Induced Friction Between AI Agents: A Search for Disruptive Solutions

Let AIs compete against each other. AI collaboration may be overrated, or rather, consensus among AIs may kill innovation.

If we're looking for disruptive and innovative AI solutions to today's problems, the biggest obstacle isn't a lack of analytical capacity, but rather that the most likely responses tend to be the most predictable, a product of their average training, and this is what any consultant would say in a first meeting.

To reach another level, we need something different. Instead of AI debate environments, such as collaborative (1,2) or adversarial debates (3,4), whose results can reveal vulnerabilities and convergence traps, as have already been identified (5), a working environment is proposed where multiple AIs run an obstacle relay race where the obstacles are sequential rejections driven by non-judgmental AI arbiter that does not evaluate quality or creativity, but only rejects outputs and measure semantic drift between successive responses. We call this methodology Sequential Stress Optimization (SSO).

The core idea of SSO is simple: instead of promoting debate between agents, which can end in predictable consensus or unproductive, token-consuming loops (5), the proposal is a sequential relay framework with obstacles. Each AI agent receives the previous result from another agent with a single instruction: beat it by divergence. This is a friction scheme built upon successive, non-qualitative rejections, executed across two distinct phases: Exploring and Polishing.

How does it work in practice?

Let's take a real engineering problem: the bottleneck in the submittals review process.

We pose the problem to the first AI agent. Its response is neither corrected nor discussed; it's picked up by the key actor in this scheme: the AI Arbiter.

The AI ​​Arbiter does not generate ideas or judge content qualitatively. Its rule is procedural: in the first phase, it rejects a percentage of the proposals submitted by AIs, let's say 70%, regardless of what each one says.

The Rejection Threshold:

The 70% rejection rate is an arbitrary benchmark set here to illustrate the procedure. In practice, this threshold is a parameter that should be adjusted based on computational cost and solution quality. It can be adjusted dynamically by the user or adaptively by the AI Arbiter (user-selectable option), if the cost outweighs the benefit of optimization (e.g., when the semantic differential decreases). A detailed analysis of the fixed and adaptive rejection rates and their results is beyond the scope of this publication, as validating them for various scenarios requires a significant compute time and token expenditure; therefore, a detailed statistical analysis is reserved for future publication.

Convergence criteria:

The first response (e.g., "hire more reviewers") is discarded. Once this result is obtained, the AI Arbiter passes it on to a second AI with an explicit instruction: "This was already rejected as obvious; find the root flaw and propose something different".

From this succession of responses passed between AI agents, it is pushed out of its statistical comfort zone, allowing potentially innovative solutions to emerge. For example: replacing sequential review with parallel review across disciplines. In successive iterations, increasingly better and more radical proposals are reached. This is no longer optimization; it's a model shift.

When to stop?

Chained rejection cannot continue indefinitely. The first stopping signal occurs when, after reaching the established rejection rate for the first phase, in this case, 70%, the semantic differential between consecutive solutions approaches a minimum. The AI Arbiter takes this signal to mean that the solution space in that direction has been exhausted. If this semantic minimum proves difficult to reach, the AI Arbiter grants an additional 5% margin of rejections threshold for this phase and treats the identified solution as optimal.

Once the first phase, Exploring, concludes, the AI Arbiter activates the second phase: Polishing. Here the goal is to refine the solution from the first phase; the solution obtained is subjected to a new sequence of rejections, now capped between 20% and 30% of everything generated. After that, the AI Arbiter again identifies the minimum semantic dispersion among the results and extracts the optimal solution from the system. This optimal, tentatively disruptive solution is what the AI Arbiter identifies as the final result.

SSO Flow Chart:

Risk and Operational Feasibility Considerations.

Sequential Stress Optimization carries practical risks. Rejecting 70% of the most probable responses as referenced above pushes agents into low-probability territory, where a wrong answer or a hallucination can be argued with full logical consistency. The result can look entirely correct without being accurate. AI Agents may also learn to disguise familiar answers in complex language to pass the Arbiter's filter without changing the underlying idea, so the first phase rejection output will need human follow-up to check substance.

The two failure modes mentioned above, hallucinations and disguised responses, point to the same gap: the Arbiter measures the semantic distance, not the meaning. To compensate for this, a new actor is added to the SSO scheme (see modified flow chart below): the Dynamic Divergence Refiner (DDR) agent. The Dynamic Divergence Refiner (DDR) performs a parallel qualitative check on each sequence. While the Arbiter measures semantic distance, the DDR verifies whether the new response remains within the problem domain and whether its divergence is substantive rather than merely lexical. When the DDR detects disguised repetition or a logically coherent but impractical answer, it vetoes the handoff and restarts the next agent from the last valid state. If drift persists across consecutive rejections, the DDR rewrites the next agent’s input by explicitly naming the repeated concept to exclude and the neglected variable to pursue.

Modified SSO Flow Chart (SSO+DDR):

The Cost Side:

Repeated rejection cycles in SSO imply repeated token generation, which increases latency, token usage, and computational cost. This is even more pronounced in the SSO-DDR scenario, where these costs increase even further. For this reason, SSO is not considered a default method. It is suitable for cases where the value of a better response justifies the additional cost: strategic decisions, not routine tasks. The SSO+DDR scenario would be reserved exclusively for the pursuit of innovation.

Conclusion:

The SSO as a problem-solving approach is not a multi-agent friction system in the traditional sense. It is not based on debate, consensus building, or adversarial interaction between agents It's an obstacle relay race involving multiple AI agents with different analytical approaches. In this race, induced friction arises from the predisposition of an Arbiter to reject the solutions submitted sequentially by each agent. Therein lies the possibility that by promoting difference, something unexpected may emerge, such as a more creative or disruptive solution.

