BRN Discussion Ongoing

[CHIPS]

Intel got another views:

1. "They still talking about Akida. Should we commercialize Loihi?"
2. "They still talking about Akida. Should we buy BrainChip?"

Oh no no nooooo ... We are not for sale!!!

 

[Frangipani]

Uni researchers in the United Arab Emirates (from New York University Abu Dhabi and Khalifa University, Abu Dhabi) have experimented with Akida for neuromorphic AI-based robotics - the field our CTO Dr. Tony Lewis is an expert in:

https://arxiv.org/pdf/2404.03325.pdf

View attachment 60342
View attachment 60343

View attachment 60344

Rachmad Vidya Wicaksana Putra and Muhammad Shafique, two of the New York University (NYU) Abu Dhabi eBrain Lab researchers, whose above paper on experimenting with Akida for neuromorphic AI-based robotics was published almost exactly a year ago, appear to be extremely enamoured with our neuromorphic processor! They co-authored another paper on AKD1000 with their colleague Pasindu Wickramasinghe that was published yesterday and that also got accepted at the International Joint Conference on Neural Networks (IJCNN), June 30th - July 5th, 2025 in Rome, Italy.


“Neuromorphic Processors
1) Overview: The energy efficiency potentials offered by
SNNs can be maximized by employing neuromorphic hard- ware processors [22]. In the literature, several processors have been proposed, and they can be categorized as research and commodity processors. Research processors refer to neuro- morphic chips that are designed only for research and not commercially available, hence access to these processors is limited. Several examples in this category are SpiNNaker, NeuroGrid, IBM’s TrueNorth, and Intel’s Loihi [4].
Meanwhile, commodity processors refer to neuromorphic chips that
are available commercially, such as BrainChip’s Akida [7] and SynSense’s DYNAP-CNN [8]. In this work, we consider the Akida processor as it supports on-chip learning for SNN fine-tuning, which is beneficial for adaptive edge AI systems [7].

(…)


E. Further Discussion
It is important to compare neuromorphic-based solutions against the state-of-the-art ANN-based solutions, which typically employ conventional hardware platforms, such as CPUs, GPUs, and specialized accelerators (e.g., FPGA or ASIC). To ensure a fair comparison, we select object recognition as the application and YOLOv2 as the network, while considering performance efficiency (FPS/W) as the comparison metric. Summary of the comparison is provided in Table I, and it clearly shows that our Akida-based neuromorphic solution achieves the highest performance efficiency. This is due to the sparse spike-driven computation that is fully exploited by neuromorphic processor, thus delivering highly power/energy-efficient SNN processing. Moreover, our Akida-based neuromorphic solution also offers an on-chip learning capability, which gives it further advantages over the other solutions. This comparison highlights the immense potentials of neuromor- phic computing for enabling efficient edge AI systems.

VI. CONCLUSION
We propose a novel design methodology to enable efficient SNN processing on commodity neuromorphic processors. It is evaluated using a real-world edge AI system implementation with the Akida processor. The experimental results demonstrate that, our methodology leads the system to achieve high performance and high energy efficiency across different applications. It achieves low latency of inference (i.e., less than 50ms for image classification, less than 200ms for real- time object detection in video streaming, and less than 1ms for keyword recognition) and low latency of on-chip learning (i.e., less than 2ms for keyword recognition), while consuming less than 250mW of power. In this manner, our design methodology potentially enables ultra-low power/energy design of edge AI systems for diverse application use-cases.”

https://arxiv.org/pdf/2504.00957v1

View attachment 81147
View attachment 81148 View attachment 81149 View attachment 81150 View attachment 81151 View attachment 81152 View attachment 81154 View attachment 81155

Another paper containing an Akida reference by two of the New York University (NYU) Abu Dhabi eBRAiN Lab researchers, Alberto Marchisio and Muhammad Shafique, came out yesterday.
It is titled “Neuromorphic Computing for Embodied Intelligence in Autonomous Systems: Current Trends, Challenges, and Future Directions”.

While Shafique (who is Director of the eBRAIN Lab), Marchisio and other colleagues had previously published in detail about evaluating Akida for neuromorphic AI-based robotics (for the first paper above in collaboration with two other Abu Dhabi-based researchers from Khalifa University), this new paper - as the title suggests - is more foundational and mentions Akida only once when listing examples of available neuromorphic hardware. It also addresses “key challenges and open research questions”.


Under ACKNOWLEDGMENT it says:

“This work was supported in parts by the NYUAD Center for Cyber Security (CCS), funded by Tamkeen under the NYUAD Research Institute Award G1104, and the NYUAD Center for Artificial Intelligence and Robotics (CAIR), funded by Tamkeen under the NYUAD Research Institute Award CG010.”

NYU Abu Dhabi - Tamkeen

Tamkeen LLC is an Abu Dhabi company that partners with leading local and international institutions to deliver projects that enrich the UAE’s social, cultural and educational landscape.

www.tamkeenuae.com

“Tamkeen is New York University’s partner in the UAE, enabling NYU Abu Dhabi’s ongoing development as a local and global center of academic excellence and community engagement.”

---

https://arxiv.org/pdf/2507.18139v1

 

[Rach2512]

vBIC: A Virtualization Framework for Brain-Inspired Chips | NeuromorphicCore.AI posted on the topic | LinkedIn

Beyond GPUs: vBIC’s Game-Changing Virtualization for Next-Gen Brain-Mimicking AI Chips A study by Dahu Feng and Rong Zhao, presented at the 2025 IEEE Electron Devices Technology & Manufacturing Conference, introduces vBIC, a virtualization framework for Brain-Inspired Chips (BICs) like...

www.linkedin.com

Your Phone Could Soon Diagnose Skin Cancer—Thanks to a New Kind of AI - NeuromorphicCore.ai

Discover how QANA, a new brain-inspired AI, is bringing fast, private, and accurate skin cancer diagnosis directly to portable devices. Learn how this innovation helps close healthcare gaps, addressing shortages of dermatologists and improving outcomes, especially for underserved communities.

neuromorphiccore.ai

 

[BrainShit]

Hailo-10H: Second-Generation AI Accelerator Now Available | Hailo posted on the topic | LinkedIn

We are very excited to announce the commercial availability of the Hailo-10H! Our second-generation AI accelerator - featuring powerful generative AI capabilities, enabling seamless execution of large language models (LLMs), vision-language models (VLMs), and other generative AI models entirely...

www.linkedin.com

How does this compare to Akida?

The Chip Hailo-10H and Akida have distinct core architectures and target somewhat different edge AI use cases, but they do compete in the broader edge AI market where power efficiency and real-time intelligence matter.

So, the Hailo-10H is more of a classic, powerful AI accelerator for complex neural networks, including generative AI at the edge, with good energy efficiency. Akida specializes in particularly energy-efficient, neuromorphic real-time AI for sensors and continuous data processing with a completely different architecture.

