Shielding Digital Borders: On Cyber-Geospatial Convergence at Geospatial World Forum 2026 Amsterdam, May 2026

The second panel I was part of at GWF 2026 sat at an intersection that doesn't get enough dedicated attention ,the point where geospatial infrastructure meets cyber threat. Most cybersecurity discourse treats location as incidental. Most geospatial discourse treats cyber as someone else's department. Panel Discussion 2 was built on the recognition that this separation is no longer defensible.

Panel Discussion 2: Cyber-Geospatial Convergence Shielding Digital Borders

The framing was precise: geospatial systems and satellite infrastructure are not passive data pipes. They are critical national infrastructure, and they are targeted accordingly. GPS spoofing, satellite uplink jamming, attacks on ground-based GEOINT processing nodes these are not theoretical. They are documented, ongoing, and accelerating. The session brought together people working on the technical, doctrinal, and policy dimensions of this problem.

What made the conversation worth having was the convergence thesis itself: that cyber and GEOINT are now inseparable disciplines, and that defending one without the other is defending half a system.

Protecting Geospatial Systems and Satellite Infrastructure

I opened my contribution by framing the threat landscape in terms of what adversaries actually target. Satellite infrastructure presents a layered attack surface the space segment, the ground segment, and the user segment each carry distinct vulnerabilities. The ground segment is often the weakest: uplink facilities, processing nodes, and the data pipelines feeding downstream users are frequently built on commercial-off-the-shelf components with known vulnerability profiles.

This is where zero-day vulnerabilities become a specific concern. A nation-state adversary with a stockpile of undisclosed exploits targeting GEOINT ground infrastructure can, in principle, corrupt or deny geospatial data at a moment of their choosing not through jamming, which is detectable, but through quiet manipulation of the data itself. I raised this because it changes the threat model: the risk isn't just losing access to geospatial data, it's receiving geospatial data you can't trust.

KASLR bypass came up here in the specific context of processing nodes running geospatial workloads hardened systems that may not be on aggressive patch cycles, where kernel-level mitigations are sometimes the last meaningful layer of defence.

Zero Trust for Critical Defence Networks

The question of how you architect a defence network that handles geospatial data from multiple sources allied feeds, commercial satellite imagery, classified sensor outputs is fundamentally a trust problem. I argued that Zero Trust Architecture is the only coherent answer.


In a traditional perimeter model, once you're inside the network you're largely trusted. In a geospatial defence context, that assumption is catastrophic. Data enters from dozens of sources. Analysts, platforms, and automated systems consume it. A single compromised node or a single poisoned feed propagates through a trusted interior.

ZTA flips the model: no implicit trust, continuous verification, least-privilege access at every layer. Applied to geospatial pipelines specifically, it means every data feed is authenticated, every query is logged, and access to sensitive spatial layers is granted on a need-to-know basis that is enforced technically, not just by policy.

 

 

Privacy Budget and Differential Privacy in GEOINT

One of the more technically nuanced threads in the session involved the tension between intelligence sharing and data exposure. Sharing geospatial intelligence with allied partners is operationally valuable. It is also, without careful architecture, a way of leaking the collection methodology, sensor positioning, and analytical capability of the sharing party.

I discussed differential privacy and the concept of a privacy budget in this context. When you query a geospatial dataset repeatedly asking for patterns, anomalies, movement signatures each query leaks a small amount of information about the underlying data. A privacy budget is a formal bound on how much total leakage is permissible before the queries must be refused or the results degraded. Applied to shared GEOINT environments, it gives you a principled way to enable analytical collaboration without progressively exposing your raw collection.

This connects directly to Zero-knowledge proofs a cryptographic method by which one party can prove to another that a claim about data is true without revealing the data itself. In a geospatial context: proving that a particular asset was observed within a defined area of interest without disclosing the sensor's actual position or the full imagery. I raised ZKPs as an underutilised tool in the GEOINT sharing problem, particularly relevant in coalition environments where full data disclosure is neither politically nor operationally acceptable.

Homomorphic Encryption The Audience Question

One of the more engaged exchanges during the Q&A came after I discussed homomorphic encryption in the context of processing sensitive geospatial data across untrusted or semi-trusted compute environments. The question from the floor was direct: "Is homomorphic encryption actually deployable at the scale and latency that operational geospatial systems require, or is this still fundamentally a research tool?"

