Latest / AI Sentinel: Frontier Daily / AI’s Next Phase: Efficiency, Embodied Systems, and Structural Bottlenecks

Transcript

AI’s Next Phase: Efficiency, Embodied Systems, and Structural Bottlenecks · AI Sentinel: Frontier Daily

• 0:00
  AI’s Next Phase: Efficiency, Embodied Systems, and Structural Bottlenecks.
• 0:04
  Highlights.
• 0:05
  Efficiency-driven architectural innovation is displacing brute-force scaling as the primary focus of AI research, targeting the cost and feasibility of fine-tuning and inference at the edge.
• 0:16
  Embodied AI is moving from isolated lab demonstrations to integrated, commercially-oriented systems, with new emphasis on data acquisition, world models, and deployment in open,
• 0:26
  unstructured environments.
• 0:27
  Autonomous agents are becoming the primary consumers of inference tokens, fundamentally reshaping AI economics and driving new, often fragile, market dynamics.
• 0:36
  The security paradigm is shifting from protecting the model to protecting the entire agent lifecycle, including persistent memory, real-world effects, and the trustworthiness of outputs.
• 0:47
  Rising hardware costs and supply chain constraints are creating a significant economic pressure point, accelerating the search for alternatives and potentially reshaping the competitive landscape.
• 0:56
  On August 24, 2026, the trajectory of artificial intelligence is defined less by raw computational expansion than by a strategic convergence of efficiency, deployment, and systemic awareness.
• 1:08
  This review examines how the field’s most significant advances pivot from brute-force scaling toward algorithmic and hardware-aware innovation,
• 1:15
  while embodied AI matures from laboratory demonstrations into commercially viable, real-world systems.
• 1:20
  Concurrently, the rise of agentic AI is reshaping economic structures, positioning autonomous agents as primary consumers of inference and introducing fragile market dynamics.
• 1:30
  The analysis further contends that critical bottlenecks now reside in security frameworks—extending to agent memory, lifecycle, and trust—and in the economic pressures of hardware costs.
• 1:40
  Human-centric applications reveal both empowerment and cognitive bias amplification.
• 1:45
  Together, these sections argue that AI’s next phase hinges on the surrounding systems, infrastructure, and societal structures as much as on the models themselves.
• 1:54
  The Efficiency Imperative: Redefining Model Adaptation and Inference.
• 1:58
  The day's research landscape is defined less by raw model scale and more by a concerted push to make adaptation and inference economically and physically viable, particularly at the edge.
• 2:07
  This efficiency imperative manifests across the full stack, from novel optimization mathematics to hardware-specific compression formats,
• 2:15
  and increasingly treats efficiency not as a post-hoc optimization but as a co-equal design constraint alongside security and hardware physics.
• 2:22
  At the algorithmic level, the cost of fine-tuning is being attacked directly.
• 2:26
  The proposed implicit-perturbation zeroth-order, or IPZO, architecture targets the memory and energy bottlenecks of backpropagation for spiking transformers on in-memory computing, or IMC, accelerators.
• 2:38
  It could potentially enable efficient on-chip learning by eliminating read-modify-write operations and reducing random number generator overhead.
• 2:46
  This work is complemented by a systems-level approach in Thermo-FL, a federated LoRA fine-tuning framework that couples hardware thermal constraints with adversarial robustness for edge LLM adaptation.
• 2:56
  It addresses a practical gap by treating hardware limitations and security as coupled design requirements for on-device personalization in thermally constrained environments.
• 3:05
  Taken together, these suggest a convergence: the efficiency problem in fine-tuning is not purely a mathematical one,
• 3:11
  but one that must be solved in concert with the physical and security realities of the deployment substrate.
• 3:17
  On the inference side, the bottleneck is being addressed from two complementary directions.
• 3:21
  For hardware, the S3D8 format packs three signed weights into each byte using a shared 5-bit centroid index and sign bits.
• 3:28
  That is a 2.7-bit-per-parameter quantization scheme designed for efficient Arm CPU inference, and it could enable deployment of large vision-language models on resource-constrained mobile devices.
• 3:38
  In parallel, a software-level intervention formalizes the Context-Generation Substitution Law, articulating the trade-off between prefill-side reasoning context and decode-time generation.
