Smaller, Smarter, Closer: The Edge of Collaborative Generative AI

Roberto Morabito; SiYoung Jang

REVIEW 2 major objections 6 minor 20 references

Smaller, Smarter, Closer: The Edge of Collaborative Generative AI

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Collaborative edge SLM agents with load-aware routing can halve cloud costs.

desk verdict A useful taxonomy of edge-cloud GenAI collaboration, but the headline cost reduction claim rests on an under-specified single synthetic experiment. read the letter →

arxiv 2505.16499 v2 pith:4TO5GANH submitted 2025-05-22 cs.DC cs.AIcs.NI

Roberto Morabito , SiYoung Jang This is my paper · ORCID

classification cs.DCcs.AIcs.NI

keywords collaborativeinferencesmalllanguagemodelsedgecomputingcloudoffloadingload-awareschedulingcapabilitymetadatastoreedge-cloudorchestrationgenerativeAI

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that generative AI should be deployed through collaborative inference: small language models running on edge devices cooperate with one another, falling back to cloud models only when the edge cannot meet quality demands. It organizes collaboration into three strategies—data, computation, and knowledge cooperation—and proposes lightweight infrastructure, centered on a Capability Metadata Store, to support them. The quantitative centerpiece is an eight-device experiment with bursty synthetic traffic comparing three request-routing policies. The paper claims that switching from random to load-aware routing halved cloud costs by minimizing rejected tasks and balancing workload. If correct, even a simple scheduling change can substantially cut cloud dependence in edge GenAI deployments.

What carries the argument

The conceptual engine is a set of three task-oriented cooperation strategies—data cooperation (sharing embeddings and metadata), computation cooperation (decomposing and distributing subtasks), and knowledge cooperation (querying peers for domain expertise)—supported by four components: a Capability Metadata Store (CMS), a decentralized repository of model, configuration, session, and task-state metadata; DNS-style semantic naming for service discovery; a Task Orchestrator that decomposes queries and assigns subtasks; and a Classifier Engine that assesses task complexity and capability matching. The experimental engine is a comparison of Random, Weighted, and Load-Aware request-allocation policies across eight off-the-shelf edge devices with per-device token and concurrency limits, priced against commercial GPT-4 API rates. Load-Aware routing—directing each request to the least-utilized node—is the mechanism that carries the cost-reduction claim.

What would settle it

Re-run the Random, Weighted, and Load-Aware policies on the same eight-device setup but with a real request trace of actual token lengths, arrival surges, and device heterogeneity, and compare cloud costs; if Load-Aware does not roughly halve cloud costs relative to Random on that trace, the paper's central quantitative claim is not generalizable.

Watch

Extended reading notes

Core claim

The central discovery is that edge and cloud should act as complements rather than competitors in GenAI inference. The paper proposes an edge-first collaboration model in which SLM-enabled agents handle requests locally, cooperate through data synchronization, computation offloading, and knowledge exchange, and redirect to cloud LLMs only when the edge cannot satisfy quality-of-experience or quality-of-service requirements. In a one-hour evaluation with eight edge devices, the paper finds that request-allocation policy matters as much as hardware: Random routing caused uneven load, rejections, and cloud redirections; Weighted routing improved on it; Load-Aware routing, which sends each request to the least-utilized node, minimized rejected tasks and halved cloud costs relative to the Random baseline. The authors frame this as evidence that orchestration, not just model size, drives the economics of edge GenAI.

Load-bearing premise

The cost-halving result rests on the assumption that the synthetic bursty workload, token-size distribution, and concurrency limits used in the one-hour experiment are representative of real edge GenAI demand; no real usage traces or sensitivity analysis back that assumption.