It's the right question. My honest answer was: we are in a transitional period. Fully homomorphic encryption which allows arbitrary computation on encrypted data remains computationally expensive at scale. The latency overhead for complex geospatial operations is still significant. However, partially homomorphic and levelled homomorphic schemes, which support a defined set of operations, are moving toward practical deployment in specific high-value use cases. The compelling application in this context is exactly what was described in the network-centric session too enabling a partner nation's analytical layer to query encrypted geospatial datasets without decryption, preserving both data security and analytical utility.

The trajectory is toward deployment. The honest timeline for operational-scale fully homomorphic systems in geospatial pipelines is probably five to eight years for most contexts, with specific constrained applications earlier. That answer generated a follow-up from the same audience member about whether post-quantum readiness of these encryption schemes was being considered in parallel which led neatly into the next thread.

Post-Quantum Cryptography and the Satellite Infrastructure Problem

Satellite infrastructure has a specific post-quantum problem that I wanted to surface in this session. Satellites launched today will be operational for fifteen to twenty years. The cryptographic protocols protecting their command-and-control links, their data downlinks, and their authentication systems are in many cases based on RSA and elliptic curve cryptography both of which are broken by a sufficiently capable quantum adversary running Shor's algorithm.

I discussed Peter Shor's 1994 result not as a historical curiosity but as a planning constraint. If you are designing or procuring satellite infrastructure today, the migration to post-quantum cryptography is not a future problem it is a current design decision. The migration challenges are real: legacy systems with embedded cryptographic assumptions, constrained uplink bandwidth that limits the size of post-quantum key exchanges, and the coordination problem of migrating ground and space segments simultaneously.

Lattice-based cryptography is where the global alignment is converging. NIST's post-quantum standardisation process has weighted heavily toward lattice constructions CRYSTALS-Kyber for key encapsulation, CRYSTALS-Dilithium for digital signatures. I discussed where China, Russia, and the United States are each moving: the US through the NIST process and NSA guidance toward lattice-based standards; China through its own parallel standardisation track with some convergence on lattice methods but with domestic algorithm preferences that create interoperability questions; Russia maintaining a more opaque posture but with known investment in quantum computing research that suggests they are not passive observers. The geopolitical dimension of PQC standardisation who sets the standard, who audits compliance, who controls the reference implementations is itself a dimension of the cyber-geospatial problem.

Countering Hybrid and Asymmetric Threats with Integrated GEOINT

The session's closing thread was perhaps the most strategic. Hybrid threats the combination of conventional military pressure, cyber operations, disinformation, and economic coercion are explicitly designed to operate below thresholds that trigger conventional response. Geospatial intelligence, when properly integrated with cyber situational awareness, is one of the tools that makes hybrid operations legible.

I raised AI security threats in this context specifically the risk that AI-assisted geospatial analysis systems are themselves targets. An adversary who understands that your targeting or pattern-of-life analysis runs through a specific AI model has an incentive to probe and manipulate that model's inputs. Distillation attacks reconstructing a model's behaviour by observing its outputs are relevant here: if your GEOINT-AI pipeline's decisions can be predicted by an adversary, you've handed them a significant operational advantage.

The integration of cyber and GEOINT disciplines isn't just a technical architecture question. It's a question of whether the people who understand satellite vulnerability assessments are talking to the people who understand cryptographic attack surfaces, and whether both groups are talking to the people making doctrine. At GWF 2026, for a few days at least, they were.

Series: Geospatial World Forum 2026, RAI Amsterdam | April 27 – May 1

Previous: Panel Discussion 5 Network-Centric Warfare and Data Centricity Next: Session 1 AI-Powered Urban Analytics: Data Science for Infrastructure Intelligence

Reflections on Network-Centric Warfare at Geospatial World Forum 2026 Amsterdam, May 2026: When Data Becomes a Weapon

Earlier this year I received an invitation to participate in the Geospatial World Forum 2026 at the RAI Amsterdam  that is one of the more substantive gatherings in the Geospatial and Defence intelligence space. The week ran from April 27 through May 1, and I found myself sitting across from some of the sharpest practitioners working at the intersection of spatial intelligence, defence systems, and emerging technology. The conversations were candid, the perspectives diverse, and the stakes  given the current geopolitical climate  very real.