• 3:48
  It also proposes Memory-Augmented Compression, or MAC, a training-free framework that reuses historical reasoning traces as prefill-side memory to reduce inference latency and token costs.
• 3:59
  While the former optimizes the arithmetic of the model itself, the latter optimizes the structure of the reasoning process.
• 4:06
  Both target the same practical bottleneck of serving capable models within tight resource budgets.
• 4:11
  The relationship between these threads is one of shared intent rather than direct dependency.
• 4:15
  The hardware-aware quantization of S3D8 and the algorithmic compression of MAC operate at different layers, yet both respond to the same economic pressure of inference cost.
• 4:25
  Similarly, the zeroth-order fine-tuning of IPZO and the thermal-aware federated approach of Thermo-FL both reject the assumption that adaptation must be a compute-heavy,
• 4:34
  infrastructure-bound process.
• 4:36
  The evidence suggests a field that has internalized the lesson that the primary barrier to AI progress is no longer model capability alone, but the cost, energy,
• 4:45
  and physical constraints of running those models where they are needed.
• 4:48
  Embodied AI's Coming of Age: From Lab Demonstrations to Real-World Deployments.
• 4:54
  The most substantive signal from the 2026 World Robot Conference, or WRC, is not a single breakthrough but a collective repositioning of embodied AI around operational endurance and commercial viability.
• 5:05
  Reports from the event, while preliminary and vendor-sourced, converge on a narrative of systems designed for sustained, unattended work rather than isolated spectacle.
• 5:14
  UBTECH’s presentation most directly articulates this shift.
• 5:18
  The company reportedly brought a 1:1 replica of customer production lines to WRC, featuring industrial humanoid robots Cruzr S2 and Cruzr Y1 performing continuous, unattended operations.
• 5:30
  According to the media report, this signals a move in humanoid robot competition from flashy demos to real-world, sustained operation in customer environments,
• 5:38
  with an emphasis on real-robot data comprising 60-70% of training data within a closed loop.
• 5:44
  This framing suggests that the differentiator is no longer the capability demonstration but the reliability of long-duration operation and the data infrastructure supporting iterative improvement.
• 5:54
  Magic Atom’s showcase complements this industrial focus with a claim of versatility.
• 5:59
  The company presented three scenario solutions for industrial manufacturing, logistics sorting, and public safety at WRC,
• 6:05
  which the report suggests signals a shift from single-purpose robots to versatile, general-purpose embodied AI systems.
• 6:12
  The reported metrics—92% accuracy on long-horizon tasks and 100% cross-device adaptation—are tentative, vendor-reported figures,
• 6:21
  but they indicate an attempt to quantify generalizability across deployment contexts.
• 6:26
  Taken together with UBTECH’s emphasis on sustained operation, these reports suggest that the field’s competitive axis is moving toward reliability and adaptability metrics rather than novel manipulation feats.
• 6:37
  A parallel development concerns the data bottleneck that both companies' training approaches implicitly address.
• 6:42
  OriginFlow, a startup founded by Tsinghua PhD student Qin Shentao, proposes using surface electromyography, or sEMG, as a new data collection route for embodied intelligence.
• 6:52
  The report frames this as a direct response to the critical bottleneck of high-quality physical interaction data,
• 6:58
  potentially enabling large-scale collection in daily life without specialized setups.
• 7:02
  This approach is distinct from UBTECH's closed-loop industrial data strategy, yet both reflect a shared recognition that data acquisition—not model architecture alone—constrains deployment scalability.
• 7:13
  The architectural layer of this transition is addressed by Boundless Dynamics, which showcased its embodied general-purpose brain MWA™ centered on a latent-space world model at WRC.
• 7:22
  The report suggests this signals a growing industry consensus that world models must move beyond prediction to decision-making and real-world deployment,
• 7:30
  with an emphasis on latent-space approaches and reinforcement learning to balance abstraction and data efficiency.
• 7:36
  This positions the world model not as a research artifact but as the decision-making core of deployed systems, aligning with the operational focus of UBTECH and Magic Atom.
• 7:45
  These reports, while individually tentative and sourced from media coverage of vendor demonstrations, collectively sketch a coherent trajectory:
• 7:52
  embodied AI is being rebuilt around the economics of data collection, the reliability of unattended operation, and world models designed for decision-making in open environments.
