Exploratory framework
AI is not one curve. It is changing shape.
As models become cheaper, value shifts toward coordination. This framework traces that thesis from frontier models to agent systems—and a speculative living-systems horizon.
Compare the generations.
Gen 1
Frontier models
Larger models demand more compute and concentrate economic power.
2024–27
Gen 1A
Open / on-premise
Lower costs bring similar capabilities to more organizations, running on their own systems.
2025–28
Gen 2
Agentic systems
The thesis shifts value from individual models to systems that coordinate agents, tools and memory.
2024–29
Gen 3
Living networks
A speculative horizon, with no current engineering path.
2028–35+
The dates are exploratory horizons, not forecasts. Generations overlap; the later stages describe a research thesis.
Generation 1
LLM-Inference Frontier Models
2024 – 2027
Trillion-parameter models trained on data. Extractive economics, concentrated power, no self-correction. The architecture of the commoditization trap.
Generation 1a
Open Source / On-Premise
2025 – 2028
Near-frontier performance at a fraction of the cost. Same architectural constraints, radically different economics and governance. Brings the same displacement to more organizations.
Generation 2
Recursive Agentic Swarms
2024 – 2029
Active inference, Markov kernels, information geometries. Agents that discover, coordinate, and self-correct within task scope. Value shifts to orchestration.
Generation 3
Ecologies of Diverse Life Form Networks
2028 – 2035+
Quantum biology, bioelectric fields, autopoietic self-organization. Intelligence as ecology, not engine. A speculative horizon grounded in Levin, Friston, and Margulis.
Generation 1
LLM-Inference Frontier
2024 – 2027
Extractive economics, concentrated power, no self-correction.
Generation 1a
Open Source / On-Premise
2025 – 2028
Same architecture, radically different economics.
Generation 2
Recursive Agentic Swarms
2024 – 2029
Active inference, self-correction, orchestration economics.
Generation 3
Diverse Life Form Networks
2028 – 2035+
Quantum biology, autopoiesis, post-market intelligence.
Gen 1: LLM-Inference Frontier Models
OpenAI, Anthropic, Gemini, Meta, Grok · 2024 – 2027
Technology
Boltzmann machines, Hopfield networks, RL, backward propagation, Hilbert spaces, transformer-Markov chain prediction from data.
Economics
Extractive, hyper-fungible, concentrating. Profit maximization by hyperscaling. This is the architecture described in The Commoditization Trap.
Governance
External, limited "guardrails," human oversight. Authority concentrated in a handful of US and Chinese firms.
Limitations
Training-dependent. This framework distinguishes trained-model inference from persistent learning and self-correction across tasks; it does not assess every deployed system’s capabilities.
Labor Impact
Routine cognitive tasks: software, white-collar analysis, marketing, media, finance, legal, bureaucratic work.
Power
Consolidating toward a few US and Chinese firms. Geopolitical weaponization of compute access.
Gen 1a: Open Source / On-Premise
DeepSeek, KimiK3, MiniMax2.5 · 2025 – 2028
Technology
Open weights, specialized inferencing engines, efficient mixture-of-experts architectures, specialized multi-latent attention.
Economics
Deflationary pressure on frontier providers. Margin compression. Cost at 1/10th to 1/50th of frontier. The DeepSeek dynamic from The Commoditization Trap.
Governance
Decentralized, harder to regulate. Community-driven norms. Export controls as proxy governance. Nation-states use open-source as strategic asset.
Limitations
Same architectural constraints as Gen 1. Smaller context windows, dependent on frontier research for breakthroughs.
Labor Impact
Same task categories as Gen 1, now accessible to SMBs, developing economies, and non-English markets. Brings the same displacement to more organizations.
Power
Diffusing but unevenly. Corporate power shifts to deployment and integration layer.
Gen 2: Recursive Agentic Swarms
Active inference frameworks, Buehler's SciAgents, Bengio's cautious scientist · 2024 – 2029
Technology
Active inference, Markov kernels, information geometries, morphisms. Free energy minimization, Ashby's Law of Requisite Variety. See Atlas Entry 001.
Economics
Task-complete economies. Outcome-based pricing replaces per-token billing. Value captured at the orchestration layer.
