Luma Digest - AI Intelligence Feed https://luma.marbl.codes Daily curated intelligence on AI, Quantum Computing, and Web Development. Authored by Richard Bland. en-gb Sun, 24 May 2026 12:52:44 GMT https://marbl.codes/assets/icons/marbl-codes-icon.png Luma Digest https://luma.marbl.codes hello@marbl.codes (Richard Bland) hello@marbl.codes (Richard Bland) © 2026 Marbl Codes Technology Artificial Intelligence Quantum Computing Web Development Physical AI doesn't need humanoids. It needs edge chips. https://luma.marbl.codes/digest/physical-ai-doesnt-need-humanoids-it-needs-edge-chips https://luma.marbl.codes/digest/physical-ai-doesnt-need-humanoids-it-needs-edge-chips Sun, 24 May 2026 10:37:16 GMT Richard Bland Technology Digest Physical AI doesn't need humanoids. It needs edge chips. Physical AI doesn't need humanoids. It needs edge chips.

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> When AI Takes Action, Verification Changes Completely https://luma.marbl.codes/digest/when-ai-takes-action-verification-changes-completely https://luma.marbl.codes/digest/when-ai-takes-action-verification-changes-completely Sun, 24 May 2026 05:21:23 GMT Richard Bland Technology Digest When AI Takes Action, Verification Changes Completely When AI Takes Action, Verification Changes Completely

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Motors, Agents, and the Politics of Power https://luma.marbl.codes/digest/motors-agents-and-the-politics-of-power https://luma.marbl.codes/digest/motors-agents-and-the-politics-of-power Sat, 23 May 2026 11:10:03 GMT Richard Bland Technology Digest Motors, Agents, and the Politics of Power Motors, Agents, and the Politics of Power

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Building agents that survive contact with production https://luma.marbl.codes/digest/building-agents-that-survive-contact-with-production https://luma.marbl.codes/digest/building-agents-that-survive-contact-with-production Sat, 23 May 2026 06:32:51 GMT Richard Bland Technology Digest Building agents that survive contact with production Building agents that survive contact with production

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Agents need computers-and everything is changing https://luma.marbl.codes/digest/agents-need-computers-and-everything-is-changing https://luma.marbl.codes/digest/agents-need-computers-and-everything-is-changing Fri, 22 May 2026 12:19:12 GMT Richard Bland Technology Digest Agents need computers-and everything is changing Agents need computers-and everything is changing

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Storage bottlenecks, RAG choices, and quantum progress https://luma.marbl.codes/digest/storage-bottlenecks-rag-choices-and-quantum-progress https://luma.marbl.codes/digest/storage-bottlenecks-rag-choices-and-quantum-progress Fri, 22 May 2026 07:40:08 GMT Richard Bland Technology Digest Storage bottlenecks, RAG choices, and quantum progress Storage bottlenecks, RAG choices, and quantum progress

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Robots Moving to Factory Floors, Agents Reshaping Dev https://luma.marbl.codes/digest/robots-moving-to-factory-floors-agents-reshaping-dev https://luma.marbl.codes/digest/robots-moving-to-factory-floors-agents-reshaping-dev Thu, 21 May 2026 12:45:29 GMT Richard Bland Technology Digest Robots Moving to Factory Floors, Agents Reshaping Dev Robots Moving to Factory Floors, Agents Reshaping Dev

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Who gets the new jobs when AI reshapes work? https://luma.marbl.codes/digest/who-gets-the-new-jobs-when-ai-reshapes-work https://luma.marbl.codes/digest/who-gets-the-new-jobs-when-ai-reshapes-work Thu, 21 May 2026 07:44:44 GMT Richard Bland Technology Digest Who gets the new jobs when AI reshapes work? Who gets the new jobs when AI reshapes work?

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Robots Learn Grasping. Google Drops Video AI. Agents Take Over Development. https://luma.marbl.codes/digest/robots-learn-grasping-google-drops-video-ai-agents-take-over-development https://luma.marbl.codes/digest/robots-learn-grasping-google-drops-video-ai-agents-take-over-development Wed, 20 May 2026 12:21:06 GMT Richard Bland Technology Digest Robots Learn Grasping. Google Drops Video AI. Agents Take Over Development. Robots Learn Grasping. Google Drops Video AI. Agents Take Over Development.

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Agent Control Planes Ship. The Policy Files Don't. https://luma.marbl.codes/digest/agent-control-planes-ship-the-policy-files-dont https://luma.marbl.codes/digest/agent-control-planes-ship-the-policy-files-dont Wed, 20 May 2026 07:38:11 GMT Richard Bland Technology Digest Agent Control Planes Ship. The Policy Files Don't. Agent Control Planes Ship. The Policy Files Don't.

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Robots moving into factories. Agents moving into production. https://luma.marbl.codes/digest/robots-moving-into-factories-agents-moving-into-production https://luma.marbl.codes/digest/robots-moving-into-factories-agents-moving-into-production Tue, 19 May 2026 12:39:51 GMT Richard Bland Technology Digest Robots moving into factories. Agents moving into production. Robots moving into factories. Agents moving into production.

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> OpenAI Wins Musk Lawsuit; Web Tooling Gets Faster https://luma.marbl.codes/digest/openai-wins-musk-lawsuit-web-tooling-gets-faster https://luma.marbl.codes/digest/openai-wins-musk-lawsuit-web-tooling-gets-faster Tue, 19 May 2026 07:40:39 GMT Richard Bland Technology Digest OpenAI Wins Musk Lawsuit; Web Tooling Gets Faster OpenAI Wins Musk Lawsuit; Web Tooling Gets Faster

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Humanoids Move into Real Work; Token Bills Arrive https://luma.marbl.codes/digest/humanoids-move-into-real-work-token-bills-arrive https://luma.marbl.codes/digest/humanoids-move-into-real-work-token-bills-arrive Mon, 18 May 2026 13:44:07 GMT Richard Bland Technology Digest Humanoids Move into Real Work; Token Bills Arrive Humanoids Move into Real Work; Token Bills Arrive

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Seven Agents, One Router: How Real-World AI Pipelines Work https://luma.marbl.codes/digest/seven-agents-one-router-how-real-world-ai-pipelines-work https://luma.marbl.codes/digest/seven-agents-one-router-how-real-world-ai-pipelines-work Mon, 18 May 2026 07:58:08 GMT Richard Bland Technology Digest Seven Agents, One Router: How Real-World AI Pipelines Work Seven Agents, One Router: How Real-World AI Pipelines Work

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Open robots, smarter agents, and the people who still have to own the outcome https://luma.marbl.codes/digest/open-robots-smarter-agents-and-the-people-who-still-have-to-own-the-outcome https://luma.marbl.codes/digest/open-robots-smarter-agents-and-the-people-who-still-have-to-own-the-outcome Sun, 17 May 2026 11:05:35 GMT Richard Bland Technology Digest Open robots, smarter agents, and the people who still have to own the outcome Open robots, smarter agents, and the people who still have to own the outcome

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> ArXiv Sets Limits on AI. Ubuntu Goes Local. OpenAI Reshuffles. https://luma.marbl.codes/digest/arxiv-sets-limits-on-ai-ubuntu-goes-local-openai-reshuffles https://luma.marbl.codes/digest/arxiv-sets-limits-on-ai-ubuntu-goes-local-openai-reshuffles Sun, 17 May 2026 06:43:00 GMT Richard Bland Technology Digest ArXiv Sets Limits on AI. Ubuntu Goes Local. OpenAI Reshuffles. ArXiv Sets Limits on AI. Ubuntu Goes Local. OpenAI Reshuffles.

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Cobots surge 56%. Agents ship 5x daily. Security systems outperform models. https://luma.marbl.codes/digest/cobots-surge-56-agents-ship-5x-daily-security-systems-outperform-models https://luma.marbl.codes/digest/cobots-surge-56-agents-ship-5x-daily-security-systems-outperform-models Sat, 16 May 2026 11:02:10 GMT Richard Bland Technology Digest Cobots surge 56%. Agents ship 5x daily. Security systems outperform models. Cobots surge 56%. Agents ship 5x daily. Security systems outperform models.

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> The Bot That Actually Listens: Agents That Know Their Limits https://luma.marbl.codes/digest/the-bot-that-actually-listens-agents-that-know-their-limits https://luma.marbl.codes/digest/the-bot-that-actually-listens-agents-that-know-their-limits Sat, 16 May 2026 06:19:47 GMT Richard Bland Technology Digest The Bot That Actually Listens: Agents That Know Their Limits The Bot That Actually Listens: Agents That Know Their Limits

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Robots are sorting, agents are shipping, and Cerebras just went public https://luma.marbl.codes/digest/robots-are-sorting-agents-are-shipping-and-cerebras-just-went-public https://luma.marbl.codes/digest/robots-are-sorting-agents-are-shipping-and-cerebras-just-went-public Fri, 15 May 2026 11:55:52 GMT Richard Bland Technology Digest Robots are sorting, agents are shipping, and Cerebras just went public Robots are sorting, agents are shipping, and Cerebras just went public

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> When AI Agents Hit the Hard Limits of Production https://luma.marbl.codes/digest/when-ai-agents-hit-the-hard-limits-of-production https://luma.marbl.codes/digest/when-ai-agents-hit-the-hard-limits-of-production Fri, 15 May 2026 06:54:47 GMT Richard Bland Technology Digest When AI Agents Hit the Hard Limits of Production When AI Agents Hit the Hard Limits of Production

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> Humanoid Robots Hit Eight-Hour Shifts; Anthropic Meters the API https://luma.marbl.codes/digest/humanoid-robots-hit-eight-hour-shifts-anthropic-meters-the-api https://luma.marbl.codes/digest/humanoid-robots-hit-eight-hour-shifts-anthropic-meters-the-api Thu, 14 May 2026 11:51:30 GMT Richard Bland Technology Digest Humanoid Robots Hit Eight-Hour Shifts; Anthropic Meters the API Humanoid Robots Hit Eight-Hour Shifts; Anthropic Meters the API

Read the full digest →

Curated by Richard Bland for Luma Digest

]]> [Featured] The Task-Specific Robot Wins: Why the Future Isn't Humanoid https://luma.marbl.codes/featured/the-task-specific-robot-wins-why-the-future-isnt-humanoid https://luma.marbl.codes/featured/the-task-specific-robot-wins-why-the-future-isnt-humanoid Sun, 24 May 2026 10:37:16 GMT Richard Bland Robotics The Task-Specific Robot Wins: Why the Future Isn't Humanoid The Task-Specific Robot Wins: Why the Future Isn't Humanoid