Hailo-10H would probably be better suited for complex AI models (e.g., LLMs), while Akida 2.0 has advantages in highly energy-efficient, adaptive sensor applications. It is optimized for higher-performance AI inference, including generative AI models (LLMs, Vision-Language Models, Stable Diffusion) and multi-modal workloads on edge devices with M.2 integration. ... they also good for ADAS, because it is automotive-qualified to AEC-Q100 Grade 2 standards and is aimed at automotive designs with 2026 start of production.

Akida focuses on ultra-low power, always-on event-based AI, priority in domains such as medical, environmental, and industrial sensors where power saving is paramount.

Both chips can be applied to real-time vision analytics tasks. While Hailo-10H handles this with high throughput and generative AI capability, Akida excels at ultra-low power, always-on vision pattern recognition.

Both target embedded systems and IoT deployments that demand AI at the edge with minimal latency and energy footprint, such as industrial monitoring, smart infrastructure, and real-time anomaly detection.

They overlap in edge vision and sensor-related AI applications but serve different levels of AI complexity and power requirements, making them complementary yet partially competing solutions for low-latency AI at the edge.

Source: https://hailo.ai/company-overview/n...-accelerator-with-generative-ai-capabilities/
Source: https://hailo.ai/products/ai-accelerators/hailo-10h-ai-accelerator/

 

[Frangipani]

Alican Kiraz has meanwhile received the AKD1000 Mini PCIe Board he had ordered to adapt SNN and GNN (= graph neural network) models for use in robotic arm and autonomous drone projects.

I’m not sure why he calls this an “AKD1000 PCIeM.2 module”, though, as this is evidently not the AKD1000 M.2 form factor that came out in January.

Saying that, we wouldn’t mind him purchasing both form factors, would we?

View attachment 88841


Google’s translation from Turkish to English:

“Hello friends! , I've received my BrainChipAkida AKD1000 PCIe neuromorphic chip, which I purchased to adapt SNN and GNN models for use in my Robotic Arm and Autonomous Drone projects.

As you know, today's AI projects require high-parameter Transformer-based models and GPUs to reduce the latency between real-time perception and decision-making. However, achieving this with a milliwatt power budget poses a critical engineering challenge for these systems.

To this end, I integrated the BrainChipAkida AKD1000 PCIeM.2 module into the RaspberryPi5's PCIeGen2x1 lane with a Waveshare PCIe x1, creating a fully edge-driven, event-driven Spiking Neural Network accelerator. This reduces average power consumption by 10x compared to traditional models and accelerates perception-to-decision latency by 5x.

In my research plan, I hope to develop excellent examples of hybrid GNN and SNN architectures using Causal Reasoning in energy-constrained edge applications. Friends pursuing related academic research can easily connect with me on LinkedIn. ”

Here is another post by Istanbul-based Alican Kiraz, Head of Cyber Defense Center @Trendyol Group, which links to a Medium article written by him titled “Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs” that gives us a little theoretical background regarding the PoC he is envisaging, for which he will be utilising his newly purchased Akida Mini PCIe Board:

TR | Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs | Alican Kiraz

TR & EN | Selamlar dostlar 🙏 uzun süredir yapay zeka karar-destek sistemleri üzerine kafa yorarken, Hibrit SNN & Causal Reasoning yapay zeka mimari ve yaklaşımıyla Askeri UAV'lerin nasıl evrilebileceği üzerine yeni makalemde teknik ve teorik düzeyde beyin fırtınası yaptım. Özellikle; -...

www.linkedin.com

2yDM?



(English version)

EN | Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs​

Alican Kiraz


Follow
7 min read
·
1 hour ago


I created it with Midjourney


Although the initial emergence and areas of interest for Unmanned Aerial Vehicles (UAVs) were diverse, their most prominent and widespread use today is undoubtedly in military applications. With the rise of Artificial Intelligence systems and technologies over the past five years, this interest has undergone a significant transformation. Platforms such as the Anduril YFQ-44 and Shield AI MQ-35 V-BAT are now capable of executing missions autonomously, being tasked and coordinated by decision-support AI systems.

Artificial Intelligence is increasingly employed in critical missions such as Intelligence, Surveillance, and Reconnaissance (ISR), autonomous navigation, and target identification. However, the deployment of such AI solutions on the field presents significant challenges — both in terms of mission effectiveness and operational security, as well as resource constraints. These limitations often lead to hesitation and caution in real-world use.


You can access the Turkish version of my article here:

TR | Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs​

İnsansız Hava Araçlarının ilk çıkış ve ilgi noktaları farklı olsada şuan dünyada en çok kullanıldıkları ve ses…​

alican-kiraz1.medium.com

One of the most critical challenges lies in the high computational and energy demands of conventional Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs), which are widely used in computer vision and autonomous navigation. Additionally, transformer-based Large Language Models (LLMs) operating in an embedded, agentic architecture for decision support further exacerbate these requirements. These model architectures typically necessitate high-performance GPUs or ruggedized onboard supercomputers integrated into the UAV system.

As you may recall, in my previous article, I examined Russia’s Geran-2 drone, which operates with an onboard Nvidia Jetson Orin module. You may also want to revisit that article for further insights.

TR | Nvidia Jetson Orin-Powered AI Kamikaze Drones: Beating GPS Jamming on the Battlefield​

Shahed-136’nın MS serisi versiyonları, İran ve Rusya arasındaki askeri iş birliğinin çıktısı olarak görülüyor. Bu…​

alican-kiraz1.medium.com

EN | Nvidia Jetson Orin-Powered AI Kamikaze Drones: Beating GPS Jamming on the Battlefield​

The MS series variants of the Shahed-136 are seen as a product of military cooperation between Iran and Russia. The…​

alican-kiraz1.medium.com

The integration of Hybrid Spiking Neural Network (SNN) and Causal Reasoning (CR) architectures offers significant advantages in terms of reduced energy consumption, lower latency for real-time decision-making, and increased system reliability. Let us now conceptualize such a system — CR-SNN — built upon a hybrid architecture combining SNN and CR.

This hybrid system leverages the biologically inspired, event-driven nature of SNNs to achieve energy efficiency and computational speed, especially when deployed on neuromorphic hardware platforms. Meanwhile, the Causal Reasoning (CR) component — grounded in Judea Pearl’s causal inference framework, including Structural Causal Models (SCMs) and do-calculus — enables UAVs to develop enhanced situational awareness in complex scenarios. It empowers autonomous systems to perform tasks such as target tracking more effectively and to produce more generalizable and interpretable outputs, even in novel situations not encountered during training.

Moreover, although current AI-based systems can exhibit high accuracy in controlled environments, they tend to lose output quality and reliability when exposed to the complexities of real-world civilian or battlefield environments. In such conditions, the probability of incorrect responses increases significantly. Especially under dynamic circumstances or in the presence of adversarial attacks targeting the UAV, reactive maneuvers become increasingly challenging.

