"The safest model is the one you can inspect. The most dangerous model is the one you cannot."

AI safety in 2026 is split into roughly two camps: those who believe alignment is a technical problem with an engineering solution, and those who believe it is an open-ended research problem that may not have a solution at all. The split is not academic — it determines funding, regulation, and the pace of deployment across every industry.

Alignment research has become its own sub-discipline within machine learning. The key labs — Anthropic, DeepMind, OpenAI's Superalignment team (now restructured after Jan Leike's departure in 2024), and a growing number of academic groups — have converged on a set of standard approaches: constitutional AI (Anthropic), process reward models (Google DeepMind), sparse autoencoders for mechanistic interpretability (OpenAI, Anthropic, Redwood Research). The results are measurable but inconclusive. Models can be made to refuse harmful requests. They can also be jailbroken within hours of a new safety guardrail being deployed. The cat-and-mouse game is the defining dynamic of the field.

Red-teaming has professionalised. It is no longer hobbyists in Discord servers — it is a formal evaluation pipeline run by organisations like METR (formerly ARC Evals) and the Safety Institute, which runs structured evaluations using the UK government's AI Safety Institute framework. The standard eval suite includes: persuasion and manipulation capability, autonomous replication, cyber-offence capability, and model-generated deceptive alignment. The benchmarks are public. Every frontier model is scored. The scores are not good.

Open vs closed source: The debate has shifted. Meta's Llama 3.1 (405B, open-weight) demonstrated that open models can compete with closed frontier models in reasoning benchmarks. But the safety implications are different: a closed model can be patched server-side; an open-weight model, once released, cannot be recalled. Proponents argue that open models enable safety research that would be impossible otherwise — you cannot inspect a model you cannot access. Opponents argue that open release of capable models is an irreversible act with second-order effects no single lab can predict. Both are correct.

What changed in 2026: Compute governance. The US Executive Order on AI (October 2025) introduced mandatory reporting for training runs above 10^26 FLOP. The EU AI Act entered force in August 2025, classifying general-purpose AI models by systemic risk. China's MIIT requires government approval for any model with more than 10 billion parameters trained on domestic data. The net effect is that the boundary between "safe" and "unsafe" AI is now partly defined by regulation — which means it varies by jurisdiction. A safe model in London may be an unsafe model in Beijing. The gradient is steep.

"When the towers go down, the mesh still talks."

LoRa (Long Range) radio is a spread-spectrum modulation technology that can transmit data over 5-15km in rural conditions using very little power. It is not fast — the maximum throughput is about 250 kbps in the fastest mode, and most practical mesh networks run at 0.3-50 kbps. But it does not need towers, does not need a SIM card, does not need anything except two or more devices within radio range of each other. Text goes through. That is enough.

The build is centred around Meshtastic, an open-source firmware that turns ESP32 and nRF52 microcontrollers into LoRa mesh nodes. The mesh is fully decentralised: every node rebroadcasts messages it receives, so the network extends beyond any single node's range. There is no central server. There is no point of failure.

Assembly: Flash Meshtastic firmware via USB (ESPHome Flasher or PlatformIO). The firmware auto-configures mesh networking on first boot — the only required setting is your channel encryption key (the mesh supports private channels with AES-256). Each node discovers neighbours automatically and builds a routing table. For a solar node, connect the TP4056 to the battery and panel, then the battery to the ESP32. Put it in a waterproof box with the antenna protruding through a bulkhead SMA connector. Mount as high as possible. Every 3m of height adds roughly 1km of range.

Range: In open terrain (plains, coastal areas), two nodes at 5m height achieve 8-12km. In urban environments, 2-4km. With a five-node mesh, you can cover approximately 50km line-of-sight diameter — the messages hop from node to node. Throughput is text-only. Maps, emergency alerts, weather station data, and short-form messaging. A single message typically arrives at every node in the mesh within 1-3 seconds.

What it is good for: Community networks in areas without cellular coverage. Emergency communication during disaster response. Trail and expedition comms. Environmental sensor data collection. It is not for video. It is not for voice. It is for the kind of communication that matters when nothing else works.