Editorial extensions

If this is right

If the experiment transfers, deployment planners can reduce cloud redirection costs without new hardware, simply by choosing a load-aware request-allocation policy.
The three cooperation strategies provide a shared vocabulary for designing multi-agent edge systems, with the Capability Metadata Store as a concrete coordination primitive.
Domain-specialized small models at 1.5B parameters can approximate a 7B general-purpose model on targeted math tasks while using about a fifth of the memory, strengthening the case for edge-first deployment.
Edge-first inference with cloud fallback preserves cloud LLM quality for hard cases while retaining latency and privacy benefits for routine cases.
At hyperscale request volumes, even a halved per-request cloud cost translates into large absolute savings.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct test would sweep the experiment's workload parameters—token-size distribution, burst amplitude, and device count—to map the conditions under which load-aware routing's cost advantage holds or disappears.
The same load-aware logic could be applied to multi-modal and image-generation inference, where token and compute costs scale differently per request.
The Capability Metadata Store's reliance on gossip-based metadata propagation raises an open scalability question: whether stale load estimates at city scale would degrade the routing advantage.
Because the cost calculation uses GPT-4 list prices, the halving result would need re-benchmarking against open-weight cloud models and changing API prices before being used for procurement decisions.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a well-organized position piece that gives a coherent taxonomy of edge-cloud cooperation for GenAI and some sensible design guidance. Its one quantitative payoff—the claim that Load-Aware scheduling halves cloud costs—is not yet supported by the evidence as presented.

What is actually new is the clean three-way split into data, computation, and knowledge cooperation, and the CMS/SLM Lake abstractions that tie them together. That framing is genuinely handy for thinking about distributed SLM systems, and the paper is honest about its scope and cites prior work fairly, including EdgeShard and collaborative multi-agent systems. The qualitative argument that edge-first with cloud fallback can address latency, privacy, and cost is consistent with the literature and well argued.

The soft spot is Section V. The experiment is a single one-hour synthetic run with bursty arrivals and token sizes from {50,100,200,300,500} weighted toward smaller sizes, but the arrival process, burst frequency and duration, per-device concurrency limits, token generation rates, and request count are not specified. There are no error bars, no repeated trials, no raw data, and no code. Figure 6 shows a single bar chart. Because of that, the "halved cloud costs" claim reads as a demonstration of a known property—load-aware routing beats random when some nodes saturate—rather than a measured result. The paper's own caveat that the results are not meant to generalize is appreciated, but the headline number is still under-anchored. This is fixable: add parameters, run repetitions with different seeds, provide sensitivity analysis, and publish the workload generator.

This is not a fatal flaw for the article's value as design guidance. The taxonomy and architectural discussion would be useful to practitioners and researchers framing edge GenAI systems. But as a measured claim, the cost reduction should be labeled preliminary or backed by more rigor.

I would send this to peer review—it is a legitimate magazine-style contribution with a credible qualitative core. The referee should ask for the experimental details or for the quantitative claim to be toned down. I would cite the taxonomy, not the cost number.

Referee Report

2 major / 6 minor

Summary. This article argues for an edge-first approach to generative AI in which small language models on edge devices collaborate with each other and fall back to the cloud only when needed. It proposes a taxonomy of data, computation, and knowledge cooperation; sketches a Capability Metadata Store (CMS) with gossip-based dissemination, semantic discovery and routing, task orchestration, and a classifier engine; illustrates two application scenarios; and reports a one-hour synthetic experiment comparing Random, Weighted, and Load-Aware request allocation over eight Jetson edge devices. The headline quantitative result is that Load-Aware scheduling "halved cloud costs" relative to the other policies (Section V, Figure 6).

Significance. If the scheduling result were reproducible, it would provide a useful, low-cost data point that simple request-routing heuristics can materially reduce cloud offload in heterogeneous edge deployments. The taxonomy and the concrete CMS entry examples are a worthwhile conceptual contribution for a magazine-style venue. The paper is also honest about its non-exhaustive framework and explicitly disclaims generalization of the experimental results. However, the quantitative claim is not yet anchored: the experiment is described too sparsely to be reproduced or interpreted, and no sensitivity analysis is given. The design claims about CMS scalability are also asserted rather than demonstrated. These issues are fixable within the article's scope by adding experimental details or by explicitly reclassifying the cost result as an illustrative anecdote.

major comments (2)

