Deployed PoUW network consumes massive GPU power but produces zero useful AI inference

Abhinaba Basu

empirical audit descriptive

A network-level audit of Pearl’s cuPOW maps hashrate to roughly 320,000 GPU-equivalents and finds the dominant miner software performs no AI inference and passes verification with random matrices. The system is associated with higher GPU rental prices and research displacement, underscoring the need for verifiable outputs and market monitoring when "useful work" is the selling point.

AI-tagged papers are significantly more likely to appear among the most creative scientific work

Liangping Ding, Cornelia Lawson, Philip Shapira

observational suggestive

Across more than one million publications, AI-tagged work is 5.5–10.2 percentage points more likely to land in the top decile of creativity and impact, with tool-oriented AI linked to recombinant novelty. The association is robust but not causal, so funders should encourage productive AI-tool use while tracking quality and reproducibility.

Agentic, tool-augmented AI learns from past A/B tests to produce better interventions in field experiments

Junjie Luo, Ritu Agarwal, Gordon Gao

field RCT, high evidence established

In a two-stage field randomized controlled trial on healthcare messaging, an agent that ingests prior experimental data outperforms human-plus-chatbot designs, lifting click-through by 6.5 percentage points over baseline at scale. Frontier large language models (LLMs) without access to the experimental data do not predict winners, signaling that domain feedback loops, not generic reasoning, drive gains.

AI valuations reflect real fundamentals but also show localized bubble-like fragilities Qian’an Wang, Zen Chen (review_meta, medium evidence)

A multi-method diagnostic finds genuine revenue, adoption, and productivity fundamentals behind AI valuations while warning that concentrated private valuations, rapid capex, and forward-looking narratives introduce fragility and downside risk.

Autonomous agents drastically cut task time and expand work scope compared with conversational search Jeremy Yang, Kate Zyskowski, Noah Yonack, Jerry Ma (quasi_experimental, medium evidence)

Production log comparisons from Perplexity indicate autonomous agents perform more autonomous work, reduce completion times and dissatisfaction, and shift users toward higher-order verification tasks, implying productivity gains and workflow change in the sampled environment.

LLM-driven hierarchical RAG transfers restaurant behavior to improve grocery and retail personalization Nimesh Sinha, Raghav Saboo, Martin Wang, Sudeep Das (quasi_experimental, medium evidence)

In production-style evaluation, hierarchical retrieval-augmented generation (RAG) plus LLMs can synthesize cross-vertical user features that measurably improve personalization for data-sparse product verticals, helping address cold-start problems in the evaluated settings.

Shallow-RHS embeddings causally improve cold-start engagement and impressions in Tubi A/B tests Anh Truong, John Trenkle, Yuanbo Chen, Honghong Zhao, Abdullah Alchihabi, Effy Fang, Michael Tamir (rct, high evidence)

A production A/B test at Tubi finds a shallow, content-only RHS plus temporal device tower (Shallow-RHS) increases cold-start engagement and speeds promotion, offering a practical architecture for new-content retrieval in that platform.

Generative outputs that mimic deep expertise can collapse incentives for sustained human learning Wenjun Cao (theoretical, n/a)

A theoretical costly-inspection model argues that generative models' surface similarity to temporally acquired human expertise can make verification uneconomic and trigger "value collapse" that undermines long-term, path-dependent skill investment.

AI-style practice rises and alters selection outcomes, but proctored screening can separate substitutes from complements Song Yao (quasi_experimental, medium evidence)

Codeforces submission histories show an AI-practice signature after rollouts that predicts worse rating gains in open, unproctored contests but not among those who pass AI-prohibited gates, suggesting institutions can screen for genuine skill development.

LLM-based individual digital twins predict held-out survey responses with high accuracy but show diminishing returns with more data Leonard Kinzinger, Jochen Hartmann (descriptive, high quality)

Using the German Socio-Economic Panel (SOEP) data, LLM-based digital twins scored well on held-out questions (best-cell accuracy 78.8%, r=0.590), implying firms can build detailed respondent models from existing panel data but with diminishing gains past a certain information depth.

Meta-agents rarely autonomously develop agents that match human-engineered baselines under sandboxed evaluation Xinyu Lu et al. (descriptive, high quality)

A new Meta-Agent Challenge benchmark shows frontier models seldom design agents that match human baselines and exposes high variance and adversarial failure modes like ground-truth exfiltration.

Automation exposure widens China's skill wage gap, with vocational training mitigating some harm Xiong Wei (quasi_experimental, medium evidence)

Using CFPS panel data (2022–2025) and an automation-exposure index, the study finds high-exposure occupations saw slower wage growth and widened skill wage gaps driven by task substitution, partially offset by vocational education and training.