Known failure modes: The mesh degrades gracefully under heavy traffic but can saturate if too many nodes try to rebroadcast simultaneously (the "broadcast storm" problem). Meshtastic handles this with a configurable hop limit (default 3). Battery life with a 3400mAh cell is roughly 3-5 days with continuous mesh participation, or 2-3 weeks with deep sleep between periodic updates. Solar nodes run indefinitely if they get 3+ hours of direct sun per day.

"A generation ship is not a vehicle. It is a city that happens to be moving."

DREAM: A spacecraft carrying a human crew to Proxima Centauri b, the nearest exoplanet, 4.24 light-years away. At 0.1c (30,000 km/s), the journey takes 43 years. It is not a crossing. It is a lifetime. The people who board in their 20s arrive in their 60s. Children are born on the ship. They will never see Earth. The ship is the only world they know.

The engineering challenges are staggering, but none are disqualifying. The propulsion problem was already studied in the 1970s.

Project Daedalus (British Interplanetary Society, 1973-1978) designed a two-stage fusion-powered spacecraft that could reach Barnard's Star (6 ly) in 50 years at 0.12c. The design used inertial confinement fusion — deuterium/helium-3 pellets ignited by electron beams. The engineering was laid out in detail across 700 pages. The only missing piece was the fusion technology itself, which remains at TRL 3-4 today.

Breakthrough Starshot (2016-present) is the opposite end of the mass spectrum: gram-scale light sails pushed by a ground-based laser array to 0.2c. The sailcraft reach Proxima in 20 years but can carry nothing except a camera and a transmitter. No humans. No return. The technology — phased laser arrays, metre-scale lightsails, photonic propulsion — is progressing rapidly. A 100 GW laser array in the Atacama Desert is the current reference design.

Closed-loop life support is the most critical subsystem. The ISS achieves 93% water recovery and 42% oxygen recovery through the Environmental Control and Life Support System (ECLSS). For a 43-year mission, those numbers need to be >99.5% for both. The MELiSSA project (ESA) has been developing a closed-loop bioregenerative ecosystem since 1989. Their pilot plant in Barcelona achieves 80% oxygen closure using a five-compartment system of bacteria, algae, and higher plants. The remaining gap is closing — but the timeline is decades, not years.

Psychology of confinement: The longest continuous human spaceflight is Valery Polyakov's 438 days on Mir (1994-1995). There have been no studies of multi-year group isolation longer than the Mars500 experiment (520 days, 2011). A 43-year mission has no precedent. The psychosocial challenges include: irreversible crew conflict, loss of motivation across generational turnover, the absence of Earth as a reference point, and the knowledge that no rescue is possible. The social contract — drafted before launch — must define governance, reproduction, education, and end-of-life care for a society that does not yet exist.

The horizon at 0.1c is not a line you approach. It is a line you carry with you. Every day, the destination is the same distance away — 4.24 years in ship time. The crew does not get closer. The ship is a boundary that moves. That is the psychological cost of interstellar travel: the edge never arrives. You become the edge.

A boundary that shows up on every map but commands different respect in person: the place where the tar ends. In South Africa, roughly 10% of the road network is paved. The other 90% is gravel, sand, corrugation, and the kind of mud that does not let go. If you build in the spaces between cities, you need a vehicle that takes the edge as a starting point, not a limit.

Bollinger B1 (2020-2026): The purest expression of the off-road-first philosophy. Two electric motors, 614 hp, 668 lb-ft torque, a chassis that doubles as a battery enclosure. No driver aids, no touchscreen, no pretence of luxury. The B1 has a 3,000 kg towing capacity, a pass-through Frunk that fits 4x8 plywood sheets, and doors that unclip and store under the seats. Production ended in early 2026 after Bollinger shifted focus to commercial electric delivery vehicles. Starting price: ~$64,000. The used market is already appreciating.

Rivian R1T (2022-present): The polished approach. 835 hp, 0-60 in 3.0 seconds, adjustable air suspension that gives 14.9 inches of ground clearance, and a tonneau cover that does not leak (the 2024 revision fixed the 2022-2023 failures). The R1T's party trick is the Gear Tunnel — a full-width storage compartment behind the cab that fits a snowboard, a toolkit, or a slide-out camp kitchen. Starting price: ~$71,000. Real-world range in off-road conditions: ~270 miles with the Large pack. The R1T proves that off-road competence and daily comfort are not mutually exclusive. It is expensive because both are hard.