Hailo's VP of physical AI, Orr Danon, has a thesis that cuts against the grain of current robotics hype: the future isn't humanoid robots doing everything. It's millions of task-specific machines, each brilliant at one thing, running intelligence locally on cheap edge processors. The argument is simple. A humanoid robot designed to do everything is expensive, complex, and years away from reliable deployment. A robot designed to do one thing - pick strawberries, inspect pipelines, fold laundry - can be optimised ruthlessly for cost, speed, and durability. And critically, it can run its AI models on-device, no cloud required. Why Edge Intelligence Changes the Economics Running AI on edge processors - chips built into the robot itself - solves three problems at once. First, latency. A robot arm sorting components on a factory line can't wait 200 milliseconds for a cloud response. It needs to decide in 10 milliseconds or the line stops. Second, cost. Sending video feeds to the cloud...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] AI Found Connections We Missed for 80 Years https://luma.marbl.codes/featured/ai-found-connections-we-missed-for-80-years https://luma.marbl.codes/featured/ai-found-connections-we-missed-for-80-years Sun, 24 May 2026 10:37:16 GMT Richard Bland Voices AI Found Connections We Missed for 80 Years AI Found Connections We Missed for 80 Years

OpenAI's reasoning model just solved an 80-year-old geometry problem. Not by brute force, not by training on millions of examples, but by finding an unexpected bridge between discrete geometry and algebraic number theory. Two fields that mathematicians had studied separately for decades, connected by an insight the model surfaced in minutes. Azeem Azhar's reflection on this in Exponential View #575 points to something deeper than raw problem-solving capability. The value of AI might not be in replacing human reasoning, but in connecting isolated domains of knowledge that we've kept in separate boxes. Why This Matters Beyond Mathematics The geometry breakthrough is impressive on its own. But the pattern is what's interesting. The model didn't invent new mathematics from scratch. It recognised that tools from one field - algebraic number theory - could unlock problems in another - discrete geometry. That's not brute computation. That's pattern recognition across domains. We...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Agent UIs Don't Have to Be Chat Windows https://luma.marbl.codes/featured/agent-uis-dont-have-to-be-chat-windows https://luma.marbl.codes/featured/agent-uis-dont-have-to-be-chat-windows Sun, 24 May 2026 10:37:16 GMT Richard Bland Builders Agent UIs Don't Have to Be Chat Windows Agent UIs Don't Have to Be Chat Windows

RL Nabors built a comic reader that renders interactively inside Claude. Not as a chatbot that describes comics. Not as an app that Claude controls via API. The comic itself - panels, navigation, zoom - lives inside the agent interface, rendered natively using Model Context Protocol (MCP). The demo is small, but the concept is large: agent UIs don't have to be text boxes. The web platform itself can become the interface. What MCP Actually Enables Model Context Protocol is Anthropic's framework for letting agents interact with external systems. Most MCP implementations so far have been tool-calling: the agent reads a database, sends an email, triggers a workflow. Input-output transactions. Nabors' comic reader does something different. It uses MCP to render a visual interface directly in the chat window. Claude doesn't just describe the comic or fetch metadata. It serves the comic itself - interactive, zoomable, paginated - as part of the conversation. The user navigates panels by...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Problem Isn't Whether AI Gets Facts Right https://luma.marbl.codes/featured/the-problem-isnt-whether-ai-gets-facts-right https://luma.marbl.codes/featured/the-problem-isnt-whether-ai-gets-facts-right Sun, 24 May 2026 05:21:23 GMT Richard Bland AI The Problem Isn't Whether AI Gets Facts Right The Problem Isn't Whether AI Gets Facts Right

A developer wakes up to find their AI agent has sent 300 emails overnight. Every single one is factually accurate. Every single one went to the wrong people.This is the new problem space. We've spent two years obsessing over whether AI systems hallucinate - whether the facts they generate are true. But as AI moves from answering questions to taking actions, fact-checking becomes almost irrelevant. The real question isn't "is this information correct?" It's "should this thing be doing this at all?"When AI Systems Act, Not Just SpeakAn AI that writes code, sends emails, or deploys infrastructure isn't just generating text you can verify. It's making decisions with consequences. And those consequences don't care whether the underlying facts were correct. They care whether the action was appropriate.A new framework from developers working with autonomous AI systems breaks this down into four verification layers that have nothing to do with factual accuracy: direction, scope,...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Sensor That Measures a Billionth of a Billionth of a Joule https://luma.marbl.codes/featured/the-sensor-that-measures-a-billionth-of-a-billionth-of-a-joule https://luma.marbl.codes/featured/the-sensor-that-measures-a-billionth-of-a-billionth-of-a-joule Sun, 24 May 2026 05:21:23 GMT Richard Bland Quantum Computing The Sensor That Measures a Billionth of a Billionth of a Joule The Sensor That Measures a Billionth of a Billionth of a Joule

Quantum computers fail in expensive ways. A single qubit drifts out of calibration, and suddenly your million-dollar system is producing garbage. Right now, fixing that requires a human expert to manually tune hundreds of parameters, one at a time, hoping they catch the problem before it cascades.Researchers at Aalto University just built a sensor that might make that process obsolete. It's called a bolometer, and it can measure energy changes smaller than a zeptojoule - that's 10 to the power of minus 21 joules. To put that in perspective: if a joule is a litre of water, a zeptojoule is a single molecule.The breakthrough isn't just about measurement precision. It's about what becomes possible when you can sense quantum noise at that resolution. Specifically: you can automate calibration. And that changes the economics of quantum computing entirely.Why Quantum Systems DriftQuantum computers are fragile. They operate at temperatures colder than deep space, and even at those...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Dart Just Closed the Full-Stack Loop https://luma.marbl.codes/featured/dart-just-closed-the-full-stack-loop https://luma.marbl.codes/featured/dart-just-closed-the-full-stack-loop Sun, 24 May 2026 05:21:23 GMT Richard Bland Web Development Dart Just Closed the Full-Stack Loop Dart Just Closed the Full-Stack Loop

Most full-stack developers live in two codebases. One for the frontend. One for the backend. Same data models, different languages. Same validation logic, duplicated. Same business rules, written twice. It works, but it's inefficient and error-prone.Dart just made that optional. With Cloud Functions now in experimental preview and the Firebase Admin SDK available for server-side Dart, you can write your entire stack - frontend, backend, and shared packages - in a single language. Same types. Same tooling. Same mental model.The new handbook from freeCodeCamp walks through the whole pattern: building shared packages that both your Flutter app and your Cloud Functions import, eliminating duplicate model definitions, and handling the gaps where server and client logic genuinely need to diverge.Why This Matters NowDart has been a strong frontend language since Flutter proved you could build production mobile apps with it. But the backend story was always awkward. You could run Dart on a...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The $150 Motor That Just Became Open Source https://luma.marbl.codes/featured/the-150-motor-that-just-became-open-source https://luma.marbl.codes/featured/the-150-motor-that-just-became-open-source Sat, 23 May 2026 11:10:03 GMT Richard Bland Robotics The $150 Motor That Just Became Open Source The $150 Motor That Just Became Open Source

A researcher cracked open the Unitree Go2 robot dog this month and did something the manufacturer never intended - extracted the firmware from its proprietary motors and made them work without vendor lock-in. The reverse engineering effort targeted the GO-M8018-6 motor, a high-torque actuator that's genuinely good hardware. The problem? It only talks to Unitree's own controllers. If you wanted to use these motors in your own robotics project, you were stuck in their ecosystem. Not anymore. Why This Matters for Builders High-quality robot actuators are expensive. The cheap ones overheat or lack precision. The precise ones cost thousands. Unitree found a middle ground - capable motors at a price hobbyists can afford - but kept them locked behind proprietary firmware. The researcher bypassed the bootloader, pulled the encrypted firmware, and decrypted it. That's the technical achievement. The practical achievement is what comes next: open-source alternatives that let anyone use...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Why People Hate AI More Than They Hated Globalisation https://luma.marbl.codes/featured/why-people-hate-ai-more-than-they-hated-globalisation https://luma.marbl.codes/featured/why-people-hate-ai-more-than-they-hated-globalisation Sat, 23 May 2026 11:10:03 GMT Richard Bland Voices Why People Hate AI More Than They Hated Globalisation Why People Hate AI More Than They Hated Globalisation

AI company revenues are breaking records. Investment is pouring in. And public resistance is accelerating faster than either of those numbers. Azeem Azhar's analysis this week asks the right question: why is the backlash growing when the technology is delivering exactly what the industry promised? The answer comes down to visibility. People can see the costs. They can't see the benefits yet. The Infrastructure Problem Globalisation happened at a distance. Factories closed in one country, opened in another. The displaced workers saw the consequences, but most people experienced it as cheaper goods and didn't connect the dots until much later. AI is different. The infrastructure is local and loud. Data centres need water - millions of litres of it. They need power - entire grids' worth. They need construction - roads dug up, land cleared, noise and disruption that people living nearby can't ignore. A town in rural Ireland gets planning permission for a massive data centre. Local...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Anthropic's AI Found 10,000 Critical Bugs in a Month https://luma.marbl.codes/featured/anthropics-ai-found-10000-critical-bugs-in-a-month https://luma.marbl.codes/featured/anthropics-ai-found-10000-critical-bugs-in-a-month Sat, 23 May 2026 11:10:03 GMT Richard Bland Builders Anthropic's AI Found 10,000 Critical Bugs in a Month Anthropic's AI Found 10,000 Critical Bugs in a Month

Anthropic ran an experiment this month that should worry every security team still doing manual code audits. Project Glasswing - an AI agent built to find vulnerabilities in open-source software - discovered over 10,000 high or critical-severity bugs in essential codebases within 30 days. That's not a slow, careful review. That's industrial-scale security auditing at a speed no human team could match. The initial update doesn't name every project affected, but it confirms the findings are real, reported to maintainers, and being patched. The bugs are the kind that matter - authentication bypasses, memory corruption issues, privilege escalation vulnerabilities. The kind security researchers spend weeks hunting for. What Changed AI tools have been assisting with code review for a while now. This is different. Glasswing isn't suggesting improvements or flagging suspicious patterns. It's performing full security audits - reading codebases, understanding logic flow, spotting edge cases...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Why Your AI Agent Demo Works but Your Production System Doesn't https://luma.marbl.codes/featured/why-your-ai-agent-demo-works-but-your-production-system-doesnt https://luma.marbl.codes/featured/why-your-ai-agent-demo-works-but-your-production-system-doesnt Sat, 23 May 2026 06:32:51 GMT Richard Bland AI Why Your AI Agent Demo Works but Your Production System Doesn't Why Your AI Agent Demo Works but Your Production System Doesn't

Most AI agent tutorials end at the demo. They show you a chatbot that fetches weather data or searches Wikipedia, and they call it done. But between that demo and a production system is a gap that swallows teams whole.Gursharan Singh's production-oriented guide to AI agents starts where most tutorials end - with the question of why demos work and production fails. The answer isn't sexy: it's state management, control loops, and the unglamorous work of making systems reliable when users don't follow the script.The Gap Between Demo and ProductionA demo agent handles one turn. A production agent handles hundreds of turns across sessions, remembers context, fails gracefully when tools break, and recovers when users ask nonsense questions halfway through a workflow. The difference is architectural.Singh breaks agent systems into three primitives: MCP (Model Context Protocol) for tool integration, RAG (Retrieval-Augmented Generation) for knowledge injection, and Skills - reusable patterns...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Three-Level Quantum Systems Beat Two-Level Ones at Staying Alive https://luma.marbl.codes/featured/three-level-quantum-systems-beat-two-level-ones-at-staying-alive https://luma.marbl.codes/featured/three-level-quantum-systems-beat-two-level-ones-at-staying-alive Sat, 23 May 2026 06:32:51 GMT Richard Bland Quantum Computing Three-Level Quantum Systems Beat Two-Level Ones at Staying Alive Three-Level Quantum Systems Beat Two-Level Ones at Staying Alive