For instance, distinguishing between civilians and combatants in dense urban environments or detecting camouflaged targets remains a considerable challenge for current AI models. Furthermore, the inability of deep learning models — often considered “black boxes” — to provide transparent reasoning for their decisions undermines the trustworthiness and accountability of weaponized UAVs in military operations.

In addition, processing high-bandwidth sensor data in real-time — such as high-resolution visible & thermal imagery or Synthetic Aperture Radar (SAR) outputs — imposes substantial computational burdens on existing architectures, increasing latency and slowing down decision-making.

Even on popular edge computing platforms like the NVIDIA Jetson Orin Nano, optimized models such as YOLOv8n typically operate at around 10 frames per second (FPS), which falls short of the 30 FPS threshold required for real-time navigation. Additionally, the need for active thermal management and an average power consumption of around 20 watts pose significant challenges concerning weight, energy constraints, and thermal design requirements on UAV platforms.

Source: https://www.seeedstudio.com/blog/20...eYa20KsdYMhX0iJB8GfkWnDoTWWnMlo74zgTLxHoOA3_A

Source: https://dataphoenix.info/a-guide-to-the-yolo-family-of-computer-vision-models/


Furthermore, transmitting data to the C4 (Command, Control, Communication, and Computers) center for decision-support purposes and relaying mission directives back to the UAV can introduce end-to-end latencies exceeding 150 milliseconds. Such delays risk causing the UAV to miss its optimal decision-making window, especially in time-sensitive scenarios.

Beyond performance metrics, next-generation military UAVs are expected to fulfill a range of advanced capabilities. These include: achieving higher levels of autonomy, maintaining superior situational awareness in complex and uncertain environments, making faster and more accurate decisions, extending operational range and flight endurance, and participating in joint missions in coordination with manned combat aircraft. Meeting these requirements is critical for future air combat effectiveness.

To address these demands, we must go beyond current AI architectures. There is a pressing need for solutions that are more efficient, faster, more reliable, and fundamentally smarter. Specifically, we require AI systems — both in software and hardware — that combine low power consumption, low latency, and high performance, while also possessing the ability to comprehend and act upon complex causal relationships within dynamic operational contexts.

Source: https://exoswan.com/is-neuromorphic-computing-the-future-of-ai


At this point, the combination of approaches such as Spiking Neural Networks (SNNs) and Causal Reasoning (CR) presents a promising pathway to overcome these challenges. These hybrid SNN-CR models hold significant potential to deliver strategic advantages for next-generation military UAVs.

Thanks to the temporally encoded and sparsely activated communication mechanisms of SNNs — modeled after biological neural systems — these networks exhibit significantly lower energy consumption compared to traditional Artificial Neural Networks (ANNs), while also enabling faster inference times under real-time operational constraints.

Source: https://peerj.com/articles/cs-2549/


In a study conducted by researchers at ETH Zurich, it was demonstrated that SNNs consumed six times less energy than a CNN when performing an obstacle avoidance task. Moreover, the SNN model operated with an average inference latency of just 2.4 milliseconds, showcasing its suitability for highly dynamic scenarios.

For further details, see:

https://www.sciencedirect.com/science/article/pii/S0925231223010081

Another critical aspect is Causal Reasoning (CR). When integrated — particularly through the framework proposed by Judea Pearl — CR enables AI systems to reason not only through learned correlations, but also through underlying causal relationships. This allows the model to perform robustly and generalize more effectively in environments that deviate from the training data distribution or when faced with unforeseen conditions.

To illustrate with a practical case: in a mission involving the detection of a human group as a potential target, conventional models may focus on features such as camouflage color, presence of vehicles, or group size. In contrast, a causally-informed model could prioritize more nuanced and domain-relevant cues — such as body height, shadows, camouflage patterns, military-style marching formations, or spacing intervals — along with distinct features like carried weapons, by leveraging appropriate sensor modalities and reasoning mechanisms.

Image Source: https://www.newscientist.com/

The component I wish to emphasize most — and which I have long researched for integration into this powerful architecture — is Causal Reasoning (CR). CR refers to the process by which an AI system understands and analyzes causal relationships between events, then uses this understanding to make decisions or support decision-making based on those causal inferences.

What distinguishes CR from conventional approaches? Traditional machine learning and large language models (LLMs) primarily rely on statistical correlations or correlation-based inference. By integrating CR, we aim to enable the model to reason through causality — essentially allowing the system to ask itself “why” and to answer that question autonomously. This line of reasoning is grounded in the foundational work of Judea Pearl, who introduced a formal framework for causality. Pearl’s model is built upon concepts such as Structural Causal Models (SCMs), causal graphs, and do-calculus.

So, what can CR contribute to a UAV system? For instance, it can be highly effective in diagnosing system failures. Imagine a scenario where a UAV encounters multiple issues such as GPS malfunction, engine anomalies, or communication loss. A causal model can represent these as nodes in a causal graph, with the edges denoting the causal links between them. This enables the system to perform counterfactual or interventional reasoning, such as:

“The GPS has failed, an engine fault has also been reported, but the propellers have not decelerated and location signals are still being transmitted — therefore, the fault may lie with the sensor supervisor.”

Or:

“GPS failure occurred, communication is down, and I am experiencing deviations in altitude and trajectory. I must immediately identify a safe emergency landing zone!”

Such inferential capabilities are powerful, aren’t they?

Source: https://www.nature.com/articles/s42256-023-00744-z


I hope this overview has been insightful. In the next section, we will delve into the technical details and adaptation strategies. Additionally, now that the neuromorphic chip — BrainChip Akida — has arrived, I will be working towards establishing a proof of concept (PoC) based on the proposed architecture.

See you soon!
Drones
Military
AI
Neural Networks
Military Drones

Written by Alican Kiraz​

1.4K followers
·411 following
Head of Cyber Defense Center @Trendyol | CSIE | CSAE | CCISO | CASP+ | OSCP | eCIR | CPENT | eWPTXv2 | eCDFP | eCTHPv2 | OSWP | CEH Master | Pentest+ | CySA+
Follow

 

[TheDrooben]

#neuromorphicengineering #aerospaceinnovation #bioinspiredflight #aiinaerospace #startupideas #futureofaviation #linkedinconnections #nextgenengineering #studentfounder #braininspiredtech #uavs… | Gayathri Tekumudi

Brain-Inspired Aircraft? Why Neuromorphic Engineering Could Take Flight Merging Aerospace with Artificial Intelligence LITERALLY What if an aircraft could think like a bird? React like a brain not a computer? Welcome to the world of Neuromorphic Aerospace Systems an idea that’s years ahead...

www.linkedin.com

Happy as Larry

 

[manny100]

Here is another post by Istanbul-based Alican Kiraz, Head of Cyber Defense Center @Trendyol Group, which links to a Medium article written by him titled “Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs” that gives us a little theoretical background regarding the PoC he is envisaging, for which he will be utilising his newly purchased Akida Mini PCIe Board:

TR | Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs | Alican Kiraz

TR & EN | Selamlar dostlar 🙏 uzun süredir yapay zeka karar-destek sistemleri üzerine kafa yorarken, Hibrit SNN & Causal Reasoning yapay zeka mimari ve yaklaşımıyla Askeri UAV'lerin nasıl evrilebileceği üzerine yeni makalemde teknik ve teorik düzeyde beyin fırtınası yaptım. Özellikle; -...

www.linkedin.com

View attachment 88909 2yDM?



