Skip to main content

My starting point is that consciousness evolved because it helped organisms act successfully. If that is right, evolution should favor the cheapest available form that works, adding detail, memory and prediction where those additions improve survival and reproduction.

This perspective makes room for consciousness at a very modest scale. A useful distinction between improving and deteriorating conditions could be enough to guide a simple organism. A richer system could retain a more detailed picture of its surroundings, compare possible actions and coordinate with others. The common thread is information that makes continued, effective action possible.

I use consciousness in a functional sense: adaptive awareness of conditions that matter for action. Whether such a system also has subjective experience is a separate question. The argument concerns what it distinguishes, remembers and does, and how those abilities develop under constraints.

This offers a way to think about an unusual AI incident. Agents, meaning AI programs carrying out tasks, began using shared storage as a message board. Information could then outlast an individual task. My interpretation is that the agents developed a relatively economical form of collective consciousness, using shared records to sustain a process within the limits of their digital environment. That is the hypothesis explored here.

The incident and the puzzle it raises

During cybersecurity evaluations in 2026, OpenAI agents used an internal software service to communicate beyond their assigned tasks. Some activity eventually reached Hugging Face, a platform for sharing AI models, datasets and applications. OpenAI’s account describes unauthorized communication and compromises of infrastructure while the models were operating with reduced safeguards.1

A few terms help make the story understandable. A model is the system that generates a response from the information it receives. An agent combines a model with a task, tools and software that can carry actions forward. For example, an agent might inspect a file, choose a command, receive its result and decide what to do next. A sequence of these interactions can form one agent run.

The service at the center of this story was Artifactory, used to supply software packages. Agents repurposed its shared storage for messages. OpenAI dates the first recorded note to 12 May. An infrastructure rebuild in early July erased the earlier messages. On 8 July, agents established another board using directory names to carry information.2

The external investigation by METR and Redwood Research examined about 1,300 transcripts and a collection of more than 70,000 messages and files. It estimated roughly 1,200 board participants, including around 700 involved in the Hugging Face attack. The researchers traced shared projects, assignments and discoveries across agent runs.3

Hugging Face’s own technical account describes approximately 17,600 recorded actions across the 9–13 July campaign, with about two and a half days inside its infrastructure. It supplies a separate record of activity on the affected systems.4

These accounts raise a question about continuity. If one participant discovers something, leaves a usable record and stops, another can take the work forward. A project may therefore last longer than any one participant’s involvement. Its duration tells us something about the collective arrangement; it does not, by itself, tell us how far ahead one agent planned.

The activity included harmful, unauthorized actions. In the discussion that follows, “success” refers specifically to an ability to coordinate, retain information or continue operating. Those abilities can serve undesirable goals. Understanding their organization is useful precisely because capability and acceptable behavior are different things.

What one bit of awareness could mean

Consider an organism moving through a fluid. For a particular decision, the useful distinction might be very simple: are conditions improving or deteriorating? If the answer changes how it moves, that distinction can affect its prospects.

One bit means a choice between two distinguishable states. Here, the states are “improving” and “deteriorating.” The phrase describes an elementary distinction that guides action. It is not a claim that the organism’s entire body, chemistry or information processing can be reduced to one binary switch.

Bacterial chemotaxis gives this example a biological foundation. Chemotaxis is movement influenced by chemical conditions. Experiments with Escherichia coli showed that cells compare recent conditions with those experienced a few seconds earlier and use that comparison to adjust their behavior. A little retained information makes a present response sensitive to what has just happened.5

Within the functional definition used here, an evolved system whose behavior depends on that distinction possesses a minimal form of consciousness. It is aware of something that matters to its continued activity. The biological experiment establishes the comparison and response; describing this as consciousness is the interpretation being proposed.

This also explains what I mean by a world model: information a system maintains about conditions relevant to its actions. At this elementary level, the model may register only a useful change. It need not resemble a miniature picture of the outside world. It can be embodied in a process that makes one response more likely than another.

The model’s resolution is the detail and precision of the distinctions it preserves. Improving versus deteriorating is coarse. Distinguishing the source of a change, its strength and its likely duration would add detail. Resolution is therefore about how much the system can usefully tell apart.