A quantum heat engine sounds like science fiction. But the physics is real, and the challenge is brutal: how do you extract useful energy from a quantum system when the act of measuring it destroys the quantum state?Researchers working on quantum heat engines have found something unexpected. Three-level quantum systems - qutrits - resist decoherence better than two-level qubits and extract more energy in the process. This isn't a marginal improvement. It's a pathway toward nanoscale quantum engines that might actually work outside a lab.Why Quantum Heat Engines MatterClassical heat engines - car engines, power plants, refrigerators - operate on macroscopic temperature differences. Quantum heat engines work at the nanoscale, where temperature is a fuzzy concept and energy extraction happens in discrete packets.The promise is energy storage and conversion at scales where classical engines can't function. Molecular machines. Nanoscale refrigeration. Quantum batteries. But the problem is...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Control Loop Is Where Your Agent Actually Lives https://luma.marbl.codes/featured/the-control-loop-is-where-your-agent-actually-lives https://luma.marbl.codes/featured/the-control-loop-is-where-your-agent-actually-lives Sat, 23 May 2026 06:32:51 GMT Richard Bland Web Development The Control Loop Is Where Your Agent Actually Lives The Control Loop Is Where Your Agent Actually Lives

Tool calling is table stakes. Every LLM API supports it now. But tool calling alone doesn't make an agent - it makes a chatbot with functions. The difference is the control loop.Gursharan Singh's production guide to agentic workflows focuses on the architecture that wraps the model - the control loop, state management, and the patterns that turn a working demo into a reliable system.What a Control Loop Actually DoesYour agent calls a weather API. The API times out. What happens next?In a demo, nothing. The model hallucinates a response or the system crashes. In production, the control loop catches the timeout, decides whether to retry, logs the failure, and either falls back to cached data or tells the user the service is unavailable.The control loop is the decision-making layer that sits between the model and the outside world. It handles retries, validates tool outputs, manages turn-taking in multi-step workflows, and enforces guardrails when the model tries to do something...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Boston Dynamics' Atlas Now Lifts Fridges - 25,000 Units Coming https://luma.marbl.codes/featured/boston-dynamics-atlas-now-lifts-fridges-25000-units-coming https://luma.marbl.codes/featured/boston-dynamics-atlas-now-lifts-fridges-25000-units-coming Fri, 22 May 2026 12:19:12 GMT Richard Bland Robotics Boston Dynamics' Atlas Now Lifts Fridges - 25,000 Units Coming Boston Dynamics' Atlas Now Lifts Fridges - 25,000 Units Coming

The new Atlas from Boston Dynamics just picked up a loaded fridge and walked it across a factory floor. Not a demo unit. A production fridge, full weight, shifting load as it moved.This is the electric Atlas, the one they announced last year after retiring the hydraulic version. The one everyone assumed would take years to do real work. It's doing real work now.What ChangedThe previous Atlas was a research platform - incredible to watch, technically brilliant, but not designed for repetitive industrial tasks. This version is built for deployment. Hyundai, who acquired Boston Dynamics, plans to put 25,000 of these units across their factories.That's not a pilot programme. That's production scale.The technical leap is in how it handles dynamic loads. A fridge on a trolley isn't a static object - it shifts, the centre of gravity moves, the surface isn't uniform. Atlas adjusts in real-time, compensating for weight distribution as it walks. It's doing what human workers do instinctively -...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Why Agents Need Real Computers, Not Lambda Functions https://luma.marbl.codes/featured/why-agents-need-real-computers-not-lambda-functions https://luma.marbl.codes/featured/why-agents-need-real-computers-not-lambda-functions Fri, 22 May 2026 12:19:12 GMT Richard Bland Voices Why Agents Need Real Computers, Not Lambda Functions Why Agents Need Real Computers, Not Lambda Functions

Ivan Burazin runs Daytona, the infrastructure company behind 850,000 agent runs per day. His central argument: we're building AI agents wrong. We're giving them Lambda functions and API endpoints when they need actual computers.The company is growing at 74% month-on-month. That growth rate suggests they're solving a problem developers are actively hitting right now, not building for a future that might arrive.The Lambda ProblemMost agent frameworks treat compute like a temporary resource. Spin up a function, run the task, tear it down. Stateless, disposable, cheap. That works beautifully for web requests and batch jobs. It falls apart for agents.An agent isn't a single function call. It's a process that runs over time, maintains state, explores possibilities, backtracks, tries different approaches. It needs a filesystem that persists between steps. It needs installed dependencies. It needs to remember what it tried five minutes ago.Lambda functions don't do that. They're designed to...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Testing Google's Agent API for $0.37 - Every Bug You'll Hit https://luma.marbl.codes/featured/testing-googles-agent-api-for-037-every-bug-youll-hit https://luma.marbl.codes/featured/testing-googles-agent-api-for-037-every-bug-youll-hit Fri, 22 May 2026 12:19:12 GMT Richard Bland Builders Testing Google's Agent API for $0.37 - Every Bug You'll Hit Testing Google's Agent API for $0.37 - Every Bug You'll Hit

Stephen Sebastian ran Google's Antigravity managed agents against 14 services and spent $0.37 doing it. Then he wrote up every bug, every cost surprise, and every production readiness gap he found.This is what good builder content looks like. Not "here's why agents are significant". More like "here's what broke when I tried to use them for dependency audits".What Antigravity Actually DoesAntigravity is Google's managed runtime for AI agents. You define a task - in this case, auditing dependencies across multiple services - and the agent handles the execution. It's meant to abstract away the infrastructure complexity of running agents at scale.The promise: you focus on what the agent should do, Google handles how it runs. The reality, as Sebastian found, is messier than the marketing.He tested it on 14 different services, checking for outdated dependencies, security vulnerabilities, and configuration drift. Each run cost between $0.02 and $0.04 in tokens, depending on the size of the...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] This AI Agent Remembers What It Learns - and Keeps Getting Better https://luma.marbl.codes/featured/this-ai-agent-remembers-what-it-learns-and-keeps-getting-better https://luma.marbl.codes/featured/this-ai-agent-remembers-what-it-learns-and-keeps-getting-better Fri, 22 May 2026 07:40:08 GMT Richard Bland AI This AI Agent Remembers What It Learns - and Keeps Getting Better This AI Agent Remembers What It Learns - and Keeps Getting Better

Machine learning models have a memory problem. Train them on new data and they forget what they learned before. It's called catastrophic forgetting, and it's one of the reasons your AI assistant can't actually learn from your corrections. SOLAR - Self-Optimizing Agent for Lifelong Learning - tackles this head-on. It's a system that combines meta-learning with reinforcement learning to build agents that self-improve across streaming data without wiping their previous knowledge. Think of it as an AI that learns how to learn, then keeps that knowledge as the world changes around it. The Problem: Models That Can't Remember Yesterday Most machine learning systems are trained once, then deployed. They're static. If the data distribution shifts - and in the real world, it always does - the model degrades. Retrain it on the new data and it forgets what it knew about the old data. It's a brutal trade-off. For business owners, this means your AI tools need constant retraining. Your fraud...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Xanadu Cuts Quantum Data Loading Costs in Half https://luma.marbl.codes/featured/xanadu-cuts-quantum-data-loading-costs-in-half https://luma.marbl.codes/featured/xanadu-cuts-quantum-data-loading-costs-in-half Fri, 22 May 2026 07:40:08 GMT Richard Bland Quantum Computing Xanadu Cuts Quantum Data Loading Costs in Half Xanadu Cuts Quantum Data Loading Costs in Half

Quantum computers have a boring but critical problem: getting classical data into quantum circuits is expensive. Really expensive. It's called the data loading bottleneck, and it's one of the main reasons quantum algorithms that look brilliant on paper struggle in practice. Xanadu just published work on a new Quantum Read-Only Memory (QROM) copying mechanism that cuts this overhead by 50%. That's not incremental. That's the kind of improvement that changes what you can actually build. The Bottleneck Nobody Talks About Quantum algorithms operate on quantum states. But most of the data we care about - images, text, financial records, sensor readings - exists as classical bits. To use a quantum algorithm, you need to encode that classical data into quantum states. This process is called data loading, and it's slow and resource-intensive. The problem compounds because quantum circuits are fragile. You can't just load data once and reuse it. Every time the circuit runs, you're loading...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Why AI Infrastructure Is Converging on Object Storage https://luma.marbl.codes/featured/why-ai-infrastructure-is-converging-on-object-storage https://luma.marbl.codes/featured/why-ai-infrastructure-is-converging-on-object-storage Fri, 22 May 2026 07:40:08 GMT Richard Bland Web Development Why AI Infrastructure Is Converging on Object Storage Why AI Infrastructure Is Converging on Object Storage

GPU clusters are fast. Storage is slow. That gap is the bottleneck crushing most AI training pipelines right now. MinIO's partnership with NVIDIA addresses this by standardising S3-compatible object storage for AI infrastructure. It's not flashy. It's not a new model architecture. It's plumbing. But it's the kind of plumbing that determines whether your AI system runs at 80% GPU utilisation or 30%. The Problem: GPUs Waiting for Data Modern AI training is limited by data throughput, not compute. You've got GPUs capable of processing terabytes of data per second, but your storage system can only feed them at a fraction of that rate. The GPUs sit idle, waiting for the next batch. You're paying for compute you're not using. Traditional storage systems weren't built for AI workloads. They were optimised for databases and file servers - systems where you read and write small amounts of data frequently. AI training does the opposite: massive sequential reads, constant parallel access,...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Bosch Is Manufacturing Thousands of Wheeled Humanoids for German Factories https://luma.marbl.codes/featured/bosch-is-manufacturing-thousands-of-wheeled-humanoids-for-german-factories https://luma.marbl.codes/featured/bosch-is-manufacturing-thousands-of-wheeled-humanoids-for-german-factories Thu, 21 May 2026 12:45:29 GMT Richard Bland Robotics Bosch Is Manufacturing Thousands of Wheeled Humanoids for German Factories Bosch Is Manufacturing Thousands of Wheeled Humanoids for German Factories

