https://alaintural.com/writing/ Engineering notes from Pixedi: what shipped, what broke, how it was measured and what it cost. en Tue, 15 Sep 2026 09:00:00 +0000 https://alaintural.com/writing/approval-is-not-publication/ https://alaintural.com/writing/approval-is-not-publication/ Tue, 15 Sep 2026 09:00:00 +0000 Shipping with agents Everyone builds the data, the agent and the approval screen. The handoff from approval to action is plumbing, and it is where throughput lives. https://alaintural.com/writing/neural-rendering-into-synthesis/ https://alaintural.com/writing/neural-rendering-into-synthesis/ Mon, 14 Sep 2026 09:00:00 +0000 Rendering and 3D Neural stages used to reconstruct a frame from sparse evidence. Generative neural rendering produces appearance that was never sampled at all. https://alaintural.com/writing/the-renderer-as-a-data-engine/ https://alaintural.com/writing/the-renderer-as-a-data-engine/ Mon, 14 Sep 2026 09:00:00 +0000 Rendering and 3D A renderer turned scenes into images. Synthetic data for physical AI makes the same pipeline a source of supervised training material. https://alaintural.com/writing/local-ai-as-a-compute-fabric/ https://alaintural.com/writing/local-ai-as-a-compute-fabric/ Sun, 13 Sep 2026 09:00:00 +0000 Standards and interfaces Local inference assumed a single machine. Spreading work across the devices on a network turns local AI into a fabric rather than an install. https://alaintural.com/writing/agent-infrastructure-as-a-category/ https://alaintural.com/writing/agent-infrastructure-as-a-category/ Sat, 12 Sep 2026 09:00:00 +0000 Agents and runtimes Agent infrastructure was something each team wrote around the model. Managed runtimes, memory and sandboxes are turning it into a product category. https://alaintural.com/writing/choosing-a-local-model-with-a-stopwatch/ https://alaintural.com/writing/choosing-a-local-model-with-a-stopwatch/ Sat, 12 Sep 2026 09:00:00 +0000 Measuring instead of guessing A local model comparison settled by running the code: the official build beat a community repack of the same model by 2.4 times, and scored higher. https://alaintural.com/writing/real-time-lighting-lost-to-baked-light/ https://alaintural.com/writing/real-time-lighting-lost-to-baked-light/ Sat, 12 Sep 2026 09:00:00 +0000 3D for the web Why real time lighting lost to baked light in a 3D product simulation: 90 seconds per option down to 3, and the tone mapping order that costs 0.17. https://alaintural.com/writing/what-broke-in-an-agency-run-by-agents/ https://alaintural.com/writing/what-broke-in-an-agency-run-by-agents/ Sat, 12 Sep 2026 09:00:00 +0000 Shipping with agents Three failures in an AI-run delivery stack: 0 of 29 posts delivered, a renderer exiting 0 on crash, an agent reporting DONE after writing nothing. https://alaintural.com/writing/no-check-command-not-automated/ https://alaintural.com/writing/no-check-command-not-automated/ Fri, 11 Sep 2026 09:00:00 +0000 Agents and runtimes Eight stages of automated production, why engines are not interchangeable, and the honest boundary of an agentic setup today. https://alaintural.com/writing/a-gate-the-model-writes/ https://alaintural.com/writing/a-gate-the-model-writes/ Thu, 10 Sep 2026 09:00:00 +0000 Shipping with agents A gate that cannot see the data, a gate that rewards producing the shape of the work, and a local counter reporting health for a dead engine. https://alaintural.com/writing/the-scoring-model-was-wrong/ https://alaintural.com/writing/the-scoring-model-was-wrong/ Tue, 08 Sep 2026 09:00:00 +0000 Measuring instead of guessing Ranking 2,163 companies surfaced two failures: throttled measurement recorded as absence, and a scoring formula that rewarded a single strong axis. https://alaintural.com/writing/building-software-got-easy/ https://alaintural.com/writing/building-software-got-easy/ Sat, 05 Sep 2026 09:00:00 +0000 Shipping with agents What building Aristo showed about the distance between shipping a working product and making it matter to a business, with the measurements behind it. https://alaintural.com/writing/capability-as-operational-risk/ https://alaintural.com/writing/capability-as-operational-risk/ Fri, 04 Sep 2026 09:00:00 +0000 Models and evaluation Model safety focused on output. A