The diversity in agent training provides the necessary variance in the explored sequential solutions. While SSO is architecturally valid with a single agent with different activated modes, its effectiveness decreases with the homogeneity of the agents involved.

On the other hand, the SSO approach identifies two likely failure modes under friction: hallucinations and disguised responses. Their occurrence depends on the complexity of the problem, and to address them, a modified SSO is conceived, with a more complex architecture that includes the Dynamic Divergence Refiner agent (DDR). This agent assists the arbiter by performing qualitative checks that the arbiter does not. The SSO is affected by the computational cost of sequential rejection. The addition of the DDR agent further increases this computational cost, making it significantly more expensive. Therefore, SSO+DDR is recommended for special cases where the goal is outside-the-box solutions.

The SSO and SS+DDR scenarios are theoretical proposals that require validation in terms of both results and costs, the latter being the main constraint. It is estimated that these tests will begin with two or three agents.

 

 References:

 

1. J. Anglen. “AI Agent Debate Systems: How Multi-Agent Collaboration Improves Decision Making”. 2026. RUH.AI.

2. E. Schepis “Democratic Multi-Agent AI: Debate-Based Consensus”. 2025. Medium.

3. J. Mishra. “When Your AI Needs an Enemy”. 2026. Medium

4. Yao et al. “Peacemaker or Troublemaker”. 2025. ArXiv.

5. Zhang et al. “Stop Overvaluing Multi-Agent Debate”, 2025. ArXiv.

Transparency Statement: The author acknowledges the use of Artificial Intelligence (AI) as an assistive tool during the research, data structuring, and content optimization process. The core concept, final review, and critical analysis remain the sole responsibility of the author.

Spanish version: https://ingconcurrente.blogspot.com/2026/06/la-friccion-inducida-entre-agentes-ia.html

The Vigilant PM and PMO: Maintaining Human Leadership and Corporate Sovereignty in the AI Era

Why AI as a strategic co-pilot cannot replace the frontline context of the Project Manager, and how the PMO must build a framework to protect both data and human teams from cold optimization.

The project management landscape is currently flooded with narratives positioning artificial intelligence as an inevitable replacement for the Project Manager, or at least a substitute for many of their key activities. While maximizing personal productivity through automated scheduling and reporting is valuable, it represents a superficial understanding of project dynamics.

The truth is that deep AI integration introduces risks that an automated system can never resolve: the threat of cold management and the long-term danger of vendor dependency. Surviving this transition requires a powerful reclamation of the PM's unique value as the frontline human filter, backed by a Project Management Office (PMO) that serves as the ultimate assurance layer for corporate sovereignty.

1. The PM as the Frontline Filter: Context Over Optimization

In an AI-assisted environment, the PM's primary responsibility shifts from data management to contextual validation. They stand as the first line of defense between raw AI output and the project team.

The PM Operational Filters:

Vetting Automated Triggers: AI models excel at tracking patterns, such as task completion rates or email response times. However, when the AI detects a deviation and flags a performance or delay risk, it operates without empathy or situational awareness. The PM must catch these alerts before they impact the team, evaluating the human reality behind the metrics.

The Context Filter: If the AI flags a resource as underperforming, the PM applies real-world understanding. Is the team member tackling an undocumented technical hurdle? Are they mentoring a junior developer? By processing the data through a filter of human insight, PM prevents automated metrics from being used to unfairly penalize personnel.

Guarding Team Dynamics: This active filtering is what prevents the PM from sliding into cold optimization, a system that treats people as variables to be adjusted rather than individuals to be understood. The PM uses AI insights as operational indicators to prompt a conversation, rather than absolute verdicts to enforce compliance.

Shielding AI Noise and Historical Bias: As AI generates continuous predictive scenarios and updates, the PM acts as a shield against alert fatigue caused by constant analytical noise. Furthermore, since AI models are trained on historical data that may contain past organizational dysfunction, such as overly aggressive baselines, the PM must filter out these historical references to prevent any ongoing project from being penalized by toxic benchmarks.

The attitude described above demands new soft skills from the project manager: the ability to effectively question data from an authoritative source and to exercise healthy skepticism. It's about reading AI alerts as clues to start a discussion, not as conclusions to execute. The PM applies judgment before action.

2. The PMO as the Assurance Layer: Strategic Governance

While the PMs handle day-to-day human interactions, they may lack the organizational authority to question the recommendations of an internally institutionalized AI or how it manages the data it uses. The PMO must step in to institutionalize and protect the PM's human-centered boundaries.

The PMO Strategic Safeguards:

Establishing Boundaries for Metrics: The PMO creates the formal rules of engagement for technology within the enterprise. It dictates exactly how data harvested by AI can be used, explicitly banning automated metrics from being connected to formal performance evaluations or HR actions without human review.

Backing Human Authority: In data-driven organizations, there is a dangerous tendency for upper management to treat automated forecasts as infallible truth. The PMO acts as an institutional buffer, validating the PM's right to override the AI projections based on qualitative team assessments.

Systemic Compliance Audits: The PMO monitors the overall health of the project ecosystem. If data shows a spike in turnover or friction in a specific department, the PMO investigates whether that project's leadership is relying too heavily on automated recommendations without applying the necessary human filter.

3. Securing Corporate Sovereignty: The Multi-Vendor Blueprint.

True human-centered management cannot exist if the organization itself is trapped by an AI vendor. The PMO must implement an enterprise architecture that treats commercial AI models as interchangeable utilities rather than indispensable partners.