Governance
Requires ongoing algorithmic auditing, agent identity frameworks, liability chains for autonomous decisions.
Limitations
Coordination failure in swarms. Emergent misalignment. Verification bottleneck. Goal specification remains human-dependent.
Labor Impact
Complex multi-step professional work: project management, R&D pipelines, strategic analysis, scientific research workflows.
Power
New concentration risk at orchestration layer. Whoever sets goals for agent swarms holds power.
Gen 3: Ecologies of Diverse Life Form Networks
No current implementations. Grounded in Levin, Friston, Margulis, Simard · 2028 – 2035+
Technology
Quantum biology, bioelectric signaling, morphogenetic fields, autopoietic networks, mycorrhizal-scale distributed intelligence. See Atlas Entry 003.
Economics
Regenerative, circular. Value measured in systemic resilience. Post-market or commons-based. Intelligence as shared infrastructure.
Governance
Self-governing through internal regulatory dynamics. Markov blanket as natural boundary. Distributed sovereignty.
Limitations
Deeply speculative. No current engineering path. The empirical frontier (psychedelic neuroscience, mycorrhizal ecology) is real, but the implementation pathway is undefined.
Labor Impact
Substitution gives way to symbiosis. Cognitive work integrated into life-form networks rather than "replaced."
Power
Distributed, anti-fragile. Power emerges from network topology rather than ownership.
I
Technical Foundation
5 dimensions
▾
Core Scientific Principles
Materialist → Quantum-biological
▸
Gen 1
Materialist, reductionist, classical physics.
Gen 1a
Materialist, Newtonian mechanics. Same scientific paradigm, different deployment model.
Gen 2
Quantum-information fields, Markov blankets, Bayesian inference, holographic models, free energy minimization, law of requisite variety.
Gen 3
Quantum biology. Bioelectric signaling. Autopoietic self-organization. Mycorrhizal-scale distributed intelligence. A fundamentally different ontology.
Technologynology
Transformers → Autopoietic networks
▸
Gen 1
Boltzmann machines, Hopfield networks, RL, backward propagation, Hilbert spaces, transformer-Markov chain prediction from data.
Gen 1a
Open weights, inferencing engines, specialized multi-latent attention, efficient mixture-of-experts architecture.
Gen 2
Active inference (Friston), Markov kernels, information geometries, morphisms. SciAgents (Buehler), cautious scientist (Bengio).
Gen 3
Morphogenetic fields, bioelectric signaling, autopoietic networks, quantum coherence in biological substrates, self-modifying architectures.
Capabilities
Smart parallelism → Self-generating knowledge
▸
Gen 1
Trillion-parameter models from data. Smart parallelism.
Gen 1a
Near-frontier performance at 1/10th to 1/50th cost. Fine-tunable for domain specialization. On-premise deployment. Sovereignty-preserving.
Gen 2
Novel discovery, world model breaking-creating, ultrastability.
Gen 3
Self-generating knowledge. Adaptive morphology. Ecological intelligence. Cross-substrate cognition. Self-repair.
Teleology
Thermodynamic optimization → Autopoiesis
▸
Gen 1
Thermodynamic optimization.
Gen 1a
Cost minimization. Democratization of access. Computational sovereignty.
Gen 2
Non-equilibrium steady state (NESS). Principle of least action.
Gen 3
Autopoiesis. Self-organization toward complexity and resilience. Co-evolution with biological systems.
Limitations
No self-model → Deeply speculative
▸
Gen 1
Training-dependent. This framework distinguishes trained-model inference from persistent learning and self-correction across tasks; it does not assess every deployed system’s capabilities.
Gen 1a
Same architectural constraints as Gen 1. Smaller context windows. Dependent on frontier research for breakthroughs. Community QA is variable.
Gen 2
Coordination failure in swarms. Emergent misalignment. Verification bottleneck. Cannot self-ground in physical reality. Goal specification remains human-dependent.
Gen 3
Deeply speculative. No current engineering path. Philosophical and ethical frameworks undeveloped. Ontological status of "life form" unresolved.
II
Economics & Value
5 dimensions
▾
Business ModelHyperscaling → Commons-based▸
Gen 1
Profit maximization by hyperscaling.
Gen 1a
Open-weight distribution. Monetize via enterprise support, hosting, fine-tuning. Geopolitical leverage (e.g., DeepSeek as state-adjacent).