(English version)

EN | Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs​

Alican Kiraz


Follow
7 min read
·
1 hour ago


I created it with Midjourney


Although the initial emergence and areas of interest for Unmanned Aerial Vehicles (UAVs) were diverse, their most prominent and widespread use today is undoubtedly in military applications. With the rise of Artificial Intelligence systems and technologies over the past five years, this interest has undergone a significant transformation. Platforms such as the Anduril YFQ-44 and Shield AI MQ-35 V-BAT are now capable of executing missions autonomously, being tasked and coordinated by decision-support AI systems.

Artificial Intelligence is increasingly employed in critical missions such as Intelligence, Surveillance, and Reconnaissance (ISR), autonomous navigation, and target identification. However, the deployment of such AI solutions on the field presents significant challenges — both in terms of mission effectiveness and operational security, as well as resource constraints. These limitations often lead to hesitation and caution in real-world use.


You can access the Turkish version of my article here:

TR | Hybrid SNN Models and Architectures for Causal Reasoning in Next-Generation Military UAVs​

İnsansız Hava Araçlarının ilk çıkış ve ilgi noktaları farklı olsada şuan dünyada en çok kullanıldıkları ve ses…​

alican-kiraz1.medium.com

One of the most critical challenges lies in the high computational and energy demands of conventional Artificial Neural Networks (ANNs) and Convolutional Neural Networks (CNNs), which are widely used in computer vision and autonomous navigation. Additionally, transformer-based Large Language Models (LLMs) operating in an embedded, agentic architecture for decision support further exacerbate these requirements. These model architectures typically necessitate high-performance GPUs or ruggedized onboard supercomputers integrated into the UAV system.

As you may recall, in my previous article, I examined Russia’s Geran-2 drone, which operates with an onboard Nvidia Jetson Orin module. You may also want to revisit that article for further insights.

TR | Nvidia Jetson Orin-Powered AI Kamikaze Drones: Beating GPS Jamming on the Battlefield​

Shahed-136’nın MS serisi versiyonları, İran ve Rusya arasındaki askeri iş birliğinin çıktısı olarak görülüyor. Bu…​

alican-kiraz1.medium.com

EN | Nvidia Jetson Orin-Powered AI Kamikaze Drones: Beating GPS Jamming on the Battlefield​

The MS series variants of the Shahed-136 are seen as a product of military cooperation between Iran and Russia. The…​

alican-kiraz1.medium.com

The integration of Hybrid Spiking Neural Network (SNN) and Causal Reasoning (CR) architectures offers significant advantages in terms of reduced energy consumption, lower latency for real-time decision-making, and increased system reliability. Let us now conceptualize such a system — CR-SNN — built upon a hybrid architecture combining SNN and CR.

This hybrid system leverages the biologically inspired, event-driven nature of SNNs to achieve energy efficiency and computational speed, especially when deployed on neuromorphic hardware platforms. Meanwhile, the Causal Reasoning (CR) component — grounded in Judea Pearl’s causal inference framework, including Structural Causal Models (SCMs) and do-calculus — enables UAVs to develop enhanced situational awareness in complex scenarios. It empowers autonomous systems to perform tasks such as target tracking more effectively and to produce more generalizable and interpretable outputs, even in novel situations not encountered during training.

Moreover, although current AI-based systems can exhibit high accuracy in controlled environments, they tend to lose output quality and reliability when exposed to the complexities of real-world civilian or battlefield environments. In such conditions, the probability of incorrect responses increases significantly. Especially under dynamic circumstances or in the presence of adversarial attacks targeting the UAV, reactive maneuvers become increasingly challenging.

For instance, distinguishing between civilians and combatants in dense urban environments or detecting camouflaged targets remains a considerable challenge for current AI models. Furthermore, the inability of deep learning models — often considered “black boxes” — to provide transparent reasoning for their decisions undermines the trustworthiness and accountability of weaponized UAVs in military operations.

In addition, processing high-bandwidth sensor data in real-time — such as high-resolution visible & thermal imagery or Synthetic Aperture Radar (SAR) outputs — imposes substantial computational burdens on existing architectures, increasing latency and slowing down decision-making.

Even on popular edge computing platforms like the NVIDIA Jetson Orin Nano, optimized models such as YOLOv8n typically operate at around 10 frames per second (FPS), which falls short of the 30 FPS threshold required for real-time navigation. Additionally, the need for active thermal management and an average power consumption of around 20 watts pose significant challenges concerning weight, energy constraints, and thermal design requirements on UAV platforms.

Source: https://www.seeedstudio.com/blog/20...eYa20KsdYMhX0iJB8GfkWnDoTWWnMlo74zgTLxHoOA3_A

Source: https://dataphoenix.info/a-guide-to-the-yolo-family-of-computer-vision-models/


Furthermore, transmitting data to the C4 (Command, Control, Communication, and Computers) center for decision-support purposes and relaying mission directives back to the UAV can introduce end-to-end latencies exceeding 150 milliseconds. Such delays risk causing the UAV to miss its optimal decision-making window, especially in time-sensitive scenarios.

Beyond performance metrics, next-generation military UAVs are expected to fulfill a range of advanced capabilities. These include: achieving higher levels of autonomy, maintaining superior situational awareness in complex and uncertain environments, making faster and more accurate decisions, extending operational range and flight endurance, and participating in joint missions in coordination with manned combat aircraft. Meeting these requirements is critical for future air combat effectiveness.

To address these demands, we must go beyond current AI architectures. There is a pressing need for solutions that are more efficient, faster, more reliable, and fundamentally smarter. Specifically, we require AI systems — both in software and hardware — that combine low power consumption, low latency, and high performance, while also possessing the ability to comprehend and act upon complex causal relationships within dynamic operational contexts.

Source: https://exoswan.com/is-neuromorphic-computing-the-future-of-ai


At this point, the combination of approaches such as Spiking Neural Networks (SNNs) and Causal Reasoning (CR) presents a promising pathway to overcome these challenges. These hybrid SNN-CR models hold significant potential to deliver strategic advantages for next-generation military UAVs.

Thanks to the temporally encoded and sparsely activated communication mechanisms of SNNs — modeled after biological neural systems — these networks exhibit significantly lower energy consumption compared to traditional Artificial Neural Networks (ANNs), while also enabling faster inference times under real-time operational constraints.

Source: https://peerj.com/articles/cs-2549/


In a study conducted by researchers at ETH Zurich, it was demonstrated that SNNs consumed six times less energy than a CNN when performing an obstacle avoidance task. Moreover, the SNN model operated with an average inference latency of just 2.4 milliseconds, showcasing its suitability for highly dynamic scenarios.