[Section V, Figure 6] The central quantitative finding that Load-Aware scheduling "halved cloud costs" is not supported by the information provided. The text states that workloads were bursty and token demands came from {50, 100, 200, 300, 500} with smaller sizes weighted more heavily, but it does not specify the arrival process, burst frequency or duration, number of requests, the per-device concurrency and token-generation limits used, the token-size weights, or the exact OpenAI GPT-4 input/output token prices. It is also unclear which baseline the "halved" comparison uses: Figure 6 compares three policies, so the factor-of-two claim could be with respect to Random, Weighted, or both. No repeated trials, error bars, or raw cost values are reported. Because this is the article's most concrete quantitative payoff and is used to motivate impact at hyperscale, the claim should either be fully specified with a reproducible scenario and sensitivity checks, or be explicitly reclassified as an illustrative example rather than a measured result. The paper's caveat that the results are not intended to generalize does not resolve this, because the sentence is still stated as a finding for the specific "in this case" experiment.
[Section III.C] The decentralized Capability Metadata Store (CMS) and its gossip-based information sharing are presented as key enabling elements, but no evidence is given that metadata dissemination overhead remains acceptable at scale; the cited reference [15] addresses distributed learning on mobile devices rather than capability metadata stores. Since the CMS underpins semantic discovery, task orchestration, and all three cooperation types, the paper should either provide a basic overhead argument or an order-of-magnitude estimate (e.g., metadata size per agent, gossip interval, convergence time), or explicitly state that CMS scalability is an open research question. As written, the architecture's feasibility is asserted rather than supported.

minor comments (6)

[Section V] Minor language: "system consistent of eight off-the-shelf edge devices" should be "system consisting of eight off-the-shelf edge devices"; also, "Jetson Orin Nano 0" contains a stray "0".
[Figure 6] The bar labels in Figure 6 appear as "Requests ProcessedRequests RejectedCloud GPT Costs" without clear separation; please check the axis or legend rendering so each bar is identified unambiguously.
[Section II.B and Figure 2] The "Smarter" and "Closer" panels of Figure 2 report accuracy and latency comparisons without indicating data sources or measurement methodology; adding a footnote with sources would help readers judge the motivation claims.
[Section III.B] The phrase "adaptability to task complexity ," has an extra space before the comma; please correct the typography.
[Section III.C] The paragraph beginning "The CMS acts as a decentralized repository..." follows the reference to Figure 4 without a clear transition; connect the figure and the CMS description more explicitly.
[References] Reference [18] is a prior work that includes one of the authors; this is not problematic, but a full citation or a brief note in the text would clarify the relation to the present article.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are either qualitative design guidance or empirically measured comparisons, with no fitted parameter, self-citation chain, or definitional equivalence driving the central result.

full rationale

This paper is a systems/position article rather than a derivation-based study, so there is no claimed chain from first principles to a prediction that could reduce to its own inputs. The central quantitative claim in Section V—that Load-Aware scheduling 'halved cloud costs' relative to Random and Weighted allocation—is presented as a measured outcome of a stated experiment, not as a quantity derived from an assumed model. The experiment uses externally sourced inputs: device token-generation and concurrency limits 'derived from benchmarking GenAI workloads [11]' and 'OpenAI API GPT-4 pricing.' No parameter is fitted to a subset of data and then renamed as a prediction, and no uniqueness or existence theorem is invoked to force a conclusion. The taxonomy of Data, Computation, and Knowledge cooperation is a classification scheme, not a result that presupposes its own conclusion. The only self-citation is reference [18], which includes co-author S. Y. Jang, and it is used to support the incidental claim that the number of AI-enabled wearables is increasing; it is not load-bearing for any central result. The paper also explicitly disclaims generalization of the Section V measurements, which further separates the measured cost reduction from any claim that it is universally forced. Under the hard rules, self-citation alone is not circularity, and no quoted equation or construction shows that any 'prediction' is equivalent to an input. Therefore the appropriate finding is no significant circularity with score 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The central design depends on several unvalidated premises: the CMS and gossip coordination overhead is acceptable, domain-specific SLMs retain enough accuracy on decomposed tasks, and the synthetic workload represents real GenAI demand. The quantitative result also depends on hand-chosen token sizes, an unspecified burst pattern, and unspecified per-device limits. These choices are not fitted to a target result, so they are not evidence of circularity, but they do limit generality.

free parameters (3)