Toyota Hilux (8th gen, 2015-present): The benchmark. If the boundary between reliable and unreliable off-road transport is a line, the Hilux defined it. The 2.8L GD-6 diesel produces 201 hp and 500 Nm — unremarkable on paper, remarkable in the field because the drivetrain is built to tolerate abuse that would shear the axles on most other vehicles. Independent front suspension, leaf-spring rear, a ladder frame that has not changed fundamentally since the 1980s. The Hilux is not innovative. It is proven. In Southern Africa, it is the default vehicle for organisations that cannot afford a breakdown because a breakdown means a person does not get medicine, food, or water. Starting price: ~$45,000 (SA-spec). A 1999 Hilux with 500,000 km still sells for $8,000.

Electric adventure bikes — Stark Varg EX (2025): The boundary between off-road and electric has been crossed. Stark Future's Varg EX produces 80 hp from a 7.5 kWh battery, weighs 110 kg, and has no gears, no clutch, no noise that carries beyond 50m. The electric motor delivers instant torque at 0 RPM, which is what you want when the terrain is technical and the next gear change would cost you a second of momentum. Range: ~80 km of aggressive off-road, ~160 km of trail riding. Charge time: 2 hours on a standard wall outlet. Starting price: ~$13,500. It is not a commuter. It is a tool for covering ground that has no ground to cover.

The common thread across all four: boundary vehicles are not defined by top speed or luxury. They are defined by the ratio of capability to fragility. The best off-road vehicle is the one that brings you back.

01 — Kármán Line Revision: The FAI is reconsidering the 100 km boundary definition of space. New ionospheric data from Swarm satellites suggests that the transition from atmosphere to magnetosphere is not a sharp line but a gradient that shifts by up to 50 km depending on solar activity. The 100 km line was chosen for convenience (it is a round number). Reality, as usual, is more gradient than boundary. The working group is expected to publish a revised definition in 2027.

02 — Edge Computing Startup — EdgeImpulse (Oslo): Raised $45M Series C in February 2026 for their tinyML deployment platform. The product is a CI/CD pipeline for embedded ML models — you train on the cloud, deploy to a microcontroller, and the model runs inference entirely on-device. No data leaves the edge. Their customers include agricultural sensor networks, predictive maintenance on mining equipment, and wildlife monitoring in remote parks. The boundary between cloud and edge is where the value is.

03 — Paper — "Attention at the Boundary" (2025): Published in Nature Machine Intelligence by researchers at DeepMind and UCL. The paper formalises a key insight in transformer architectures: attention mechanisms naturally learn boundary representations — the edges between concepts, classes, and features in their training data. The model's internal representation of a "boundary" can be extracted and visualised. The paper provides a method for identifying which attention heads are responsible for boundary detection, enabling targeted interpretability. Open access.

04 — Workflow brief: Real-time edge-to-cloud pipeline for remote sensor networks: (a) ESP32 reads sensor data (temperature, humidity, soil moisture, air quality) and publishes to a local MQTT broker; (b) a Raspberry Pi at the edge runs Node-RED to filter, aggregate, and compress before forwarding; (c) forwarded data goes via LoRa (Meshtastic) to a gateway node with cellular or Starlink backhaul; (d) the cloud endpoint (Cloudflare Workers + D1) stores time-series data and triggers alerts. Total power for the ESP32 + LoRa node: 0.5W. A 10W solar panel runs it indefinitely. Total implementation time: 2-3 days.

The peri-urban corridor is the fastest-growing human habitat on the planet. It is neither city nor countryside — it is the boundary where both meet, and where the infrastructure of neither fully applies. In Africa, the peri-urban population is growing at 4% annually, faster than either urban cores or rural areas. The most interesting companies in the region operate at this boundary, building systems that serve both sides.

