🚨AI-Generated Videos : Can We Really Tell If an AI-Generated Video Is Fake? 🎥

As use of major LLMs become more widespread, we have been seeing a crazy influx of AI-generated content on our feeds. It used to be fairly easy to spot - with a character sporting six fingers or something phasing through a solid object; but with recent improvements, these videos are getting harder and harder to tell.

This new paper, published mid-August 2026, is an interesting read, specifically surveying crisis-themed content, titled:

Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination.

They systematically evaluate how well detectors perform on crisis-themed content, how generation conditions affect detectability, how humans perceive authenticity, or whether detectors remain reliable after videos are shared and altered during social dissemination.

📦 Core Contribution: RA-Bench Dataset

The authors also introduce RA-Bench, a large-scale benchmark for AI-generated video detection that uses real videos as anchors, matching real-world crisis scenarios:

• Total videos: 17,886

• Real anchors: 1,830 authentic crisis/event videos across 10 social-risk categories

• Generated clips: 16,056 synthetic clips created from:

  • ✅ 4 open-source generators

  • ✅ 5 closed-source generators

• Designed to reflect real-world crisis contexts rather than generic footage

🔍 They then tested different types of AI detectors, including specialised deepfake detectors and multimodal AI models, like CNNSpot, UnivFD and ReStraV .

🚨The results were sobering.

1. Current detection methods were not reliable for crisis scenarios.

2. No single type of detector consistently worked across different AI generators. A detector might perform well on videos from one generator but struggle with another.

3. Impact of social media dissemination: Uploading, compressing, resizing, and reposting a video can destroy the subtle clues detectors rely on, making fake videos even harder to spot.

4. Videos that were most convincing to humans were also harder for AI detectors to identify.

This is especially concerning since crisis events are high-stakes environments where misinformation can escalate rapidly, highlighting the need for detectors robust to evolving video generators.

Beyond AdamW: The Next Generation of Optimizers for Training AI Models 📈

For years, AdamW has been the undisputed workhorse of large language model (LLM) pretraining. 🖥️

AdamW is an element-wise, first-order optimizer that updates every parameter independently: simple, reliable and easy to scale. It powers many frontier models today. However, as researchers push toward larger models, longer training runs, and massive batch sizes, AdamW is beginning to show its weaknesses.

NVIDIA researchers found at global batch sizes approaching 100 million tokens per step, AdamW begins to lose effectiveness. hence a new study from NVIDIA explores a fundamental question:

Can we build optimizers that are smarter than AdamW and allow future AI models to train faster, more efficiently, and more reliably?

The New Contenders: Muon and SOAP 🚀

Unlike AdamW, which treats parameters independently, Muon and SOAP introduce structure-aware optimization, methods that use information about the geometry and relationships within neural network weights.

🌀 Muon — Spectral Orthogonalization: Making Gradient Updates More Balanced

• Matrix-aware updates: Optimizes the geometry of gradient updates rather than scaling parameters independently like AdamW.

• Balanced learning directions: Uses Newton-Schulz iterations to orthogonalize momentum updates and balance singular values.

• Memory efficient: Removes the need for AdamW-style second-moment statistics, reducing optimizer memory.

• Better scaling: Maintains stable training at larger batch sizes and model scales.

🌐 SOAP — Preconditioned Optimization with Adam-Style Adaptivity

• Uses parameter structure: Applies Shampoo-inspired Kronecker preconditioning to capture row and column correlations in weight matrices.

• Better optimization space: Rotates gradients into an eigenbasis where directions become far more effective to update.

• Adam-compatible: Performs adaptive updates like AdamW but in a more informative coordinate system.

• Second-order benefits: Gains curvature awareness without the prohibitive cost of full second-order optimization.

📌 This shift towards new structure-aware optimization approaches would allow for:

1. More efficient AI development

2. Lower computational costs

3. More capable future models

👉 To make this transition possible, NVIDIA has released open-sourced implementations in Megatron-LM and a standalone Emerging-Optimizers library to be available to the public.