For further details, see:

https://www.sciencedirect.com/science/article/pii/S0925231223010081

Another critical aspect is Causal Reasoning (CR). When integrated — particularly through the framework proposed by Judea Pearl — CR enables AI systems to reason not only through learned correlations, but also through underlying causal relationships. This allows the model to perform robustly and generalize more effectively in environments that deviate from the training data distribution or when faced with unforeseen conditions.

To illustrate with a practical case: in a mission involving the detection of a human group as a potential target, conventional models may focus on features such as camouflage color, presence of vehicles, or group size. In contrast, a causally-informed model could prioritize more nuanced and domain-relevant cues — such as body height, shadows, camouflage patterns, military-style marching formations, or spacing intervals — along with distinct features like carried weapons, by leveraging appropriate sensor modalities and reasoning mechanisms.

Image Source: https://www.newscientist.com/

The component I wish to emphasize most — and which I have long researched for integration into this powerful architecture — is Causal Reasoning (CR). CR refers to the process by which an AI system understands and analyzes causal relationships between events, then uses this understanding to make decisions or support decision-making based on those causal inferences.

What distinguishes CR from conventional approaches? Traditional machine learning and large language models (LLMs) primarily rely on statistical correlations or correlation-based inference. By integrating CR, we aim to enable the model to reason through causality — essentially allowing the system to ask itself “why” and to answer that question autonomously. This line of reasoning is grounded in the foundational work of Judea Pearl, who introduced a formal framework for causality. Pearl’s model is built upon concepts such as Structural Causal Models (SCMs), causal graphs, and do-calculus.

So, what can CR contribute to a UAV system? For instance, it can be highly effective in diagnosing system failures. Imagine a scenario where a UAV encounters multiple issues such as GPS malfunction, engine anomalies, or communication loss. A causal model can represent these as nodes in a causal graph, with the edges denoting the causal links between them. This enables the system to perform counterfactual or interventional reasoning, such as:


Or:


Such inferential capabilities are powerful, aren’t they?

Source: https://www.nature.com/articles/s42256-023-00744-z


I hope this overview has been insightful. In the next section, we will delve into the technical details and adaptation strategies. Additionally, now that the neuromorphic chip — BrainChip Akida — has arrived, I will be working towards establishing a proof of concept (PoC) based on the proposed architecture.

See you soon!
Drones
Military
AI
Neural Networks
Military Drones

Written by Alican Kiraz​

1.4K followers
·411 following
Head of Cyber Defense Center @Trendyol | CSIE | CSAE | CCISO | CASP+ | OSCP | eCIR | CPENT | eWPTXv2 | eCDFP | eCTHPv2 | OSWP | CEH Master | Pentest+ | CySA+
Follow

Great post Frangipani.
" Similarly, components previously unseen in Iranian Shahed models — such as the Jetson Orin AI computer and several infrared (IR) camera modules recently announced by Russia — have been integrated into the MS series. This level of sophistication also necessitates the formation of a global procurement network: U.S.-made Nvidia chips and Japanese-manufactured camera sensors have reportedly reached Iran and Russia through indirect channels. In fact, reports from 2023 mention that over $17 million worth of Nvidia products entered Russia via gray-market routes, including Jetson Orin modules."

How long before Russia, Iran, China etc manage to get their hands on AKIDA one way or another. Pretty easy.
The US DOD would be aware of this. They would not want to be going up with traditional AI against Neuromorphic Edge AI.
Easy to see why the DOD is transitioning to Neuromorphic Edge AI.
US software quality will be the lynchpin for AKIDA success.
If it comes to AKIDA Vs AKIDA in battlefield the best software will win out.,ie software produced by RTX, Lockheed-Martin etc.
AKIDA is the Brains and the software does the thinking.

 

[Rach2512]

Why AI won't replace humans: Energy efficiency and the future of work | Joris Bastien posted on the topic | LinkedIn

This is why humans won't be replaced so easily by AI. This is about energy efficiency. The human brain is incredibly more efficient than any AI farm with NVidia chips, and costs much less in energy to get to similar results. (although I know a few whose energy consumption is a total waste 😁)...

www.linkedin.com

 

[Frangipani]

No doubt Frontgrade Gaisler will be spruiking GR801 - their first neuromorphic AI solution for space, powered by Akida - during their upcoming Asia Tour through India, Singapore and Japan:

#gaisler2025asiatour #spacetech | Frontgrade Gaisler

We are touring across Asia this year! Frontgrade Gaisler is glad to engage directly with partners, agencies, and customers across India, Singapore, and Japan. You can find us here: 🇮🇳 India July 21-22: Ahmedabad July 23: Bangalore 🇸🇬 Singapore July 25 🇯🇵 Japan July 28 - August 1: SPEXA –...

www.linkedin.com

Yesterday, Frontgrade Gaisler visited ST Engineering Satellite Systems (SatSys) on the Singapore leg of their 2025 Asia Tour.

SatSys is a joint venture between ST Engineering, DSO National Laboratories and NTU (Nanyang Technological University).

#gaislerasiatour2025 #spacecomputing | Frontgrade Gaisler

#GaislerAsiaTour2025 is now in Singapore! Our team is in Singapore today to visit ST Engineering and connect with leading experts in aerospace and defense. Thank you to everyone for the warm welcome. It's been a pleasure to exchange ideas and explore opportunities for collaboration. We’re...

www.linkedin.com

ST Engineering - Wikipedia

en.m.wikipedia.org

ST Engineering’s largest shareholder with about 50.9%* of total issued shares as at 31 December 2024 is Temasek Holdings, a global investment company wholly owned by the Singapore Government.

*Total shareholdings of Temasek Holdings and Vestal Investments.

FAQ | ST Engineering

Have a question on ST Engineering regarding the stock or dividends? Have it answered here in our list of frequently asked questions. Read answers to frequently asked questions relating to bond redemption here.

www.stengg.com

Temasek Holdings - Wikipedia

en.m.wikipedia.org

Satellite Systems | ST Engineering

We have complete manufacturing and assembly capabilities dedicated for satellite design, testing and manufacturing. Learn more.

www.stengg.com

About Us | Space & Satellite Solutions Provider Singapore | SatSys

www.satsys.com.sg

Other SatSys partners are STE-GI (ST Engineering Geo-Insights Pte. Ltd.), an end-to-end satellite downstream services provider as well as CRISP, the NUS (National University of Singapore) Centre for Remote Imaging, Sensing and Processing, which was established with funding from A*STAR, Singapore’s Agency for Science, Technology and Research.

 

[FuzM]

Good thing that Singapore is aware of Brainchip for a few years now.

Page 14

Infocomm Media Development Authority https://www.imda.gov.sg PDF ANNEXES A-8

 

[Bravo]

 

[cosors]

Could this be interesting or is it impossible because 2nm is too small?