Enterprise architecture for AI:

The Threat of Captive Data: When an enterprise feeds its historical project records, proprietary risk logs, and unique operational methodologies into a closed, single-vendor AI cloud, it risks losing its operational independence. If that AI vendor increases prices significantly or changes their compliance terms, the organization cannot easily walk away because its institutional intelligence is locked in the vendor's ecosystem.

The Vendor-Agnostic AI Layer as a Shield: The Sovereign PMO mandates a strict separation between corporate data and the AI models that process it. By implementing a Vendor-Agnostic AI Layer, such as an Agnostic RAG architecture, enterprise data remains securely isolated internally, routed through independent middleware rather than vendor sites.

The Operational Kill Switch: A primary AI vendor can fail the organization in different ways: a prolonged outage, a sudden change in terms of service, or a security breach. Regardless of the cause, the PMO holds the capability to disconnect that AI model and activate an alternative commercial or open-source model running on its own infrastructure. The frontline PM and their teams continue working without operational interruption, insulated from external vendor AI volatility.

Summary: Minimizing AI risk

This layered approach transforms human-centered management from a corporate aspiration into a predictable, structured governance model.

By defining the PM as the operational filter and the PMO as the strategic guardian, the AI-powered organization achieves the ideal balance. AI becomes a powerful engine for predictive and strategic analytics, integrated into a people-controlled framework. In this strategic AI integration scheme, the PM provides the context, and the PMO ensures the system is secure. Sovereignty over data, people, and work remains where it belongs: within the organization.

Transparency Statement: The author acknowledges the use of Artificial Intelligence (AI) as an assistive tool during the research, data structuring, and content optimization process. The core concept, final review, and critical analysis remain the sole responsibility of the author.

Spanish version: https://ingconcurrente.blogspot.com/2026/06/el-pm-y-la-pmo-vigilantes-como-mantener.html

When AI is a Strategic Copilot: The New Role of the Sovereign PMO

Towards a management model with integrated AI and a PMO acting as vigilant support to guarantee security, avoid vendor dependency, and promote e long-term responsible governance of AI in project management

Project management is changing. It is no longer just about using artificial intelligence to automate tasks. Today, AI can serve as a strategic ally in project management: reading behavioral signals in teams, anticipating tension or burnout before it escalates, or identifying friction in client relationships before it becomes a complaint.

These benefits are clear. However, this advanced use creates dependency. If a project manager makes decisions based on an external AI provider, the organization assumes real risks: unilateral price changes, modifications to terms of service, or supply interruptions due to technical failures, commercial, or regulatory decisions in the provider's jurisdiction.

It is precisely this strategic use of AI that demands a new role for the Project Management Office (PMO). The PMO must act as a Vigilant Support: responsible for ensuring operational continuity, proper governance, corporate security, and data sovereignty against technology providers. This role would then be called: The Vigilant PMO: AI-Sovereign Support for Project Governance.

To achieve this, the PMO relies on three practical pillars:

1. Supervision of predictive management centered on people. 

The premise here is that AI can detect, for example, behavioral anomalies in teams or client relationships, and unseen operational risks. However, the responses and corrective actions must be human, not AI-driven. Instead of waiting for a problem to escalate into a conflict or a team member's departure, the project manager can use AI data to identify tensions before they worsen. Here, the vigilant PMO acts as the guarantor that this human-centric approach remains the standard practice. 

Specifically, if AI detects changes in communication patterns or unusual delays in tasks, the project manager intervenes with concrete actions: adjusting workloads, improving cross-functional communication, or redefining scopes. Technology aims to protect the team and the client, not to control them. The PMO periodically verifies that this process is being followed.

2. Ensuring technological independence.

Relying on a single provider is a risk the PMO must actively manage. The solution is to implement an architecture where the data processed by AI for project management is separated from the AI model processing it. 

In practice, this means critical project intelligence, such as lessons learned, risk logs, and internal policies, is stored in a company-controlled environment, while the layer connecting this data to the AI remains independent. If a current provider changes its terms, raises prices, or experiences downtime, the PMO can seamlessly switch the underlying AI model (for example, to a locally hosted open-source solution) without disrupting the project manager's daily workflow. This can be achieved by implementing a Vendor-Agnostic AI Layer (leveraging an Agnostic RAG architecture). In plain terms, it decouples the organization's data from any single provider, language model (LLM), or database. Ultimately, this technological sovereignty empowers the PMO to swap AI vendors without interrupting project momentum, radically reducing the risk of vendor lock-in.

3. Long-term risk governance as the foundation of responsible AI adoption.

The PMO must apply a pragmatic approach to technology strategy, recognizing that AI has limitations and cannot always explain how it reaches a conclusion. 

Therefore, it establishes clear governance rules: critical decisions regarding scope, budget, or risk are not fully delegated to AI. Resources are allocated not only to implement new tools, but also to back them up, audit their results, and keep the team trained to question automated recommendations. Project stability is prioritized over technological advantage, applying a principle of long-term risk prevention. 

Effective Altruism provides a concrete decision filter here: it asks not just whether a decision produces immediate benefits, but whether it might cause large-scale, long-term harm that is easy to overlook. Applied to AI in project management, this means the PMO must evaluate low-probability, high-impact risks, such as a provider shutdown that short-term efficiency metrics tend to hide.

Conclusion: 

Using AI as a strategic ally requires maturity in project management. The competitive advantage does not lie in having the most advanced tool, but in having a PMO that ensures the responsible use of technology. As vigilant support, the PMO guarantees that AI contributes to management and organizational objectives, protecting both people and the company from unnecessary dependencies, and ensuring that a human-centered approach remains the standard practice. 

 

Transparency Statement: The author acknowledges the use of Artificial Intelligence (AI) as an assistive tool during the research, data structuring, and content optimization process. The core concept, final review, and critical analysis remain the sole responsibility of the author.