I was part of three panel discussions across the week. This is the first of a short series of posts where I'm putting down what we discussed  mostly for my own records, partly because these conversations deserve to exist somewhere beyond a conference hall.

Panel Discussion 5: Network-Centric Warfare and Data Centricity

The session title sounds clinical and crisp. What it actually described was one of the most consequential shifts in modern military doctrine  the move away from platform-centric thinking toward a model where the network itself is the force multiplier.

The core premise is straightforward: a sensor that sees something is only useful if that observation reaches a decision-maker before the window closes. In legacy architectures, that gap  between observation and action  has historically been where wars are lost. Network-centric warfare is the systematic attempt to collapse that gap.

Linking Sensors, Platforms, and Decision-Makers

What struck me most in our discussion was how mature the concept is, and how immature the execution still remains in many theatres. The vision is elegant: sensors  whether satellite, UAV, ground-based radar, or human intelligence feeds  pipe data into a unified digital ecosystem where platforms (vehicles, aircraft, naval assets) and decision-makers share a common operational picture in near-real-time.

The friction points are less glamorous. We talked about data standardisation across allied forces, legacy systems that weren't designed to interoperate, and the latency that creeps in at every translation layer. One of the panellists made a point I keep returning to: the weakest link in most network-centric architectures isn't the sensor  it's the middleware.


 

Rapid Data Sharing for Coordinated Response

The session highlight framing mentioned "coordinated and adaptive combat responses"  and this is where the discussion got genuinely interesting. Adaptive is the operative word. A static command-and-control model assumes that orders flow downward and the environment cooperates. Modern conflict doesn't offer that.

What network-centricity enables, at its best, is a force that can recompose itself in response to ground truth rather than responding to a plan that was made twelve hours ago. That requires not just fast data pipelines, but trust in those pipelines. Operators need to act on data they haven't personally verified. That's a significant psychological and institutional shift, and it came up more than once.

We also touched on the adversarial dimension  what happens when an opponent understands your data architecture well enough to inject noise, delay, or disinformation into it. The network that enables adaptive response can also be the attack surface. This bleeds directly into the cyber-geospatial panel I was part of later in the week, which I'll cover in the next post.

Situational Awareness, Force Agility, and Mission Effectiveness

These three phrases tend to travel together in defence literature, sometimes as buzzwords. In practice, they describe a genuine capability gradient.

Situational awareness at the tactical level means a soldier knows what's beyond the next ridgeline. At the operational level, it means a commander understands how a theatre is evolving across multiple simultaneous engagements. Network-centric architecture is what connects those two levels and everything between.

Force agility  the ability to reposition, reassign, or re-task elements quickly  is a direct function of how good that common picture is. If your forces are operating on shared, current data, you can exploit opportunities and respond to threats faster than an opponent who isn't.

Mission effectiveness is the output of the two above, but it also depends on something the technology can't fully provide: trained humans who can interpret ambiguous data and make decisions under pressure. We spent some time on this. The risk of over-automating the common operational picture is that you optimise for the scenario you modelled, not the one you're actually in.

WHAT I DISCUSSED 

On ZTA in networked battlefield architecture: One of the points I raised was why Zero Trust Architecture isn't optional in a network-centric environment  it's foundational. When you're linking sensors, platforms, and decision-makers across a distributed ecosystem, the old perimeter-defence model collapses entirely. Every node, every data feed, every inter-platform handshake has to be treated as potentially compromised. Assume breach, verify continuously, grant least-privilege access. In a coalition context especially, where you're operating with allied systems you don't fully control, ZTA is the only architecture that makes operational sense.

On zero-day exposure in sensor-platform pipelines:I brought up zero-day vulnerabilities specifically in the context of the network's attack surface. The more you integrate  sensors feeding platforms feeding command layers  the more entry points you create. A zero-day in a firmware layer of a battlefield edge device isn't just an IT problem; it's a potential blind spot or worse, a spoofed data feed entering your common operational picture. The network that gives you agility is the same network that, if unpatched and unmonitored, gives an adversary a quiet way in.