AdamW is unlikely to disappear and will remain a staple for small-to-medium training runs. Its simplicity and maturity make it an excellent choice for many applications.

╰┈➤ˎˊ˗ But for frontier-scale training? It still remains to be seen, where it'll go next.

Vision Mamba: A New Architecture Changing Computer Vision 🚀

For years, computer vision has relied on two dominant architectures: CNNs and Transformers. >>> Enter Mamba, built on State Space Models (SSMs).

> CNNs: Fast, but limited local field → struggle with long-range dependencies

> Transformers: Great global context, but high complexity → slow, high memory use

> Mamba: → delivers linear complexity + global modeling

Read more on this: Vision Mamba: A Comprehensive Survey and Taxonomy

⚙️ Key Technical Innovations

1. Selective SSM (S6) Mechanism

• Parameters B, C, Δ are input-dependent (time-varying, not fixed)

• Dynamically updates hidden state → focuses only on relevant features

• Achieves linear complexity O(L) — speed and memory scale proportionally with input size

• Mamba-2 further optimizes with vectorized computation and GPU-friendly design, matching Transformer hardware efficiency

2. Adapting SSMs to 2D Visual Data

• Vim: Adds bidirectional scanning and positional awareness to overcome unidirectional bias

• VMamba: Introduces Cross-Scan Strategy (SS2D) — traverses images in four directions to turn 2D grids into ordered sequences without breaking spatial structure

• Later variants: Local, atrous, and deformable scanning to balance fine detail and global context

📊 Key Advantages

Unlike older sequence models, Vision Mamba does not treat every part of the image equally. Instead, it uses its selective mechanism to prioritize meaningful features — edges, textures, and objects — while compressing or discarding irrelevant background.

1. Efficiency: 2–5× faster inference, lower FLOPs and memory footprint than comparable Transformers

2. Scalability: Linear complexity works smoothly for sequences of 10,000+ tokens without performance collapse

3. Flexibility: Easily hybridized with CNNs to retain strong local feature extraction, or paired with attention layers for maximum expressiveness

4. Performance: Matches or exceeds state-of-the-art results across classification, segmentation, restoration, and detection

📌 Where It’s Applied

• High/Mid-level: Classification, detection, segmentation, video understanding

• Low-level: Restoration, denoising, super-resolution

• 3D: Point clouds, reconstruction, volumetric medical data

• Vertical domains: Medical imaging, remote sensing, multimodal vision-language

🔭 Current Limitations & Future Directions

• Scanning dependency: Performance is sensitive to scanning order; predefined paths may not always match complex scene structure

• Local detail gap: Pure SSMs sometimes lack fine-grained detail compared to CNNs

• Stability: Larger pure Mamba models can face training instability, though hybrid designs mitigate this

Moving forward, the focus is on adaptive scanning, tighter integration with attention, and better pretraining strategies to scale Vision Mamba into a true foundation backbone.

Vision Mamba: A Comprehensive Survey and Taxonomy

🗺️ ABot‑Earth 0.5: Building the World in 3D with AI

Turn any ordinary satellite photo into a detailed, fly‑through 3D model of the Earth — in less than 10 minutes per square kilometre.

That is exactly what ABot‑Earth 0.5 , developed by Alibaba’s AMAP team, just published this month. The novel generative model formulated directly with the 3D Gaussian Splatting (3DGS) representation. It changes how we map, view, and interact with our planet.

🔍 Zoom from Space down to street level

• ABot‑Earth uses only standard satellite images as input.

• It generates 1 km² of detailed 3D terrain in under 10 minutes, and has consistent geometry and textures that match real‑world physics.

• It already covers 300+ cities across 190+ countries, and can fill in areas where no 3D scans exist at all.

• Using the LOD Quadtree Tile Hierarchy tiles, it has 6 levels of built‑in detail.