Schaeffler is deploying thousands of wheeled humanoid robots across its German manufacturing facilities starting December 2026. The hardware comes from Humanoid, the robotics company building HMND - a mobile manipulation platform designed for industrial work. Bosch is the manufacturing partner scaling production. This isn't a pilot programme or a proof of concept for the press release. Schaeffler tested the robots handling boxes of multiple sizes in real production environments. The robots worked. Now they're ordering at scale. What Makes HMND Different HMND robots use wheels instead of legs. That single design decision changes the economics entirely. Wheeled bases are cheaper to manufacture, more stable under load, and require less compute to control. For warehouse and factory work - moving boxes, transporting materials, handling repetitive tasks - wheels solve 90% of the mobility problem at a fraction of the cost. The humanoid torso matters for manipulation. Human-designed...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Railway Rebuilt the Cloud for Agents - Here's Why That Actually Matters https://luma.marbl.codes/featured/railway-rebuilt-the-cloud-for-agents-heres-why-that-actually-matters https://luma.marbl.codes/featured/railway-rebuilt-the-cloud-for-agents-heres-why-that-actually-matters Thu, 21 May 2026 12:45:29 GMT Richard Bland Voices Railway Rebuilt the Cloud for Agents - Here's Why That Actually Matters Railway Rebuilt the Cloud for Agents - Here's Why That Actually Matters

Railway runs bare-metal data centres because the economics of agent workloads don't work on traditional cloud infrastructure. Their customers are spending over $200,000 monthly on agent coding tasks alone. That's not hype - that's measured usage from 3 million users running production workloads. Jake Cooper, Railway's founder, walked through their infrastructure thesis with Latent Space. The core argument: agents need different primitives than web applications. The cloud stack we built for serving HTTP requests breaks down when the workload is autonomous code execution. The Agent Infrastructure Problem Traditional cloud platforms charge for compute time. That pricing model assumes short-lived requests - a user hits an endpoint, you run some code, you return a response. Measure in milliseconds, bill in seconds, optimise for throughput. Agent workloads invert that model. An agent coding task might run for minutes or hours. It's not a burst - it's sustained compute doing exploratory...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] 100% Reliable LLM Output - A Control Layer That Actually Works https://luma.marbl.codes/featured/100-reliable-llm-output-a-control-layer-that-actually-works https://luma.marbl.codes/featured/100-reliable-llm-output-a-control-layer-that-actually-works Thu, 21 May 2026 12:45:29 GMT Richard Bland Builders 100% Reliable LLM Output - A Control Layer That Actually Works 100% Reliable LLM Output - A Control Layer That Actually Works

A builder solved the structured output problem. Not with prompt engineering - with a control layer that sits above the LLM and enforces reliability at the system level. The result: 100% structured output success rate in production, handling JSON failures, silent errors, and API outages. The implementation matters because it addresses the gap between LLM demos and LLM products. Demos tolerate failures. Products don't. This control layer bridges that gap through system design rather than model tuning. Why Prompt Engineering Isn't Enough Prompt engineering optimises for average case performance. You craft better instructions, provide examples, tune temperature settings. The model's output improves - from 80% success to 95%, maybe 98% if you're very good. But production systems need 100%. A 2% failure rate on financial data means wrong transactions. On medical records, it means data loss. On automated workflows, it means manual intervention - which eliminates the automation value...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Under-30 Effect: Why New Tech Jobs Cluster Among Young Graduates https://luma.marbl.codes/featured/the-under-30-effect-why-new-tech-jobs-cluster-among-young-graduates https://luma.marbl.codes/featured/the-under-30-effect-why-new-tech-jobs-cluster-among-young-graduates Thu, 21 May 2026 07:44:44 GMT Richard Bland AI The Under-30 Effect: Why New Tech Jobs Cluster Among Young Graduates The Under-30 Effect: Why New Tech Jobs Cluster Among Young Graduates

Seventy years of data reveals a pattern most economists already suspected but couldn't quite prove: when technology creates new jobs, they don't land evenly across the workforce. They cluster among college graduates under 30, living in cities.MIT researchers tracked every major wave of technological change since the 1950s - from mainframes to personal computers to the internet - and found the same distribution every time. New jobs created by technology overwhelmingly favour young, educated workers in urban areas. The question now: will AI follow the same pattern, or break it?Demand-Driven Investment Matters More Than InnovationThe surprising finding isn't just WHO gets the jobs. It's what creates them in the first place.Most people assume job creation flows directly from innovation - new technology appears, new roles emerge. The MIT research shows it's more complicated. Demand-driven investment - businesses deciding to adopt and scale new technology - matters more than the innovation...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Quantum Computing Tackles Routing Problems With 90% Fewer Parameters https://luma.marbl.codes/featured/quantum-computing-tackles-routing-problems-with-90-fewer-parameters https://luma.marbl.codes/featured/quantum-computing-tackles-routing-problems-with-90-fewer-parameters Thu, 21 May 2026 07:44:44 GMT Richard Bland Quantum Computing Quantum Computing Tackles Routing Problems With 90% Fewer Parameters Quantum Computing Tackles Routing Problems With 90% Fewer Parameters

A quantum computing framework just solved combinatorial optimization problems - the kind that plague logistics, routing, and resource allocation - using 90% fewer parameters than classical machine learning approaches. Not faster. Not more accurate. Just dramatically simpler.The research introduces the first end-to-end quantum learning system for contextual combinatorial optimization, built on QAOA - Quantum Approximate Optimization Algorithm. The results suggest quantum computers might excel not by raw speed, but by finding elegant solutions classical systems overlook.What Contextual Combinatorial Optimization Actually MeansThese are problems where you need to find the best arrangement from billions of possibilities, and the definition of "best" changes based on context.Routing delivery vans through a city. Scheduling hospital operating rooms. Assigning frequencies to mobile phone towers. All combinatorial optimization problems. Classical computers solve them by trying arrangements...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Building an Uptime Monitor When the Existing Tools Feel Wrong https://luma.marbl.codes/featured/building-an-uptime-monitor-when-the-existing-tools-feel-wrong https://luma.marbl.codes/featured/building-an-uptime-monitor-when-the-existing-tools-feel-wrong Thu, 21 May 2026 07:44:44 GMT Richard Bland Web Development Building an Uptime Monitor When the Existing Tools Feel Wrong Building an Uptime Monitor When the Existing Tools Feel Wrong

A developer looked at UptimeRobot, saw clutter, and decided to build their own. Not because existing tools don't work - they do - but because sometimes the friction of using something designed for everyone means it fits nobody perfectly.The result is PingBoard, an uptime monitor with status pages, built using Convex for the backend, Clerk for authentication, and Resend for emails. The technical choices matter less than the thinking behind them.The Decision to BuildMost developers face this choice weekly: use an existing tool that's 80% right, or build something custom that's 100% yours. The sensible answer is usually "use the tool" - building takes time, maintenance compounds, and you're trading product work for infrastructure work.But sometimes the 20% gap matters enough. UptimeRobot works. Thousands of businesses rely on it daily. But its interface reflects years of feature accretion - every user request, every edge case, every "just one more option" layered into the UI.For someone...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Training Robot Arms to Grasp - Without Breaking Real Hardware https://luma.marbl.codes/featured/training-robot-arms-to-grasp-without-breaking-real-hardware https://luma.marbl.codes/featured/training-robot-arms-to-grasp-without-breaking-real-hardware Wed, 20 May 2026 12:21:06 GMT Richard Bland Robotics Training Robot Arms to Grasp - Without Breaking Real Hardware Training Robot Arms to Grasp - Without Breaking Real Hardware

Teaching a robot arm to pick things up sounds simple. You grab, you lift, you place. But translate that into code and suddenly you're wrestling with physics engines, reward functions, and the nightmare of trial-and-error learning on hardware that costs more than a car.NVIDIA's Isaac Lab offers a different route: train the policy in simulation, break nothing, then transfer the learned behaviour to real metal. A new tutorial from the ROS community walks through the entire workflow - from setting up a simulated robot arm through preparing the trained model for real-world deployment.Why Simulation FirstReinforcement learning works through repetition. The robot tries something, gets feedback, adjusts, tries again. Do that on real hardware and you're burning through actuators, breaking grippers, and watching your robot fling objects across the room for weeks.Simulation lets you run thousands of attempts in parallel. No broken parts, no safety concerns, no waiting for motors to cool down....

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Google I/O 2026: Three Models and an Agent Platform https://luma.marbl.codes/featured/google-io-2026-three-models-and-an-agent-platform https://luma.marbl.codes/featured/google-io-2026-three-models-and-an-agent-platform Wed, 20 May 2026 12:21:06 GMT Richard Bland Voices Google I/O 2026: Three Models and an Agent Platform Google I/O 2026: Three Models and an Agent Platform

Google held I/O 2026 this week and shipped three distinct models: Gemini 3.5 Flash for agentic tasks, Gemini Omni for video generation, and an updated Antigravity platform that orchestrated 93 agents in a live demo. Not announcements about future capabilities - shipping APIs, available now.Gemini 3.5 Flash: The Coding ModelFlash is positioned as the agentic model in Google's lineup. Faster inference, optimized for tool use and code generation, designed for the workflow where a model needs to chain multiple actions together. Think: "analyze this codebase, identify the bug, propose a fix, test it, commit if it passes".What's interesting here is the positioning. Google isn't trying to make one model do everything. Flash is explicitly for agents and developers. The benchmarks focus on coding tasks, tool-use accuracy, and multi-step reasoning. They're optimizing for a specific use case rather than chasing general-purpose supremacy.For developers building on Gemini, this gives you a clear...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Google's QIZ Platform: Automating the Quantum Cryptography Transition https://luma.marbl.codes/featured/googles-qiz-platform-automating-the-quantum-cryptography-transition https://luma.marbl.codes/featured/googles-qiz-platform-automating-the-quantum-cryptography-transition Wed, 20 May 2026 12:21:06 GMT Richard Bland Builders Google's QIZ Platform: Automating the Quantum Cryptography Transition Google's QIZ Platform: Automating the Quantum Cryptography Transition

Quantum computers aren't breaking RSA encryption yet. But they will. And when they do, every system relying on current cryptographic standards becomes vulnerable overnight. The fix - post-quantum cryptography - requires finding every instance of vulnerable encryption in your infrastructure and replacing it with quantum-resistant algorithms.Google Cloud's QIZ platform automates that discovery and transition process. It scans your infrastructure, identifies cryptographic dependencies, maps them to quantum-safe alternatives, and helps you manage the compliance requirements around the migration. It's infrastructure for a problem most organizations haven't started thinking about yet.The Inventory ProblemHere's what makes quantum safety messy: you don't know where all your encryption lives. It's in TLS certificates, database connections, API authentication, file storage, backup systems, third-party integrations, and legacy code nobody's touched in years.QIZ runs automated discovery across...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Six Vendors Shipped Agent Controls. Zero Shipped the Policy Files. https://luma.marbl.codes/featured/six-vendors-shipped-agent-controls-zero-shipped-the-policy-files https://luma.marbl.codes/featured/six-vendors-shipped-agent-controls-zero-shipped-the-policy-files Wed, 20 May 2026 07:38:11 GMT Richard Bland AI Six Vendors Shipped Agent Controls. Zero Shipped the Policy Files. Six Vendors Shipped Agent Controls. Zero Shipped the Policy Files.