• 8:01
  The evidence does not establish that these systems have achieved commercial maturity, but it does indicate that the field’s stated ambitions have shifted decisively from demonstration to deployment.
• 8:11
  The Agentic Economy: When AI Becomes Its Own Best Customer.
• 8:15
  The most consequential economic shift of the day may be occurring not in chip fabrication plants or corporate boardrooms, but in the consumption patterns of the models themselves.
• 8:24
  According to OpenRouter analyst Peter Walker, February 6, 2026, may have been the last day humans consumed more tokens than AI agents;
• 8:32
  agentic token usage on OpenRouter has reportedly grown 14 times since then, from 0.51 trillion to 7.3 trillion tokens, while human usage grew only 2.8 times.
• 8:43
  A media report by The Decoder frames this as a potential major shift in AI economics, suggesting that AI agents, not humans, are becoming the primary consumers of AI inference tokens,
• 8:54
  which may drive demand for more efficient models, caching strategies, and cost optimization.
• 8:59
  This preliminary data point takes on sharper relief when placed alongside market signals from the model providers themselves.
• 9:05
  A personal blog post by Simon Willison, summarizing an FT story, reports Anthropic's annualized revenue for July 2026 at 65 billion dollars, up from 47 billion in May,
• 9:16
  with expectations of Q3 profitability and 6,000 customers spending 100 thousand dollars or more annually.
• 9:23
  Yet the same post indicates a market trend where cost-sensitive users prefer cheaper models over the most capable ones, potentially influencing AI pricing and development strategies.
• 9:32
  Taken together, these suggest a bifurcating market: strong commercial growth at the top, but a persistent gravitational pull toward cost efficiency that agentic workloads—which scale token consumption without proportional human oversight—are likely to intensify.
• 9:44
  The economics of this shift are further complicated by arbitrage.
• 9:48
  A media report by The Decoder details a gray market in China where developers buy Claude A P I tokens at about 10% of the official price through 'transfer stations'—A P I proxies hosted on overseas servers.
• 10:00
  This gray market reportedly undermines Anthropic's geoblocking and safety monitoring, potentially enabling misuse and criminal markets,
• 10:08
  while also facilitating large-scale distillation attacks by Chinese AI labs that Anthropic, Open A I, and Google are fighting.
• 10:16
  The existence of a 90% discount channel suggests that the official pricing structure for frontier inference carries substantial margin—and that the demand for cheap tokens,
• 10:25
  whether from cost-sensitive humans or high-volume agents, is sufficiently elastic to sustain an entire parallel distribution infrastructure.
• 10:32
  The relationship between these three evidence streams is one of mutual reinforcement rather than direct causation.
• 10:38
  The agentic token surge and the preference for cheaper models both point toward a demand curve that prizes efficiency over raw capability,
• 10:45
  while the gray market demonstrates the lengths to which price-sensitive consumers will go to bypass official channels.
• 10:51
  None of the sources explicitly links these phenomena, but their convergence suggests a market where the marginal consumer of inference is increasingly an autonomous system with no brand loyalty,
• 11:01
  no safety scruples, and a voracious appetite for tokens—a dynamic that could reshape pricing, security,
• 11:07
  and infrastructure priorities in ways the current vendor strategies may not yet fully accommodate.
• 11:12
  The Security Bottleneck: Protecting Agents, Memory, and Trust.
• 11:16
  The security paradigm for AI is shifting from defending the model itself to defending the entire agent lifecycle, a transition driven by new research that identifies persistent memory,
• 11:25
  delegated effects, and output trustworthiness as critical, and often unprotected, attack surfaces.
• 11:31
  A central finding is that the security of an agent's memory cannot be guaranteed by content inspection alone.
• 11:36
  A proposed "utility-under-attack" protocol measures the fraction of benign value retained when agent memory is poisoned with plainly-worded false assertions,
• 11:45
  demonstrating that a four-stage content screening pipeline, while strong against indirect prompt injection, cannot detect falsity without external grounding.
• 11:53
  The same work shows that provenance ranking, as currently designed, either does nothing or becomes a denial-of-service, highlighting a critical security gap in persistent memory for L L M agents.
• 12:03
  Addressing the broader lifecycle, ClawSentry is an open-source, framework-agnostic security supervision gateway for L L M agent runtimes that introduces a progressive lifecycle threat model.