Token demand set = 50, 100, 200, 300, and 500 tokens; weighted toward smaller requests
Chosen by hand to reflect "common usage patterns" with no real trace or sensitivity analysis, and it directly shapes cloud offload cost calculations in Section V.
Bursty workload arrival pattern = not specified in paper
The one-hour evaluation uses "periodic surges in user requests," but the distribution and intensity are unstated, so the measured rejection and cost outcomes are not reproducible or generalizable (Section V).
Per-device token generation and concurrency limits = not specified in paper
Each device has predefined limits "reflecting constraints derived from benchmarking GenAI workloads [11]," but the exact values are not given, and they determine how often requests overflow to the cloud (Section V).

assumptions (3)

ad hoc to paper A decentralized Capability Metadata Store synchronized over gossip protocols can keep agent capabilities current with acceptable overhead in real deployments.
Introduced in Section III-C as a key enabler of coordination; no implementation, measurements, or failure analysis is provided, yet the cooperation strategies rely on it.
domain assumption Small language models running on edge devices can maintain sufficient accuracy on decomposed subtasks for the collaborative system to meet quality-of-experience goals.
Section II-B and the application scenarios assume domain-specific SLMs achieve accuracy comparable to larger models, citing external benchmarks rather than validating the multi-agent decomposition.
domain assumption The synthetic workload and cost model in Section V are representative of real edge GenAI demand.
The evaluation uses token sizes weighted toward smaller requests and GPT-4 pricing, but no real user traces or sensitivity analysis support this representativeness.

invented entities (2)

Capability Metadata Store (CMS)
purpose: A decentralized registry of agent capabilities and configuration metadata used to enable dynamic discovery and coordination.
Described at length in Section III-C and Figure 4, but not implemented or evaluated; it is a proposed architectural component with no falsifiable handle outside the paper.
SLM Lake
purpose: A pool of small language model instances from which agents can draw, shown in the conceptual framework of Figure 4.
Appears in Figure 4 but is not defined or evaluated in the text; it is a label for an unvalidated system component.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Smaller, Smarter, Closer: The Edge of Collaborative Generative AI." pith.science (2026). https://pith.science/paper/4TO5GANH

@misc{pith2026250516499,
  author       = {Pith},
  title        = {Pith review of: Smaller, Smarter, Closer: The Edge of Collaborative Generative AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4TO5GANH}},
  note         = {Machine review of arXiv:2505.16499}
}

read the original abstract

The rapid adoption of generative AI (GenAI), particularly Large Language Models (LLMs), has exposed critical limitations of cloud-centric deployments, including latency, cost, and privacy concerns. Meanwhile, Small Language Models (SLMs) are emerging as viable alternatives for resource-constrained edge environments, though they often lack the capabilities of their larger counterparts. This article explores the potential of collaborative inference systems that leverage both edge and cloud resources to address these challenges. By presenting distinct cooperation strategies alongside practical design principles and experimental insights, we offer actionable guidance for deploying GenAI across the computing continuum.

Figures

Figures reproduced from arXiv: 2505.16499 by the authors.

**Figure 2.** Smaller, Smarter, Closer: (1) Progression towards smaller language models with comparable or improved efficiency (top); (2) Improved accuracy of smaller models as measured by Pass@1 scores (middle); and (3) Infrastructurerelated latency comparisons across global locations for major AI systems, highlighting geographical disparities in access speed (bottom). models vary in size, with the smallest, Qwen2.5, featuring… view at source ↗

**Figure 3.** Task-oriented cooperation types: Data, Computation, and Knowledge cooperation strategies for multi-agent collaboration. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: On top, our conceptual framework for enabling collaboration among distributed SLM-enabled agents. On the bottom, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: A collaborative language model inference scenario [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Impact of request allocation methods on the number [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 12 canonical work pages

[15]

Distributed learning on mobile devices: a new approach to data mining in the internet of things,

X. Zhang, X. Zhu, W. Bao, L. T. Yang, J. Wang, H. Yan, and H. Chen, “Distributed learning on mobile devices: a new approach to data mining in the internet of things,” IEEE Internet of Things Journal, vol. 8, no. 13, pp. 10 264–10 279, 2020

work page 2020
[1]

