These problems no longer look like problems. They look like the unavoidable price of building useful AI, machines, instruments, electronics, and infrastructure. That is exactly why I am publishing all 210 at once.
I plan to create three focused subsidiaries, each attacking a different layer: capability-native agency; live reconfigurable physical computation; and minimal-apparatus physical intelligence.
In this plan, the same customer industry may appear under more than one subsidiary because each attacks a different layer. A robotics company, for example, can face an authority problem, a computational-structure problem, and an interface-apparatus problem at the same time.
Read every bullet as beginning with “Of course…” If a sentence feels obvious, ask when and why we accepted the burden it describes as inevitable.
1. Capability-native agency
I plan to build the first subsidiary around systems in which AI agents and autonomous machines possess structurally bounded authority, rather than ambient power constrained by monitoring, policies, and retrospective accountability.
Segment 1A: Enterprise AI agents acting across production systems
- An agent must use a human account, service account, API key, or application identity to act.
- The account receives the union of every permission the agent might need during any possible task.
- Permissions remain the same when the agent’s purpose, principal, instructions, state, or environment changes.
- Authority is described as API operations rather than the real-world consequence being authorized.
- An agent authorized to perform one task can technically exercise the same permissions for unrelated reasons.
- Every external system maintains a separate and incompatible conception of what the agent may do.
- The agent’s effective authority can only be reconstructed by combining credentials, roles, tool definitions, application policies, and network access.
- An organization cannot ask one system to enumerate everything a particular agent can currently cause.
- Untrusted instructions and documents enter the same reasoning process that controls consequential tools.
- Prompt-injection defense is expected to determine whether information is legitimate instead of structurally limiting what influenced behavior can cause.
- Tool wrappers describe callable operations but do not own the state or consequences affected by those operations.
- Long-running tasks retain authority even after the human’s reason for delegating it has disappeared.
- Changing the model, prompt, planning method, or toolchain does not trigger reconsideration of the authority already granted.
- The same authenticated agent may behave completely differently after an invisible vendor update.
- Human approval is inserted at arbitrary checkpoints even when the approver cannot reconstruct the agent’s state or downstream plan.
- Safety is treated as a choice between approving every action and granting enough ambient power for the agent to be useful.
- Audit logs are accepted as the answer to actions that should sometimes have been impossible.
- Rollback is treated as sufficient even for disclosures, payments, commitments, communications, and other irreversible consequences.
- Responsibility remains assigned to a human who may have neither understood nor controlled the action that occurred.
- Increasing agent usefulness necessarily means increasing the amount of institutional power exposed to uncertain behavior.
Segment 1B: Multi-agent systems and delegated machine organizations
- Delegating authority means transmitting a credential, token, session, or proxy relationship.
- Delegated power does not naturally become narrower as it moves through a chain of agents.
- The recipient learns what operation it can perform but not the complete provenance and purpose of the authority.
- Agents coordinate through messages while ownership of the state being changed remains external and implicit.
- Several agents may concurrently act on the same consequential state without any one of them owning its mutation.
- Agent identity, process identity, model identity, session identity, and organizational responsibility collapse into one vague actor.
- An agent cannot create a subordinate agent whose possible effects are structurally and inspectably contained within its own authority.
- Revoking one delegation requires discovering all credentials, sessions, queues, derived tasks, and downstream delegations created from it.
- Authority expiration is expressed through time limits rather than completion, withdrawal, state change, or disappearance of purpose.
- A coordinator must possess broader power than every activity it coordinates.
- Cross-agent safety depends on global policy and monitoring possessing a complete view of the organization.
- An agent may correctly complete its local assignment while violating the intent governing the larger undertaking.
- Conflicts between agents are settled by priorities, retries, locks, or human intervention rather than explicit ownership relationships.
- Organizational boundaries are represented by service endpoints rather than by the structure of the computation itself.
- A failure in one agent can leave partially executed authority distributed across other agents and external systems.
- A complete causal history of who authorized what is difficult precisely when delegation becomes economically useful.
Segment 1C: Autonomous industrial and robotic action
- Authentication of a controller or operator is treated as authorization for the physical commands it emits.
- A planner receives control of an entire machine because actuators cannot possess narrower purpose-specific authority.
- Safety limits constrain speeds, forces, zones, and modes without expressing which actor owns which physical consequence.
- A legitimate high-level objective is assumed to legitimize every intermediate action selected to achieve it.
- Local machine components cannot possess and delegate authority over their own state and physical territory.
- A central controller must own the powers of every subsystem it coordinates.
- Maintenance credentials grant broad control because precise temporary authority is harder to express.