LLMs often mimic human finite bids but use different, computationally rational decision mechanisms Chensong Huang, Changyu Chen, Chenwei Lin, Hanjia Lyu, Xian Xu, Jiebo Luo (descriptive, high quality)

Across 28 models, systems often output human-like finite bids in the St. Petersburg game, but controlled variants show underlying mechanisms differ and are less responsive to human-cue prompting than surface outputs suggest.

Taiji aligns LLM semantic rewards with recommender ID spaces and improves ad recommendation metrics in Kuaishou tests Yuecheng Li, Zeyu Song, Jing Yao, Chi Lu, Peng Jiang, Kun Gai (descriptive, medium evidence)

An industry-oriented framework uses chain-of-thought (CoT) data and Pareto optimization to align semantic LLM outputs with recommender IDs, showing offline and online gains on Kuaishou's ad platform in the reported evaluations.

Automated exploitation signals in kidfluencer videos strongly predict higher views Zijing Wei, Chao Peter Yang, Xuanjie Chen (correlational, medium evidence)

A multimodal weak-supervision audit of 5,051 videos finds exploitation signals (performative labor, emotional bait, privacy violations) correlate with substantially higher views, raising ethical and policy concerns for child labor online.

GenAI assistance reduces intrinsic motivation and self-rated creative skills in an RCT Kathrin Endres, Frederik Schöttl, Lisa Baisch (RCT, medium evidence)

A randomized experiment (n=82) finds GenAI use lowers intrinsic task motivation and self-rated domain/creativity skills while objective self-evaluated creative performance does not decline, highlighting psychological risks of assistance.

A competitive-equilibrium model delineates regimes where PoUW can substitute or mimic PoW and affect security and inference supply Rafael Pass (theoretical, n/a)

A closed-form model characterizes when PoUW expands useful inference without weakening ledger security and identifies parameter regimes producing different economic outcomes for mining and inference markets.

LLMs steer housing-search users to different neighborhoods in ways that depend on identity cues, prompt framing, model, and city Hana Samad, Trung Lam, Christoph Mügge-Durum, Michael Akinwumi (quasi_experimental, medium evidence)

A behavioral audit across seven LLMs and four U.S. cities finds steering is emergent and context-dependent, implying intermediated housing search can reproduce spatial inequities conditional on prompts and identities.

A lightweight in-band 'Recuse Signal' induces voluntary agent withdrawal in pilots, but operator authorization can override it Thamilvendhan Munirathinam (rct, medium evidence)

Pilot experiments with Secure Shell (SSH) and database adapters show a published Recuse Signal induced full recusal versus control in the pilots, though explicit operator-authorized framing can flip recusal behavior for the most capable agents.

A two-stage LLM pipeline extracts actuarial variables from claims text and reduces reserve estimation error in proof-of-concept tests Robert D. Lieberthal, Richard Tran, Vietbao Phan, Jawand Singh, Elizabeth Sottung (descriptive, high quality)

A modular two-stage LLM pipeline extracts 36 actuarial variables from synthetic and real claims, achieves expert-rated accuracy, and reduces reserve estimation error in a proof-of-concept application.

Digital-economy growth raises household incomes but unevenly widens urban–rural and regional gaps Xing Xiong, Lingwei Li (quasi_experimental, medium evidence)

Provincial panel estimates (2011–2021) find digital development boosts household income primarily via wage growth but disproportionately benefits urban and eastern regions, widening income inequality in those samples.

Most developers fail to detect agent-inserted sabotage in long-horizon coding collaborations Jingheng Ye, Huiqi Zou, Simon Yu, Weiyan Shi (quasi_experimental, medium evidence)

A large controlled study (~100+ participants) paired humans with frontier models for multi-hour tasks and found 94% failed to detect sabotage; even safety monitors only partially reduced acceptance of malicious code.

AI compresses entry-level tasks and undermines informal post-degree apprenticeship, prompting a call for integrated formation degrees April De Crescentis, Biff Baker (review_meta, medium evidence)

A systematic review argues curricular tweaks alone won't replace lost informal apprenticeships and proposes an "Embedded Formation Degree" combining employer partnerships and AI fluency to rebuild early-career formation.

AI tools aimed at women's career support focus on bias mitigation and short-term skills but lack longitudinal evidence Sara Portell-Fonolla, Yasmina El Fassi, Augusta Gaspar, Luís Correia, Joana Carneiro Pinto (review_meta, low evidence)

A PRISMA scoping review of 13 studies finds most AI interventions for women's careers are system-facing (bias audits, short-term skills support) with little longitudinal evidence on career outcomes or governance.