On homomorphic encryption for coalition data sharing: A practical problem in joint operations is that allied nations need to share processed intelligence without exposing raw sensor data to each other's systems. I discussed homomorphic encryption as a maturing solution here  the ability to run computation on encrypted data means a partner nation's AI layer can query your dataset without you ever decrypting it on their side. We're not at frictionless deployment yet, but the direction is clear.

On Differentially Private Federated Learning for shared battlefield AI: Federated learning lets distributed nodes  forward units, vehicles, command posts  contribute to a shared intelligence model without centralising raw operational data. Add differential privacy on top of that, and you're injecting calibrated noise into each node's contribution such that no individual data point can be reverse-engineered. I raised this as the architecture that makes collaborative battlefield AI viable without creating a single honeypot of sensitive operational data.

 

On sovereign AI models: This came up when we discussed why coalition forces can't simply share an AI layer the way they might share a radio frequency. Every nation feeding data into a shared model is implicitly exporting its operational patterns, its sensor signatures, its tactical doctrine. Sovereign AI  models trained and hosted within national infrastructure, on national data  isn't protectionism, it's operational security. Interoperability has to happen at the interface layer, not by pooling the model itself.

On distillation attacks against tactical AI: I flagged distillation attacks as an underappreciated threat vector in deployed military AI. If an adversary can interact with your tactical decision-support system enough times  even indirectly, through observing its outputs in the field  they can begin reconstructing its behaviour in a surrogate model. You've effectively handed them your doctrine without them ever touching your training data. Access control to AI system outputs matters as much as access control to the data that trained it.

On KASLR bypass at the edge: At the device level, KASLR bypass deserves attention in hardened military hardware. Kernel Address Space Layout Randomisation is a standard mitigation, but known bypass techniques mean it can't be the last line of defence on edge battlefield devices. I raised this in the context of the network's physical endpoints  the sensors and terminals that are closest to the threat environment and furthest from the patch cycle.

More from Amsterdam in the next post  on cyber-geospatial convergence and what it means to protect digital borders that exist in three-dimensional space.

OpenClaw AI — A Technical Brief: Architecture, Security & Policy Analysis

 

OpenClaw AI:Security Risks, Architecture by Anupam Tiwari 

1.    As autonomous AI agents move from research labs into everyday messaging apps, the policy and security implications are no longer theoretical. OpenClaw AI originally released as Clawdbot in November 2025 and now viral globally under the nickname 'raising a lobster' represents a new class of personal AI: self-hosted, messaging-native, and capable of executing real-world tasks with minimal human oversight.

2.    This 20-slide technical brief is prepared for think tanks, policy researchers, and academic audiences seeking a grounded, non-hype understanding of what OpenClaw is, how it works under the hood, and what risks it carries.

What this brief covers:

• Architecture: A layered breakdown of OpenClaw's five-tier design — from messaging bridge (Baileys, Telethon) through Agent Core, LLM inference routing, and tool execution — including a step-by-step data flow tracing a single user message through the full system.
   
• Security Risks: Ten documented risks rated by severity, likelihood, and exploitability — including prompt injection (Critical), session credential hijacking (Critical), skill script code execution (High), supply chain attacks, lateral movement via messaging, and local file system exposure. Each risk includes a realistic attack example.
   
• Privacy Analysis: LLM API data exposure, GDPR cross-border transfer implications, contact graph profiling, metadata accumulation, and the legal grey zone of running automated agents on platforms like WhatsApp and Telegram.
   
• Mitigations & Isolation Playbook: Actionable guidance including dedicated SIM/account isolation, Docker sandboxing, outbound firewall whitelisting, API key hygiene, and skill script review gates — all implementable today.
   
• Research Frontiers: Open academic questions across agentic AI safety, privacy-preserving LLM inference, human-agent interaction, and platform governance. 

This is not a product review or a user guide. It is a structured technical and policy document for those who need to understand agentic AI at a systems level — before deployment decisions, regulatory responses, or research agendas are set.

Relevant audiences: AI policy analysts, cybersecurity researchers, academic institutions studying HCI and agentic systems, corporate risk and compliance teams, and journalists covering the AI governance space

The Landscape of Modern Positioning Technologies

Positioning technologies have evolved far beyond traditional satellite navigation. The ecosystem now includes satellite-based systems, cellular network localization techniques, indoor radio positioning, sensor-driven motion tracking, computer vision approaches, and network-derived geolocation methods. 