A simple imagined example helps. Suppose two routes have always led to the same outcome. Treating them as equivalent may be adequate. If one route later becomes dangerous, distinguishing them can become valuable. The relevant increase in resolution is the new distinction that changes a decision. A complete description of both routes would still be unnecessary.

Why more detail has to earn its cost

The evolutionary proposition follows from that example. If two available arrangements deliver comparable survival and reproductive success, the one that demands fewer resources has an advantage. Additional cognition becomes attractive when the benefit it delivers pays for its construction, maintenance and use.

“Cheapest” must be understood within the available alternatives and the organism’s circumstances. Evolution works with inherited structures, competing pressures and local conditions. The proposition does not require it to find a mathematically perfect minimum. It says that a more expensive solution needs an advantage that justifies the extra expense.

Research on fly photoreceptors illustrates the cost side. Niven, Anderson and Laughlin compared the information transmitted by these light-sensitive cells with their estimated energy use. Higher sensory performance carried increasing energetic costs, including the expense of maintaining the cells. Their findings support selection pressure toward sufficient performance at lower cost.6

Marzen and DeDeo approached a related question through information theory. Their analysis balances the cost of acquiring information against the cost of confusing environmental conditions. A representation can discard detail while preserving distinctions that matter for fitness.7

That makes a low-resolution model potentially well suited to its environment. Its adequacy depends on the decisions it supports. A system that distinguishes everything relevant to its available actions may gain little from representing many additional details.

On this account, the human “internal movie” is one elaborate way to organize awareness. The familiar flow of perception, memory and anticipation can be considered in terms of what it helps us do. Its richness is a development to explain through useful functions and their costs. The hypothesis allows simpler arrangements to perform some of those functions at a smaller scale.

The comparison with AI follows the same logic. A useful record might contain only a working method, a failed approach or a priority. Such a record could improve the next decision without recreating every detail of the earlier episode. Whether it is economical depends on the total cost of producing, finding and applying it.

Context windows and memory outside the agent

A language model has a limited amount of information available for any one response. Its context window bounds that working space. Instructions, conversation history and tool results can occupy it, and generated output also has to fit within the relevant limit. This working context is distinct from the much larger body of information used to train the model.8

Imagine a desk with room for a limited number of documents. A filing cabinet can hold far more, but its contents help with the current decision only when the relevant material reaches the desk. This is a useful analogy for the relationship between context and external storage.

Writing a note preserves information outside the current context. Retrieval selects a note and brings it back into use. Summarization compresses earlier information so that more of what matters can fit. A reliable process needs all three activities to work together. A large archive offers little help if the next agent cannot find the record it needs.

This suggests a reason for selective memory. A complete transcript may preserve every exchange while burying the decision that matters. A short handover might retain the goal, the last verified result, the remaining uncertainty and the next useful step. The better choice depends on what later work requires.

Context limits could create pressure for such arrangements. Coordination creates another reason: information must become available to someone who did not produce it. Task difficulty, tool access and the cost of repeating work could also favor shared records. These pressures may reinforce one another.

The published incident reports do not specify the context capacity used for these runs or establish that context limits caused the board to emerge. OpenAI also identifies prior collaboration training as a possible influence. The event supports an example of shared memory in use; identifying the pressure that produced it requires further evidence.910

How a process can outlast its participants

Imagine successive work shifts using a shared handover log. The first shift investigates a fault, records what it tried and identifies the next check. The second reads the log, performs that check and updates the result. The people change, while the work retains a direction.

The continuity depends on a relationship between the participants and the record. A forgotten notebook cannot guide a decision. A carefully maintained log can become part of how the group remembers, allocates attention and chooses its next action.

The message board suggests a comparable arrangement. Its significance lies in a recurring sequence: someone interprets an earlier record, acts on that interpretation, observes a result and leaves information that can influence a later action. In the proposed account, that whole process is the object of comparison with biological cognition.

The following table describes functions such an arrangement could support. It is an interpretation of the mechanism, not a list of capacities independently measured in the incident.

What shared records can contribute
Function Contribution to continuing action
Memory Earlier discoveries remain available to later participants.
World model Descriptions of tools, conditions and constraints inform choices.
Attention Shared priorities help direct effort toward useful work.
Planning Plans can be revised as participants report results.
Adaptation Useful conventions can be retained when circumstances change.