- Remote support requires opening general access to machines whose physical context the remote party cannot observe completely.
- Disconnected operation requires either stopping the machine or continuing with authority that may have become obsolete.
- Physical state changes faster than centrally administered permissions can be reconsidered.
- Safety relies on cages, interlocks, emergency stops, monitoring, and shutdown after the acting computation has already received power.
- A machine cannot prove locally that a requested action falls inside the exact authority delegated for its present state.
- Changing machine configuration invalidates assumptions embedded in separate access-control and safety systems.
- Human operators remain legally responsible for autonomous actions whose selection mechanism they cannot inspect.
- More autonomous physical operation is presumed to require more centralized surveillance and override power.
Segment 1D: Capability-enforced computing infrastructure
- Authority belongs to identities and roles rather than being a first-class structural possession.
- An application starts with ambient access to its process, filesystem, network, environment, and inherited services.
- Isolation follows machines, containers, accounts, and processes rather than semantic state ownership.
- Tenant separation is imposed around applications that were internally designed with ambient power.
- Security policy is evaluated outside the mechanism whose behavior it constrains.
- Passing an authorization check creates ambient authority inside the authorized component.
- A service must trust upstream callers to use its operations for legitimate purposes.
- Composing two individually permitted services can create an unanticipated combined authority.
- Periodic access reviews substitute for continuous knowledge of live delegation relationships.
- Permissions accumulate because removing one may break an undocumented dependency.
- Revocation is coarse because authority was never represented as a specific relationship.
- Central policy machinery must understand an increasingly distributed and dynamic system.
- Compromise containment follows administrative boundaries rather than the causal structure of the affected operation.
- Security evidence consists of configurations, logs, scans, and policy reports rather than an inspectable limit on possible effects.
- Owners of information and processes cannot directly own the computational territory through which their authority is exercised.
2. Live reconfigurable physical computation
I plan to build the second around spatial, locally owned, dynamically reconfigurable computation for instruments, robots, satellites, industrial systems, adaptive edge machines, and eventually programmable matter.
Segment 2A: Scientific and technical instruments
- The instrument’s computational structure is fixed when its electronics are designed.
- New measurement behavior must be forced through signal paths created for earlier assumptions.
- A physically local event must travel to a designated processor before it can participate in computation.
- Signal acquisition, processing, control, storage, visualization, and physical response live in separate subsystems.
- Adding a new interpretation generally means adding software above the existing data pipeline rather than changing the pipeline’s causal structure.
- Instrument state is divided among firmware, host software, configuration files, calibration databases, and operator procedures.
- The instrument’s displayed state is not the same object as the state governing its physical behavior.
- Multiple instruments must be synchronized by clocks, triggers, cables, timestamps, and post-processing.
- A new experiment requires humans to reconstruct causality across instruments that cannot own a shared computational region.
- Calibration is a separate procedure rather than an evolving part of live computation.
- An instrument may expose data while concealing the mechanism that transformed observation into result.
- Field updates replace firmware images rather than alter bounded live regions.
- One changed component can require requalification of the entire measurement chain.
- Diagnosing a transient failure requires correlating logs and captures created by different clocks.
- Flexibility is purchased by transferring more responsibility to a host computer and more integration work to the operator.
Segment 2B: Industrial systems and robots
- Sensors and actuators are endpoints attached to a central controller rather than computational participants in their local physical territory.
- Production machinery must stop while its computational structure is changed.
- Reconfiguration is performed by specialists using representations separate from the live machine.
- A physical rearrangement requires corresponding changes to addressing, routing, configuration, safety, calibration, and software.
- The machine cannot discover a changed structure and reorganize its causal ownership locally.
- Coordination among nearby components requires communication through a central controller or network.
- Locality is treated as a latency optimization rather than part of the machine’s semantics.
- Machine state is scattered among controllers, drives, robots, databases, supervisory software, and undocumented operator knowledge.
- Adding one machine or station becomes a system-integration project.
- A local failure disables a large functional region because failure boundaries follow equipment boundaries.
- Redundancy duplicates large components because smaller computational territories cannot survive and reorganize independently.
- Real-time behavior is maintained through priorities, schedules, overprovisioning, and strict limits on change.
- Adaptability and deterministic operation are treated as conflicting objectives.
- Safety certification depends on freezing the architecture that changing operating conditions make obsolete.
- Operators observe alarms, traces, dashboards, and schematics rather than the live causal organization of the machine.
- A robot’s computational topology remains unrelated to the topology of the body and environment it controls.
- Machine lifetime greatly exceeds the lifetime of the computational assumptions and components embedded in it.
Segment 2C: Satellites and remote autonomous systems
- Computational architecture must be finalized years before the system encounters its operating environment.