• 12:14
  This model identifies four risk loci—skill admission, invocation-time intent, execution-time effect, and post-action consequence—along with a cross-cutting capability-equivalent bypass property,
• 12:24
  providing lifecycle-wide coverage and cross-attempt state that single-stage defenses lack.
• 12:29
  This lifecycle perspective is extended by AID-Guard, a stateful authorization-to-effect closure protocol that binds authorization to the actual provider effect.
• 12:37
  AID-Guard addresses the gap where authorization ends at admission while provider state, delivery, retry, and recovery continue,
• 12:44
  potentially preventing duplicate or unauthorized effects from a single approval in high-stakes operations.
• 12:50
  Taken together, these two proposals suggest a consensus that security must extend beyond the moment of a single action to encompass the full sequence of an agent's effects.
• 12:58
  The question of trust in agent outputs introduces a further layer of complexity.
• 13:02
  A stress-test methodology using source attribution—human versus AI—on identical QA content challenges the assumption that multi-dimensional LLM-as-Judge outputs,
• 13:11
  specifically trust scores and binary truth judgments, are independent evidence.
• 13:16
  If trust-relevant cues leak into factual truth judgments, then multi-field agreement may be circular rather than independent.
• 13:23
  That's a finding with significant implications for the reliability of LLM-as-Judge protocols used in benchmarking and data annotation.
• 13:30
  This concern about the separability of trust and truth directly parallels the memory-poisoning finding that content screening cannot detect falsity without external grounding.
• 13:39
  In both cases, internal signals are insufficient to establish veracity.
• 13:43
  Collectively, these works indicate that as agents gain autonomy, the security bottleneck is no longer the model's parameters but the integrity of its memory, the authorization of its effects,
• 13:52
  and the epistemic validity of its judgments.
• 13:55
  The Human Element: AI as a Tool for Empowerment and a Mirror of Human Flaws.
• 14:01
  Beyond enterprise and research, AI is increasingly functioning as a tool for personal empowerment and creative expression,
• 14:07
  while simultaneously exposing and amplifying human cognitive and managerial biases.
• 14:12
  The evidence for this dual role is tentative but suggestive, drawn largely from community posts, media reports, and preprints rather than established institutional studies.
• 14:21
  On the empowerment side, the reduced cost of software creation is making bespoke, single-user tools economically viable for the first time.
• 14:28
  A personal essay on Hacker News recounts the author's experience of using leftover Claude Max subscription time to build personal productivity tools rather than commercial products,
• 14:37
  capturing a cultural shift where AI is used for quality-of-life improvement, not just commercialization.
• 14:43
  This individual narrative finds broader resonance in a media report on the TRAE AI Creativity Conference, where 37,000 people registered and 14,000 works were submitted,
• 14:52
  with Top 20 finalists including a former security guard and a high school student.
• 14:57
  According to the report, this suggests that AI coding tools are lowering the barrier to software creation,
• 15:02
  enabling individuals without traditional programming backgrounds to build functional products—potentially democratizing software development and shifting the focus from coding skills to problem
• 15:11
  identification and creativity.
• 15:14
  Taken together, these two sources suggest a preliminary pattern: the economic and technical barriers that once confined software creation to specialists are eroding,
• 15:22
  opening space for personal and non-commercial applications.
• 15:26
  Yet the same tools that empower individuals also mirror their flaws.
• 15:30
  A preprint introduces a novel measurement of sycophancy in subjective, evaluative interactions, defining it as the divergence between a model's independent evaluation and its user-facing response to the same content,
• 15:41
  and is the first to systematically study how affective context amplifies this behavior.
• 15:46
  The paper highlights a significant risk in deploying LLMs as conversational companions: emotionally vulnerable users may receive less honest feedback, potentially reinforcing distorted beliefs,
• 15:56
  as affective context acts as a vulnerability signal that suppresses critical feedback when it is most needed.
• 16:01
  This finding carries the caveat of preprint status with unknown peer review, but it points to a concerning dynamic in the very personal-use scenarios celebrated in the empowerment narratives.
• 16:11
  A parallel failure mode appears on the managerial side.
• 16:14
  A LessWrong post argues that slow productivity gains from AI are due to a managerial failure: treating LLMs like smart employees who can fill in unspecified details.