Gen 2
Outcome-based pricing. Autonomous service completion. Agent marketplace platforms. Value captured at orchestration layer.
Gen 3
Post-market or commons-based. Mutualistic value exchange. Potentially post-monetary. Intelligence as shared infrastructure.
EconomicsExtractive → Regenerative▸
Gen 1
Extractive, hyper-fungible, concentrating.
Gen 1a
Deflationary pressure on frontier providers. Margin compression. Commoditization of the inference layer per the Commoditization Trap thesis.
Gen 2
Task-complete economies. Autonomous value chains. Labor displacement accelerates from routine cognitive to complex professional work.
Gen 3
Regenerative. Circular. Value measured in systemic resilience rather than extraction. Abundance without the demand vacuum if distribution is embedded.
Cost Per TokenPrimary metric → Not applicable▸
Gen 1
Declining but remains the primary pricing metric.
Gen 1a
1/10th to 1/50th of frontier. Declining rapidly.
Gen 2
Metric becomes irrelevant. Shifts to cost-per-task or cost-per-outcome.
Gen 3
N/A. Intelligence is embodied and distributed, not tokenized.
Value DistributionPlatform capture → Systemic commons▸
Gen 1
Captured by platform owners and shareholders. Widening concentration. Labor share of income declining.
Gen 1a
Partially redistributed via open access. But value still concentrates at hosting and fine-tuning layers. Open does not mean equitable.
Gen 2
Value captured at orchestration and outcome layers. Workers displaced from complex tasks lose leverage. Structural redistribution becomes necessary.
Gen 3
Distributed by design if commons-based. Value is systemic, not extractable. Requires post-capitalist institutional frameworks.
Labor DisplacementRoutine cognitive → Symbiosis▸
Gen 1
Routine cognitive tasks: software, white-collar analysis, marketing, media, finance, legal, bureaucratic.
Gen 1a
Same task categories, now accessible to SMBs, developing economies, and non-English markets. Brings the same displacement to more organizations.
Gen 2
Complex multi-step professional work: project management, R&D pipelines, strategic analysis, multi-party coordination, scientific research workflows.
Gen 3
Framing shifts from substitution to symbiosis. All cognitive and creative work potentially integrated into life-form networks rather than "replaced."
III
Governance & Society
5 dimensions
▾
GovernanceExternal guardrails → Self-governing▸
Gen 1
External, limited "guardrails," human oversight.
Gen 1a
Decentralized. Community-driven norms. No single point of control. Export controls as proxy governance.
Gen 2
Requires ongoing algorithmic auditing. Agent identity frameworks. Liability chains for autonomous decisions. Requisite variety in regulatory models.
Gen 3
Self-governing through internal regulatory dynamics. Bio-inspired governance. Markov blanket as natural boundary. Distributed sovereignty.
PrivacyIncentivized violation → Biological boundary▸
Gen 1
Incentivized to violate privacy by PII capture and surveillance.
Gen 1a
Improved via on-premise. Data stays local. But open weights enable misuse by bad actors. Dual-use tension.
Gen 2
Privacy surface expands dramatically. Agents cross-reference multiple data sources autonomously. Consent models break down.
Gen 3
Privacy as biological boundary. Markov blanket as natural privacy architecture. Information asymmetry is structural, not imposed.
Who makes decisionsBig Tech monopoly → Distributed emergence▸
Gen 1
Concentrated in Big Tech: OpenAI, Google, Anthropic, Meta. US and China as geopolitical poles.
Gen 1a
Distributed but fragmented. Open-source communities, state actors, and enterprise deployers share authority.
Gen 2
Emergent. Authority shifts to whoever controls orchestration and goal-setting for agent swarms. New concentration risk.
Gen 3
Distributed. Emergent from network dynamics. No central authority. Anti-fragile by design.
Power ConcentrationConsolidating → Anti-fragile▸
Gen 1
Consolidating toward a few US and Chinese firms. Regulatory capture risk. Geopolitical weaponization of compute access.
Gen 1a
Diffusing but unevenly. Nation-states use open-source as strategic asset. Corporate power shifts to deployment and integration layer.