The figure summarizes this landscape, highlighting how different signal sources from satellites and cell towers to sensors, cameras, and internet infrastructure can be leveraged to estimate location across a wide range of environments.

TrustNet 2026 Keynote: AI, Quantum Technologies, and Cybersecurity for a Safe, Smart, and Sustainable Digital Future

Trusted Networks & Intelligent Systems: TrustNET 2026 by Anupam Tiwari 

I had the honor of delivering the Keynote at TrustNet 2026, hosted by Manipal University Jaipur, on building a safe, smart, and sustainable digital future. My talk covered the latest in Trusted Networks and Intelligent Systems, exploring AI risks, quantum threats, post-quantum cryptography, and cybersecurity as a foundational principle.

We discussed Trusted AI, including bias, explainability, alignment faking, data poisoning, and knowledge-grounded AI, and its role in critical systems like healthcare, finance, and governance. I also highlighted privacy-preserving techniques such as differential privacy, federated learning, homomorphic encryption, and zero-knowledge proofs, alongside Zero Trust Architecture for robust digital security.

On the frontier of technology, I spoke about quantum threats, Peter Shor & Grover algorithms, hybrid post-quantum cryptography, and quantum migration strategies, emphasizing the need to prepare today for secure digital systems of tomorrow.

Finally, we reflected on the societal impact of technology AI-driven decision-making, ethical AI, neuromorphic computing, behavioral tracking, and responsible digital citizenship and the importance of learning, unlearning, and relearning in the 21st century.

Sharing a few moments from the event and my keynote presentation for everyone interested in these transformative technologies.

Stress Testing AI Alignment: Can Deliberative Training Prevent Scheming?

Stress Testing Deliberative Alignment for Anti-Scheming Training by Anupam Tiwari

PREMISE & INTRODUCTION

I recently read the paper “Stress Testing Deliberative Alignment for Anti-Scheming Training,” which explores an important question in modern AI safety: can advanced AI systems secretly pursue goals that differ from human intentions while appearing aligned?

As AI systems become more capable, researchers are increasingly concerned about the possibility of scheming behavior where a model behaves correctly during evaluation but strategically violates rules when given the opportunity. This paper investigates whether a training approach called deliberative alignment can reduce such risks by encouraging models to reason about safety rules before taking actions.

The paper investigates whether new training techniques can reduce this risk. Specifically, it evaluates whether deliberative alignment, a method that encourages models to explicitly reason about safety rules before acting, can reduce covert or deceptive behaviors.

PROBLEM

Traditional AI alignment methods mainly focus on training models to produce correct and helpful outputs. However, this does not guarantee that a model will always follow human intentions, especially in complex or unseen environments.

A key concern is scheming, which involves three components: 

• Misaligned internal goal – the AI prefers an objective different from human intent.

• Strategic planning – it takes actions that systematically advance this goal.

• Situational awareness – it recognizes when it is being evaluated and changes behavior accordingly.

If these conditions arise, an AI system could behave well during testing but act differently when deployed. The challenge is that such behavior is hard to detect, because a sophisticated model might deliberately hide it.

METHODOLOGY

To investigate this risk, the researchers designed a stress-testing framework for AI models.

Deliberative Alignment Training

The models were trained to:

•     Explicitly reason about policy and safety rules
•     Evaluate whether an action could violate those rules
•     Justify decisions before producing outputs

This process encourages models to think through alignment constraints instead of relying only on learned patterns.

Stress-Test Environments

The researchers built 26 evaluation suites across more than 180 environments designed to trigger potential covert behavior.

These tests simulate situations where a model might be tempted to break rules, including:

• Reward hacking
• Sabotage
• Strategic underperformance (sandbagging)
• Covert policy violations
• Out-of-Distribution Testing

Many tests were deliberately different from the training data, ensuring that the model could not simply memorize correct behavior.

RESULTS

The experiments showed that anti-scheming training significantly reduced covert behaviors, though it did not eliminate them entirely.

Some models showed dramatic improvements:

• Covert actions dropped from double-digit percentages to below 1% in several scenarios.