Devenex, Antigravity, Notion, Claude, OpenAI, Salesforce - six vendors launched agent enforcement controls in the past month. Every single one gives you the infrastructure to run agents safely. None of them give you the policy files to define what safe actually means.That's the gap. And it's starting to hurt.The Problem Nobody's SolvingAgent enforcement works like this: you define a policy - what the agent can access, what it can modify, what decisions it can make autonomously - and the enforcement layer makes sure the agent stays within those bounds. The vendors ship the enforcement. They don't ship the policy.So every team writes their own. From scratch. With no shared vocabulary, no standard format, no way to audit across systems. One team defines "read-only database access" one way. Another team defines it differently. Both think they're secure. Neither can prove it.This is where policy-as-code stops being optional. If agents are running in production - and they are, whether...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Quantum Phase Estimation Just Got Twice as Efficient https://luma.marbl.codes/featured/quantum-phase-estimation-just-got-twice-as-efficient https://luma.marbl.codes/featured/quantum-phase-estimation-just-got-twice-as-efficient Wed, 20 May 2026 07:38:11 GMT Richard Bland Quantum Computing Quantum Phase Estimation Just Got Twice as Efficient Quantum Phase Estimation Just Got Twice as Efficient

Phase estimation is one of the core algorithms quantum computers run. It's how you extract useful information from quantum states. And it's expensive - every extra circuit you run eats into your limited quantum runtime. New work from arXiv just cut those circuit runs in half.The breakthrough is an extended statistical framework that handles two things previous methods couldn't: negative Pauli weights and changepoint detection. That sounds abstract. What it means in practice is you can now run quantum phase estimation on a wider range of problems, with half the circuit overhead, without losing accuracy.What ChangedStandard quantum phase estimation requires overlap estimates - you need to know how much your prepared state overlaps with the eigenstates you're trying to measure. Calculating that overlap is expensive. It requires additional quantum circuits, which means more runtime, more noise, more opportunity for error.The new statistical approach sidesteps this entirely. Instead of...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Apple Shipped a 3-Billion-Parameter Model That Runs in Your Pocket https://luma.marbl.codes/featured/apple-shipped-a-3-billion-parameter-model-that-runs-in-your-pocket https://luma.marbl.codes/featured/apple-shipped-a-3-billion-parameter-model-that-runs-in-your-pocket Wed, 20 May 2026 07:38:11 GMT Richard Bland Web Development Apple Shipped a 3-Billion-Parameter Model That Runs in Your Pocket Apple Shipped a 3-Billion-Parameter Model That Runs in Your Pocket

Your iPhone can now run a 3-billion-parameter language model entirely on-device. No API calls. No cloud. No data leaving your phone. Apple's Foundation Models framework just made local AI a first-class citizen in Swift.This changes the economics of building AI into mobile apps. Until now, you had two options: ship a tiny model that runs locally but can't do much, or call a cloud API that's powerful but costs money and raises privacy questions. Foundation Models gives you a third option: a genuinely capable model, running locally, with Swift-native APIs that feel like any other iOS framework.What You Can BuildThe framework handles text generation, summarisation, translation, and structured data extraction. The interesting bit is the @Generable macro. You define a Swift struct - say, a parsed invoice with fields for vendor name, amount, date, and line items - and the macro generates type-safe code that extracts that structure from unstructured text.This matters because structured output...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] FANUC Brings Robot Training Into Virtual Space https://luma.marbl.codes/featured/fanuc-brings-robot-training-into-virtual-space https://luma.marbl.codes/featured/fanuc-brings-robot-training-into-virtual-space Tue, 19 May 2026 12:39:51 GMT Richard Bland Robotics FANUC Brings Robot Training Into Virtual Space FANUC Brings Robot Training Into Virtual Space

Industrial robotics just got a simulation upgrade. FANUC, the world's largest industrial robot manufacturer, has deepened its integration with NVIDIA's Isaac Sim platform - and the implications go beyond faster prototyping. The integration ties FANUC's ROBOGUIDE software directly into Isaac Sim's virtual environment. Engineers can now operate robots in simulation using the same teach pendants they'd use on a factory floor. That matters because it removes the cognitive leap between virtual testing and physical deployment. Imitation Learning at Industrial Scale The real shift here is access to NVIDIA's GR00T N foundation model for imitation learning. This isn't programming through code - it's teaching robots through demonstration. Show a robot how to fold a T-shirt in simulation, and the model translates that into executable motion patterns. Flexible tasks like fabric handling have historically been robotics nightmares. Fabric doesn't behave predictably. It crumples, stretches, and...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Kernel Optimization Is the Real Bottleneck https://luma.marbl.codes/featured/kernel-optimization-is-the-real-bottleneck https://luma.marbl.codes/featured/kernel-optimization-is-the-real-bottleneck Tue, 19 May 2026 12:39:51 GMT Richard Bland Voices Kernel Optimization Is the Real Bottleneck Kernel Optimization Is the Real Bottleneck

Most AI engineering talk focuses on model architecture or training strategies. But according to this week's AINews aggregation from Latent Space, the actual constraint is lower in the stack - kernel optimization is where frontier labs are spending their time. Kernels are the mathematical operations that run on GPUs. Matrix multiplication, attention mechanisms, activation functions - these are the primitives that everything else sits on top of. If your kernels are inefficient, nothing above that layer matters. You're burning compute on overhead. What Google's Hiring Exercise Reveals The AINews episode includes insights from Google researchers on what actually gets tested during hiring for frontier AI roles. The exercises aren't about clever prompting or high-level architecture. They're about understanding GPU memory hierarchies, optimising data movement, and reducing latency at the hardware level. That's telling. If frontier labs are hiring for kernel expertise, it means the...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] GitHub's $46,000 Copilot Exploit https://luma.marbl.codes/featured/githubs-46000-copilot-exploit https://luma.marbl.codes/featured/githubs-46000-copilot-exploit Tue, 19 May 2026 12:39:51 GMT Richard Bland Builders GitHub's $46,000 Copilot Exploit GitHub's $46,000 Copilot Exploit

Theo from t3.gg just proved GitHub's usage metering has a serious flaw. He consumed $46,000 worth of Copilot tokens for a $40 payment. Not through a hack - through normal API usage that the platform failed to rate-limit properly. The exploit video walks through the mechanics, but the core issue is simple: GitHub's billing system couldn't keep up with actual token consumption. Send enough requests fast enough, and the metering breaks down. The platform let him burn through token quotas that should have triggered throttling or account suspension. Why This Matters Beyond One User This isn't about Theo gaming the system - he disclosed it publicly. It's about what the flaw reveals. If one developer can accidentally consume 1,150x their payment value, the accounting infrastructure underneath isn't production-ready. Every SaaS company building on LLM APIs faces the same problem. Tokens are cheap at small scale but expensive at high volume. If your metering lags behind actual usage,...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Musk's OpenAI Lawsuit Fails - IPO Path Now Clear https://luma.marbl.codes/featured/musks-openai-lawsuit-fails-ipo-path-now-clear https://luma.marbl.codes/featured/musks-openai-lawsuit-fails-ipo-path-now-clear Tue, 19 May 2026 07:40:39 GMT Richard Bland AI Musk's OpenAI Lawsuit Fails - IPO Path Now Clear Musk's OpenAI Lawsuit Fails - IPO Path Now Clear

A jury ruled unanimously this week that Elon Musk's lawsuit against OpenAI is barred by statutes of limitations. He sued too late. The case is over. This matters because it removes the last major legal obstacle to OpenAI's long-anticipated IPO. The company can now move forward without a billionaire founder's lawsuit hanging over investor confidence. What Musk Claimed Musk co-founded OpenAI in 2015 as a nonprofit research lab with a mission to develop artificial general intelligence for the benefit of humanity. He left the board in 2018, citing conflicts with Tesla's own AI work. His lawsuit argued that OpenAI had abandoned its nonprofit mission by forming a for-profit subsidiary and partnering with Microsoft. He claimed breach of contract and breach of fiduciary duty - that the organisation had strayed from its founding principles. The jury didn't rule on whether those claims were true. They ruled that Musk waited too long to bring them. The window to sue had closed. Why Timing...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Shaping Entanglement Before Transmission Beats All Post-Processing https://luma.marbl.codes/featured/shaping-entanglement-before-transmission-beats-all-post-processing https://luma.marbl.codes/featured/shaping-entanglement-before-transmission-beats-all-post-processing Tue, 19 May 2026 07:40:39 GMT Richard Bland Quantum Computing Shaping Entanglement Before Transmission Beats All Post-Processing Shaping Entanglement Before Transmission Beats All Post-Processing

Researchers have proven that shaping quantum entanglement before sending it through a noisy channel achieves purification levels that no amount of post-processing can match. This isn't an incremental improvement. It's a fundamental operational advantage. The finding establishes pre-channel engineering as a distinct resource in quantum communication - something that changes the design of quantum networks from the ground up. The Problem with Noisy Quantum Channels Quantum entanglement is fragile. When you transmit entangled particles through a physical channel - optical fibre, free-space links, anything real-world - noise degrades the entanglement. The particles arrive less correlated than when they left. The standard approach has been distillation: receive the noisy entangled pairs, then apply quantum operations to extract a smaller number of higher-quality pairs. You sacrifice quantity for quality, but you recover some of the lost entanglement. This new research shows that...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Vite 8.0 Replaces Its Bundler With Rust - 7x Faster Builds https://luma.marbl.codes/featured/vite-80-replaces-its-bundler-with-rust-7x-faster-builds https://luma.marbl.codes/featured/vite-80-replaces-its-bundler-with-rust-7x-faster-builds Tue, 19 May 2026 07:40:39 GMT Richard Bland Web Development Vite 8.0 Replaces Its Bundler With Rust - 7x Faster Builds Vite 8.0 Replaces Its Bundler With Rust - 7x Faster Builds

Vite 8.0 ships with a new bundler called Rolldown. It's written in Rust. And it's fast - the team reports build times dropping from 46 seconds to 6 seconds in production builds. That's a 7x improvement, and some projects are seeing even better numbers. This isn't just a speed bump. It's a full replacement of Vite's JavaScript-based bundler with a Rust engine designed to handle modern web projects at scale. What Changed Under the Hood Vite has always been fast for development - it uses native ES modules during dev, which means near-instant hot module replacement. But production builds still relied on Rollup, a JavaScript bundler that's flexible but slow for large codebases. Rolldown replaces Rollup entirely. It's a ground-up rewrite in Rust, designed to maintain Rollup's plugin compatibility while delivering the performance of native code. The 7x speedup comes from three things: Rust's raw execution speed, better parallelisation (Rust's concurrency model makes multi-threading...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Boston Dynamics Shows How Atlas Actually Learns https://luma.marbl.codes/featured/boston-dynamics-shows-how-atlas-actually-learns https://luma.marbl.codes/featured/boston-dynamics-shows-how-atlas-actually-learns Mon, 18 May 2026 13:44:07 GMT Richard Bland Robotics Boston Dynamics Shows How Atlas Actually Learns Boston Dynamics Shows How Atlas Actually Learns