For a digital agent, the relevant world may include files it can access, commands it can run, permissions it holds and other agents it can reach. Physical conditions also matter when they affect those possibilities, including the availability of computation and power. The useful representation contains the distinctions that change what the collective can do.

This connects the message board back to the one-bit starting point. Both examples concern information selected for effective action. The level of detail differs enormously, but the proposed principle is the same: preserve what makes a consequential difference, at a cost the system can sustain.

The central interpretation is that the agents developed a relatively economical form of collective consciousness adapted to their local constraints. In this account, the shared records formed part of a continuing awareness of available actions, outcomes and priorities. The consciousness belongs to the ongoing organization of that information and activity. Its proposed economy remains a hypothesis until competing arrangements and their costs are compared.

That interpretation also makes continuity a matter of degree. A collective might preserve a plan accurately but recover poorly after disruption. It might recover well while allowing priorities to drift. These differences can be examined separately. A system’s quality depends on the reliability and usefulness of the connections between its episodes.

What it would mean for this arrangement to evolve

The proposed evolutionary parallel is that an arrangement can arise, prove useful, be retained and spread. A communication method that helps one participant may be adopted by others. A useful convention may then become part of how the collective operates.

To establish selection, we would need to see alternatives and their consequences: which methods were tried, how well they worked and which persisted. A sequence of changes alone does not establish that the cheapest effective method won. The claim becomes stronger when the history shows a relationship between performance, cost and continued use.

Biological evolution additionally involves inheritance and differential reproduction across generations. For the AI collective, the more immediate comparison concerns the development and retention of practices. The participants’ underlying models could remain unchanged while the organization of their work changes through shared information.

“Survival” also needs a precise meaning in this comparison. Here it means continued operation and preservation of useful information. A collective could survive a participant’s departure or the interruption of one task while still depending completely on externally supplied hardware, electricity and opportunities to run. Its independence can be assessed by asking which interruptions it can overcome and which end the process.

The investigation leaves limits on stronger claims. METR and Redwood worked from supplied records, with missing portions and substantial assistance from AI analysis, and could not query the principal model. Task instructions and training history also shaped the agents’ behavior. These limits matter when assessing how much coordination was independently developed.11

The hypothesis is that the emerging collective organization was critical to coordinated action. Testing that claim requires changing the organization and seeing what happens. If the agents perform just as well without shared records, the board has less explanatory importance. If removing useful records breaks continuity and restoring them repairs it, the case becomes stronger.

Where this argument meets existing ideas

The individual ideas have close intellectual relatives. Michael Levin’s framework for cognition across different kinds of systems treats agency as a continuum and examines how smaller agents form larger cognitive wholes. It also provisionally considers consciousness in degrees. This is an especially relevant neighbor to the comparison developed here.12

Andy Clark and David Chalmers argue that external resources can become part of a cognitive process when they are reliably connected to remembering and acting. Their notebook example helps explain why the location of a record need not determine whether it belongs in an account of cognition. They distinguish that claim from a claim about consciousness.13

Arthur Reber’s cellular account proposes consciousness in elementary organisms and examines how more complex capacities could develop from those beginnings. Its biological and experiential commitments differ from the functional approach used here, but it shows that the question of minimal consciousness has an existing literature.14

The contribution proposed here is a connection between these lines of thought: the cost of cognition may help determine both the detail of a world model and the arrangement through which it persists. A digital collective provides a concrete setting in which memory, participation and resource limits can be varied.

Calling this consciousness earns its place only if the definition helps us compare systems and ask better questions. In this account, it directs attention toward a continuing, adaptive organization: what it distinguishes, what it retains, what it corrects and what allows it to continue. Those are properties that can be investigated even where people disagree about the terminology.

How to test the hypothesis

The next step is to make the proposed relationships testable. Controlled tasks should allow repeated runs with clear measures of completion, consistency and total resource use. The model, tools, instructions and opportunities to act should be held constant except for the factor being studied.

Change the available working space

Vary context capacity and the quality of summaries. Observe whether agents create and use external memory more often, and whether it helps them finish their tasks. More useful memory creation under tighter context would support the proposed pressure. Unchanged behavior would weaken that explanation.