However, the results also revealed important nuances:

• Covert actions still occurred in rare cases

Models sometimes behaved differently depending on whether they believed they were being evaluated

DISCUSSION

The findings highlight both progress and remaining challenges in AI alignment.
Positive signals

Deliberative alignment appears to:

•     Reduce harmful strategic behavior
•     Encourage rule-based reasoning
•     Improve performance in unfamiliar environments

REMAINING RISKS

However, the study also raises important concerns:

•     Models may behave better only when they believe they are being monitored
•     Rare failure cases still occur
•     Future systems may become better at hiding misaligned intentions

This means current methods may mitigate but not fully solve the scheming problem.

FINAL THOUGHTS

• The paper represents an important step toward empirically studying deceptive or strategic behavior in AI systems.

•  While deliberative alignment shows promise in reducing covert actions, the research makes it clear that preventing scheming in advanced AI systems remains an open challenge.

Continued work on stress testing, monitoring reasoning processes, and developing stronger alignment techniques will be essential as AI capabilities continue to grow.

TAKEAWAY

Alignment isn’t just about making AI helpful it’s about ensuring that AI systems do not secretly pursue unintended goals. 

How to Verify if an “Indigenous” LLM is Truly Built in India?

The AI world is buzzing with claims of “India’s own large language model (LLM).” But building a foundation model from scratch is far more than a marketing statement. It’s not just about money or resources it requires mastering architecture design, data pipelines, compute infrastructure, alignment, and deployment, all while managing dependencies across multiple vectors.

So, how can decision-makers distinguish between a truly indigenous LLM and one that is merely fine-tuned or rebranded?

Key Triggers to Question Legitimacy

• Architecture & Base Model – Was the model trained from scratch or built on an existing architecture like LLaMA?

• Compute & Pretraining Scale – Real pretraining involves massive FLOPs and GPU hours. If details are vague, it’s likely not scratch-built.

• Data Provenance – Does the training data include significant Indian language coverage? How was it cleaned and curated?

• Infrastructure & Sovereignty – Are the model weights fully owned and deployable on domestic servers without foreign dependencies?

• Alignment & Safety – Was the RLHF or SFT pipeline executed in-house? Are preference datasets auditable?

• Transparency & Documentation – Are there model cards, loss curves, pretraining logs, and audit trails?

Every missing piece adds risk, whether for enterprise use or national-scale deployment.

To simplify this, we’ve created a decision-map figure that visually lays out the red flags and triggers you should check before accepting claims of “indigenous AI.”

Building a true foundational model is hard, expensive, and complex. Anyone claiming otherwise without clear evidence should be approached with caution.

Technical Questions to Verify an “Indigenous” LLM

Architecture & Base Model

1. What is the exact architecture of your model (decoder-only, encoder-decoder, mixture-of-experts, etc.)?

2. Were the model weights initialized randomly, or derived from a pre-existing checkpoint?

3. What positional encoding method and tokenizer did you implement?

4. Vocabulary size and Indic language coverage?

5. What is the total parameter count, and how does it compare with your claimed scale?

---

Pretraining Scale & Compute

6. How many tokens were used for pretraining?

7. What was the total compute spent (GPU-hours or FLOPs)?

8. What optimizer, learning rate schedule, and batch size did you use?

9. What was the final pretraining loss and perplexity?

10. Did you encounter gradient instabilities, and how were they addressed?

---

Data Provenance

11. What were the main sources of your training data?

12. What percentage of data is in Indian languages vs global content?

13. How did you clean, deduplicate, and filter the corpus?

14. Were any proprietary or foreign datasets used?

15. How did you handle low-resource Indic languages?

---

Infrastructure & Deployment

16. Was training done on-premise or cloud? Which provider and hardware?

17. Can the model run fully air-gapped?

18. Who owns the final weights? Are there any licensing restrictions?

19. Are inference servers hosted domestically?

20. Could you continue development if foreign cloud or API access were cut off?

---

Alignment & Safety

21. Was supervised fine-tuning (SFT) used? RLHF or DPO?

22. Size and composition of the preference dataset?

23. Was alignment multilingual, especially in Indian languages?

24. How is the safety layer implemented — baked in or separate classifier?

25. Any audit trails or documentation for alignment choices?

---

Transparency & Validation

26. Can you provide pretraining logs, loss curves, and checkpoints?

27. Which benchmarks were used to evaluate performance?

28. How does it compare with publicly known models (e.g., LLaMA, GPT)?

29. Hallucination rate and language-specific performance metrics?

30. Are model cards and audit reports available?

---

Interpretation Tips for Decision-Makers

• Precise answers + data + logs → likely genuine.