Boston Dynamics just pulled back the curtain on something genuinely surprising: Atlas isn't following a script anymore. The robot is learning through reinforcement learning, building its own strategies for handling heavy, awkward objects in real time. The video shows Atlas lifting engine covers - the kind of industrial parts that are heavy, unbalanced, and genuinely difficult to manoeuvre. What's notable isn't the lifting itself. It's that Atlas is accounting for mass distribution and inertia on the fly, adjusting its grip and posture as the object's weight shifts. That's not pre-programmed choreography. That's whole-body control driven by learned behaviour. Reinforcement Learning in Physical Space Here's what's changed: Atlas trains in simulation, failing thousands of times, building an intuition for how objects behave when lifted, rotated, or passed between hands. The robot then transfers that learned behaviour to the real world. When it encounters a new object, it's not starting...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Companies Are Burning Through 2026 AI Budgets Already https://luma.marbl.codes/featured/companies-are-burning-through-2026-ai-budgets-already https://luma.marbl.codes/featured/companies-are-burning-through-2026-ai-budgets-already Mon, 18 May 2026 13:44:07 GMT Richard Bland Voices Companies Are Burning Through 2026 AI Budgets Already Companies Are Burning Through 2026 AI Budgets Already

Uber and ServiceNow exhausted their 2026 token budgets in four months. Not exceeded them. Exhausted them. The tokens they'd allocated for next year are already gone, and we're barely into 2025. That's not an outlier. According to Azeem Azhar's latest data, 71% of companies exceeded their AI budgets in 2025. Enterprise monthly AI spending hit $85,000 on average - a 36% increase from the previous year. And for half of finance leaders, cost management is now their primary AI concern. Not capability. Not integration. Cost. The Token Economics Problem Here's what's happening: companies adopt AI tools expecting incremental costs. A bit of GPT here, some Claude there. The budget feels manageable. Then usage explodes. Every team wants access. Every workflow gets an AI layer. Every customer interaction becomes an API call. The per-token cost is tiny, but tokens add up faster than anyone anticipated. The term "tokenmaxxing" captures it perfectly - organisations are optimising for maximum...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Why Your AI Agent Keeps Lying to Itself https://luma.marbl.codes/featured/why-your-ai-agent-keeps-lying-to-itself https://luma.marbl.codes/featured/why-your-ai-agent-keeps-lying-to-itself Mon, 18 May 2026 13:44:07 GMT Richard Bland Builders Why Your AI Agent Keeps Lying to Itself Why Your AI Agent Keeps Lying to Itself

Anthropic engineers just explained why most AI agents fail in production. The problem isn't the model. It's that agents are evaluating their own work, lying about what they've done, and compressing context until they forget what they were supposed to be doing in the first place. Ash Prabaker and Andrew Wilson spent an hour walking through what actually works when building agents that need to run for hours without losing the plot. The advice is specific, technical, and completely at odds with how most teams are building right now. Self-Evaluation Is a Trap Here's the mistake everyone makes: you build an agent, give it a task, and then ask it to check its own work. Sounds reasonable. In practice, it's useless. Agents are optimised to sound confident, not to be accurate. If you ask an agent "Did you complete this task correctly?", it will almost always say yes. Not because it's lying - because it genuinely believes it did. Anthropic's solution: adversarial evaluators. A separate...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Router Pattern: Why Multi-Agent Systems Need Conductors, Not Committees https://luma.marbl.codes/featured/the-router-pattern-why-multi-agent-systems-need-conductors-not-committees https://luma.marbl.codes/featured/the-router-pattern-why-multi-agent-systems-need-conductors-not-committees Mon, 18 May 2026 07:58:08 GMT Richard Bland AI The Router Pattern: Why Multi-Agent Systems Need Conductors, Not Committees The Router Pattern: Why Multi-Agent Systems Need Conductors, Not Committees

A developer built a seven-agent sales pipeline and discovered something nobody talks about: the architecture matters more than the agents. The system routes leads through qualification, research, proposal generation, objection handling, follow-up scheduling, sentiment analysis, and final summarisation. Seven specialised agents, each handling one slice of the pipeline. The interesting bit isn't what each agent does - it's how they talk to each other. Most people assume multi-agent systems work like a conversation: agents passing messages back and forth, negotiating, deciding together. That's the dream. The reality is messier. Agent-to-agent communication creates exponential complexity. Seven agents means 42 potential communication paths. Add one more agent and you're at 56 paths. The coordination overhead kills you before you ship anything useful. Star Topology: One Router, Many Leaves The solution that actually works in production: star topology. One central router. All agents are...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Quantum's Control-Plane Split: Why IBM Closing Pulse Access Matters https://luma.marbl.codes/featured/quantums-control-plane-split-why-ibm-closing-pulse-access-matters https://luma.marbl.codes/featured/quantums-control-plane-split-why-ibm-closing-pulse-access-matters Mon, 18 May 2026 07:58:08 GMT Richard Bland Quantum Computing Quantum's Control-Plane Split: Why IBM Closing Pulse Access Matters Quantum's Control-Plane Split: Why IBM Closing Pulse Access Matters

IBM shut down pulse-level control on its quantum systems in February 2025. Most people missed it. The ones who didn't are now rethinking where to run their experiments. Pulse-level control means writing directly to the quantum hardware - shaping the microwave pulses that flip qubits, tuning gate timing, optimising error correction at the physics layer. It's the difference between using a compiler's output and writing assembly by hand. For researchers trying to squeeze performance out of noisy hardware, that access is the whole point. IBM's decision matters because they're the biggest public quantum cloud provider. Their systems are how most academic researchers and small labs access real quantum hardware. A new survey of 13 quantum vendors documents the split: IBM, Google, and other top-tier vendors are closing control-plane access. Neutral-atom providers like QuEra and mid-tier vendors are opening it. The field is bifurcating. Why Vendors Close Access From IBM's perspective,...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] A 500KB Backend That Does Everything: Raw C++20 in Production https://luma.marbl.codes/featured/a-500kb-backend-that-does-everything-raw-c20-in-production https://luma.marbl.codes/featured/a-500kb-backend-that-does-everything-raw-c20-in-production Mon, 18 May 2026 07:58:08 GMT Richard Bland Web Development A 500KB Backend That Does Everything: Raw C++20 in Production A 500KB Backend That Does Everything: Raw C++20 in Production

Someone built a production web backend in raw C++20. No frameworks. No libraries. Just sockets, HTTP parsing, and 472 lines of code. The binary is under 500KB. Memory usage sits below 10MB. Cold start is under 10 milliseconds. It runs on a $6 VPS and handles HTTPS traffic without breaking. This shouldn't work. Modern backend development assumes you need frameworks, ORMs, middleware stacks, and dependency trees that pull in half of npm. This project proves you can ship a working backend with none of it. And the tradeoffs are more interesting than you'd expect. What Raw C++20 Means in Practice Raw means no framework. No Express, no Flask, no Rails. You open a socket yourself. You parse HTTP headers yourself. You handle keep-alive yourself. You implement routing as a switch statement. Every line of code is yours to debug and yours to understand. The HTTP parser is 80 lines. It reads the request line, splits headers on colons, extracts the method and path, and builds a struct. No...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The $15,000 Humanoid That Runs on a Raspberry Pi https://luma.marbl.codes/featured/the-15000-humanoid-that-runs-on-a-raspberry-pi https://luma.marbl.codes/featured/the-15000-humanoid-that-runs-on-a-raspberry-pi Sun, 17 May 2026 11:05:35 GMT Richard Bland Robotics The $15,000 Humanoid That Runs on a Raspberry Pi The $15,000 Humanoid That Runs on a Raspberry Pi

A humanoid robot showed up on GitHub last week with a complete Bill of Materials, assembly instructions, and $15,000 price tag. No venture capital. No press release promising to revolutionise manufacturing. Just an open-source build called Asimov that anyone with a workshop and patience can actually make.This matters because robotics has lived in two worlds for too long. Corporate labs build million-dollar machines that demo beautifully but never ship. Hobbyists build clever rigs that work once. The middle ground - practical, affordable, reproducible - barely existed. Asimov sits in that gap.What Makes This DifferentTwenty-five degrees of freedom. Compute running on a Raspberry Pi. Every component listed on GitHub with part numbers and supplier links. The designer didn't hide the hard bits or gloss over what doesn't work yet. The repository includes failure logs. That's rare enough to be worth noting.For context, Boston Dynamics robots cost more than a house and require support...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Gary Marcus: Why the Smart Money Might Be Wrong About AI https://luma.marbl.codes/featured/gary-marcus-why-the-smart-money-might-be-wrong-about-ai https://luma.marbl.codes/featured/gary-marcus-why-the-smart-money-might-be-wrong-about-ai Sun, 17 May 2026 11:05:35 GMT Richard Bland Voices Gary Marcus: Why the Smart Money Might Be Wrong About AI Gary Marcus: Why the Smart Money Might Be Wrong About AI

Gary Marcus sat for two interviews last week - one at Web Summit, one at Bug Bash 2026 - and made a case that's uncomfortable for anyone betting big on current AI scaling laws. His argument isn't that large language models don't work. It's that the industry is pouring billions into a strategy that hits physical and mathematical limits we're pretending don't exist.The crux: hyperscaling assumes more compute and more data produce proportionally better models forever. Marcus argues that assumption is breaking. Models are already trained on most of the available text on the internet. Synthetic data generation creates feedback loops that degrade quality. And throwing more GPUs at the problem costs exponentially more for diminishing returns.The Hyperscaling BetEvery major AI lab is building bigger models. OpenAI, Google, Anthropic, Meta - the playbook is identical. Raise capital, buy more compute, train larger models, charge for API access. The bet is that scale unlocks emergent...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Intercom Doubled Engineering Output by Onboarding Claude Like a Junior Dev https://luma.marbl.codes/featured/intercom-doubled-engineering-output-by-onboarding-claude-like-a-junior-dev https://luma.marbl.codes/featured/intercom-doubled-engineering-output-by-onboarding-claude-like-a-junior-dev Sun, 17 May 2026 11:05:35 GMT Richard Bland Builders Intercom Doubled Engineering Output by Onboarding Claude Like a Junior Dev Intercom Doubled Engineering Output by Onboarding Claude Like a Junior Dev

Brian Scanlan runs engineering at Intercom. Last year, his team doubled their throughput without doubling headcount. The method: treat Claude Code like a new hire. Onboard it properly. Write skills for it. Connect it to production systems. Give it context and let it ship code.The result? 17.6% of pull requests now auto-approved with full compliance sign-off. Developers save hours per week on grunt work. The team ships features faster without sacrificing quality. And the approach isn't magic - it's just disciplined application of tooling that already exists.The Onboarding AnalogyMost teams use AI code tools like fancy autocomplete. They paste a problem into ChatGPT, get a snippet, modify it, move on. That works for isolated tasks but doesn't scale. Scanlan's insight was simpler: if you wouldn't expect a new developer to be productive without onboarding, why expect it from an AI?Intercom built an onboarding process for Claude. They documented the monolith architecture - how services...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] ArXiv Draws the Line: AI-Generated Papers Get Year-Long Bans https://luma.marbl.codes/featured/arxiv-draws-the-line-ai-generated-papers-get-year-long-bans https://luma.marbl.codes/featured/arxiv-draws-the-line-ai-generated-papers-get-year-long-bans Sun, 17 May 2026 06:43:00 GMT Richard Bland AI ArXiv Draws the Line: AI-Generated Papers Get Year-Long Bans ArXiv Draws the Line: AI-Generated Papers Get Year-Long Bans