Spanish version: https://ingconcurrente.blogspot.com/2026/06/cuando-la-ia-es-copiloto-estrategico-el.html

Artificial Intelligence and Quantum Systems

The convergence of Artificial Intelligence (AI) and Quantum Systems (QS) is currently one of the most active areas of research, with published and verifiable results. This convergence has progressed toward synergy between the two fields, with two distinct directions: first, where Artificial Intelligence as a tool helps control and optimize Quantum Systems, and second, where Quantum Systems as computational tools help Artificial Intelligence advance. Both directions present concrete progress and equally concrete limitations, which this document attempts to describe free from technical jargon or unfounded extrapolations.

I. First approach: Artificial Intelligence (AI) as a tool to help Quantum Systems (QS).

Here we find, among other things, the following:

1. Artificial Intelligence and Quantum Computers (AI+QC)

Experimental quantum systems involve extremely complex and precise physical functionalities. This is stated in general terms to avoid delving into quantum physics, which is beyond the scope of this publication. Consequently, attempting to manually manage these physical functionalities, which are also understood to be highly sensitive to the environment, is highly impractical. This is known, broadly speaking, as the problem of achieving optimal quantum control.

Here, the terms quantum coherence and decoherence (the opposite) become prominent. To simplify, coherence can be understood as the extremely short and sensitive state in which a quantum computing unit, known as a qubit, maintains valid, useful, or coherent information. Therefore, any quantum protocol must be executed within its coherence window, and this is where AI capabilities come into play: AI optimizes operations within the coherence window, assisting in calibration, adaptive control, and real-time error mitigation. In essence, AI acts as a control system that maintains the delicate quantum state necessary for quantum computers to function reliably.

2. Quantum Zeno and Anti-Zeno Effects (1,2)

The quantum Zeno effect is real and well-documented. It is based on the fact that measuring a state in a quantum system at a sufficiently high frequency promotes the suppression of its transitions to other states. Its counterpart, the Anti-Zeno effect, is also verified: there is a regime of intermediate frequencies in which measurements accelerate the transition rate between quantum states, rather than suppressing it.

Current applications of the Zeno effect aim to characterize the noise affecting a qubit and, using the ultrafast capabilities of AI, achieve quantum processor calibrations.

AI intervenes here with algorithms and recurrent neural networks that accelerate all aspects of the measurement process, reducing calibration time from hours to minutes.

II. Second approach: Quantum Systems (QS) as support for advanced Artificial Intelligence (AI).

The relationship between Quantum Computing and Artificial Intelligence is still a highly theoretical scenario, with little experimentation, but very promising, and already has emerging practical applications (3,4). We also find here the emerging field known as Quantum Machine Learning or QML, (5) and it is one of the most profound technological synergies at the moment.

This Artificial Intelligence and Quantum Computing convergence is currently trending towards the following areas:

1. Exponential parallel processing resulting from the quantum properties of superposition and entanglement. This would allow AI models to be trained with volumes of variables and dimensions that would collapse the memory of any classical supercomputer. A quantum AI could evaluate thousands of data scenarios simultaneously.

2. Hyperfast Combinatorial Optimization. Here we find the development of quantum algorithms or optimization algorithms, which would allow significant reductions in AI training times.

3. Hidden Pattern Recognition. Here we are faced with the idea that quantum computers, in theory, would be exceptionally capable of identifying complex correlations, thanks to quantum entanglement, where the state of one particle instantaneously alters another. Because of this, a quantum AI could detect patterns or vulnerabilities in computer systems, with revolutionary applications in cryptography and cybersecurity. This is a matter of great importance and also a potential threat to the security of the entire current computing environment.

4. Hybrid Systems. Here we see intermediate quantum hardware, where classical computers handle the interface and basic processing, and send the heavy load of calculation and interpretation to a quantum coprocessor to solve it in a fraction of the time.

Summary:

The current intersection of AI and quantum physics yields real value in verifiable fields: the optimal control of qubits through reinforcement learning and the calibration of quantum noise using algorithms. In all these cases, AI operates as a classical tool that optimizes parameters acting on quantum systems through conventional physical instruments.

Author’s Note: The content of this publication was developed with input and support from Artificial Intelligence in data acquisition, content structuring, and analysis.

Spanish version: https://ingconcurrente.blogspot.com/2026/06/inteligencia-artifical-y-sistemas.html

References:

1. S. Greenfield, A. Kamal, and others. A unified picture for quantum Zeno and anti-Zeno effects- a review. (2025). Arxiv. Cornell University.

https://arxiv.org/abs/2506.12679

2. A.G. Kofman,  G. Kurizki. Frequent observations accelerate decay: The anti-Zeno effect (2001), Arxiv. Cornell University.

https://arxiv.org/abs/quant-ph/0102002

3. Google. Meet Willow, our state-of-the-art quantum chip (2024).

https://blog.google/technology/research/google-willow-quantum-chip/

4. Google. M. Ivezic. Verifiable Quantum Advantage” on Willow Quantum Chip (2025).

https://postquantum.com/quantum-research/googles-quantum-advantage

5. Y. Du, X. Wang, N.Guo, and others. Quantum Machine Learning: A Hands-on  Tutorial for Machine Learning Practitioners and Researchers (2025). Arxiv, Cornell University.

https://arxiv.org/abs/2502.01146

UNEXPLORED AI FIELDS

From How can AI help to work faster and more efficiently? to How could AI move the boundaries of what is physically and intellectually possible?