✅ Fast, low‑cost, and everywhere

Traditional 3D mapping needs expensive planes, LiDAR scanners, and months of work — and still only covers major cities. ABot‑Earth can generate 3D terrain for any spot on Earth, even remote or poorly mapped regions, at a tiny fraction of the usual cost.

Whether you’re curious to see what a remote area looks like in 3D, planning a project, or just love exploring our world, this is your new window to the Earth.

🤖 More than just pretty pictures

These models are simulation‑ready: PERFECT for:

• Training drones for autonomous delivery

• Planning cities - infrastructure design, and traffic simulation

• Generating instant 3D views for disaster response.

Where would you zoom in first? 🌍

>>> Check it out at: ABot Earth Studio · 即刻生成你的星球

Thought this may be of interest!

🦷 Can Your Smartphone Spot Cavities Before You Do? The Future of Dental Care Fits in Your Pocket

Tooth decay (dental caries) remains one of the most widespread chronic health issues worldwide, affecting people of every age group. Yet millions of people, especially those living in rural areas, only seek help once the pain becomes too much to handle. By then, damage is often severe, costly, and painful to treat.

📖 Read More on the recent report on global oral health by WHO here: ISBN: 978-92-4-006148-4 Global oral health status report: towards universal health coverage for oral health by 2030

BUT What if we could change that with something almost everyone already owns: a smartphone? 📱

Traditionally, dentists rely on visual checks, dental tools, and X-rays to find decay. But as we know, these methods only work well in skilled hands with good equipment.

A recent systematic review highlights how artificial intelligence (AI) combined with smartphone imaging are becoming a powerful tool for early cavity detection. Deep learning models—particularly YOLO variants, DenseNet, and MobileNetV3—are achieving impressive diagnostic accuracy when paired with smartphone-captured dental images.

Some standout findings:

• YOLOv4 hit 99% sensitivity and 94% specificity for caries detection

• DenseNet201 achieved 93% accuracy in classifying lesion severity

• MobileNetV3 delivered 90% accuracy while processing images in just 6 seconds with 90% accuracy — fast enough for real-time use

• A 2D-3D hybrid CNN reached 96.4% accuracy with 99.1% specificity, all while remaining portable and affordable

Even more exciting? One study found that parents could capture usable images of their children's teeth using smartphones, scoring high on usability scales. This opens doors for home-based screening and early childhood caries prevention in communities where dental visits are rare.

Smartphone-based AI tools could allow:

• Parents to take photos of their children’s teeth at home

• Community health workers to screen patients in rural areas

• Early detection of cavities before pain begins

• Faster referral to dentists when treatment is actually needed

Of course, at the moment, challenges remain in detecting very hidden lesions and along with dealing with limited datasets, but still the direction is clear.

👉 The authors of this review then point to exploring next-generation approaches like Vision Transformers, MedSAM segmentation models, and federated learning, as our next step forward, with much hope that this can be implemented in the real-world soon.

💬 Would you feel comfortable using an app to check your teeth or your child’s teeth at home? Share your thoughts below!

🤖 AI in Cervical Cancer Screening: China is Leading the Way with 24 AI-Products Approved.

Cervical cancer is largely preventable, yet it remains a leading cause of cancer death among women in low-resource settings.

A new systematic review of 35 studies paints a fascinating picture: 21 distinct AI-assisted cervical cancer screening technologies are now in play, with 24 products focusing on AI-assisted cytology examination already approved by China's NMPA.

🔬 Two Main Camps: Cytology vs. Colposcopy

AI-assisted cytology (17 technologies)

• In hospital settings, sensitivity ranges from 67.5% to 100% and specificity from 9.9% to 99.8%, with some technologies exceeding 90% overall accuracy.

• In community screening populations, the numbers tighten: 83.0–100.0% sensitivity and 74.2–99.9% specificity. Most studies report faster slide-reading times and improved pathologist performance.

AI-assisted colposcopy (4 technologies)

• As a standalone screening tool for high-grade lesions (CIN2+), sensitivity and specificity swing wildly: 43.6–95.5% and 51.8–93.9%, respectively.