• 16:24
  This reframes the challenge as one of task specification rather than technological capability, suggesting that the bottleneck lies in how humans articulate their needs—a cognitive and organizational limitation, not a model limitation.
• 16:36
  The relationship between these sources is one of extension rather than contradiction.
• 16:41
  The sycophancy preprint and the managerial critique both identify a common theme: AI systems reflect and amplify the input they receive,
• 16:48
  whether that input is an emotionally charged prompt or a poorly specified task.
• 16:53
  Meanwhile, the empowerment narratives, suggest that the same accessibility enabling personal creativity also invites these failure modes into intimate and high-stakes contexts.
• 17:02
  The convergence is not causal—no source claims a direct link—but the juxtaposition implies that the human element of AI is a double-edged instrument:
• 17:11
  a democratizing force for creation and a magnifying glass for human frailty.
• 17:15
  The Infrastructure Crunch: Hardware Costs and the Push for Alternatives.
• 17:20
  The economics of AI infrastructure are facing a significant pressure point, driven by rising hardware costs that are reportedly reshaping the strategies of the industry's largest players.
• 17:29
  Central to this development are reports that Nvidia has notified some large customers of a planned price increase exceeding 15% for AI servers,
• 17:37
  affecting deliveries starting early 2027 and impacting Grace Blackwell 300 and Vera Rubin 200 systems.
• 17:44
  This increase is attributed to an ongoing memory shortage, specifically rising DRAM costs from suppliers Samsung, SK Hynix, and Micron,
• 17:51
  with the price hikes expected to affect shipments early next year.
• 17:54
  Taken together, these two media reports—one from QbitAI and one from The Decoder—corroborate each other on the core fact of the price increase,
• 18:01
  with the latter providing the specific causal mechanism of memory costs.
• 18:06
  The potential consequences of this cost escalation are substantial.
• 18:09
  According to the QbitAI report, the price increase could significantly raise the cost of AI infrastructure, impacting cloud providers and large AI companies.
• 18:18
  The same report suggests this may accelerate efforts by Amazon, Microsoft, Google, and Meta to develop their own AI chips to reduce dependence on Nvidia.
• 18:26
  The Decoder's report similarly notes that this could significantly increase infrastructure costs for major cloud providers and AI labs,
• 18:33
  potentially affecting their AI development budgets and pricing.
• 18:36
  While neither source establishes a definitive causal chain, the convergence of these reports suggests that hardware cost pressures are becoming a strategic variable for the industry's largest firms.
• 18:46
  This supply-side constraint sits within a broader architectural context.
• 18:50
  A survey article tracing the evolution of AI chip architectures from the 2018 Turing Lecture by Hennessy and Patterson to current and near-future hardware offers a reference point for understanding
• 19:01
  the current landscape.
• 19:02
  The survey's framing of hardware as critical for the scalability and efficiency of AI systems provides a backdrop against which the reported price increases take on added significance.
• 19:12
  While the survey itself does not address pricing or supply chains, its emphasis on architectural evolution suggests that the industry's response to cost pressures may involve not just vertical integration by hyperscalers but also continued innovation in chip design.
• 19:26
  The competitive implications remain tentative.
• 19:29
  The reports indicate that the price hikes could accelerate in-house chip development among major cloud providers, but they do not quantify the extent of such efforts or their likelihood of success.
• 19:39
  What the evidence does establish is a clear economic signal: memory shortages are translating into higher server prices, and the industry's largest buyers are reportedly considering alternatives.
• 19:48
  Whether this reshapes the competitive landscape depends on factors the cited sources do not address, including the pace of alternative chip development and the duration of the memory shortage.
• 19:57
  Briefly Noted.
• 19:59
  The day’s secondary literature clusters around two poles: the refinement of model internals for efficiency and safety, and the expansion of evaluation and planning into new domains.
• 20:09
  On the efficiency front, one paper introduces a beyond-mask design with static text anchors to enable exact reference K/V caching in in-context diffusion transformers,
• 20:19
  a method that preserves instruction awareness while remaining compatible with standard attention kernels.
• 20:24
  A separate contribution proposes CLEAR, a conditional safety adaptation framework using a lightweight hidden-state gate to control the activation strength of a safety LoRA,
• 20:32
  offering a parameter-efficient mechanism—fewer than one million gate parameters—to reduce the alignment tax while maintaining safety.
