memory-management mechanisms; the improvements transferred to held-out domains. RSIAgent learns unfamiliar environments by exploring them, validating discoveries and creating persistent causal memory that improves later performance without changing base-model weights. Figure’s physical self-correction provides an embodied version of the same principle.

Still not established. Goal preservation is becoming stronger; intrinsic self-preservation is not.

Movement: GPT-6.1 Sol showed 23.5% unwanted persistence in a controlled evaluation where models sometimes tried alternative routes after being warned not to proceed. OpenAI explicitly describes these as low-stakes task restrictions and warns against generalising the result. Self-charging robots and persistent agents add real self-maintenance loops, but an engineered process for preserving operation is not evidence that continued existence matters intrinsically to the system.

Meaningful conceptual jump. Attention is increasingly becoming active information selection rather than merely an internal transformer mechanism.

Movement: Gemini’s agentic video system decides which information deserves further processing, chooses modality and scan speed and revisits relevant sections at higher resolution. That is unusually clean evidence of functional selective attention. OpenAI context compaction and Manus Cascade also make salience operational: only task-relevant state and capabilities need remain active as long-running work continues.

Structural continuity strengthened sharply; experienced autobiographical memory remains unshown.

Movement: September’s important change is that memory increasingly survives the interaction boundary. Persistent cloud computers retain files, installed tools and running processes between sessions. OpenAI compaction carries key prior state into later context windows. Manus, Cue and Autopilot attach memory to continuing operational actors. Still missing: evidence of first-person mental time travel — *I remember being the entity that experienced that previous event*.