• But when the AI is used in physician-assist mode, and sensitivity jumps to 95.1–97.5% while boosting consistency among less experienced colposcopists.

🌍 Meanwhile, the US Has One — But It's a Good One

Across the globe, the FDA has cleared only one AI-based system for cervical cytology screening: the Hologic Genius Digital Diagnostics System.

This standalone system represents the shift toward fully digital pathology workflows.

Here's the engineering flex: Instead of squinting at a single flat plane under a microscope, this system captures 14 focal planes in a single scan, building a 3D volumetric map of every cell on the slide. A deep learning algorithm then hunts through this digital depth, ranks the most suspicious cells, and serves them up as a curated gallery for the pathologist.

In FDA clinical trials across 4 sites with 1,994 slides, it demonstrated a statistically significant 7.5% improvement in sensitivity for HSIL+ and a 28% reduction in false negatives compared to manual microscopy. A real-world validation on 890 Pap tests confirmed it holds up outside the lab.

🔑 What this means going forward

AI is no longer just experimental. It’s becoming a structural component of cervical cancer prevention. The next phase will likely hinge on integrating these strengths: scalability, accuracy, and real-world validation.

👉 Check out this other review to get a comprehensive view on AI in cervical cancer screening.

🚨 The Sepsis Prediction Paradox: Our AI Models Are Getting More Explainable—But Are They Explaining the Right Things?

New systematic review reveals a critical gap between what ML models highlight and what clinicians actually need to see.

A PRISMA-guided review of 37 studies (2019–2025) offers encouraging news — along with other concerns:

📈 The Good News: Explainability Is Surging

Adoption of explainable ML methods is accelerating dramatically with ~67% greater odds per year. By 2023–2025, techniques like SHAP and LIME were substantially more common than in 2019–2021. SHAP dominated, used in 74% of explainability studies.

Better yet, explainability-friendly studies scored higher on methodological quality (82.6% vs. 76.9%), suggesting transparency and rigor increasingly go hand-in-hand.

⚠️ The Concerning Gap: Models Explain Vitals, Not Key Biomarkers

Across studies reporting feature importance, Heart rate topped the charts, appearing as a top-5 predictor in 11 of 25 studies. Temperature and respiratory rate also similarly dominated.

Yet clinically crucial sepsis biomarkers were strikingly absent:

• C-reactive protein (CRP): Used in only 4 of 37 studies

• Procalcitonin (PCT): Used in only 1 of 37 study

This is alarming because CRP and PCT are among the most extensively studied and clinically utilized sepsis biomarkers in practice. This led the authors to highlight a striking issue in "data architecture".

Complete CRP and PCT records in public datasets were rare: sometimes missing and were often times than not, never consistently sampled.

🏗️ Structural Barriers Only Run Deeper

• Reproducibility: Only 27% of papers shared code; 22% released datasets

• Generalizability: 27% used local-only data; 38% lacked external validation

• Real-world evidence: Only 2 of 37 studies were prospective—meaning 95% trained on retrospective data that can't establish causal relationships

In reality, 17 of 37 studies analyzed the same PhysioNet Challenge 2019 dataset, potentially creating an illusion of independent progress.

💡 The Path Forward

This review newly quantifies two parallel trends: rising explainability and persistent biological relevance gaps. Closing them requires:

1. Prospective designs capturing CRP, PCT, and other biomarkers at clinically meaningful time points

2. Explanations reflecting pathophysiology: not just generic feature rankings, but insights clinicians can act on

3. Transparency as standard: code and data sharing must become routine, not exceptional

4. Clinician-validated interpretability: SHAP values are only useful if they improve bedside decision-making

🔬 Bottom Line

We're at an inflection point. Sepsis prediction models are getting more transparent, but their explanations remain constrained by what the health records conveniently capture.

The gap between algorithmic performance and biological relevance won't close with better neural networks alone—it requires reimagining how we collect and validate clinical data for Machine Learning.