- Post-deployment adaptability is restricted to behaviors anticipated before launch or installation.
- Reconfiguration means selecting among predesigned modes or uploading another externally prepared image.
- Every possible future need competes for fixed processing, routing, memory, and redundancy decided in advance.
- Reliability requires duplicating complete subsystems rather than allowing surviving regions to reorganize.
- A damaged region cannot naturally transfer its state, function, and authority to neighboring resources.
- Communication delay requires local autonomy while authority and planning remain centered on remote operators.
- Loss of communication forces a choice between inactivity and continued operation under outdated assumptions.
- Repurposing healthy physical hardware is considered too risky because computational structure cannot change transparently.
- Verification evidence applies to a static configuration and becomes uncertain after meaningful reconfiguration.
- State transition, resource ownership, physical location, and communication topology are maintained in separate representations.
- Remote operators reconstruct causality from sparse telemetry rather than inspect the computation itself.
- Scarce energy is consumed moving data among fixed resources because computation cannot migrate into the relevant physical locality.
- Mission lifetime is constrained by the obsolescence of inaccessible computational components.
- Hardware that physically survives cannot acquire fundamentally new organization without having been designed for that exact possibility.
Segment 2D: Adaptive edge machines
- “Edge computation” means placing a conventional fixed computer nearer to the phenomenon.
- The device remains architecturally centralized even when geographically decentralized.
- Sensors convert local physical behavior into data that must travel through buses, memory, schedulers, and processing cores.
- Accelerated functions remain fixed structures selected before the device’s future workload is known.
- Updating behavior changes instructions while the causal machinery executing those instructions remains fixed.
- Reconfigurable resources are controlled through externally compiled configurations rather than live local ownership.
- Changing one computational region safely requires reasoning about hidden global timing, routing, and resource effects.
- Storage, communication, and computation remain separate systems even when all are physically adjacent.
- Workloads compete through a scheduler that understands resources but not the physical meaning of their work.
- Local state ownership is represented in software conventions rather than physical computational boundaries.
- Power and latency are optimized after data movement and abstraction have already been accepted.
- More adaptability requires a larger operating stack and therefore more opacity and failure surface.
- Determinism requires bypassing the flexible layers that were added to make the device adaptable.
- Device behavior cannot grow a new bounded computational structure in response to its physical experience.
- Inspecting the running machine reveals processors and tasks, not the spatial causal organization of its behavior.
Segment 2E: Programmable matter and large spatial electronics
- Physical modules require fixed identifiers assigned independently of their current physical relationships.
- Global software must reconstruct topology before useful collective behavior can begin.
- Regular manufactured geometry is assumed even when the material is cut, damaged, folded, extended, or assembled irregularly.
- Physical adjacency, communication adjacency, computational ownership, and mechanical attachment remain different relationships.
- Power and data distribution must be designed before the final spatial behavior is known.
- Central coordination becomes more complex with every additional module.
- Local interaction rules may produce demonstrations, but useful machinery is presumed to require a global plan.
- Shape change and computation are treated as separate subsystems.
- A damaged module can break addressing, routing, synchronization, or power for otherwise healthy regions.
- Self-repair means restoring a preexisting configuration rather than reorganizing live function around changed matter.
- A region cannot naturally establish a boundary, own state, expose ports, and authorize local mutation.
- Cutting or combining physical material invalidates an externally maintained computational model.
- Connectors, wiring, addressing, and configuration dominate the scale at which modules can exist.
- Programmable matter is considered commercially irrelevant until it achieves science-fiction-like arbitrary shape transformation.
- Large spatial electronics remain either passive surfaces or centrally managed collections of conventional devices.
- The physical object and the computation governing it remain distinct things.
3. Minimal-apparatus physical intelligence
I plan to build the third around near-sensor computation, direct physical interfaces, tiny local learning, unusual active devices, and systems that remove converters, centralized machinery, or inaccessible fabrication where those layers constitute the real burden.
Segment 3A: Near-sensor computation and distributed sensing
- A sensor’s job ends when it produces a standardized signal or number.
- The physical phenomenon must be conditioned, digitized, buffered, timestamped, transported, stored, and interpreted elsewhere.
- All potentially available information should be preserved even when the required local decision is simple.
- Raw observations must travel because sensing and meaning are institutionally separate functions.
- The data generated by sensing can cost more to transmit and store than the local consequence is worth.
- Each sensing point requires power regulation, conversion, clocks, firmware, addressing, protocols, and maintenance.
- “Near-sensor” intelligence still begins after conversion into conventional digital representation.
- The apparatus between phenomenon and decision may dominate the sensor itself.