The internet of things in the era of generative ai: Vision and challenges,

X. Wang, Z. Wan, A. Hekmati, M. Zong, S. Alam, M. Zhang, and B. Krishnamachari, “The internet of things in the era of generative ai: Vision and challenges,” IEEE Internet Computing , vol. 28, no. 5, pp. 57–64, 2024

work page 2024
[2]

Gpt-4o system card,

OpenAI, “Gpt-4o system card,” arXiv preprint arXiv:2410.21276, 2024

arXiv 2024
[3]

Estimating the carbon footprint of bloom, a 176b parameter language model,

A. S. Luccioni, S. Viguier, and A.-L. Ligozat, “Estimating the carbon footprint of bloom, a 176b parameter language model,” Journal of Machine Learning Research , vol. 24, no. 253, pp. 1–15, 2023

2023
[4]

(2024, Dec.) Incident report: Service down on december 12, 2024

OpenAI. (2024, Dec.) Incident report: Service down on december 12, 2024. Accessed on 2024-12-18. [Online]. Available: https: //status.openai.com/incidents/ctrsv3lwd797

work page 2024
[5]

Impact of genai on mobile network traffic,

Ericsson, “Impact of genai on mobile network traffic,” Ericsson Mobility Report , Nov. 2024. [Online]. Avail- able: https://www.ericsson.com/en/reports-and-papers/mobility-report/ articles/genai-impact-on-mobile-network-traffic

work page 2024
[6]

Revisiting edge ai: Opportunities and challenges,

T. Meuser, L. Lov ´en, M. Bhuyan, S. G. Patil, S. Dustdar, A. Aral, S. Bayhan, C. Becker, E. De Lara, A. Y . Dinget al., “Revisiting edge ai: Opportunities and challenges,” IEEE Internet Computing, vol. 28, no. 4, pp. 49–59, 2024

work page 2024
[7]

The future of consumer edge-ai computing,

S. Laskaridis, S. I. Venieris, A. Kouris, R. Li, and N. D. Lane, “The future of consumer edge-ai computing,” IEEE Pervasive Computing , 2024

work page 2024

Show all 20 references

[8]

Awq: Activation-aware weight quanti- zation for on-device llm compression and acceleration,

J. Lin, J. Tang, H. Tang, S. Yang, W.-M. Chen, W.-C. Wang, G. Xiao, X. Dang, C. Gan, and S. Han, “Awq: Activation-aware weight quanti- zation for on-device llm compression and acceleration,” Proceedings of Machine Learning and Systems , vol. 6, pp. 87–100, 2024

2024
[9]

Llm-pruner: On the structural pruning of large language models,

X. Ma, G. Fang, and X. Wang, “Llm-pruner: On the structural pruning of large language models,” Advances in neural information processing systems, vol. 36, pp. 21 702–21 720, 2023

2023
[10]

Minillm: Knowledge distillation of large language models,

Y . Gu, L. Dong, F. Wei, and M. Huang, “Minillm: Knowledge distillation of large language models,” in The Twelfth International Conference on Learning Representations, 2024

2024
[11]

Melting point: Mobile evaluation of language transformers,

S. Laskaridis, K. Katevas, L. Minto, and H. Haddadi, “Melting point: Mobile evaluation of language transformers,” in Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, 2024, pp. 890–907

2024
[12]

Velo: A vector database- assisted cloud-edge collaborative llm qos optimization framework,

Z. Yao, Z. Tang, J. Lou, P. Shen, and W. Jia, “Velo: A vector database- assisted cloud-edge collaborative llm qos optimization framework,” arXiv preprint arXiv:2406.13399 , 2024

2024 arXiv
[13]

Edgeshard: Efficient llm inference via collaborative edge computing,

M. Zhang, X. Shen, J. Cao, Z. Cui, and S. Jiang, “Edgeshard: Efficient llm inference via collaborative edge computing,”IEEE Internet of Things Journal, 2024

2024
[14]

Enabling synergistic knowledge sharing and reasoning in large language models with col- laborative multi-agents,

A. Das, S.-C. Chen, M.-L. Shyu, and S. Sadiq, “Enabling synergistic knowledge sharing and reasoning in large language models with col- laborative multi-agents,” in 2023 IEEE 9th International Conference on Collaboration and Internet Computing (CIC) . IEEE, 2023, pp. 92–98