🧠 AI in Rehabilitation: The Hype vs. The Reality

New umbrella review of 32 systematic reviews by Abdalla et al. (2026) —here's what actually works and what's just noise.

🔑 The One Clear Win

Post-stroke upper limb recovery is thus far, the only area with reproducible evidence. A major network meta-analysis (Zhu et al., 2023) spanning 101 publications found robotic training + VR improves activity-level outcomes.

But here's the catch, any gains on impairment and daily independence vanish when assessors are blinded and practice dose is matched (Antoni et al., 2025)

• Low back pain: AI-assisted physiotherapy also show no significant advantage over usual care (Kapil et al., 2025)

This simply means that the AI 'advantage' in its current state would be better described as 'comparable but not better' than conventional therapy.

⚠️The Brutal Lab-to-Clinic Drop

In the real world, conditions are messy: different hospitals, different equipment, different patient populations - Causing the actual implementation of AI to see some setbacks.

• Brain-computer interfaces: ~99% offline accuracy → ~50% online in actual patients (Gutierrez-Martinez et al., 2021).

• Computer vision for movement tracking? Falls apart under real-world conditions (Sardari et al., 2023).

💡 The Bottom Line for AI in Rehabilitation

Only stroke imaging AI is deployment-ready today. Everything else? Can be used as capacity extenders, but not pure replacements.

The writers close off with a demand for newer AI papers to provide Proof of meaningful functional gains, external validation, and equity-by-design before any adoption.

🤖 AI Wearables for Managing Diabetes ~

In this paper - Fraser et al., 2025 - Researchers reviewed 60 studies on AI-powered wearables - continuous glucose monitors (CGMs), for Type 2 diabetes care. The tech is advancing fast—but major gaps remain before it reaches your doctor's office.

What's New:

Smarter Glucose Prediction: Deep learning models,particularly LSTMs and transformers, can predict blood sugar swings 30-120 minutes ahead of time by spotting patterns humans miss. One model achieved R² = 0.989 accuracy, essentially forecasting your glucose curve in real-time.

AI That Acts, Not Just Predicts: We're moving from forecasting to actual intervention. Reinforcement learning models, like Warren et al., 2024 now can suggest insulin doses autonomously, with results showing HbA1c drops by 1.3%.

Non-Invasive Hacks: Researchers are ditching finger pricks entirely—using smartphone photoplethysmography (PPG), ECG signals, even tear fluid infrared sensors to estimate glucose. One PPG-based deep learning model reached 90.6% accuracy for diabetes detection.

The AI Toolkit Deep-Dive:

>> LSTMs & GRUs (45% of studies): These recurrent neural networks are the workhorses—they process CGM time-series data sequentially, remembering past glucose values to predict future trends.

>> Transformers: They use "attention mechanisms" to weigh which past data points matter most — handling longer prediction windows across diverse populations better than LSTMs.

>> Temporal Fusion Transformers: Combine static data (age, BMI) with time-varying data (glucose, heart rate) in one model. Emerging but promising for personalized predictions.

>> Multi-Agent Reinforcement Learning (MARL): Multiple AI agents compete to identify which patient features (lab values, meds, demographics) most predict adverse glycemic events in hospitals. One study achieved 92.8% precision for hypoglycemia detection.

>> XGBoost & Random Forests: Traditional machine learning still dominates when doctors need interpretability. XGBoost ranked features by importance (SHAP values), showing clinicians why the AI flagged a patient as high-risk.

>> Deep-Ensemble Learning: Stacks multiple neural networks together—CNNs for pattern recognition, BiLSTMs for sequence memory, meta-learners for combining outputs.

Despite the great advances made, there are still some hard truths to tackle:

• 60% use "black box" models that doctors can't explain - the pattern is unknown.

• Only 7% report race/ethnicity data.

• Most studies are tiny (median 150 people).

• No long-term follow-ups exist - What would we see in 1 year's time?

The authors close off the review --- The tech works in pilots but we need diverse data, transparent models, and real-world validation before this becomes standard care.