The current progress of AI allows us to examine some unexplored areas that require more than just continuous improvement. To address that, the following could be reconsidered:

1. Rethinking the computational substrate, that means moving away from traditional computer hardware to change how the data is processed, and this is happening at least across three main frontiers: The Hardware & Physics Level (near memory processing): moving processing logic directly into or right next to memory chips, The Neuromorphic & Biological Level: building hardware that mimics the human brain's physical structure rather than standard logic gates, and The Theoretical Level (non turing models): exploring alternative computational systems that do not rely on the discrete, step by step algorithmic rules, such as quantum computers or intelligent matter that physically senses and adapts to its environment.

2. Rethinking the architecture of dynamic learning.

3. Rethinking the interplay between the laws of physics and the nature of intelligence.

Based on the above, the unexplored frontiers of AI lie in changing how it interfaces with physics, biology, and human consciousness, shifting it from advanced software to a real thinking element.

It should be noted that today's suppliers & users involved in AI  are focused on generative AI (text writing, image creation, programming) and predictive analysis (trend forecasting, fraud detection), and most have been using AI as a support tool or as a skilled assistant.

Below are some unexplored ways to further leverage AI capabilities.

1. Physical neural networks that learn through material self-organization. (1,2)

Neural networks are simulated on digital hardware. An almost untouched frontier is morphological computation: letting physical systems, such as photonic chips,  memristive fabrics (that integrate computing and data storage directly into wearable clothing or flexible materials, using memristors to function as both artificial synapses and memory), and chemically active droplets, morph their internal structure to directly embody a learned function. Instead of running an algorithm on a GPU, a physical forcing (light, chemical gradient, vibration) would be applied and let the material achieve energy minimization naturally  (references report photonic chips that can perform matrix multiplications at the speed of light and with a fraction of the energy, because the physics of interference and diffraction already computes the transformation). The learning occurs by tuning the material properties (phase shifters, memristors) rather than crunching numbers. The result would be a quasi thinking object made of matter that is inference, with a very low energy requirement to maintain itself. Early work in physical reservoir computing touches this, but training a material via dissipative adaptation remains underexplored for AI.

2. Biologically Integrated AI (Organoid Intelligence). (3,4)

Massive silicon chips have been built to run AI, consuming a lot of electricity. The next unexplored frontier would be Organoid Intelligence (OI, interfacing AI software directly with living, lab-grown human brain cells (organoids).​ Instead of simulating neural networks on a computer, researchers are starting to plug silicon computing power into real biological neurons. True biological computing operates on a fraction of the energy a data center requires.

This approach presents a promising scenario for AI development, but it also raises important ethical considerations regarding the relationship between humans and AI that need to be debated.

3. Thermodynamic computing and the free energy harvesting. (5,6)

AI requires an external power source. An unexplored approach would be to build an AI as a dissipative structure that sustains its own computational state by harvesting environmental free energy (thermal gradients, chemical potentials).

4. Reverse-Engineering Human Cognitive Barriers. (7)

AI needs training to simulate human thought, but the frontier would be using AI to debug human cognitive flaws. Because AI doesn't think like a human, it can identify where our evolutionary biology trips us up. In other words, AI could address a problem and isolate precise patterns of collective cognitive bias or emotional blindness in real time. It would not only provide answers but also act as a psychological mirror, showing how and when a failure to see reality arises. 

Special attention should be paid to the warning that AI could analyze a psychological profile, predict behavior, and manipulate human emotions, in addition to the potential creation of its own AI biases.

5. Unsupervised Scientific Discovery (The "Blind" Scientist). (8,9)

​AI is used to solve defined problems, using collected data. AI has not been allowed to run completely wild on raw, unprocessed data to find laws of physics that no one even knows exist.

Scenario: If AI were requested to operate without the theoretical support of available human knowledge and were allowed to construct its own framework of reality from scratch to address a problem, it might discover alternative mathematical frameworks or hidden dimensions of physics that human cognitive biases have prevented us from perceiving. 

This is promising, but it is also an ethical threat to be wary of.  

6. Encoding intelligence in static, nonexecutable structures. (10)

The user always runs AI as a special software. But consider a static, 3D printed object that computes through its physical geometry. This is a computational metamaterial that classifies images via wave interference. While diffractive neural networks are currently being tested in labs, the potential to evolve these structures for arbitrary cognitive tasks, particularly through in situ material deposition, remains as an open frontier.

7. Collective intelligence from Swarm mechanics. (8,9)

Multi agent AI is about digital agents. What if the agents were short scale physical AI elements with minimal individual computing units, but their collective mechanical interactions formed a distributed computing fabric? The physical interactions become the computation, and the swarm thinks by reorganizing its shape. This is a kind of tangible cellular automaton where problem solving is achieved through embodiment, not symbolic messaging. Very little work has tried to program intelligence into the morphology of a swarm’s physical interactions.

8. Edge Native Hyper Localized Swarm Intelligence. (11)

​AI today relies on centralized cloud servers. The unexplored shift would be moving toward massive, decentralized swarms of tiny, low power AI instances operating on the edge (inside everyday objects) that communicate peer to peer without ever touching the internet.

9. Causal models learned from real time physics. (12,13)

AI learns correlations from static datasets. A different approach would be to create an autonomous AI agent that builds its own causal model by interacting with the physical world from its inception, using structured neural networks and active inference. The power of AI would not come from the data and how it is managed, but from the continuous, real time sensorimotor interaction between the user and the AI. This would be an AI where the learning algorithm and capabilities are developed together, with the ultimate goal of substantiating results without the need for human labeled input.

As in points 2, 4, and 5 above, the implicit ethical risk must be considered, which in this case refers to AI with freedom of judgment.  The aim of this concern is to ensure that AI will always be aligned with human needs and privacy.

