Members-Only
Recent Talks & Demos are for members only
You must be an AI Tinkerers active member to view these talks and demos.
Miruvor: Neuromorphic AI Memory
Explore a brain-inspired neuromorphic AI memory system offering low latency, energy efficiency, and post-training learning, replacing vector embeddings for faster associative recall.
Miruvor AI is replacing Vector embeddings and Similarity search with a Brain Inspired Neuromorphic way to have AI Memory that is extremely low latency, energy efficient and allows to learn from Memories post training. By storing Memories into a Live Neural network and retrieval by Pattern recognition and reactivation, we can get rid of the latency and energy cost of similarity search. This is how the brain parallelizes associative memory allowing it to think and remember in milliseconds. We’ve innovated an architecture and algorithms to use local plasticity rules to learn from new memories and have temporal context, as opposed to using a database query for AI Memory. This is what will translate onto Edge AI and Neuromorphic Chips powering Physical AI as well. I’d be happy to present my Research Paper , architecture diagrams, technical pipelines on the same, show how these memories are ingested and associated. I’d be happy to talk all about the implementation, relavence to Drones, Robotics and how this can help AI Agents and more. We can show our MVP, how the API works, how this can be integrated within Agentic Workflows and used by Agentic AI Startups as well! Self learning AI can start here. We’ve benchmarked this on the locomo Benchmark and achieved 88% accuracy and 50ms average retrieval speed.
Neuromorphic Context Engineering activates spiking neural circuits for adaptive long-term memory.
- TrueNorthIBM’s brain-inspired CMOS chip delivering 1 million programmable neurons and 256 million synapses on a 70-milliwatt power budget.IBM TrueNorth redefines edge computing by mimicking the human brain's neural architecture. This 28nm CMOS chip integrates 4,096 neurosynaptic cores to support 1 million programmable neurons and 256 million synapses. Performance is high: it executes 46 billion synaptic operations per second per watt. Efficiency is the priority: the chip consumes just 70 milliwatts during real-time operation (roughly the power density of a hearing aid). By ditching the traditional clock for an asynchronous, event-driven design, TrueNorth excels at sensory processing tasks like multi-object tracking and speech recognition. It represents a massive leap for low-power AI: specifically designed for DARPA’s SyNAPSE program requirements.
- Edge AIAI processing executed directly on endpoint devices (e.g., IoT sensors, smart cameras), eliminating cloud latency for real-time action.Edge AI shifts machine learning inference from the centralized cloud to the network's edge: devices like autonomous vehicles and industrial IoT sensors. This localized processing cuts network latency to near-zero, enabling sub-millisecond decision-making critical for safety systems and predictive maintenance. Specific advantages include enhanced data privacy (keeping sensitive data local for compliance) and reduced bandwidth costs, as only insights, not raw data, are transmitted. The technology leverages specialized hardware (like NVIDIA Jetson) to run lightweight models, ensuring high-speed, reliable operation even with intermittent connectivity.
- Neuromorphic ChipsNeuromorphic chips are brain-inspired processors: they use Spiking Neural Networks (SNNs) and event-driven computation to achieve ultra-low-power, high-speed AI at the edge.This technology re-architects computing, moving past von Neumann bottlenecks by mimicking biological neurons and synapses. Chips like Intel’s Loihi 2 (1 million neurons, 128 cores) and IBM’s retired TrueNorth (1 million neurons, 256 million synapses) leverage asynchronous, event-driven processing: computation only occurs when data arrives (a ‘spike’). This model delivers orders-of-magnitude energy efficiency improvements over conventional GPUs or CPUs, operating at power levels as low as ~1W for Loihi 2. Neuromorphic chips are purpose-built for real-time, ultra-low-power applications, specifically robotics, autonomous systems, and always-on edge AI where latency and power consumption are critical factors.
