An Alternative Trajectory for Generative AI

Abstract

The generative artificial intelligence (AI) ecosystem is undergoing rapid transformations that threaten its sustainability. As models transition from research prototypes to high-traffic products, the energetic burden has shifted from one-time training to recurring, unbounded inference. This is exacerbated by reasoning models that inflate compute costs by orders of magnitude per query. The prevailing pursuit of artificial general intelligence through scaling of monolithic models is colliding with hard physical constraints: grid failures, water consumption, and diminishing returns on data scaling. This trajectory yields models with impressive factual recall but struggles in domains requiring in-depth reasoning, possibly due to insufficient abstractions in training data. Current large language models (LLMs) exhibit genuine reasoning depth only in domains like mathematics and coding, where rigorous, pre-existing abstractions provide structural grounding. In other fields, the current approach fails to generalize well. We propose an alternative trajectory based on domain-specific superintelligence (DSS). We argue for first constructing explicit symbolic abstractions (knowledge graphs, ontologies, and formal logic) to underpin synthetic curricula enabling small language models to master domain-specific reasoning without the model collapse problem typical of LLM-based synthetic data methods. Rather than a single generalist giant model, we envision "societies of DSS models": dynamic ecosystems where orchestration agents route tasks to distinct DSS back-ends. This paradigm shift decouples capability from size, enabling intelligence to migrate from energy-intensive data centers to secure, on-device experts. By aligning algorithmic progress with physical constraints, DSS societies move generative AI from an environmental liability to a sustainable force for economic empowerment.

Figures (6)

• Figure 1: A conceptual workflow for domain-specialized reasoning with LLMs. Stages may be combined or iterated.
• Figure 2: Overview of the GraphMERT framework. (I): Unified representation (Ic) combines syntactic knowledge from text (Ib) with semantic examples that include domain-specific relations from a seed KG (Ia). (II): GraphMERT is trained on semantic examples unified with their syntactic context (IIa). It then predicts novel semantic completions, using their syntactic information as context (IIb). An LLM helps refine the linguistic structure of raw triples proposed by GraphMERT(III). After filtering them by similarity to the source context and dropping duplicate triples, we obtain the final KG.
• Figure 3: Overview of the SFT+RL pipeline. While SFT enables axiomatic grounding, the KG-path-derived reward signal provides the process supervision required for compositional reasoning.
• Figure 4: A possible agentic workflow based on a DSS society. A central orchestrator (a fine-tuned SLM) routes tasks to a library of modular DSS specialists. The architecture supports local edge deployment for energy efficiency and integrates a continuous "AI Scientist" closed-loop where automated physical/digital experiments continually update the shared system memory.
• Figure 5: AI lifecycle energy accounting with normalized training baseline. The gray horizontal line denotes one-time training energy, normalized to $E_{\mathrm{train}}=1$, whereas the blue curve shows cumulative inference energy $E_{\mathrm{inf,cum}}(t)$ over the deployment lifecycle (equivalently, cumulative query volume). The crossover point $t_{\mathrm{cross}}$ occurs when $E_{\mathrm{inf,cum}}=E_{\mathrm{train}}$, at which inference accounts for 50% of total lifecycle energy. The threshold $t_{90}$ occurs when $E_{\mathrm{inf,cum}}\approx 9E_{\mathrm{train}}$, implying inference contributes 90% of lifecycle energy. The region between $t_{\mathrm{cross}}$ and $t_{90}$ is a transition zone (inference = 50-90% of lifecycle energy). The shaded region beyond $t_{90}$ indicates the inference-dominant regime. The curve shape is illustrative; the exact trajectory depends on request growth and per-query compute over time.