Thoughts, experiments, and critical analysis on artificial intelligence, cognitive architectures, and the technologies shaping our future. Where deep thinking meets deeper code.
Senua AI's Measure held Meta's open physics engine to the laws of physics: twelve departures from the right answer, each traced to its cause, and seven patches that correct them. Twenty-three pre-registered studies, and every rule, run and file public on GitHub. What comes next for SuperDex: training data at volume and the gap to real hardware.
Read the full storySenua AI's Measure held Meta's open physics engine to the laws of physics: twelve departures from the right answer, each traced to its cause, seven patches that correct them, and twenty-three pre-registered studies with every rule, run and file public on GitHub.
We tapped Meta's contact physics and found its friction could not hold a grasp: Senua recovered the hidden friction law from sensor data alone. After the fix it drove the simulation live, one millisecond at a time, and kept a slippery, fragile wine glass in hand every time.
Australia's National AI Plan runs 37 pages and never uses the word architecture. It cites ten sources, none peer-reviewed, and its two biggest commercial claims come from a data centre broker and a tech lobby group. A close reading of what the document actually commits to.
CSIRO is right that AI bias is a chain of human decisions. But its case against technical fixes lists only the three repairs one architecture allows, then calls that the limit of technical work. A reply, and an argument for funding the architectures nobody has measured.
Most boardrooms think AI means ChatGPT. It is at least two very different kinds of machine intelligence, good at almost opposite jobs. One needs a data centre; the other runs on a chip that costs less than lunch. A plain-English guide, grounded only in peer-reviewed research.
Meta just bet a whole simulator on the idea that contact is the frontier of robotics. The physics shipped; the part that feels did not. What a week-old release gets right, where the black box sits, and how causation streams could measure the sim-to-real gap instead of discovering it.
The pitch promises local, private, real-time, capable AI all at once; the architecture quietly picks a subset. The prefill cost, the cloud offload, the smarter client that still phones home, the memory wall, and the matmul floor no quantization removes, then the architecture that does not need any of it.
A Stanford paper took the gradients out of training stateful models, forward passes only, to 4.3 billion parameters. Causal-state induction takes them out of learning entirely. Two routes past the economics of attention, one destination.
Robots can act but cannot feel. The sensors are cheap and proven; the gap is the reader. Why touch is a change-stream problem, and why causal-state cognition on a sixty dollar computer is the right layer to read it.
The industry has concluded the physical world is the next dataset. The premise is right; the bet is wrong. Why capture-and-train repeats the text-era mistake at higher stakes—and what the mathematics of causal states says instead.
The energy identity 2ν∫Ω dt ≤ E(0) gives a finite dissipation budget, not a uniform enstrophy bound. Under three cascade hypotheses the active shells are bounded; explicit smooth fields — a spectral gap, vortex blobs, a shear flow — show those hypotheses are not properties of smooth flows.
Two kinematic facts about triads on Z³, an explicit single-triad field with no phase cancellation, and inviscid truncations going to the Lee–Kraichnan equilibrium. The lattice constrains the interaction set; it does not bound a solution.
The Mertens amplitude sum Σ 2/|ρn ζ′(ρn)| with ζ′ computed at each of Odlyzko’s 2,001,052 zeros crosses 1 at K = 58 and 2 near K = 1,040, grows consistently with the conjectural (log T)5/4 law, and its zeros are recoverable from M(x)/√x by Fourier analysis.
Liquid neural networks, the sieve, the conduit, the scaffold, and how the boundary between code and model dissolves when everything is dual32. The system adapts continuously. The scaffold watches.
One data type — dual32. Native machine code. Register pairs. 1.6x faster than optimised C, not because the code is better, but because Simplex knows what a dual32 is. C doesn't.
The same scaffold array that diagnosed a fluid dynamics solver also diagnoses neural network learning failures. One tool, two settings: truncation levels for a fluid, learning rates for a gate.
Scaffold arrays run one initial datum through several Galerkin truncations and compare. The comparison found a missing −i factor in a real-coefficient triad kernel; the corrected kernel conserves energy, matches an FFT solver to 2.8×10−17, and shows inviscid truncations thermalising.
Simplex v0.14.0 introduces a hybrid classical-quantum bridge. Cost-aware QPU dispatch to Amazon Braket, IBM Quantum, and Azure. Variational algorithms powered by dual numbers. Quantum uncertainty as epistemic beliefs.
Every outstanding task closed. 162/162 tests passing. Complex numbers, matrix algebra with quantum gate constructors, complete dual numbers, HTTP client, JSON parser, SLM native bindings, and five critical compiler fixes.
From module system to 154/154 tests passing. Cross-module imports, neural gate codegen, f64 type-aware arithmetic, forward-mode AD, contract verification, and a 19,589-line runtime.
Introducing Self-Learning Annealing and Edge Hive. Hyperparameters that tune themselves through meta-gradients using dual numbers, and local-first AI that runs entirely on-device.
Announcing Neural Gates and Real-Time Learning. A fundamental shift where code logic learns from data, models adapt during runtime, and systems get smarter the longer they run. No retraining required.
A complete guide to training domain-specific specialists for Simplex. Infrastructure selection, AWS provisioning, LoRA fine-tuning scripts, and the roadmap from Python to pure Simplex.
Google's RLM handles 10M+ token contexts through recursive decomposition. It's clever engineering—and a solution to a problem Simplex solved through cognitive architecture, not context workarounds.
An honest assessment of where Simplex stands today and the roadmap to Neural IR and Neural Gates. What we have, what's missing, and what comes next.
Simplex v0.5.0 is now open source under MIT. Introducing the Anima—a cognitive soul for AI agents—and revolutionary per-hive SLM provisioning where specialists share a single language model.
A comprehensive technical deep dive into the Simplex toolchain. Compiler internals, bytecode format, VM architecture, language specification, actors, CHAI specialists, and Mnemonic extensions with code examples.
An executive overview of the Simplex programming language toolchain. How a self-hosted compiler, package manager, documentation generator, virtual machine, and language server work together.
A comprehensive survey of the AI agent tooling landscape in 2026. Comparing Simplex against Dana, Mojo, Letta, Akka, and the Python framework ecosystem.
A critique of the agentic AI paradigm — backed by Stanford/Harvard research, Gartner predictions, and expert analysis from Gary Marcus and Yann LeCun.
Formal operational semantics for Simplex cognitive constructs — how memory, beliefs, actors, and agents execute with mathematical precision.
How emergence happens in Mnemonic Hives — stigmergic coordination, swarm intelligence, and the conditions under which agent collectives develop genuine collective intelligence.
How Mnemonic Hives learn from every interaction through knowledge distillation, LoRA fine-tuning, and intelligent compression — without catastrophic forgetting.
A deep dive into the epistemic architecture for Mnemonic Hives — how to design agents that track confidence, revise beliefs rationally, and distinguish facts from opinions.
A response to Dries Buytaert's investment thesis — challenging the notion that orchestration platforms are the future of AI coordination.
How we're extending Simplex with first-class support for beliefs, goals, intentions, and autonomous behavior — the technical foundation for true AI agents.
An introduction to cognitive architecture for AI systems — why LLMs need more than context windows, and how to build agents that actually remember.
The theory behind Cognitive Hive AI and the early iterations of Simplex — why swarms of specialists will outperform monolithic models.