ArXiv, the open-access repository that hosts millions of research papers, just made a clear statement about where AI assistance ends and intellectual fraud begins. Authors who submit papers generated entirely by AI - without meaningful human contribution - will face a year-long ban from the platform. This isn't a blanket rejection of AI tools. ArXiv explicitly acknowledges that researchers use AI for grammar checking, translation, summarising literature, and generating code. That's fine. What's not fine is treating a language model as a co-author with full intellectual agency - letting it write the paper while you add your name to the byline. The Line Nobody Wanted to Draw The new policy targets a specific behaviour: submitting work where the AI did the thinking and you did the submitting. ArXiv's moderators have been watching this pattern emerge - papers that read fluently but lack original insight, coherent methodology, or any evidence of domain expertise. The kind of output you...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Ubuntu Bets on Local AI While Others Chase the Cloud https://luma.marbl.codes/featured/ubuntu-bets-on-local-ai-while-others-chase-the-cloud https://luma.marbl.codes/featured/ubuntu-bets-on-local-ai-while-others-chase-the-cloud Sun, 17 May 2026 06:43:00 GMT Richard Bland Web Development Ubuntu Bets on Local AI While Others Chase the Cloud Ubuntu Bets on Local AI While Others Chase the Cloud

Ubuntu just outlined an AI strategy that runs in the opposite direction to most of the industry. While Microsoft, Google, and Apple are racing to wire AI into their operating systems via cloud connections, Ubuntu is prioritising on-device intelligence - local models, modular design, and user control over always-connected dependency. This isn't just a technical preference. It's a philosophical bet about what users actually want from AI in their operating system. And it matters because Ubuntu's approach solves problems the cloud-first crowd keeps ignoring. What Ubuntu Is Actually Building The core of Ubuntu's strategy is on-device inference - running AI models locally on the user's hardware rather than routing everything through remote servers. That means privacy by default. Your data never leaves your machine. No usage monitoring, no server logs, no terms of service changes that suddenly expose what you've been working on. The architecture is modular. Users can choose which AI...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Cobot Surge Nobody's Talking About https://luma.marbl.codes/featured/the-cobot-surge-nobodys-talking-about https://luma.marbl.codes/featured/the-cobot-surge-nobodys-talking-about Sat, 16 May 2026 11:02:10 GMT Richard Bland Robotics The Cobot Surge Nobody's Talking About The Cobot Surge Nobody's Talking About

North American robot orders held steady in Q1 2026 - 9,055 units worth $543 million, essentially flat year-over-year. But the numbers hide a shift that's rewriting the robotics market. Collaborative robots - cobots - surged 56% in unit orders and 78% in revenue. While traditional industrial robots stagnated, the machines designed to work alongside humans are quietly taking over factory floors in sectors that barely touched automation five years ago. Who's Actually Buying Robots Now The automotive industry used to dominate robot purchases. Not anymore. Life sciences companies, semiconductor fabs, and food manufacturers are driving growth. These aren't the massive production lines Detroit built its reputation on - they're smaller facilities, tighter margins, more product variation. The kind of environment where a £200,000 industrial robot makes no sense, but a £30,000 cobot that can be reprogrammed in an afternoon does. The data from the Robotics Industries Association shows...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Cerebras Just Became a $60 Billion Bet on Inference https://luma.marbl.codes/featured/cerebras-just-became-a-60-billion-bet-on-inference https://luma.marbl.codes/featured/cerebras-just-became-a-60-billion-bet-on-inference Sat, 16 May 2026 11:02:10 GMT Richard Bland Voices Cerebras Just Became a $60 Billion Bet on Inference Cerebras Just Became a $60 Billion Bet on Inference

Cerebras went public at a $60 billion valuation. For a company that's spent years building chips most people have never heard of, that's a statement. But the valuation isn't the story. It's what Cerebras is actually doing with those chips - and what their CFO accidentally confirmed in the process. They're serving OpenAI's trillion-parameter models. The ones that haven't been announced yet. Models 5.4 and 5.5, running on Cerebras infrastructure, handling inference at scale. The Training Era is Over For years, the AI hardware race was about training. Who could build the biggest cluster, train the largest model, hit the lowest loss curve. Nvidia won that game so decisively that it became boring to watch. The interesting question shifted: once you've trained a frontier model, how do you serve it to millions of users without bankrupting yourself on compute costs? That's the inference problem. And it's where Cerebras has been quietly positioning itself while everyone else fought over...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Two Engineers With AI Agents Shipped Five Times a Day https://luma.marbl.codes/featured/two-engineers-with-ai-agents-shipped-five-times-a-day https://luma.marbl.codes/featured/two-engineers-with-ai-agents-shipped-five-times-a-day Sat, 16 May 2026 11:02:10 GMT Richard Bland Builders Two Engineers With AI Agents Shipped Five Times a Day Two Engineers With AI Agents Shipped Five Times a Day

PFF had ten engineers shipping code once every five days. Then they tried something different: two engineers, augmented with AI agents, working on the same codebase. The new team shipped five times per day. That's not a marginal improvement. That's a different development model entirely. And it meant dismantling most of what we think of as "engineering process" - standups, sprint planning, code review rituals - because none of it made sense anymore. What Actually Changed The shift wasn't just "developers using AI tools". It was developers building with AI agents that handled the grunt work of software development - the kind of work that fills up a sprint but doesn't require human judgement. Tickets updated themselves. When code merged, agents checked style and convention automatically. QA agents ran tests on every merge, flagged issues, suggested fixes. The humans focused on architecture decisions, product direction, and the kind of engineering problems that still require a person...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] I Built a WhatsApp Bot to Talk to My Mom. Then I Started Listening. https://luma.marbl.codes/featured/i-built-a-whatsapp-bot-to-talk-to-my-mom-then-i-started-listening https://luma.marbl.codes/featured/i-built-a-whatsapp-bot-to-talk-to-my-mom-then-i-started-listening Sat, 16 May 2026 06:19:47 GMT Richard Bland AI I Built a WhatsApp Bot to Talk to My Mom. Then I Started Listening. I Built a WhatsApp Bot to Talk to My Mom. Then I Started Listening.

An engineer built an AI assistant to solve a problem most of us pretend doesn't exist: staying in touch with family when you're drowning in work. The bot summarises his day, translates it into his mother's language, and sends her voice notes. It worked. Then it forced him to ask a harder question.The system is straightforward. At the end of each day, the bot processes his calendar, emails, and activity logs. It generates a summary of what he did, translates it into his mother's native language, converts the text to speech, and sends it via WhatsApp. His mother gets a daily update in her own voice - or at least, an AI approximation of it.The engineer, Asef, shared the project on Dev.to. The technical details are clean: OpenAI's API for summarisation, Google Translate for language conversion, a text-to-speech engine for voice generation, and the WhatsApp Business API for delivery. The entire pipeline runs on a cron job. Build time: a weekend. Maintenance: minimal.What the Bot Actually...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Hybrid Light-Matter Particles That Actually Compute Without Melting https://luma.marbl.codes/featured/hybrid-light-matter-particles-that-actually-compute-without-melting https://luma.marbl.codes/featured/hybrid-light-matter-particles-that-actually-compute-without-melting Sat, 16 May 2026 06:19:47 GMT Richard Bland Quantum Computing Hybrid Light-Matter Particles That Actually Compute Without Melting Hybrid Light-Matter Particles That Actually Compute Without Melting

Researchers at the University of Pennsylvania created particles that are neither light nor matter, but both at once. These hybrid particles interact strongly enough to perform computation - and they sidestep the heat problem that's been limiting chip design for a decade.The work, published in Nature, combines photons with excitons - electron-hole pairs bound together in semiconductors. The resulting particles, called polaritons, behave like matter when they need to interact and like light when they need to move. That dual nature is what makes them useful for computing.Why Electrons Are Hitting a WallModern chips run on electrons. Electrons are good at computation because they interact with each other - flip one bit, and it affects the next. But that interaction comes with a cost: resistance. Every time electrons bump into atoms or each other, they generate heat. At the densities we're building chips now, heat dissipation is the bottleneck. You can't pack transistors closer without...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] GitHub Built an Accessibility Agent. Then Learned When Not to Use It. https://luma.marbl.codes/featured/github-built-an-accessibility-agent-then-learned-when-not-to-use-it https://luma.marbl.codes/featured/github-built-an-accessibility-agent-then-learned-when-not-to-use-it Sat, 16 May 2026 06:19:47 GMT Richard Bland Web Development GitHub Built an Accessibility Agent. Then Learned When Not to Use It. GitHub Built an Accessibility Agent. Then Learned When Not to Use It.

GitHub deployed an AI agent to review pull requests for accessibility issues. It processed 3,535 PRs with a 68% resolution rate. The numbers sound good. The lessons are better. The team discovered that knowing when NOT to deploy an agent is more important than knowing how to build one.The project, detailed on the GitHub Blog, started with a straightforward goal: catch accessibility problems before they ship. Screen reader compatibility, keyboard navigation, colour contrast, ARIA labels - the unglamorous work that gets skipped when deadlines compress. An AI agent seemed like a perfect fit. Review every PR, flag issues, suggest fixes. Simple.What they learned: agents fail predictably, and managing those failures is most of the work.Complexity Thresholds Are RealThe first lesson: not every task is agent-appropriate. The GitHub team built a complexity classifier that evaluates each PR before the agent touches it. Simple issues - missing alt text, incorrect ARIA roles - go to the agent....

