TL;DR The interesting question in 2026 is not “can a local model do this”, it is “which jobs should you give it”. My stack: Ollama for inference, Letta for persistent agent memory, Obsidian as the second brain, Home Assistant for the physical world, and a small router that decides where each thought goes. Three jobs are the sweet spot for local: inbox triage, note enrichment, and routine automation. Each one is repetitive, private, and tolerant of a bit of latency. Two jobs are still worth handing to a frontier cloud model: anything novel-and-hard, and anything where you want the best draft on the first attempt. The bit nobody talks about is the router. The model is not the product. The thing that decides which model gets which job is the product. Why Local Got Interesting For years the answer to “should I run an LLM locally” was “no, just use the API”. The API was cheaper, faster, smarter, and you did not have to think about VRAM. The only reason to go local was privacy, and most people did not actually care about privacy enough to give up the quality gap. ...

Most people writing about AI risk in 2026 are recent arrivals. Roman Yampolskiy is not. He has been making the same argument - that advanced AI systems may be fundamentally uncontrollable - since before the field of AI safety had a settled name, which is partly because he is the one who gave it that name. Whether you find his conclusions alarmist, prescient, or somewhere in between depends mostly on how you read the gap between current systems and the ones he writes about. This post is an attempt to lay out the man, the argument, and the reasons it deserves more than a dismissal. ...

TL;DR Low Earth orbit (LEO) in 2026 contains roughly 40,000 tracked objects larger than 10cm and an estimated 1 million pieces of debris between 1cm and 10cm. Most of it is dead satellites, spent stages, and fragments from past collisions or anti-satellite tests. The economics are a textbook tragedy of the commons. Each launch operator captures the upside of putting hardware in orbit. The cost of debris is shared across every other operator and every future mission. There is no global price on creating debris. The risk is non-linear. Kessler syndrome - a cascade where collisions create more debris that triggers more collisions - is not a hypothetical. We are already in the early stages in some altitude bands. The fix is also a textbook commons solution: price the externality. End-of-life deorbit requirements, debris remediation funds, transparent collision-avoidance markets, and active debris removal services. Some of this exists. Most of it is undersupplied. The realistic 2026 picture: not yet a crisis, on a trajectory that becomes one within a decade if nothing changes, with the most useful policy interventions being the ones that price debris creation directly rather than relying on goodwill. The Numbers Order-of-magnitude figures, drawn from ESA’s space debris office and LeoLabs tracking data, as of 2026: ...

TL;DR The self-hosted vs managed decision in 2026 is genuinely different from the same decision in 2022. The math has shifted in three directions: cloud egress costs, AI workload economics, and self-hosted tooling maturity. Managed remains the right default for most teams. The thing that has changed is that the threshold at which self-hosting becomes worth considering has dropped. Workloads that were obviously managed in 2022 are genuine 50/50 calls in 2026. The most important shift is that self-hosting is no longer synonymous with on-premises. Modern self-hosting often means renting bare-metal in a colocation, running your own clusters in a hyperscaler, or using sovereign cloud providers - all with different economics. For specific categories - AI inference at scale, data egress-heavy workloads, predictable steady-state compute, regulated environments - self-hosting now wins on cost more often than people assume. The honest framing: managed is the right default; self-hosting is the right minority case; the minority is bigger than it used to be. Why This Decision Got Harder For most of the 2010s the answer was easy. Managed services were cheaper than self-hosting once you priced in operational overhead. The cloud providers competed aggressively. Self-hosting was for the regulated, the eccentric, and the very large. ...