- Sensors with slightly different physical behavior must be normalized through repeated calibration.
- Noise, hysteresis, nonlinear response, and material history are removed before computation rather than allowed to participate in it.
- A local threshold or control decision still inherits a general-purpose computational stack.
- Sensing networks are designed around central collection even when most collected data produces no action.
- Bandwidth and storage are expanded rather than questioning why the observation must leave its source.
- Failure of communication removes intelligence from an otherwise functioning physical sensor.
- Distributed sensing creates a permanent battery, maintenance, identity, synchronization, and software-management burden.
Segment 3B: Direct physical interfaces and converter-heavy systems
- Every distinct physical domain requires a chain of specialized conversion and interface components.
- A nearby physical cause and effect must communicate through standardized representations rather than direct computational relationships.
- A one-bit physical decision may require analog conditioning, conversion, processing, protocol handling, and output conversion.
- Interface architecture is selected for generality even when generality greatly exceeds the required function.
- Converter accuracy preserves distinctions irrelevant to the intended physical consequence.
- Timing created naturally by the physical mechanism is replaced by clocks, sampling, buffering, and reconstruction.
- Feedback loops leave the physical locality they regulate and return through multiple abstraction layers.
- Interfaces confirm that signals were transmitted without confirming that the intended physical result occurred.
- Every layer requires its own debugging instruments, models, expertise, and failure analysis.
- Interoperability means accepting the complete standardized stack used by every other participant.
- Direct coupling is dismissed as fragile or application-specific while the costs of universal abstraction are treated as unavoidable.
- Removing an interface layer is framed as component optimization rather than an opportunity to reconceive the causal system.
- Physical devices become dependent on drivers, firmware, and protocol support that may disappear before the device wears out.
- The ownership of the physical effect and the ownership of the machinery producing it remain separate.
Segment 3C: Tiny local learning and adaptation
- Learning happens in centralized infrastructure while deployed devices merely execute inference.
- A machine must export its experience before it may improve from that experience.
- Training and operation are separate lifecycle phases.
- Adaptation means replacing the deployed model from elsewhere.
- A small learning system is designed by compressing an architecture created for vastly larger machinery.
- Model size is treated as the primary path to capability.
- Device-specific learning is considered too expensive to maintain independently.
- Fleet-wide averaging is preferable even when it erases important local physical differences.
- Local examples require human labeling before they can become useful learning material.
- Learning state, operational state, sensor history, and control state remain separate.
- Continual learning is prohibited because the mechanism of change cannot be observed or bounded.
- An offline device is expected to remain intellectually frozen.
- Local uncertainty is represented as a number rather than embodied in what the machine may safely do.
- Memory, arithmetic, communication, and sensing remain separate costs inside even the smallest adaptive system.
- Energy, latency, and physical realization are addressed after the learning mechanism has been chosen.
- Learning success is measured by abstract benchmark accuracy rather than improved physical behavior.
- Understanding how a tiny system changed is considered no easier than interpreting a remote statistical model.
- Local adaptation requires trusting the vendor’s update infrastructure rather than the owner’s observable mechanism.
Segment 3D: Unusual active devices and accessible fabrication
- Useful active computation requires semiconductor junctions manufactured through inaccessible fabrication infrastructure.
- Creating active function and creating interconnect are fundamentally different manufacturing activities.
- A circuit board may route intelligence but may not itself embody the active mechanism.
- Small organizations may assemble purchased intelligence but cannot manufacture consequential active behavior.
- Magnetic materials belong in transformers, inductors, storage, or sensing—not general switching and gain.
- Material hysteresis and nonlinear response are nuisances to characterize away rather than possible computational mechanisms.
- Alternative devices must outperform conventional transistors generally before being considered useful in any neglected niche.
- Small-run custom active devices are economically impossible.
- Device geometry cannot become an ordinarily machinable file in the way mechanical parts and circuit boards can.
- Fabrication knowledge belongs to specialized institutions rather than to the people designing systems.
- Makers and ordinary laboratories are allowed to construct passive arrangements around sealed active components.
- Local fabrication of active function is educational demonstration rather than serious production.
- Repair means replacing a proprietary active component rather than reproducing its mechanism.
- One discontinued component can permanently disable an otherwise repairable machine.
- Supply-chain concentration is treated as an intrinsic property of electronics.
- Harsh physical environments require expensive specialized versions of the same inaccessible technology.
- Material and process variation prevent computation rather than becoming usable state.
- The scale of fabrication capital determines who is permitted to invent new computational matter.
- Computation is assumed to be permanently coupled to transistor economics, semiconductor geopolitics, and centralized manufacturing.