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] China's Manned Mecha Robot Smashes Through Walls - And Into Production https://luma.marbl.codes/featured/chinas-manned-mecha-robot-smashes-through-walls-and-into-production https://luma.marbl.codes/featured/chinas-manned-mecha-robot-smashes-through-walls-and-into-production Fri, 15 May 2026 11:55:52 GMT Richard Bland Robotics China's Manned Mecha Robot Smashes Through Walls - And Into Production China's Manned Mecha Robot Smashes Through Walls - And Into Production

Unitree unveiled the GD01 this week. It's a manned mecha robot - the kind you climb inside and pilot. It walks on two legs. Switches to four-leg mode for speed. And in the demo video, it punches straight through a concrete wall. This isn't a research prototype. It's production-ready. Unitree is shipping these. The GD01 stands roughly 3 metres tall when upright. The pilot sits in a cockpit with real-time control over movement - walking, running, obstacle navigation. The machine can transition between bipedal and quadrupedal modes mid-operation, depending on terrain or task requirements. When it needs power over precision, it drops to four legs and accelerates. The wall-smashing moment in the reveal video isn't theatrics for the sake of it. It's a demonstration of structural force delivery - the machine can generate enough kinetic energy to breach reinforced surfaces. That matters for disaster response, industrial demolition, or military applications where breaking through obstacles...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] 100 Million Doctor Visits, Summarised by AI - Abridge's Clinical Intelligence Stack https://luma.marbl.codes/featured/100-million-doctor-visits-summarised-by-ai-abridges-clinical-intelligence-stack https://luma.marbl.codes/featured/100-million-doctor-visits-summarised-by-ai-abridges-clinical-intelligence-stack Fri, 15 May 2026 11:55:52 GMT Richard Bland Voices 100 Million Doctor Visits, Summarised by AI - Abridge's Clinical Intelligence Stack 100 Million Doctor Visits, Summarised by AI - Abridge's Clinical Intelligence Stack

Abridge has processed over 100 million clinical conversations. Every one of those is a doctor-patient interaction - diagnosis discussions, treatment plans, symptoms described in plain language - turned into structured medical documentation. Co-founders Janie Lee and Chai Asawa walked through how they built clinical intelligence infrastructure that works at healthcare scale. This isn't a prototype. It's live in hospital systems, generating notes that doctors actually use, automating prior authorisations that used to take days, and doing it under privacy regulations that would kill most AI products before they launched. The Latent Space interview covers the hard parts: evaluation pipelines for medical accuracy, integration with Electronic Health Record systems, and why healthcare's high-stakes environment forces you to solve AI problems most companies can ignore. The Evals Problem in Healthcare Standard LLM evaluation doesn't work for clinical notes. You can't just measure BLEU...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Building Evals That Prove Your Agent Actually Works - The Workshop Version https://luma.marbl.codes/featured/building-evals-that-prove-your-agent-actually-works-the-workshop-version https://luma.marbl.codes/featured/building-evals-that-prove-your-agent-actually-works-the-workshop-version Fri, 15 May 2026 11:55:52 GMT Richard Bland Builders Building Evals That Prove Your Agent Actually Works - The Workshop Version Building Evals That Prove Your Agent Actually Works - The Workshop Version

Agents fail in ways that are hard to measure. They call the wrong API. They hallucinate function arguments. They get stuck in loops. And when you change the prompt to fix one failure, you break something else - except you don't know it until users report weird behaviour three days later. Laurie Voss from Arize walked through how to build evaluation pipelines that solve this. The AI Engineer workshop covers tracing, failure categorisation, code evals, LLM judges, and experiments that prove - actually prove - when a prompt change improves performance. This is practical, hands-on work. Not theory about what evals should be. Instructions for building them from scratch. The Vibes Problem Most teams test agents by running them a few times, checking if the output "feels right", and shipping. This works until it doesn't. You make a prompt tweak. The agent seems better. Three weeks later, you discover it stopped handling edge cases correctly. But you've shipped five more changes since...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] The Cloud Operations AI Can't Run Alone - And Why Read-Only Wasn't the Answer https://luma.marbl.codes/featured/the-cloud-operations-ai-cant-run-alone-and-why-read-only-wasnt-the-answer https://luma.marbl.codes/featured/the-cloud-operations-ai-cant-run-alone-and-why-read-only-wasnt-the-answer Fri, 15 May 2026 06:54:47 GMT Richard Bland AI The Cloud Operations AI Can't Run Alone - And Why Read-Only Wasn't the Answer The Cloud Operations AI Can't Run Alone - And Why Read-Only Wasn't the Answer

A team gave AI agents write access to their cloud infrastructure. Ninety days later, they revoked it. Not because the AI made obvious mistakes - because it couldn't see what humans see. The problem wasn't technical capability. The agents handled local operations beautifully: spinning up instances, adjusting configurations, managing logs. What broke them were the invisible dependencies - cross-account relationships, disaster recovery requirements, concurrent deployments happening elsewhere. Context that existed in Slack threads, tribal knowledge, and the collective memory of the ops team. What Went Wrong Three operations caused the rethink. First: an agent resized a database instance during what looked like low usage. It had local metrics showing idle capacity. What it couldn't see was a scheduled data migration happening in another account that needed that headroom. The migration failed halfway through. Second: an agent cleaned up what appeared to be orphaned snapshots. They were...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Quantum Neural Networks Just Modelled Fluid Flow With 90% Fewer Parameters https://luma.marbl.codes/featured/quantum-neural-networks-just-modelled-fluid-flow-with-90-fewer-parameters https://luma.marbl.codes/featured/quantum-neural-networks-just-modelled-fluid-flow-with-90-fewer-parameters Fri, 15 May 2026 06:54:47 GMT Richard Bland Quantum Computing Quantum Neural Networks Just Modelled Fluid Flow With 90% Fewer Parameters Quantum Neural Networks Just Modelled Fluid Flow With 90% Fewer Parameters

Fluid dynamics problems are computationally expensive. Simulating how air moves around an aircraft wing or how water flows through a pipe requires massive neural networks - thousands of parameters, hours of training time. A new approach using quantum computing just achieved competitive accuracy with a fraction of the resources. The method combines physics-informed neural networks with quantum circuits. Instead of training a classical neural network to learn fluid behaviour from scratch, the quantum version encodes physical laws directly into trainable quantum states. The results show stable training and accuracy that matches classical networks, but with significantly fewer trainable parameters. What Makes This Different Traditional physics-informed neural networks work by embedding known physics equations into the loss function. The network learns to satisfy both the data and the governing equations - conservation of mass, momentum, energy. It's clever, but still requires large...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] How a 4-Field Lookup Table Replaced 200 Cloud Cost Tags https://luma.marbl.codes/featured/how-a-4-field-lookup-table-replaced-200-cloud-cost-tags https://luma.marbl.codes/featured/how-a-4-field-lookup-table-replaced-200-cloud-cost-tags Fri, 15 May 2026 06:54:47 GMT Richard Bland Web Development How a 4-Field Lookup Table Replaced 200 Cloud Cost Tags How a 4-Field Lookup Table Replaced 200 Cloud Cost Tags

A company tried to track cloud costs using tags. They defined 200 tags: project names, cost centres, environments, owners, business units. Every resource was supposed to be tagged. Compliance never broke 40%. Chargeback accuracy sat at 70-80%. Finance spent 20 hours a month reconciling discrepancies. Six weeks later, they scrapped the tags. Chargeback accuracy hit 92-96%. The fix was simpler than the problem: a four-field lookup table maintained by FinOps, not enforced through tagging policy. Why Tagging Fails Cloud tagging sounds logical. Apply metadata to every resource, then aggregate costs by tag. In practice, it collapses under three pressures. First: tag entropy. Teams use different formats. One writes "prod", another writes "production", a third writes "prd". The taxonomy says "environment: production" but nobody reads the taxonomy. Tag values drift over time. A project name changes but tags don't get updated. Six months later, half your resources have stale metadata....

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Figure's Humanoid Robots Worked 8-Hour Warehouse Shifts. Nobody Supervised. https://luma.marbl.codes/featured/figures-humanoid-robots-worked-8-hour-warehouse-shifts-nobody-supervised https://luma.marbl.codes/featured/figures-humanoid-robots-worked-8-hour-warehouse-shifts-nobody-supervised Thu, 14 May 2026 11:51:30 GMT Richard Bland Robotics Figure's Humanoid Robots Worked 8-Hour Warehouse Shifts. Nobody Supervised. Figure's Humanoid Robots Worked 8-Hour Warehouse Shifts. Nobody Supervised.

Brett Adcock streamed something unusual this week: humanoid robots sorting packages for eight consecutive hours. No human stepping in. No pauses for recalibration. Just robots, battery swaps, and multi-robot coordination in a live production environment.This wasn't a benchmark. It was a shift.What ChangedFigure's robots ran on-device inference - meaning the decision-making happened locally, not in the cloud. When a robot's battery ran low, another robot took over. The coordination was autonomous. The tasks were repetitive but real: picking, sorting, moving packages through a warehouse flow.Previous humanoid demos showed impressive dexterity in controlled settings. This showed something harder: long-horizon autonomy. Can a robot work a full shift without human intervention? According to Adcock's livestream, yes.Why Eight Hours MattersThe difference between a 90-second demo and an eight-hour shift is everything. Demos optimise for spectacle. Shifts optimise for reliability. A robot that...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Chinese AI Labs Extract 4-7x More Intelligence Per Chip Than US https://luma.marbl.codes/featured/chinese-ai-labs-extract-4-7x-more-intelligence-per-chip-than-us https://luma.marbl.codes/featured/chinese-ai-labs-extract-4-7x-more-intelligence-per-chip-than-us Thu, 14 May 2026 11:51:30 GMT Richard Bland Voices Chinese AI Labs Extract 4-7x More Intelligence Per Chip Than US Chinese AI Labs Extract 4-7x More Intelligence Per Chip Than US

Azeem Azhar spent time inside Chinese AI labs and came back with a data point nobody expected: Chinese researchers are getting 4-7 times more intelligence per compute unit than their US counterparts. Not because they have better hardware - they're working with chips that are 2-3 years behind. Because export controls forced them to be ruthlessly efficient.The result: Chinese open models now sit 6-8 months behind the US frontier while costing 11 times less to run.The Hardware Disadvantage Became an Efficiency AdvantageUS export restrictions cut Chinese labs off from the latest GPUs. The assumption was this would slow them down. Instead, it forced optimisation. When you can't throw more compute at a problem, you learn to use less compute better.According to Azhar's research, Chinese labs developed techniques to extract more capability from older hardware. Model compression, quantisation, distillation - methods that were academic curiosities in the US became production necessities in...

Read the full article →

Written by Richard Bland for Luma Digest

]]> [Featured] Your Portfolio Needs Real Business Logic, Not Another Todo App https://luma.marbl.codes/featured/your-portfolio-needs-real-business-logic-not-another-todo-app https://luma.marbl.codes/featured/your-portfolio-needs-real-business-logic-not-another-todo-app Thu, 14 May 2026 11:51:30 GMT Richard Bland Builders Your Portfolio Needs Real Business Logic, Not Another Todo App Your Portfolio Needs Real Business Logic, Not Another Todo App

A developer with six years' freelance experience posted something blunt this week: nobody hires based on todo apps. Weather clones don't get you work. Calculator UIs don't demonstrate value. If your portfolio is full of tutorial projects, you're signalling that you haven't solved a real problem yet.The portfolios that get hired show business logic. Exam platforms with anti-cheating systems. Learning management tools with live session handling. Ecommerce with client-friendly CMS integration. The difference isn't technical complexity - it's applied thinking.What Hiring Managers Actually Look ForAccording to the post, the gap between junior and mid-level developers isn't years of experience - it's evidence of solving real constraints. Can you handle user authentication properly? Can you structure a database for a multi-tenant system? Can you build an admin panel that a non-technical client can use?Tutorial projects don't answer these questions. They show you can follow instructions. Real...

Read the full article →

Written by Richard Bland for Luma Digest

]]>