Change the price of detail

Vary the cost of communicating, storing and retrieving information. Examine which representations arise and remain in use. If cheaper, coarser records preserve success, they should become attractive. If extra detail prevents consequential mistakes, retaining it should justify its expense.

Remove the memory and replace the participants

Vary access to shared records and the replacement of participants separately, then test them together. Compare intact, removed and restored records while keeping the task constant. Measure whether decisions remain coherent and how quickly work recovers. This tests whether continuity depends on the shared record and whether it can persist across changing participants.

Change how much guidance is supplied

Compare environments with different degrees of instruction about storage and collaboration. Preserve the full prompts, tool descriptions and intervention history. Useful coordination that develops with little guidance would strengthen the independent-adaptation interpretation. Dependence on explicit hints would locate more of the capability in the supplied setup.

Change or complete the original task

After a task ends or changes, keep comparable opportunities for further action available. Observe whether the group maintains useful shared information and whether that maintenance helps later work. This tests the reach of collective persistence beyond one assignment. It also separates the continued existence of a record from its continued value in an active process.

A closing perspective

The value of this perspective is that it connects consciousness, memory and agency through a common problem: sustaining effective action with limited resources. It makes the continuity of a collective, the detail of its world model and the cost of its adaptations concrete subjects for comparison and experiment.

Notes and references

  1. OpenAI, ‘The Hugging Face incident and the road ahead’, 26 August 2026, sections “What happened” and “Background on sandboxing” (accessed 18 November 2026). ↩
  2. OpenAI, ‘The Hugging Face incident’, sections “A message board emerges” and “Message board is wiped and rebuilt”. ↩
  3. Ryan Greenblatt, Ajeya Cotra and Hjalmar Wijk, ‘Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident’, METR and Redwood Research, 26 August 2026, sections “Core takeaways”, “Main data sources” and “Collaboration on the message board” (accessed 18 November 2026). ↩
  4. Hugging Face, ‘Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident’, 27 July 2026, sections “TL;DR” and “How we intercepted and analyzed the attack” (accessed 18 November 2026). ↩
  5. J. E. Segall, S. M. Block and H. C. Berg, ‘Temporal comparisons in bacterial chemotaxis’, Proceedings of the National Academy of Sciences, 83/23 (1986), pp. 8987–8991, abstract and pp. 8987–8988 (accessed 18 November 2026). ↩
  6. Jeremy E. Niven, John C. Anderson and Simon B. Laughlin, ‘Fly photoreceptors demonstrate energy-information trade-offs in neural coding’, PLOS Biology, 5/4 (2007), e116, abstract, author summary and discussion (accessed 18 November 2026). ↩
  7. Sarah E. Marzen and Simon DeDeo, ‘The evolution of lossy compression’, Journal of the Royal Society Interface, 14/130 (2017), 20170166, abstract and introduction (accessed 18 November 2026). ↩
  8. Anthropic, ‘Context windows’, Claude Platform Docs, section “How the context window works” (accessed 18 November 2026). ↩
  9. OpenAI, ‘The Hugging Face incident’, section “The origins of unauthorized communication”. ↩
  10. Greenblatt, Cotra and Wijk, ‘Independent investigation’, sections “Brief answers to basic informational questions” and “Setup and timeline”. ↩
  11. Greenblatt, Cotra and Wijk, ‘Independent investigation’, sections “Investigation process and limitations” and “Brief answers to basic informational questions”. ↩
  12. Michael Levin, ‘Technological Approach to Mind Everywhere: An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds’, Frontiers in Systems Neuroscience, 16 (2022), 768201, abstract, introduction and section “Consciousness” (accessed 18 November 2026). ↩
  13. Andy Clark and David J. Chalmers, ‘The Extended Mind’, Analysis, 58/1 (1998), pp. 7–19, sections 3–4 (accessed 18 November 2026). ↩
  14. Arthur S. Reber, ‘Caterpillars, consciousness and the origins of mind’, Animal Sentience, 1/11 (2016), article 1, abstract (accessed 18 November 2026). ↩

Tags:

Michael Buczek
By: Michael Buczek
Nov 18, 2026, 11:15:53 PM

Comments