model that acts turns capability into an operational surface with access, staging and revocation to manage. https://alaintural.com/writing/the-scene-graph-as-an-api/ https://alaintural.com/writing/the-scene-graph-as-an-api/ Fri, 24 Jul 2026 09:00:00 +0000 Rendering and 3D Scene graphs were built for DCC tools. Agents reading prims, bindings and edit targets make the graph an interface rather than an internal format. https://alaintural.com/writing/from-clips-to-motion-priors/ https://alaintural.com/writing/from-clips-to-motion-priors/ Thu, 23 Jul 2026 09:00:00 +0000 Rendering and 3D Character animation was built from authored or captured clips. Learned motion priors change the question from which motion to which constraints. https://alaintural.com/writing/materials-as-learned-programs/ https://alaintural.com/writing/materials-as-learned-programs/ Wed, 22 Jul 2026 09:00:00 +0000 Rendering and 3D Materials are authored from texture maps and shader graphs. Learned appearance models turn them into small programs evaluated at render time. https://alaintural.com/writing/procedural-beyond-geometry/ https://alaintural.com/writing/procedural-beyond-geometry/ Thu, 16 Jul 2026 09:00:00 +0000 Rendering and 3D Procedural systems were about building geometry. Physics, sound and object data moving into the same node graph makes that description incomplete. https://alaintural.com/writing/the-unit-of-ai-work-is-the-task/ https://alaintural.com/writing/the-unit-of-ai-work-is-the-task/ Mon, 29 Jun 2026 09:00:00 +0000 Agents and runtimes The chat turn was the unit of AI work. Agents that run for hours make the task the unit that has to be tracked, resumed and evaluated. https://alaintural.com/writing/engines-as-operating-systems/ https://alaintural.com/writing/engines-as-operating-systems/ Wed, 24 Jun 2026 09:00:00 +0000 Rendering and 3D Engines were measured by lighting, world size and physics. The next question is which system operates a world once it already exists. https://alaintural.com/writing/the-model-as-a-replaceable-backend/ https://alaintural.com/writing/the-model-as-a-replaceable-backend/ Thu, 11 Jun 2026 09:00:00 +0000 Standards and interfaces Provider choice used to decide message format, tool schema and deployment. A common interface turns the model into a backend that can be swapped. https://alaintural.com/writing/the-harness-is-part-of-capability/ https://alaintural.com/writing/the-harness-is-part-of-capability/ Fri, 17 Apr 2026 09:00:00 +0000 Agents and runtimes Agent performance gets discussed as a property of the model. Context policy, tool design and recovery logic are part of what capability means. https://alaintural.com/writing/physics-engines-as-trainable-parts/ https://alaintural.com/writing/physics-engines-as-trainable-parts/ Thu, 19 Mar 2026 09:00:00 +0000 Rendering and 3D Forward simulation predicts the next state. Differentiable simulation turns the engine into a component that can be tuned against a measured outcome. https://alaintural.com/writing/the-agent-needs-an-environment/ https://alaintural.com/writing/the-agent-needs-an-environment/ Fri, 13 Mar 2026 09:00:00 +0000 Agents and runtimes Tool calling let models reach outside the prompt. Sustained work needs a filesystem, a workspace and state that survives between steps. https://alaintural.com/writing/coding-agents-as-computer-workers/ https://alaintural.com/writing/coding-agents-as-computer-workers/ Mon, 09 Feb 2026 09:00:00 +0000 Agents and runtimes Code was a good first environment for long-running agents because results are verifiable. The same properties are appearing in other kinds of work. https://alaintural.com/writing/open-standards-outlive-models/ https://alaintural.com/writing/open-standards-outlive-models/ Thu, 11 Dec 2025 09:00:00 +0000 Standards and interfaces Model generations turn over quickly. The interfaces between models, tools and data have a longer useful life than the models they were built against. https://alaintural.com/writing/colour-management-is-a-contract/ https://alaintural.com/writing/colour-management-is-a-contract/ Thu, 20 Nov 2025 09:00:00 +0000 Rendering and 3D Colour management looks like display settings. An explicit working space makes it a contract between every stage of a production