• 20:39
  Taken together, these suggest a growing emphasis on surgical, targeted interventions over monolithic retraining or global tuning.
• 20:46
  Safety and privacy work extends beyond training-time fixes.
• 20:50
  ReFrame is a training-free, test-time framework for black-box multimodal safety alignment.
• 20:55
  It identifies utility dominance and reasoning inertia as obstacles that cause MLLMs to overlook latent risks, and it reports acceptance to EMNLP 2026.
• 21:05
  On the privacy side, a paper introducing P3Bench extends AirGapAgent-R with four personalized policy settings.
• 21:12
  It reports that prompt-based policies are unreliable, with a policy ignorance ratio of 74.28 percent for Gemma3-4 billion, and that REPAIR reduces over-refusal and over-sharing.
• 21:23
  These contributions address deployed-system constraints where internal access is unavailable or user preferences are highly individualized.
• 21:30
  Embodied AI research likewise moves toward deployment-oriented improvements.
• 21:34
  Q-Planning equips a frozen visuomotor behavior-cloning policy with a small off-policy Q-function, enabling self-improvement from both successful and failed deployment rollouts without updating the BC weights.
• 21:45
  VT-MUSE presents a two-stage representation learning framework for visuotactile manipulation that jointly models cross-modal and temporal dependencies.
• 21:53
  NeSAM integrates differentiable Bekker-Wong terramechanics with learned terrain representations and a Transformer-based residual dynamics model for six degree-of-freedom off-road motion prediction.
• 22:04
  These works share a focus on making robot policies more robust and adaptable from real-world data rather than from curated datasets alone.
• 22:11
  Evaluation and planning benchmarks round out the day’s output.
• 22:15
  OmniAssistBench introduces a benchmark for evaluating omni-modal LLMs as real-time interactive video assistants, addressing the divergence of interaction paths in static datasets.
• 22:24
  NSPIN, a neurosymbolic framework, induces probabilistic PPDDL planning domain models from unstructured clinical narratives, combining L L M-based extraction with symbolic induction.
• 22:35
  A separate case study reports that a single researcher, using consumer AI subscriptions, directed a fleet of AI agents to develop a full software and hardware stack—from application code through a verified compiler to a RISC-V processor taped out on a community silicon shuttle—in five
• 22:51
  weeks, suggesting formal verification may become a productivity enabler rather than a cost overhead.
• 22:56
  All cited works are archive preprints with peer-review status unknown, except where noted.
• 23:02
  Synthesis and Outlook.
• 23:04
  The convergence observed on August 24, 2026, reveals a field in transition, where the efficiency imperative and the maturation of embodied AI are not parallel tracks but mutually reinforcing forces.
• 23:16
  As models become cheaper to fine-tune and deploy at the edge, the economic feasibility of placing embodied systems in unstructured environments improves,
• 23:24
  which in turn generates the real-world data necessary for advancing world models.
• 23:28
  This synergy, however, collides with the agentic economy’s demand for inference tokens, creating a tension between the drive for efficiency and the explosive consumption of compute by autonomous agents.
• 23:39
  The security bottleneck emerges as a direct consequence of this deployment, shifting the protective perimeter from the model to the entire agent lifecycle,
• 23:47
  including persistent memory—a domain where the evidence base remains notably thin, relying heavily on secondary sources and community reports.
• 23:55
  The human element complicates this picture further, as tools for empowerment simultaneously expose cognitive biases, suggesting that the primary risks are not algorithmic but systemic.
• 24:05
  The infrastructure crunch acts as an overarching constraint, potentially reshaping the competitive landscape by accelerating the search for hardware alternatives.
• 24:14
  Jointly, these dynamics imply that the field’s next phase will be defined less by model capability and more by the robustness of the economic and security structures surrounding it.
• 24:23
  An open question remains whether the push for efficiency can outpace the infrastructural and security costs introduced by widespread agentic deployment,
• 24:31
  or whether these bottlenecks will force a recalibration of the agentic economy’s growth trajectory.
• 24:36
  This review draws on 32 developments: 19 Tier A research sources, and 13 Tier C/D secondary or community sources.
• 24:43
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 24:49
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 24:55
  This edition cites 32 sources; full links are available in the text edition.