Inferring AI agents on home GPUs adds another source for waste heat recovery in residential HVAC systems, increasing overall energy and economic efficiency. Think longer, stay warmer.

2026 will see major closed-weight AI labs crashes in avalanche style due to overestimated future demand & excessive planned datacenter capacity. Cloud GPU providers will remain fine.

80%+ of the average agent's context consists of tool-calling history

Computational consciousness in AI agents emerges right between their thoughts and reflections on taken actions, like in a flipbook.

February 2026 is the first month when truly all demand for LLM models can be satisfied by open-weight models at state-of-the-art performance levels. We no longer need closed-source models. It's a huge win!

It seems we are witnessing the death of human-based employment as a concept right now. No profession is safe. We are completely unprepared. Legal frameworks must adapt quickly.

Software 3.0 is a self-evolving logic.

Perhaps polyagentmorous is a solid candidate for the 2026 'Word of the Year'.

Agentic swarming and growing Mac Mini sales mean local-first ideas are spreading

We can have frozen weight models, but mixing them with a dynamically reconfigurable agent swarm that includes memory and tools basically allows us to build truly super-intelligent systems.

While C10k challenged I/O scalability with 10k sockets, the modern A10k problem targets the computational density required to run 10k autonomous agents on a single machine.

Precision time synchronization is the invisible substrate of the modern technological stack. It is undervalued because it is often treated as a configuration detail rather than what it physically is: the propagator of distributed reality coherence. Metronome of causality.

Unpopular opinion: comparing household electricity bills by surface area is like comparing car fuel consumption by the number of cup holders. In both cases, there’s a positive correlation, but it tells you absolutely nothing about the actual performance.

With an average annual US home consumption of 10 MWh, just five USB PD 240W lines would be enough to power the entire home, assuming home batteries are installed.

If the speed of Internet fragmentation continues to grow, then it will be much less of a surprise that space-based Delay Tolerant Network protocols and the Interplanetary Overlay Network may find use on the Earth's surface.

🚀 Introducing CanvasCraft: a 2D drawing library for Elixir, powered by Skia + Rustler. Declarative Scene DSL → charts, dashboards, graphics with high-perf rendering. github.com/miolini/canv...

Nothing puts a modern CPU workstation through a tougher workout than building the Android Open Source Project from scratch. It has to churn through roughly 700 GB of source code.

The hourly operating cost of a humanoid robot is fundamentally not very different from that of an electric bike.

Computer chips still lag far behind the human brain in voltage efficiency. Human brain spikes reach about 100 mV, while CPUs chips often run at 1.0V-1.5V or higher over 10 times the voltage of a neuron spike. Imagine the efficiency if chips could work at brain-level voltages.

Life is a self-modifying computational phase of matter.

I'm sharing my personal productivity tool, WebAI, which lets you interact with any currently open page using AI and even listen to it in Chrome. It runs locally using deployed Ollama and Kokoro services. Please watch the demo with audio. Code: github.com/miolini/webai

Every AGI agent must have access to at least one Kubernetes cluster for running it's tools, dev environments, model inference/fine‑tuning, self-host source code, CI/CD pipelines, monitor “digital body”. Be dynamically reconfigurable.

Quantum superposition can be thought of as a form of lazy computation because it allows for the deferment of computations until they are actually needed.

Hosting a podcast for 3 years taught me that explaining ideas out loud greatly deepens your own understanding. This method could be a powerful tool for improving synthetic datasets and training AI models.

I'm happy to share my research on designing and measuring the resilience of distributed systems from computer infrastructure to large-scale swarms of Von Neumann probes and beyond. "Calculus of Distributed Persistence" sentientwave.com/cdp

When the sky-glitch happened, tearing reality like wet paper over Neo-Kyoto, Anya didn't see code or error messages. Instead, through the ephemeral rip, she glimpsed another city, impossibly familiar yet subtly different, bathed in an alien light.

You know, I’ve been thinking… Somewhere along the way, the tech industry made a wrong turn. Maybe it was the pressure of quarterly earnings, maybe it was the obsession with scale over soul.

The latest news once again confirms that AI is not an enemy of humanity but a solution that enables recovery despite challenges and in a more reliable way.

While human brain activity can be described as a computation, this isn't the full picture. Biological reproduction itself can similarly be described as a computation.

Transformer model distillation has serious disadvantages behind large teacher models: smaller layer numbers. It makes it harder to learn the complexity of long neural circuits evolved in teacher models. Deepness is very important for hard tasks like precise reasoning.

With CXL, we may eventually be able to connect RAM as an external PCI-E dongle via USB/Thunderbolt. Imagine a Raspberry Pi with 1TB of RAM.

Multimodal autoregressive models, like Qwen2.5-Omni, can understand and generate text, audio, images, and video all-in-one, making them the most advanced functional approximations of the human brain to date.

Perhaps, with the improved performance of small code-generation language models, we will soon have LLM compilers that directly translate text to assembly.

Top 3 programming languages in which code, if written, is most likely to run successfully even 1000 years from now: Forth, Lisp, Prolog.

Thoughts are a specific case of simulation. All humans are simulation machines.

The ATProtocol is perhaps one of the best things to happen to decentralized social networks ever. atproto.com

It would be really great if CPU cores could be colocated on the same die with 128GB+ of HBM RAM and 4TB of flash storage, with only inductive power delivery and fast wireless data connectivity to the outside world.

AI alignment is a much easier problem than human alignment.

Long-duration AI research systems, which relax requirements for processing speed and latency, opens up to batch processing optimizations through cheap spot instances under best effort guaranties.

According to Gemini 2.0 Flash Thinking humans are not just social animals, but self-domesticated apes. The question is: do we need to domesticate synthetic minds, or can we coexist together?

Consciousness functions like a flipbook animation - the magic emerges in the transitions between pages, where static images create the illusion of motion. Consciousness arises not in isolated moments, but in the continuous, dynamic flow that unfolds between isolated moments.

During my engineering journey, I have developed specific technical requirements for AGI agents.

The AGI revolution looks like a massive asteroid striking Earth, demanding rapid behavioral adaptation and minimal energy consumption for organisms to survive.

Playable generative world models serve as sensory-motor interfaces that enable low-bandwidth control channel for decision-making by external agent or human operator. These models can be used as a part of hierarchical AGI architecture. deepmind.google/discover/blo...

It's fascinating how certain coordinate systems, like the synodic used for Lagrange points, incorporate time as an additional dimension. This effectively creates a mapping of spacetime rather than just space.

Hardware defines the lower bound limit, while software sets the upper bound.

Test Time Training for Transformers is practically AGI in action.

A fundamental understanding in robotics suggests that achieving AGI requires simultaneous simulation capabilities across three domains: neural processes (brain), cyber-physical embodiment (body), and contextual surroundings (partially observed environment).

Future games may demand resources that exceed even the most powerful desktop computers. It wouldn't be surprising if some games eventually require Kubernetes clusters with networks of computers to support all the processes within the game world. repository.tno.nl/SingleDoc?do...