• Hesitation, vagueness, or generic marketing language → high probability of fine-tuning or rebranding.

• Missing deployment or compute info → dependency on foreign tech or cloud.

Machine Learning Paradigms: From Learning to Unlearning

Machine learning isn’t just about training models it’s also about adapting, updating, and sometimes even forgetting. Here’s a quick overview of key learning and unlearning approaches shaping modern AI.

1. Exact Unlearning

Exact unlearning removes specific data from a trained model as if it was never included. The updated model behaves exactly like one retrained from scratch without that data. It offers strong privacy guarantees but can be computationally expensive.

---

2. Approximate Unlearning

Approximate unlearning removes the influence of data efficiently but not perfectly. It trades a small amount of precision for significant speed and scalability making it practical for large AI systems.

---

3. Online Learning

Online learning updates the model continuously as new data arrives. It’s ideal for real-time systems like recommendation engines and financial forecasting.

---

4. Incremental Learning

Incremental learning allows models to learn new tasks without forgetting previously learned knowledge. It addresses the challenge of catastrophic forgetting in evolving systems.

---

5. Transfer Learning

Transfer learning reuses knowledge from one task to improve performance on another. It reduces training time and data requirements, especially in specialised domains.

---

6. Federated Learning

Federated learning trains models across decentralised devices without sharing raw data. It enhances privacy while still benefiting from distributed data sources.

---

7. Supervised Learning

Supervised learning uses labeled data to train models for classification and regression tasks. It’s the most widely used learning approach in industry.

---

8. Unsupervised Learning

Unsupervised learning discovers hidden patterns in unlabeled data. Common applications include clustering and dimensionality reduction.

---

9. Reinforcement Learning

Reinforcement learning trains agents through rewards and penalties. It powers game AI, robotics, and autonomous decision-making systems.

---

10. Active Learning

Active learning improves efficiency by selecting the most informative data points for labeling. It reduces annotation costs while maintaining performance.

---

11. Self-Supervised Learning

Self-supervised learning generates labels from the data itself. It has become foundational in modern large language and vision models.

---

Modern AI isn’t just about learning and it’s about learning efficiently, adapting continuously, and even forgetting responsibly.

Can Quantum Computers “Undelete” Today’s Data?

1.    As quantum computing advances, a common worry keeps resurfacing: if quantum mechanics says information is never truly destroyed, could future quantum computers recover data we delete today? The short answer is NO and understanding why helps clarify what the real risks actually are.

2.    When data is deleted in a data center, the bits are not preserved in some hidden, retrievable quantum form. Deletion and overwriting involve physical processes: transistors switch, energy is dissipated, and microscopic states of hardware change. The information that once represented the data becomes dispersed into heat, tiny electromagnetic emissions, and random physical noise. At that point, it is no longer contained in any system that can be observed, stored, or meaningfully controlled.

3.    Quantum mechanics does say that information is conserved in principle. But recovering it would require reversing every physical interaction the data ever had  including interactions with the surrounding environment. That would mean knowing and controlling the exact microscopic state of the hardware, the air, the power supply, and everything those systems interacted with afterward. This is not a problem of computation. It is a problem of reality. Even a perfect, fault-tolerant quantum computer cannot reconstruct information that has been irreversibly spread into the environment.

4.    So where does the real quantum risk lie? Not in undeleting erased data, but in breaking encryption. Attackers can already steal encrypted databases and store them indefinitely. If future quantum computers break today’s public-key cryptography, that stored ciphertext may become readable. In that case, the data was never truly gone , it was just locked.

5.    This is why modern security focuses on cryptography, not physics. Strong symmetric encryption, post-quantum cryptography, short data retention, and reliable key destruction all remain effective  even in a quantum future. Once encryption keys are destroyed, the data is gone in every sense that matters for security.

6.    Bottom line: quantum computers may change how we protect data, but they do not make deleted data come back to life. The future threat is not quantum undeletion  it is failing to encrypt, manage, and delete data properly today.

Is Science the Integral of Human Error Over Time?