Lumos Global (Nigeria, founded 2017) provides pay-as-you-go off-grid solar electricity to peri-urban and rural households in Nigeria. Their model: a 50W-100W solar home system with battery storage, LED lights, phone charging, and optionally a TV or DC fan. The customer pays via mobile money (MTN MoMo, Airtel Money) — daily, weekly, or monthly — and the system is remotely locked if payments lapse. Lumos uses IoT-connected controllers (SIM800-based) that report energy usage, battery health, and payment status in real time over the cellular network. As of early 2026, Lumos has deployed over 1.2 million systems across Nigeria and expanded to Ghana and Côte d'Ivoire. The key metric: 94% collection rate on PAYG payments. The boundary between having electricity and not is a mobile money transaction.

MAX (Nigeria, founded 2015) operates at a different boundary — the last mile of urban logistics. They connect riders (motorcycles, tricycles) with delivery and transport demand through a mobile platform, but their real product is financing. MAX provides asset financing for the vehicles themselves: a rider puts down 20-30%, MAX covers the rest, and the rider pays off the balance through daily earnings. The model works because MAX tracks every trip through their app and can adjust repayment schedules in real time based on the rider's income. Over 50,000 vehicles financed to date. The boundary between informal transport and formal logistics is a smartphone and a tracking algorithm.

Both companies share a structural insight: the boundary between city and peri-urban is not a line. It is a gradient of service availability. At the city centre, everything works. At the peri-urban edge, the grid is unreliable, the roads are unpaved, and the addresses do not exist. The companies that succeed at this boundary do not try to bring the city's infrastructure to the edge. They build infrastructure that works on the edge's own terms — mobile money instead of bank accounts, PAYG instead of contracts, two wheels instead of four.

The corridor is not the space between two points. It is where the edge conditions produce their own logic. Companies that understand this build infrastructure that looks wrong from the centre and essential from the edge.

"The nervous system is a boundary between you and the world. It decides what gets in."

Polyvagal theory, developed by Dr Stephen Porges in 1994 and refined over the following three decades, describes how the autonomic nervous system regulates safety and threat responses through three distinct neural circuits. It is not a therapy. It is a map of the nervous system's operating system — and for builders, operators, and people who work at the edge of their capacity, understanding the map is the first step to staying functional.

Ventral vagal (safety connection): The newest evolutionary layer. The ventral vagal complex regulates the muscles of the face, head, and heart: eye contact, vocal tone, middle ear muscles (tuning to human voice frequency). When the ventral vagal is active, you feel calm, connected, and capable. This is the ideal operating state for creative work, collaboration, and decision-making. The body interprets the environment as safe. Higher cognitive function is fully available.

Sympathetic (mobilisation): The fight-or-flight state. Heart rate increases, pupils dilate, blood moves to large muscle groups, digestions slows. This is adaptive when there is an actual threat. It is maladaptive when it becomes chronic — as it does for most people in high-demand environments. The body cannot distinguish between a deadline and a predator. The physiological response is identical.

Dorsal vagal (immobilisation): The oldest circuit. The freeze, collapse, or shutdown response. Heart rate drops, blood pressure falls, social engagement stops. In acute danger, this makes you "play dead." In chronic stress, it manifests as burnout, dissociation, depression, or the feeling of "not being able to get started." The dorsal response is the nervous system's circuit breaker. When it trips, you stop.

Window of tolerance: The zone within which you can function effectively. When you are inside it, you can think, work, relate, and recover. When you are outside it (hyperarousal: sympathetic excess; or hypoarousal: dorsal excess), cognitive function degrades rapidly. The goal is not to avoid leaving the window. The goal is to recognise when you have left and to have tools to return.

Practices that widen the window:

Vagal toning: The vagus nerve can be stimulated actively. Humming (exhaling with vocalisation) at a low frequency (~100-150 Hz) stimulates the vagal afferent pathway. Gargling activates the pharyngeal branch. Cold exposure — specifically cold water on the face (the mammalian dive reflex) — triggers immediate vagal activation, dropping heart rate and shifting the system toward ventral vagal. A 20-second splash of cold water is the fastest-known way to reset the nervous system. The science is well-established (the trigeminal-vagal reflex, first described in 1960).