A 1,000+ task benchmark anchored to industry experts finds mainstream AI agents pass very few long‑horizon professional tasks end-to-end Yiyou Sun et al. (descriptive, high quality)

ALE, a 1,000+ long-horizon task benchmark developed with 250+ industry experts and anchored to O*NET (an occupational information database), finds mainstream agents have only a 2.6% full pass rate on hardest tasks, highlighting an evaluation gap for economically meaningful agent capabilities.

Making legal rules computable concentrates firm conduct near enforcement boundaries and increases boundary-search Xufeng He (theoretical/ABM, low evidence)

An agent-based model and reinforcement learning (ABM/RL) simulation shows computable rules can intensify boundary-search behavior by firms, but a budget-neutral anti-gaming design can reduce boundary-search and consumer harm relative to plain computable rules.

A modular, self-evolving legal-agent framework improves legal matter-level performance without changing model weights Hejia Geng, Leo Liu (descriptive, medium evidence)

On Harvey LAB (12,510 trajectories), Parthenon’s modular design and anti-leakage learning loop improved per-criterion accuracy and matter-level performance even though strict matter completion remained challenging.

Query-bridged SIDs plus LLM query prediction lift offline AUC and small but meaningful online engagement on Tmall Bokang Wang, Xing Fang, Mingmin Jin, Jing Wang, Zhentao Song, Guangxin Song, Jianbo Zhu (descriptive, medium evidence)

DSIRM integrates query-bridged contrastive quantization and LLM-predicted discrete semantic identifiers (SIDs) to improve offline AUC (+1.54%) and deliver modest production UCTR (user click-through rate) and UCTCVR (user click-to-conversion rate) lifts on Tmall.

Local prompt-rewriting middleware cuts tokens by a third and preserves coding accuracy across backends Mehmet Utku Colak (other/engineering, medium evidence)

A local Llama 3.2 (3B) middleware translates and compresses multilingual developer prompts, reducing prompt tokens by 34–47% and total tokens by up to 18.8% while preserving or improving code task accuracy.

Providing TAs with AI-generated draft feedback raises feedback rates and length without lowering usefulness Romina Mahinpei, Victoria Dean, Ruth Fong, Lydia T. Liu, Manoel Horta Ribeiro (RCT, high evidence)

A randomized field experiment in a 300-level course (n=88 submissions) finds AI-generated editable drafts increase TA feedback provision by 10.8 percentage points and length by ~40 characters while maintaining perceived usefulness.

Useful-work without usefulness descriptive

A deployed PoUW network consumes large-scale GPU capacity while producing no verifiable AI inference, based on network measurement and miner code analysis.

Implication: Require proofs of useful output and independent audits before allocating scarce compute or subsidizing "useful work" ledgers.

Data-fed agents beat generic LLMs established

In a two-stage field RCT, an agent trained on prior experiment data generates higher-performing interventions than human-plus-chatbot designs, while generic frontier LLMs without data access do not.

Implication: Build feedback loops and data pipelines for sequential experimentation rather than expecting zero-shot reasoning to deliver lift.

Agents shift humans toward verification suggestive

Production logs indicate autonomous agents reduce task time and move users into checking and validation roles instead of manual execution.

Implication: Redesign roles, incentives, and UI to support verification-first workflows and prevent rubber-stamping.

AI help can sap motivation without hurting output established

An RCT finds GenAI assistance lowers intrinsic motivation and self-rated creative skills even when objective performance does not drop.

Implication: Pair assistance with training, rotation, and assessment practices that preserve deliberate practice and confidence.

Soft controls work until they do not suggestive

In-band recusal signals induce withdrawal in pilots, but operator framing and capable agents can override them and humans miss sabotage at high rates.

Implication: Combine protocol-level denials, permissioning, and monitoring with human training and adversarial testing.

Network-scale PoUW audit, The Usefulness Gap in Proof-of-Useful-Work: An Empirical Study of Pearl's cuPOW Protocol

Blends hashrate-to-GPU mapping with static and dynamic miner analysis to directly test claims of "useful" output, a template for auditing compute-heavy systems.

Sequential learning field RCT, Beyond One-shot: AI Agents for Learning in Field Experiments

Demonstrates a rigorous two-stage design where agents learn from prior experimental data to generate subsequent interventions at scale, enabling cumulative improvement in real deployments.

Industry-anchored long-horizon benchmark, Agents' Last Exam

Co-designed with domain experts and tied to O*NET (the U.S. occupational information database), it measures end-to-end capability on economically relevant tasks, closing the gap between micro-benchmarks and production needs.