"Google Tensor
Pixel 11’s Tensor G6 will reportedly use 2nm process at TSMC, no longer lagging behind
| Jun 24 2025 - 5:30 am PT

According to a new report, Google is planning to jump to TSMC’s 2nm process for Tensor G6 faster than it has in prior years.

The first Tensor chip in 2021’s Pixel 6 was built on a 5nm process at Samsung, with the following year’s Tensor G2 chip using the same process despite the same year’s Snapdragon 8 Gen 2 moving to a 4nm process. It took Google a year to catch up, with Tensor G3 being the first on a 4nm process, and the current Tensor G4 following suit despite current Qualcomm and MediaTek chips being built on a 3nm process. Google’s move to TSMC later this year with Tensor G5 will mark the transition to a 3nm process.

But, according to a new report, Google isn’t looking to lag behind next year.

It’s being reported (via analyst @dnystedt) that Google will leverage TSMC’s 2nm process for Tensor G6 on the Pixel 11 series in 2026. If true, that marks a shift in how Google has kept up with the competition, as even the next Snapdragon 8 Elite chipset is likely going to still be using a 3nm process. Somewhat unbelievably, that could mean that Google could beat its competitors to the punch in this regard, as a 2nm Snapdragon chip wouldn’t arrive until a few months after Tensor G6. Notably, though, there is word that Qualcomm might use TSMC for a 3nm version of the upcoming Snapdragon flagship, while Samsung may produce a 2nm version specific to Galaxy devices.

The focus, for now, should absolutely be on Google’s initial switch to TSMC’s 3nm process later this year with Tensor G5, but it’s interesting nonetheless to see that Google is looking for further advancements this quickly."

Pixel 11's Tensor G6 will reportedly use 2nm process at TSMC, no longer lagging behind

According to a new report, Google is planning to jump to TSMC’s 2nm process for Tensor G6 faster than it...

9to5google.com

 

[Diogenese]

This bloke is a professor at Carnegie Mellon and founded spin-off, Efficient Computer, which claims 100 * energy efficiency.

Brandon LuciaBrandon Lucia • 3rd+Premium • 3rd+Professor at Carnegie Mellon University, Co-Founder & CEO at Efficient ComputerProfessor at Carnegie Mellon University, Co-Founder & CEO at Efficient Computer2mo • 2 months ago • Visible to anyone on or off LinkedIn
Follow

I recently spoke to Embedded Insiders about how Efficient Computer is shaping the future of computing - for AI/ML, but also much more. We discuss the criticality of energy efficiency and the absolute need for general-purpose architectures, as opposed to over-specialized hardware. Generality keeps programmers happy, supports innovation, and actually, somewhat surprisingly, *increases* end-to-end efficiency (ask me about Amdahl's Law!)

We also talked about the future of the computing workforce and the next generation of computer engineers. Implicit in that discussion was the idea that major initiatives like the National Science Foundation are so crucial to shaping future scientists, startup endeavors like Efficient, and generally to creating trillions in economic value.

If you're interested in the energy-efficient future of computing, hit me up. We are doing something unique at Efficient Computer and I'd love to tell you more. Thanks again to Embedded Insiders for the great discussion!

This is their patent:

US2025190189A1 ENERGY-MINIMAL DATAFLOW ARCHITECTURE WITH PROGRAMMABLE ON-CHIP NETWORK 20220922

Disclosed herein is a co-designed compiler and CGRA architecture that achieves both high programmability and extreme energy efficiency. The architecture includes a rich set of control-flow operators that support arbitrary control flow and memory access on the CGRA fabric. The architecture is able to realize both energy and area savings over prior art implementations by offloading most control operations into a programmable on-chip network where they can re-use existing network switches.

US army has rights to the invention.

The patent dates from 2022 so its development overlapped with the Buainchip University pilot at Carnegie:

BrainChip Launches University AI Accelerator to Empower Innovators

BrainChip launches University AI Accelerator Program, empowering next-generation innovators with access to Akida neuromorphic technology and tools.

brainchip.com

Laguna Hills, Calif. – August 16, 2022 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of neuromorphic AI IP, is bringing its neuromorphic technology into higher education institutions via the BrainChip University AI Accelerator Program, which shares technical knowledge, promotes leading-edge discoveries and positions students to be next-generation technology innovators.

The Program successfully completed a pilot session at Carnegie Mellon University this past spring semester and will be officially launching with Arizona State University in September. There are five universities and institutes of technology expected to participate in the program during its inaugural academic year. Each program session will include a demonstration and education of a working environment for BrainChip’s AKD1000 on a Linux-based system, combining lecture-based teaching methods with hands-on experiential exploration.

The invention is above my paygrade, but it seems to be like removing traffic lights and allowing the program steps to weave their way through the process using inherent collision-avoidance code. Because it does use code, I would guess its slower than Akida/TENNs.

 

[Diogenese]

This bloke is a professor at Carnegie Mellon and founded spin-off, Efficient Computer, which claims 100 * energy efficiency.

Brandon LuciaBrandon Lucia • 3rd+Premium • 3rd+Professor at Carnegie Mellon University, Co-Founder & CEO at Efficient ComputerProfessor at Carnegie Mellon University, Co-Founder & CEO at Efficient Computer2mo • 2 months ago • Visible to anyone on or off LinkedIn
Follow

I recently spoke to Embedded Insiders about how Efficient Computer is shaping the future of computing - for AI/ML, but also much more. We discuss the criticality of energy efficiency and the absolute need for general-purpose architectures, as opposed to over-specialized hardware. Generality keeps programmers happy, supports innovation, and actually, somewhat surprisingly, *increases* end-to-end efficiency (ask me about Amdahl's Law!)

We also talked about the future of the computing workforce and the next generation of computer engineers. Implicit in that discussion was the idea that major initiatives like the National Science Foundation are so crucial to shaping future scientists, startup endeavors like Efficient, and generally to creating trillions in economic value.

If you're interested in the energy-efficient future of computing, hit me up. We are doing something unique at Efficient Computer and I'd love to tell you more. Thanks again to Embedded Insiders for the great discussion!

This is their patent:

US2025190189A1 ENERGY-MINIMAL DATAFLOW ARCHITECTURE WITH PROGRAMMABLE ON-CHIP NETWORK 20220922

View attachment 88977



Disclosed herein is a co-designed compiler and CGRA architecture that achieves both high programmability and extreme energy efficiency. The architecture includes a rich set of control-flow operators that support arbitrary control flow and memory access on the CGRA fabric. The architecture is able to realize both energy and area savings over prior art implementations by offloading most control operations into a programmable on-chip network where they can re-use existing network switches.

US army has rights to the invention.

The patent dates from 2022 so its development overlapped with the Buainchip University pilot at Carnegie:

BrainChip Launches University AI Accelerator to Empower Innovators

BrainChip launches University AI Accelerator Program, empowering next-generation innovators with access to Akida neuromorphic technology and tools.

brainchip.com

Laguna Hills, Calif. – August 16, 2022 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of neuromorphic AI IP, is bringing its neuromorphic technology into higher education institutions via the BrainChip University AI Accelerator Program, which shares technical knowledge, promotes leading-edge discoveries and positions students to be next-generation technology innovators.