Summarizing:

Taken together, the fields mentioned above show an AI that would be unlike anything that exists so far.

This picture is concrete:

.- A photonic neural network learns by settling into a lowest energy state.

.- Organoid intelligence leans toward causal reasoning. Pair it with a causal world model, and you might get a learning system.

.- Edge native swarms as highly efficient physical nodes could be deployed en masse. They would process sensory data on the spot, exchange causal features, and form a global perceptual grid that is self-organized.

This isn't just about making AI faster. It's about changing the relationship between intelligence, matter, and environment.

The paradigm shift: moving away from asking How can AI help to do a job faster and more efficiently? and towards How can AI expand the limits of what is physically and intellectually possible?

References:

1. Self-organising Memristive Networks as Physical Learning Systems (2025). Arxiv, Cornell University.

https://arxiv.org/abs/2509.00747

2. K. S. Riley, S. Koner, J. Osorio, and others. "Neuromorphic metamaterials for mechanosensing and perceptual associative learning". (2022). Arxiv, Cornell University.

https://arxiv.org/abs/2203.10171

3. Smirnova, L., Caffo, B.S., and others (2023). "Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a dish". Frontiers.

https://www.frontiersin.org/journals/science/articles/10.3389/fsci.2023.1017235/full

4. Cai, H., Guo, F., and others (2023). "Brain organoid reservoir computing for artificial intelligence". Nature Electronics.

https://www.researchgate.net/publication/376413285_Brain_organoid_reservoir_computing_for_artificial_intelligence

5. Laurent Caraffa. BEDS: Bayesian Emergent Dissipative Structures. A Formal Framework for Sustainable Digital Twins and Continual Learning Systems. (2026). Univ. Gustave Eiffel, French National Institute of Geographic and Forest Information, Ministry of Ecological Transition, France. Arxiv.

https://arxiv.org/html/2601.02329v1

6. G. Auti, H. Daiguji, G. Tanaka. "Hebbian Physics Networks: A Self-Organizing Computational Architecture Based on Local Physical Laws". (2026). Arxiv, Cornell University.

https://arxiv.org/abs/2507.00641

7. G. Cauwenberghs. "Reverse engineering the cognitive brain" (2013). PNAS.

https://www.pnas.org/doi/10.1073/pnas.1313114110

8. Brunton, S.L., Proctor, J.L., & Kutz, J.N. (2016). "Discovering governing equations from data by sparse identification of nonlinear dynamical systems (SINDy)." Proceedings of the National Academy of Sciences (PNAS). 

https://www.pnas.org/doi/10.1073/pnas.1517384113.

9. Discovery of Physics From Data: Universal Laws and Discrepancies (2020). Frontiers.

https://www.frontiersin.org/journals/artificial-intelligence/articles/10.3389/frai.2020.00025/full

10. G. Tanaka, T. Yamane, J. B. Héroux, and others. "Recent advances in physical reservoir computing: A review" (2019). ScienceDirect.

https://www.sciencedirect.com/science/article/pii/S0893608019300784

11. Decentralized Multi-Agent Swarms (DMAS) Architecture Research (2026). "Decentralized Multi-Agent Swarms for Autonomous Grid Security in Industrial IoT: Consensus-based Approach". Arxiv, Cornell University.

https://arxiv.org/abs/2601.17303

12. A. Sharma, A. Gupta, C. Wang. "Inducing Causal World Models in LLMs for Zero-Shot Physical Reasoning" (2025). Arxiv.

https://arxiv.org/html/2507.19855v4

13. AIT Staff Writer. "The Physics of Intelligence: Can AI Systems Develop an Internal Model of Reality?" (2026). AIThority.

https://aithority.com/ait-featured-posts/the-physics-of-intelligence-can-ai-systems-develop-an-internal-model-of-reality/

AI-based Simulation of Coagulation, Flocculation, and Sedimentation for Oil & Gas Influents

AI-based simulations of coagulation, flocculation, and sedimentation processes for Oil & Gas influents increasingly use machine learning models, particularly artificial neural networks and fuzzy systems, to predict and control removal efficiency for various contaminants with different chemical reagents. These AI models can handle diverse influent characteristics (turbidity, pH, suspended solids) and optimize dosages for commonly used chemicals, such as poly aluminum chloride, ferric chloride, and various flocculants, based on real-time and historical data.

These AI simulations work based on neural network models previously trained on operational data sets (coagulant dosages, influent and effluent quality, pH, temperature, turbidity, etc.) to capture the nonlinear relationships governing the effectiveness of each stage. The models can forecast the outcome of coagulation and flocculation (e.g., effluent turbidity, contaminant removal rate) when different chemicals and doses are used, enabling optimization of both process efficiency and chemical consumption.

Predictive AI can recommend dosage adjustments for chemicals such as aluminum and iron salts or organic polymers, tailored to specific influent load, minimizing overdosing and underdosing risks.

Benefits and Outcomes:

.- Machine Learning–driven dosing leads to more consistent removal of suspended and dissolved contaminants, reducing variability due to fluctuating influent quality. 

.- AI-optimized systems regularly report chemical cost reductions (10–25%) while maintaining or improving effluent quality and regulatory compliance.

.- The use of multilayer perceptron neural networks has demonstrated high predictive accuracy (R² > 0.96) in simulating coagulation-flocculation processes, outperforming traditional rule-based or fuzzy regression approaches.

Current advanced approaches combine neural networks with genetic algorithms to find the most efficient parameter settings (e.g., coagulant dosage, mixing time) for different oil and gas influents and treatment objectives.

Most research and pilot applications use AI to evaluate the performance of standard chemicals for water treatment in the O&G industry (e.g., poly aluminum chloride, ferric chloride, cationic/anionic polymers) under different influent properties, enabling rapid simulation and process tuning.