TL;DR With the open table format wars largely settled, the strategic fight in 2026 has moved up to the catalog layer - the system that manages tables, namespaces, governance, and access. Four credible open or open-ish catalogs are now in serious play: Unity Catalog (Databricks), Polaris (Snowflake), Apache Gravitino (Datastrato/community), and Project Nessie (Dremio/community). All four implement the Iceberg REST catalog spec to varying degrees, which means clients can talk to them through a common protocol. The differentiation has moved to governance, multi-tenancy, lineage, federation, and developer experience. Unity is the most production-mature and the most coupled to Databricks. Polaris is the cleanest open implementation of the REST spec. Gravitino is the most ambitious in scope - aiming to catalog non-table assets too. Nessie is the most opinionated about Git-style branching for data. The winning catalog will probably not be a single project. It will be the protocol (Iceberg REST) plus multiple compliant implementations plus federation between them. That is the picture 2026 ends with. Why The Catalog Layer Matters Now A table format defines how data is laid out on disk. A catalog defines: ...

A few weeks ago, four astronauts came home from the Moon for the first time since 1972. Artemis II splashed down on April 11, 2026, after a nine-day flight that took its crew further from Earth than any human has ever travelled - 252,756 miles, a new record set by Reid Wiseman, Victor Glover, Christina Koch, and Jeremy Hansen. It is the clearest signal yet that human spaceflight has stopped being a thing of the past and started being a thing of the near future again. But the headline mission is only one piece of a much larger picture. The decade we are living in is shaping up to be the most consequential one for crewed space travel since Apollo - and unlike the 1960s, this time it is not a single government driving it. ...

TL;DR 2026 is the inflection point for humanoid robotics - real customers like BMW, GXO, and Mercedes-Benz are paying for deployments, not just watching demos Hardware is no longer the bottleneck; the constraints have shifted to physical training data, unstructured-task autonomy, and production supply chains The economics work today for two-to-three shift warehouse operations via Robots-as-a-Service contracts at roughly USD 30-50K per year Production volumes still lag announcements by 3-5x - Unitree is likely the 2026 volume leader, not Tesla or Figure The form factor wins where environments are human-shaped and mixed-use; wheeled robots remain cheaper in purpose-built facilities For most of the last decade, humanoid robotics looked like a category that would always be three years away. Demos were impressive, factory floors stayed empty, and serious analysts pointed to bipedal locomotion, dexterous manipulation, and the price of high torque-density actuators as reasons the form factor would lose to wheeled and fixed-arm systems for any real industrial work. ...

TL;DR One year after Polkadot 2.0 shipped its three flagship pieces - Agile Coretime, Elastic Scaling, and Asynchronous Backing - the picture is mixed but mostly positive. What worked: core prices collapsed, network utilization roughly doubled, and the barrier to entry is now hundreds of dollars instead of millions. New teams are shipping that would never have run a crowdloan. What did not: the secondary market for cores is thin, bulk sales are dominated by a small set of repeat bidders, and the developer story for “buy a core and ship something” is still rougher than it should be. The honest verdict: Agile Coretime delivered on the economics. It did not deliver on the user-experience promise. Polkadot 2.0 is a better foundation than Polkadot 1.0 by every measurable metric, but the application layer is still where the network has to prove itself. Where We Were A Year Ago Last September I wrote a plain-English explainer of Agile Coretime. The pitch was simple: stop selling parachain slots like reserved parking spaces and start selling them like a parking meter. Pay for what you use, when you use it. Resell what you do not. ...

If you have spent any time around electronic music in the last decade, you have probably seen the letters MPE written on the side of a controller and not thought too much about them. The acronym sounds like a feature bullet. It is not. It is a quiet but fundamental reframing of what an electronic instrument can do, and once you have spent serious time playing one, going back to a fixed-velocity keyboard feels like trading a touch screen for a number pad. ...

By mid-2026, the “in-the-box” vs “out-of-the-box” debate has fundamentally shifted. We no longer argue about analog warmth or filter aliasing - neural synthesis has made those distinctions almost invisible to the ear. The new battleground is cognitive load, and that is where dedicated hardware sequencers are quietly winning ground back. As I argued in The Automation Paradox, once AI can generate a passable 16-bar loop in seconds, the human’s job shifts to curation and intent. A hardware sequencer is the most direct tool we have for enforcing that intent. ...