The Program successfully completed a pilot session at Carnegie Mellon University this past spring semester and will be officially launching with Arizona State University in September. There are five universities and institutes of technology expected to participate in the program during its inaugural academic year. Each program session will include a demonstration and education of a working environment for BrainChip’s AKD1000 on a Linux-based system, combining lecture-based teaching methods with hands-on experiential exploration.

The invention is above my paygrade, but it seems to be like removing traffic lights and allowing the program steps to weave their way through the process using inherent collision-avoidance code. Because it does use code, I would guess its slower than Akida/TENNs.

Oh, of course!

https://www.efficient.computer/resources/the-real-reason-processors-waste-energy-and-how-we-fixed-it

Modern computing systems face a fundamental trade-off: achieving extreme energy efficiency often comes at the cost of general-purpose programmability. Today, we're introducing one of our key innovations in the Electron E1 processor that helps reconcile this tension, called Non-Uniform Processing-Element Access (NUPEA). This work was just published at this year's International Symposium on Computer Architecture, our field's top academic conference.

To understand why NUPEA is needed, it's important to recognize that in modern architectures, data movement—not computation—is the dominant bottleneck for energy, performance, and scalability. Thus, achieving high efficiency is synonymous with keeping the bulk of the computational work as close as possible to the memory that is accessed.
Today’s systems, including CPUs, GPUs, and other programmable accelerators, address this challenge by using distributed data memories amongst processing elements (PEs) [refer to figure of NUMA]. The idea is to predetermine which program data should be mapped to which memory, as well as how tasks should be assigned to PEs near each corresponding memory. This approach is widely known as Non-Uniform Memory Access (NUMA), as different memories have different (i.e., non-uniform) access times to different processors.

NUMA alone fails to balance efficiency and generality. For all but the most simple programs, determining how to co-schedule data and tasks is a nearly impossible compiler and systems problem. Workloads with irregular computational patterns (e.g., sparse ML models) are too hard to analyze, and even with perfect information it might not be possible to map data to significantly reduce data movement. Another common approach is to shift this burden to the programmer using domain-specific languages or low-level APIs. However, this sacrifices generality and limits usability.
NUPEA is an alternative to NUMA that comes from a simple insight: while analyzing and manipulating data is unduly difficult, analyzing and manipulating instructions is trivial for today’s compilers. Thus, rather than distributing memories, NUPEA instead gives PEs varying proximity to memory [refer now to the figure for NUPEA]. The effect is that some PEs are near-memory – thus more efficient — and other PEs are farther away. Efficient’s effcc Compiler can identify the critical instructions (e.g., that fire most often or are on the critical path) and place them closer to memory, significantly reducing data movement. Other, less critical instructions can be placed further away without a significant effect on energy or performance. Additionally, by limiting the fraction of near-memory PEs, NUPEA architectures can use lightweight, energy-efficient memory access networks.
NUPEA is a step toward smarter hardware-software co-design—offering a practical path to energy-efficient, general-purpose computing. In short, NUPEA reduces communication energy by moving the most important instructions to a limited set of near-data PEs. It works without any programmer involvement or profiling, works on any programming language, and does not require the data access pattern to be regular or analyzable – thus simultaneously achieving efficiency and general purpose programmability.

 

[MDhere]

Oh no no nooooo ... We are not for sale!!!

Well with an intel shareprice of $21 usd they do need to pull a rabbit out of their hats and I think Brn holds plenty of hidden rabbits.

Arm is $163 usd, Amd is $167usd, Qualcomm is $158 usd, Nvidia is $173usd Probably the race is on to which big guy attempts to acquire Brainchip.

Somehow with intels shareprice of $21usd I don't think they can afford us but probably do want to make an effort before the other big players do.

That's my thoughts anyway on that subject.

 

[smoothsailing18]

Has it been historically possible for a share price to go from 20 cent to $ 8.00 in one year ?.

 

[MDhere]

Has it been historically possible for a share price to go from 20 cent to $ 8.00 in one year ?.

interesting question... well Qualcomm was $3.99 in Jan 1999 and went to $75.50 in January 2020 so if you /that by 10 .39 to $7.50 its not undoable.... lol also this is usd

I reckon though that would be a very tall order to fill to reach $8.00 anytime soon.... I rather think half or even a quarter of that would be more like it... but as I pointed out it's not undoable. lol

 

[Jefwilto]

Has it been historically possible for a share price to go from 20 cent to $ 8.00 in one year ?.

AI Overview

Yes, it is highly plausible for an Australian company's share price to increase from $0.20 to over $8.00 in a single year, although it is not a common occurrence. Such a dramatic rise would represent a significant gain for investors, potentially a 40-fold increase in value. While not impossible, this would require exceptional circumstances and a strong positive catalyst for the company.

Here's why:

• Significant Growth Potential:
  A company starting at $0.20 per share may be a small or emerging entity with substantial growth potential. If the company achieves significant milestones, such as securing a major contract, launching a successful product, or experiencing positive market trends, its share price can increase rapidly.
• Market Sentiment and Speculation:
  Positive news or market trends can drive significant investor interest and speculation, pushing the share price higher.
• Company-Specific Factors:
  A company's performance, including its revenue, profits, and overall financial health, plays a crucial role in its share price. Strong financial performance can lead to increased investor confidence and demand for the stock, driving up the price.
• External Factors:
  The broader economic environment, industry trends, and even government policies can all influence a company's performance and, consequently, its share price.
• Lack of Historical Precedent:
  While such a drastic increase is possible, it is not a typical or frequent occurrence. It would likely involve a company experiencing exceptional circumstances and outperforming expectations by a wide margin.

In conclusion, while a share price increase from $0.20 to over $8.00 in a year is possible for an Australian company, it would be a remarkable event driven by exceptional performance and favorable market conditions.

 

[IloveLamp]

Apologies if posted already

EDGE AI FOUNDATION

We're the EDGE AI FOUNDATION, formerly known as tinyML Foundation! We are a global non-profit community focused on innovation, collaboration, advocacy, and education for efficient, affordable, and scalable edge AI technologies. Our mission? Create a highly active community of knowledge sharing...

m.youtube.com

Great to see this panel moderated by Michael Kuptz and insights from Kurt Busch from Syntiant and Steve Brightfield from Brainchip - check out the whole session at our youtube channel - https://lnkd.i | Pete Bernard

Great to see this panel moderated by Michael Kuptz and insights from Kurt Busch from Syntiant and Steve Brightfield from Brainchip - check out the whole session at our youtube channel - https://lnkd.in/gjKDph2B

www.linkedin.com

 

[Frangipani]

LinkedIn Login, Sign in | LinkedIn

Login to LinkedIn to keep in touch with people you know, share ideas, and build your career.

www.linkedin.com

Extremely interesting comment by Alf Kuchenbuch in response to a post by STEMIX.TECH, a newly-founded Romanian “deep-tech and innovation-focused company specializing in AI-centric hardware and software products” that has plans to launch an AI-powered kids’ toy.