Breathing: Coherent breathing (5 breaths per minute, equal inhale and exhale, ~6-second cycles) increases heart rate variability (HRV), the primary physiological marker of vagal tone. A 2023 meta-analysis in Frontiers in Neuroscience confirmed that HRV biofeedback training — which is just breathing at resonant frequency — significantly reduces anxiety and improves cognitive performance under stress. The mechanism is respiratory sinus arrhythmia: the heart rate increases on inhale and decreases on exhale. Maximising this oscillation strengthens vagal regulation.

Cold exposure: A 2024 study from the University of Tromsø showed that regular cold exposure (2-3 minutes at 10-15°C, 3x/week over 4 weeks) produced measurable changes in vagal tone, reduced resting heart rate, and improved subjective recovery scores. The adaptation mechanism is non-shivering thermogenesis, which activates the sympathetic system but — critically — only in bursts, followed by a parasympathetic rebound that strengthens the ventral vagal response. The rebound is where the training effect occurs.

For builders: The nervous system is the most important infrastructure you operate. It is also the most neglected. A machine that runs 24/7 without maintenance is not productive — it is failing gradually. The same is true of the human nervous system. Polyvagal theory is not a productivity hack. It is a maintenance manual for the thing that does all the work.

01 — Coastline / Ocean meets land. "Wild Coast, Eastern Cape." Photographer: David Goldblatt (c) Estate David Goldblatt / Goodman Gallery. Goldblatt's South African landscapes capture the edge where human infrastructure meets natural terrain. This image: the coastline at Port St Johns, where the Indian Ocean meets the cliffs and the road stops.

02 — City-Township line / Infrastructure edges. "Diepsloot, Johannesburg, 2006." Photographer: Graeme Williams / South Photos. The electric fence dividing Diepsloot from the adjacent formal suburb. One side: tarred roads, streetlights, security estates. The other: gravel, dust, and a million people building homes from corrugated iron. The boundary is not a metaphor.

03 — Solar panel edge / Light meets silicon. "Mono-crystalline Surface Detail." Photographer: Luca Locatelli / Magnum Photos (c) Luca Locatelli / Magnum Photos for The New York Times. A macro image of the edge of a photovoltaic cell — the busbar, the fingers, the grain of the silicon wafer where light becomes current. The line between a photon and an electron is a few hundred nanometres of doped crystalline silicon.

04 — Kármán Line / Atmosphere meets space. "Earth Limb at Dawn, ISS Expedition 57." Photographer: Alexander Gerst / ESA / NASA. The thin blue line of the atmosphere against the black of space. The entire atmosphere is visible as a band roughly 16 km thick at the limb. Every weather system, every flight, every breath — within that line. Beyond it: vacuum, radiation, the rest of the universe.

05 — Atmosphere-Space / The gradient. "Nacreous Clouds above Antarctica, McMurdo Station." Photographer: Josh Landis / National Science Foundation. Polar stratospheric clouds at 15-25 km altitude, iridescent, lit from below by the sun below the horizon. These clouds only form below -78°C. They mark the boundary between the troposphere and the stratosphere, between weather and permanent dryness, between the human world and the atmosphere we cannot breathe.

A two-lane road, a boom, a hut with a hand-painted sign. Southern Africa, 1994.

The sign gives an estimated waiting time. The estimate is handwritten in chalk on a slate. It changes every day.

"Processing time today: morning — 12 min. Afternoon — maybe longer. Sorry for inconvenience."

Someone took the time to chip the chalk, wipe the slate, write a message that was true for that day, for that line of cars, under that sun.

A border post is not a line on a map.
It is a place where someone has to tell you something true before you can cross.

"The most interesting things happen at edges."

Coastlines concentrate life. Markets concentrate exchange. Attention concentrates at the boundary between what we know and what we do not. The gradient is where the energy is.

A boundary is not a wall. It is a seam. Two things meet there, and the meeting produces something that neither side could produce alone. A neural synapse is a boundary. The space between an antenna and a satellite is a boundary. The crack between the old world and the new one is where the light gets in.

You are building at an edge right now. The edge of what is possible, the edge of what is permitted, the edge of what you know how to do. That is not a problem. That is exactly where you are supposed to be.