- RAGRAG (Retrieval-Augmented Generation) is the GenAI framework that grounds LLMs (like GPT-4) on external, verified data, drastically reducing model hallucinations and providing verifiable sources.RAG is a critical GenAI architecture: it solves the LLM 'hallucination' problem by inserting a retrieval step before generation. A user query is vectorized, then used to query an external knowledge base (e.g., a Pinecone vector database) for relevant document chunks (typically 512-token segments). These retrieved facts augment the original prompt, providing the LLM (e.g., Gemini or Llama 3) the specific, current, or proprietary context required. This process ensures the final response is accurate and grounded in domain-specific data, avoiding the high cost and latency of full model retraining.
- LoihiIntel's neuromorphic research chip utilizes asynchronous spiking neural networks to deliver brain-inspired computing at extreme energy efficiency.Loihi 2 advances neuromorphic engineering via the Intel 4 process: it packs 1 million programmable neurons and 120 million synapses into a single chip. This architecture processes information only when data spikes occur (asynchronous execution), which cuts power consumption significantly compared to traditional silicon. Developers use the open-source Lava framework to run real-time workloads like gesture recognition and robotic path planning. With 128 cores per chip, Loihi 2 achieves up to 10x faster processing than its predecessor while maintaining a minimal energy footprint.
- Loihi 2Intel's second-generation neuromorphic research chip (Loihi 2) scales brain-inspired AI with 10x faster processing and 15x greater resource density.Loihi 2 leverages the Intel 4 process node to pack 1 million programmable neurons into a 31 square millimeter die. This asynchronous spiking neural network (SNN) architecture eliminates global clocks to minimize power consumption (often by orders of magnitude compared to traditional GPUs). It integrates 128 neuromorphic cores and supports the open-source Lava software framework for cross-platform development. Researchers use it for real-time robotics, tactile sensing, and gesture recognition: tasks requiring high-speed inference at the extreme edge.
- SpiNNakerA million-core neuromorphic supercomputer designed to simulate biological neural networks in real time.Steve Furber’s team at the University of Manchester built this neuromorphic powerhouse using 1,036,800 ARM968 processors. The architecture mimics the brain’s massive parallelism by prioritizing packet-based communication over traditional memory access. It currently simulates 1 billion neurons (roughly 1% of the human brain) while consuming a fraction of the power required by standard supercomputers. Researchers deploy SpiNNaker for large-scale brain modeling, low-latency robotic control, and testing asynchronous spiking neural networks.
- BrainScaleSBrainScaleS is a mixed-signal neuromorphic platform that emulates biological neural networks at speeds up to 10,000 times faster than real time.BrainScaleS (Heidelberg University) emulates brain functions using mixed-signal CMOS technology. The architecture employs analog circuits for membrane dynamics and digital pulses for spike communication (spatio-temporal events). Each BrainScaleS-2 chip integrates 512 neurons and 131,072 synapses, operating at acceleration factors up to 10,000 times real time. This speed enables researchers to compress a day of biological learning into under nine seconds. With on-chip microprocessors for plasticity, the system provides a high-bandwidth platform for studying synaptic changes and efficient edge-AI applications.
Related projects
Memory AI
Boston
The talk explains a memory‑AI system that stores user interactions securely via encryption, using vector search, AI chat…
AI safety in healthcare
Vienna
Learn how to design safer, more deterministic AI for complex industries like healthcare. Discover methods for evaluating AI…
Stateful AI Without Vendor Lock-In
Amsterdam
See a live demo of Intangle, a stateful AI system featuring a self-learning memory layer, secure memory graphs,…
DeepCache - An Agentic Framework for Neural Compression
Toronto
The talk explains how to attach complexity metadata to agent messages, use it for compression, verify integrity, predict…
Neural Probing LLMs for 'Immediate Inference'
Montreal
This talk details using novel LLM heads to measure internal states, accelerating inference by bypassing token generation, a…
Build bulletproof generative AI applications
Berlin
Learn how to build efficient generative AI applications using caching, fact-checking, and Retrieval Augmented Generation with practical examples…