These AI simulations can help reduce the need for traditional jar-testing by providing reliable virtual assessments of process outcomes under a range of influent scenarios.

In summary, AI simulations of coagulation, flocculation, and sedimentation offer the water treatment industry powerful tools for process optimization and decision support across a wide variety of influent conditions and chemical regimes.

The benefits of AI simulation for specific O&G effluents include: 

1. Reduces reliance on costly and time-consuming jar tests by providing virtual experimentation under various chemical and inlet conditions.

2. Enables rapid scenario analysis to optimize treatment chemicals and process parameters, tailoring them to the unique characteristics of each oil and gas effluent.

3. Facilitates proactive process adjustments in response to inlet fluctuations, improving regulatory compliance and reducing operating costs.

In practice, combining detailed effluent characterization with AI-driven predictive modeling creates a powerful tool set for optimizing coagulation, flocculation, and sedimentation in oil and gas wastewater treatment using widely adopted chemical reagents.

The best-performing machine learning models for simulating coagulation in O&G water treatment processes typically include Artificial Neural Networks (ANN), Random Forests (RF), Support Vector Machines (SVM), and Adaptive Neuro-Fuzzy Inference Systems (ANFIS). Among these, ANN models generally show superior accuracy and precision in predicting process outcomes like coagulation efficiency, turbidity removal, and optimal chemical dosing.

Performance Highlights:

-. Artificial Neural Networks (ANN): Most commonly used due to their strong ability to model nonlinear relationships in complex wastewater data. ANN models achieve high coefficients of determination (R² close to 0.96 or higher) and low prediction errors in coagulation-flocculation simulations.

-. Random Forest (RF): Performs well with high accuracy and robustness, especially in handling noisy or complex datasets. Sometimes used to rank input features that influence coagulation processes significantly.

-. Support Vector Machines (SVM): Effective in classification and regression tasks within coagulation studies, but often slightly less accurate than ANN for continuous outcome predictions.

-. Adaptive Neuro-Fuzzy Inference System (ANFIS): Combines neural networks with fuzzy logic, providing a good fit and interpretability, excelling in process optimization scenarios such as electrocoagulation or chemical dosing.

-.  Hybrid models and optimization algorithms (e.g., Genetic Algorithms, Particle Swarm Optimization): Often coupled with ANN or ANFIS for fine-tuning parameters and maximizing removal efficiency.

ANN and ANFIS are frequently considered top choices for O&G coagulation simulations due to their high predictive accuracy and flexibility in handling nonlinear process behavior, whereas RF and SVM provide complementary strengths in feature importance and classification tasks.

Comparing Artificial Neural Networks (ANN), Random Forests (RF), Support Vector Machines (SVM), and K-Nearest Neighbors (K-NN) based on their typical performance in predicting and simulating coagulation process metrics:

-. ANN excels in accuracy and modeling nonlinearities common in coagulation but is computationally intensive and less interpretable.

-. RF offers robustness to noise and useful interpretability through feature importance, but is slightly less accurate than ANN.

-. SVM is a solid option for well-defined classification/regression tasks, but less flexible for complex coagulation kinetics.

-. K-NN is easy to implement and interpret, but struggles with large, noisy, or high-dimensional datasets typical in water treatment.

In short, ANNs and RFs are usually preferred in coagulation process modeling when accuracy and robustness are priorities, while SVMs and K-NNs can be useful in simpler or complementary functions.

PLC and AI Controllers (Artificial Intelligence) in Industrial Automation

 PLCs and AI Controllers (Artificial Intelligence) represent two distinct approaches to industrial process automation and control. Selecting the most suitable approach will depend on the user's current needs. While both aim to optimize processes, their capabilities, applications, and architectures have key differences, which are summarized below:

 1.       Core Functionality

·       PLCs:

- Traditional PLCs execute deterministic, rule-based logic (e.g., ladder logic) for tasks like conveyor control or motor sequencing. They excel in real-time reliability and safety-critical operations due to their predictable response times.

- Its response capacity is limited to scenarios foreseen in its programming, without autonomous adaptation to unscheduled changes.

·       AI Controllers:

- These systems leverage machine learning (ML), predictive analytics, and adaptive algorithms to optimize processes dynamically. They handle nonlinear systems (e.g., chemical processes) by adjusting real-time parameters based on sensor data.

 2. Decision-making and adaptability

 ·       PLCs:

- Work with pre-programmed logic and fixed sequences, executing repetitive tasks based on defined rules. For example, controlling the synchronization of an assembly line or activating emergency stops.

·       AI Controllers:

- Make autonomous decisions for unforeseen scenarios based on historical patterns, as well as on digital twins, improving efficiency and reducing downtime.

- Use machine learning algorithms to analyze data in real time and dynamically optimize processes without manual reprogramming.

Key algorithm references:

-. LSTM neural networks for compressor failure prediction, increasing accuracy in early detection

-. Distributed Machine Learning and Federated Learning for multi-plant predictive maintenance to reduce data transfer.

-. Q-learning adaptive control in adjusting control elements (e.g. pneumatic valves)

 3.       Handling complexity and non-linear processes

 ·       PLCs:

- Efficient in structured and repetitive processes, such as motor control or assembly line synchronization. Programs are based on standard languages ​​(ladder logic) and can integrate modules for specific tasks (e.g. PID control). (2)

- Limited in managing non-repetitive tasks or those that require continuous adjustments, such as polishing parts with dimensional variations.

·       AI Controllers:

- Capable of managing non-linear and dynamic systems. For example, in production lines where demand fluctuates, they automatically adjust routes using data from IoT sensors.