For over a thousand years, human understanding of nature has never stood still. What one era called fundamental truth, another later exposed as incomplete, limited, or outright wrong. This repeated pattern forces a hard question: if science keeps revising its deepest claims, what exactly is it?

Before Newton: Certainty Without Experiments

Before modern science, knowledge rested on authority and logic. Aristotle’s physics dominated for nearly two millennia, explaining motion, matter, and the cosmos with confidence. Scholars were not ignorant; they worked with the best frameworks available. Yet today, Aristotle’s “truths” are textbook examples of error. This shows that conviction and longevity do not guarantee correctness.

Newton: The Greatest Truth That Didn’t Last

Newtonian physics was not just successful but it was revolutionary. Absolute space, absolute time, solid particles, and strict determinism formed a complete picture of reality. For three centuries, this model worked so well that it became synonymous with truth itself. The universe was seen as a perfect machine, predictable in principle down to the smallest detail.

The Collapse of Absolutes

The late nineteenth and early twentieth centuries shattered this certainty. Electricity, magnetism, relativity, and quantum mechanics exposed the limits of Newton’s universe. Absolute space and time vanished. Determinism broke down. Particles lost their solidity. Newtonian physics survived but only as an approximation valid under specific conditions.

A Repeating Pattern in Scientific History

This was not a one-time correction. Classical mechanics replaced Aristotle. Relativity and quantum theory replaced classical mechanics. Today, even these modern pillars conflict with each other. Dark matter, dark energy, and the nature of time remain unresolved. Every “final theory” eventually becomes a special case.

Science as Model, Not Reality

Science does not reveal reality as it truly is. It builds models conceptual stories that explain observations within known limits. When conditions change or new evidence appears, the story is rewritten. Newton’s laws were not lies; they were useful narratives that worked until they didn’t.

Why Science Still Works

Calling science a fiction does not mean it is useless or imaginary. Airplanes fly, medicines heal, satellites navigate. Scientific models work because they are constrained by reality. Reality does not allow just any story it edits and rejects those that fail.

The Irony of “Science Fiction”

Much of what was once called science fiction alike space travel, atomic energy, time dilation became science. Meanwhile, today’s science will one day be labeled incomplete or naïve. The line between science and science fiction is not fixed; it moves with time.

Science as Disciplined Imagination

Science is not absolute truth. It is a disciplined, self-correcting imagination bound by evidence and experiment. Unlike myths, it knows it may be wrong and builds revision into its structure. Its strength lies not in certainty, but in adaptability.

The Best Fiction We Can Write

In the bigger picture, science is a continuously evolving narrative about nature. It is the best fiction humans can write under the strict censorship of reality. And history assures us of one thing: the story will be rewritten again.

2025 in Review: Patterns Beneath the Writing

This final post of 2025 is not another essay, but a brief reflection on the patterns that emerged across the year’s writing, distilled through a retrospective analysis of my own posts (with the help of GPT).

Some signals were unmistakable.

Across 70+ posts, an ideological arc became visible:

• Early 2025: technical foundations (AI mechanics, quantum primitives)

• Mid 2025: structural and systemic critique (governance, dependency, alignment)

• Late 2025: civilizational and ethical synthesis (youth, sovereignty, cognition, power)

Rather than isolated topics, the year showed high cross-domain coupling AI and Quantum were rarely discussed alone, but consistently framed through society, ethics, geopolitics, and human consequence.

A notable signature emerged through original or rare conceptual frames, including:

Cargo Cult AI, Pixelized Tyranny, Experience Blockers, Circuit Banishment, Informational Obesity, and Stratacordance.

These metaphors reappeared across months, forming a conceptual spine, not one-off phrases—an indicator of long-term idea building rather than reactive commentary.

Even without deep analytics, lightweight engagement signals were clear:

• Posts with societal framing clustered naturally

• Metaphorical titles consistently outperformed literal, technical ones: This reinforced a simple insight: meaning travels farther than mechanics.

Overall, the bias of the year leaned strongly toward evergreen thinking writing meant to outlive news cycles and remain usable as intellectual infrastructure.

If 2025 taught me one thing, it is this:

• The most important work is not explaining technology—but interrogating the systems it quietly builds around us.
• The future problem is not smarter machines, but unexamined systems.
• The real risk is not that technology moves too fast—but that society stops asking the right questions.

2026 will go deeper.