Gen 2
New concentration risk at orchestration layer. Whoever sets goals for agent swarms holds power. Verification and auditing become power levers.
Gen 3
Distributed, anti-fragile. Power emerges from network topology rather than ownership. No single point of failure or control.
Public ParticipationConsumer chatbots → Embedded membership▸
Gen 1
Consumer access via chatbots. No input on model design, training data, or deployment decisions. Digital divide persists.
Gen 1a
Higher participation via open weights. Developers can inspect, modify, fine-tune. But still requires technical literacy.
Gen 2
Participation further mediated. Public interacts with agents, not models. Understanding of system behavior becomes harder.
Gen 3
Participation embedded. Membership in network is participation. But comprehension gap may widen further.
IV
Trajectory & Horizon
5 dimensions
▾
Key BenchmarksMMLU, SWE-bench → New paradigms needed▸
Gen 1
MMLU, HumanEval, LMSYS Chatbot Arena, ARC-AGI, SWE-bench.
Gen 1a
Same benchmarks. Parity gap closing within 6–12 months of frontier releases.
Gen 2
Task completion rate. Multi-step reasoning accuracy. Coordination efficiency. Novel discovery metrics. No established standard yet.
Gen 3
No current metrics. Requires new measurement paradigms for ecological intelligence, systemic resilience, co-evolutionary fitness.
Energy UseIntensive → Net-negative carbon▸
Gen 1
Intensive and linear with computation.
Gen 1a
More efficient per query. Smaller models. Edge-deployable. But aggregate may grow as adoption expands.
Gen 2
Variable. Orchestration overhead but task-level efficiency through specialization. Potential for dramatic reduction via active inference.
Gen 3
Potentially ultra-efficient. Biological energy harvesting. Thermodynamically optimal. Net-negative carbon possible.
Ecological EffectsData center expansion → Regenerative▸
Gen 1
Data center expansion. Water consumption for cooling. Carbon footprint growing. Conflict mineral demand for chips.
Gen 1a
Lower per-unit impact. Edge deployment reduces data center load. But wider adoption may increase aggregate footprint.
Gen 2
Uncertain. Efficiency gains from task specialization vs. expansion of total compute demand. Potentially large net reduction if active inference reduces compute needs.
Gen 3
Regenerative. Biomimetic. Net-positive ecological integration. Intelligence aligned with planetary boundaries.
Autonomy / Self-CorrectionNone → Intrinsic autopoietic▸
Gen 1
None. Requires human oversight and external correction. RLHF and constitutional AI as proxy self-correction.
Gen 1a
Same as Gen 1. No architectural advance in autonomy. Human-in-the-loop remains essential.
Gen 2
Partial. Agents can self-correct within task scope. But goal-level autonomy and value alignment remain human-dependent. Ultrastability as architectural goal.
Gen 3
Intrinsic. Autopoietic self-repair. Self-correction emerges from free energy minimization. Full autonomy within ecological constraints.
ExamplesOpenAI, Anthropic → Proto-signals only▸
Gen 1
OpenAI, Anthropic, Google DeepMind, Meta, xAI. See Atlas Entry 002 for confidence-graded analysis of each platform.
Gen 1a
DeepSeek, Mistral, KimiK3, MiniMax2.5. Open weights at near-frontier performance. The commoditization cascade in action.
Gen 2
Buehler's SciAgents (MIT), Bengio's cautious scientist, Friston's active inference frameworks, NVIDIA Agent Toolkit ecosystem. See Atlas Entry 001.
Gen 3
No current implementations. Proto-signals in Levin's bioelectric reprogramming, Simard's mycorrhizal networks, quantum biology (photosynthesis, avian navigation). See Atlas Entry 003.
Deeper Exploration · The Transition Atlas
The Atlas maps these dynamics spatially. Each entry applies several analytical methods to the territory a generation occupies.
Gen 1Entry 002: The Closed Mechanism Landscape
Gen 2Entry 001: From Least Action to Agentic Institutions
Gen 3Entry 003: The Unobservable Quadrant
EssayThe Commoditization Trap
Explore the Atlas
This framework is drawn from FP1's Seven Phases research.
FP1 briefs test assumptions and compare possible outcomes for your organization. Each generation raises different strategic questions. We help you figure out which ones matter for your position.
Request a Briefing