STEMIX.TECH

stemix.tech

They seem to have been in contact for a while, given the reply by STEMIX.TECH?!

Of course, as a start-up founded in 2024, they are not in our target group of companies that would be able to afford signing an IP license with us.
But their vision of “becoming a leading provider of LLM edge devices” sounds as if Akida 2.0 would be a much better fit than any AKD1000/1500 chips?

So is this possibly a case of a collaboration involving a potential software license of TENNs only?! After all, you don’t necessarily need Akida for running TENNs!



View attachment 66457

View attachment 66458
View attachment 66462
View attachment 66459

Until February, both STEMIX.TECH co-founders Alf Kuchenbuch addressed in his comment used to be with CyberSwarm, another AI start-up from Romania that has previously been discussed here on TSE. One as Project Manager, the other one as CTO. So both of them are already familiar with neuromorphic technology, although CyberSwarm uses analog architecture rather than digital.

View attachment 66485
View attachment 66486

View attachment 66487
View attachment 66488


Looks as if they recently met up with Alf Kuchenbuch at Hardware Pioneers Max in London:

View attachment 66490

Here is some more info on the AI toy they are planning on launching:



View attachment 66491


So STEMIX.TECH’s AI toy will definitely have a connection to the cloud, as “parents can access a web portal to view all questions, answers, and interactions their child has with the toy.”

Some parents will like this idea, but I personally think it opens up a can of privacy and data protection issues.

Just over a year ago, I spotted an intriguing comment by Alf Kuchenbuch under a STEMIX.TECH LinkedIn post. The founders of this Romanian startup “specializing in AI-centric hardware and software products” had previously worked for CyberSwarm, a neuromorphic startup with roots and an R&D centre in Romania as well, but headquartered in San Mateo, CA.

(…)

While we never heard anything concrete with regards to an actual collaboration between BrainChip and STEMIX.TECH in the ensuing months, I did notice a number of LinkedIn ‘likes’ being exchanged, including this recent one: Their CTO Ionuț Moldovanu gave Alf Kuchenbuch’s Edge AI Milan 2025 post a , in which our VP of Sales, EMEA had mentioned “validating our approach with Akida2 with TENNs, which we will start rolling out for evaluation in summer 2025”:

So it appears our companies are still interested in each other, if not already secretly an item.

Although the AI toy project sounded pretty advanced in development at the time (“We’re about to launch…”), STEMIX.TECH never again referred to it on LinkedIn.

Is the Bucharest-based startup still working on it in secrecy?
Or have they meanwhile shelved that project, which would actually be somewhat surprising, given what their CTO stated in this video uploaded to YouTube on 17 June 2024, less than a month before Alf Kuchenbuch made the above comment, whose content suggested they had already been in contact for some time.

“STEMIX.TECH D1 demo video - Voice AI development kit - Use the full power of LLM with our low power development board”:




From 3 min onwards, Ionuț Moldovanu says “Basically you could run a lot of other apps, like voice-to-voice - we’re gonna do a demo with that one as well. You could also tie sensors to it, so the generation could be made accordingly to what the sensors see, sensor values [?]. You can embed it in wearables, you can embed it in industrial equipment, in toys. Actually toys are one of our best choices right now for integrating this device, because smart toys with connections to LLM could be really interactive and engaging for users.”


I pointed out at the time that the envisaged smart toy appeared to be conceptualised as cloud-based, which made me wonder how our company would fit into this, since we keep on spruiking the benefits of privacy and security thanks to on-device processing without having to send data to and from the cloud. Remember, we’re talking about an electronic children’s toy here, where data protection should be of paramount importance!

Also, our potential partner was a yet unknown Romanian startup. It seemed implausible that they would have the budget to take out a regular IP license with us, although our business model had evolved into just that - IP licensing... That’s why I wrote at the time: “So is this possibly a case of a collaboration involving a potential software license of TENNs only?! After all, you don’t necessarily need Akida for running TENNs!”

Fast forward a year and see our BrainChip team buzzing with excitement about Akida GenAI and LLMs on the Edge. This made me revisit that July 2024 STEMIX.TECH post and speculate whether the reason why the smart toy hasn’t yet been launched may possibly have to do with our company convincing the STEMIX.TECH team at the time to shelve their original concept of utilising LLM cloud models, and instead offer to work with them on “LLMs on the Edge” in the true sense of the word: “Small Large Language Models” that can fit on a chip inside the toy and function without any internet connection, hence keeping conversations between children and their AI toys private and safe, all within the family’s four walls.

Is BrainChip possibly working with STEMIX.TECH in the form of a joint partnership (which would circumvent any upfront licence fee) to demonstrate what is possible with regards to LLMs at the Edge?

As I said, all this is pure speculation only. Maybe the project simply got shelved - be it temporarily or indefinitely.


Shortly after the LinkedIn post about the AI toy, STEMIX.TECH announced a partnership with Butonul Roșu - Niciodată Singur (https://butonulrosu.ro/, Butonul Roșu meaning “Red Button”) on the development of an “LTE-M SOS enabled bracelet or keychain focused on safety and health monitoring”.

What piqued my interest about this device - given I was aware of BrainChip’s prior intention of working on another project with STEMIX-TECH - is the sentence “The IoT-LLM technology we are developing will be part of the next wave of energy efficient, AI integrated hardware and software products. It represents a new generation of infrastructure that can analyse and “think” about physical data in ways we never imagined before.”

According to a LinkedIn post I stumbled across yesterday, the health wearable is apparently going to be launched under the name STEMIXGuardian. The first few lines in Romanian translate as “Thank you, Liviu! Thank you, Ionut! Stemix, or a reason not to leave Romania permanently:”

Other than that, STEMIX.TECH’s focus seems to have shifted to industrial applications :

#appliedai #aiinindustry #iot #smartfactories #stemixtech #industrialinnovation #digitaltransformation | STEMIX.TECH

𝗖𝗼𝗻𝗻𝗲𝗰𝘁𝗶𝗻𝗴 𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗮𝗹 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝘂𝗻𝗶𝘁𝘀 𝘁𝗼 𝘁𝗵𝗲 𝗳𝘂𝘁𝘂𝗿𝗲 Artificial intelligence (AI) is often discussed in abstract or futuristic terms, nonetheless at Stemix.Tech we believe in Applied AI Solutions - meaning intelligent solutions that improve real-world operations, now. We don't just use AI — we...

www.linkedin.com

STEMIX.TECH - AI Operations Platform

Applied AI Solutions for industrial data collection and analysis

www.stemix.tech

So nothing concrete to see here for now, but I shall continue to keep an eye on this “Applied AI Solutions” startup…