2023
[16]

Pag-llm: Paraphrase and aggregate with large language models for minimizing intent classification errors,

V . Yadav, Z. Tang, and V . Srinivasan, “Pag-llm: Paraphrase and aggregate with large language models for minimizing intent classification errors,” in Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2024, pp. 2569– 2573

2024
[17]

Attacking and protecting data privacy in edge–cloud collaborative inference systems,

Z. He, T. Zhang, and R. B. Lee, “Attacking and protecting data privacy in edge–cloud collaborative inference systems,” IEEE Internet of Things Journal, vol. 8, no. 12, pp. 9706–9716, 2020

2020
[18]

Collaborative inference via dynamic composition of tiny ai accelerators on mcus,

T. Gong, S. Y . Jang, U. G. Acer, F. Kawsar, and C. Min, “Collaborative inference via dynamic composition of tiny ai accelerators on mcus,” arXiv preprint arXiv:2401.08637 , 2023

2023 arXiv
[19]

Edge intelligence: Paving the last mile of artificial intelligence with edge computing,

Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelligence: Paving the last mile of artificial intelligence with edge computing,” Proceedings of the IEEE , vol. 107, no. 8, pp. 1738–1762, 2019

2019
[20]

Efficient contextual llm cascades through budget-constrained policy learning,

X. Zhang, Z. Huang, E. O. Taga, C. Joe-Wong, S. Oymak, and J. Chen, “Efficient contextual llm cascades through budget-constrained policy learning,” in The Thirty-eighth Annual Conference on Neural Information Processing Systems , 2024

2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [15]

Distributed learning on mobile devices: a new approach to data mining in the internet of things,

X. Zhang, X. Zhu, W. Bao, L. T. Yang, J. Wang, H. Yan, and H. Chen, “Distributed learning on mobile devices: a new approach to data mining in the internet of things,” IEEE Internet of Things Journal, vol. 8, no. 13, pp. 10 264–10 279, 2020

work page 2020

[2] [1]

The internet of things in the era of generative ai: Vision and challenges,

X. Wang, Z. Wan, A. Hekmati, M. Zong, S. Alam, M. Zhang, and B. Krishnamachari, “The internet of things in the era of generative ai: Vision and challenges,” IEEE Internet Computing , vol. 28, no. 5, pp. 57–64, 2024

work page 2024

[3] [2]

Gpt-4o system card,

OpenAI, “Gpt-4o system card,” arXiv preprint arXiv:2410.21276, 2024

arXiv 2024

[4] [3]

Estimating the carbon footprint of bloom, a 176b parameter language model,

A. S. Luccioni, S. Viguier, and A.-L. Ligozat, “Estimating the carbon footprint of bloom, a 176b parameter language model,” Journal of Machine Learning Research , vol. 24, no. 253, pp. 1–15, 2023

2023

[5] [4]

(2024, Dec.) Incident report: Service down on december 12, 2024

OpenAI. (2024, Dec.) Incident report: Service down on december 12, 2024. Accessed on 2024-12-18. [Online]. Available: https: //status.openai.com/incidents/ctrsv3lwd797

work page 2024

[6] [5]

Impact of genai on mobile network traffic,

Ericsson, “Impact of genai on mobile network traffic,” Ericsson Mobility Report , Nov. 2024. [Online]. Avail- able: https://www.ericsson.com/en/reports-and-papers/mobility-report/ articles/genai-impact-on-mobile-network-traffic

work page 2024

[7] [6]

Revisiting edge ai: Opportunities and challenges,

T. Meuser, L. Lov ´en, M. Bhuyan, S. G. Patil, S. Dustdar, A. Aral, S. Bayhan, C. Becker, E. De Lara, A. Y . Dinget al., “Revisiting edge ai: Opportunities and challenges,” IEEE Internet Computing, vol. 28, no. 4, pp. 49–59, 2024

work page 2024

[8] [7]

The future of consumer edge-ai computing,

S. Laskaridis, S. I. Venieris, A. Kouris, R. Li, and N. D. Lane, “The future of consumer edge-ai computing,” IEEE Pervasive Computing , 2024

work page 2024