- Integrates digital twins to simulate scenarios and optimize processes before physically implementing them.

- Applies artificial vision for quality inspection, detecting defects in products with greater precision than traditional methods.

- Learn from data generated by IoT sensors, allowing production routes to be optimized or speeds to be adjusted according to demand.

 4.       Key Advantages 

PLCs	AI Controllers
Deterministic performance (critical for safety) (1), (2)	Adaptive decision-making (e.g., self-learning control loops) (3), (4)
Ease of programming (ladder logic) (1)	Predictive capabilities (e.g., maintenance, energy optimization) (3), (5)
Proven reliability in harsh environments (2)	Handling complex data (e.g., anomaly detection via computer vision) (3), (9)

 5.       Maintenance and resilience

 ·       PLCs:

- Requires scheduled maintenance and manual fault diagnosis. Its robustness makes it resistant to harsh environments (dust, vibrations), but it relies on human intervention to resolve unforeseen problems.

·       AI Controllers:

- Implement self-diagnosis, advanced predictive maintenance, and proactive correction. For example, AI systems that identify wear and tear on machines and schedule maintenance before failures occur. They automatically correct operating parameters, reducing the dependence on specialized technicians, as algorithms can "learn" from past operations and adjust parameters automatically. 

6.       Applications

 ·       PLCs:

 - Fixed automation tasks (e.g., conveyor belts, assembly lines).

 - Safety interlocks and emergency shutdowns. (1), (2)

·       AI Controllers:

- Predictive maintenance: Analyzing vibration or temperature trends to preempt failures. (3), (5)

- Process optimization: Dynamically adjusting parameters in chemical reactors for energy efficiency. (3), (4)

- Human-machine collaboration: NLP interfaces for voice-controlled adjustments. (2), (3)

7.        Technological Synergies

The future lies in hybrid systems combining PLCs with AI:

- Edge computing: PLCs process local data for low-latency control, while AI models run on edge devices for real-time analytics. (2), (6), (7)

- Digital twins: Virtual replicas simulate scenarios (e.g., equipment failure) to refine PLC logic and AI predictions. (2), (4)

- 5G/IIoT: High-speed connectivity enables distributed AI controllers to coordinate with PLCs across facilities. (6), (7) 

8.       Scalability and costs

 ·       PLCs:

- Moderate initial costs, ideal for specific applications. However, scaling functions require additional hardware (e.g., I/O modules) .

- Cost-effective in stable processes where investment in AI is not justified.

·       AI Controllers:

- Higher initial investment, but they offer scalability through software updates. For example, adding an optimization algorithm without changing hardware.

 9.       Challenges 

·       PLCs:

- Limited scalability and integration with modern IT systems (e.g., MES/ERP). (2)

- Inflexibility in adapting to dynamic processes. (2)

·       AI Controllers:

 - Data dependency: Requires high-quality, labeled datasets, which legacy systems often lack. (2), (3)

 - Cybersecurity risks: Increased attack surfaces due to interconnected systems. (8-7),

 - Skill gaps: Engineers need expertise in both PLC programming and ML tools. (3), (5)

- Possible hallucinations in AI responses due to insufficient model training, poor data retrieval, or other underlying deficiencies in the AI input data.

10.   Future Trends

• Autonomous factories: AI-driven PLCs will enable self-optimizing production lines with minimal human input.
• AI code generation: Tools like Gemini 2.0 Flash automate ladder logic creation, cutting programming time by 30%. (1)
• Ethical AI: Transparent decision-making frameworks to ensure accountability in autonomous systems. (2), (4)
• Sustainable automation: AI prioritizes energy efficiency and waste reduction (e.g., Gallium Nitride "GaN"-based components). (4), (5) 

11.   Conclusion

PLCs remain indispensable for deterministic tasks, while AI controllers unlock adaptability and predictive power. PLCs are ideal for harsh environments and critical applications that require stability and real-time responses and remain relevant in low-risk, critical applications. On the other hand, AI controllers revolutionize the industry with adaptability, predictive analytics, and the ability to manage complexity. The choice depends on factors such as the required flexibility, budget, and the need for integration with advanced technologies. Companies adopting hybrid systems will lead in efficiency, sustainability, and resilience.

 

References:

1. PLC Programming: Traditional vs AI -Which Wins?                             https://accautomation.ca/plc-programming-traditional-vs-ai-which-wins/         
2. Ai Plc Smart Industrial Zones.                                                                        https://zeroinstrument.com/ai-plc-smart-industrial-zones/
3. How AI can be applied to program PLCs in industrial automation.                 https://antomatix.com/how-ai-can-be-applied-to-program-plcs-in-industrial-automation/
4. From optimization to autonomy - Top five manufacturing automation trends for 2025 from OMRON.                                                                                                 https://industrial.omron.eu/en/news-discover/blog/from-optimization-to-autonomy-top-five-manufacturing-automation-trends-for-2025-from-omron
5. Industrial Automation Trends 2025.                                           https://www.piglerautomation.com/industrial-automation-trends-2025/
6. 8 Key Industrial Automation Trends in 2025.                              https://www.rockwellautomation.com/en-us/company/news/the-journal/8-key-industrial-automation-trends-in-2025.html
7. Top 10 Industrial Automation Trends in 2025.                                                             https://www.startus-insights.com/innovators-guide/industrial-automation-trends/
8. Top 10 Industrial Automation Trends to Watch in 2025.                https://industrialautomationco.com/blogs/news/top-10-industrial-automation-trends-to-watch-in-2025
9. The Future of Industrial Automation.               https://www.industrialautomationindia.in/articles/industrial-automation-trends-2025-ai-ml-smart-manufacturing