pipeline. https://alaintural.com/writing/the-best-model-may-be-a-router/ https://alaintural.com/writing/the-best-model-may-be-a-router/ Mon, 11 Aug 2025 09:00:00 +0000 Models and evaluation Model choice used to be a developer decision. A real-time router makes the policy that decides how a task is solved the thing worth evaluating. https://alaintural.com/writing/world-models-are-not-game-engines/ https://alaintural.com/writing/world-models-are-not-game-engines/ Fri, 08 Aug 2025 09:00:00 +0000 Rendering and 3D A generative model can produce an interactive world in real time. Authoring, persistent state and inspectable failure are what production still needs. https://alaintural.com/writing/memory-as-a-system-capability/ https://alaintural.com/writing/memory-as-a-system-capability/ Mon, 26 May 2025 09:00:00 +0000 Models and evaluation Memory needs scope, provenance and expiry, not just capacity. It is becoming a declared system capability rather than a side effect of a long prompt. https://alaintural.com/writing/coding-agents-change-the-unit/ https://alaintural.com/writing/coding-agents-change-the-unit/ Mon, 19 May 2025 09:00:00 +0000 Agents and runtimes Autocomplete is judged at the line. When an agent takes a task end to end, the unit of software work changes and so does how it must be reviewed. https://alaintural.com/writing/agents-need-protocols/ https://alaintural.com/writing/agents-need-protocols/ Fri, 11 Apr 2025 09:00:00 +0000 Agents and runtimes Agent architecture is usually discussed from the inside out. Agent-to-agent communication across teams and platforms is a separate engineering problem. https://alaintural.com/writing/agents-need-a-runtime/ https://alaintural.com/writing/agents-need-a-runtime/ Thu, 13 Mar 2025 09:00:00 +0000 Agents and runtimes Once a task runs longer than a few turns, state, retries, handoff and inspection stop being prompt problems and become runtime problems. https://alaintural.com/writing/reasoning-is-not-the-only-path/ https://alaintural.com/writing/reasoning-is-not-the-only-path/ Fri, 28 Feb 2025 09:00:00 +0000 Models and evaluation Reasoning is a real direction and not the only one. Contextual understanding often decides whether a production request is carried out correctly. https://alaintural.com/writing/rendering-as-a-reconstruction-stack/ https://alaintural.com/writing/rendering-as-a-reconstruction-stack/ Thu, 09 Jan 2025 09:00:00 +0000 Rendering and 3D Rendering increasingly means reconstruction: sparse physical samples plus motion data turned into a final frame by learned stages, with new failure modes. https://alaintural.com/writing/agent-reliability-is-evaluation/ https://alaintural.com/writing/agent-reliability-is-evaluation/ Mon, 16 Dec 2024 09:00:00 +0000 Agents and runtimes Agent failures get diagnosed at the prompt layer. Knowing a task completed, and detecting failure before the agent builds on it, is an evaluation problem. https://alaintural.com/writing/mcp-over-another-model-release/ https://alaintural.com/writing/mcp-over-another-model-release/ Fri, 29 Nov 2024 09:00:00 +0000 Standards and interfaces Capability is easy to see and integration is not. MCP addresses the boundary that decides whether a capable model is useful inside a real company. https://alaintural.com/writing/computer-use-missing-layer/ https://alaintural.com/writing/computer-use-missing-layer/ Mon, 28 Oct 2024 09:00:00 +0000 Standards and interfaces Integration assumes an API exists. Older desktop tools expose only part of their function, which is why screen-level control became a serious layer. https://alaintural.com/writing/inference-time-compute-scaling/ https://alaintural.com/writing/inference-time-compute-scaling/ Thu, 19 Sep 2024 09:00:00 +0000 Models and evaluation Scaling used to mean larger models and more training compute. Time spent at inference is a separate axis, and it changes how results must be reported. https://alaintural.com/writing/the-final-pixel-and-the-renderer/ https://alaintural.com/writing/the-final-pixel-and-the-renderer/ Thu, 15 Aug 2024 09:00:00 +0000 Rendering and 3D The renderer has been the final authority on appearance. That boundary is moving, and structural correctness and perceptual fidelity are splitting apart. https://alaintural.com/writing/open-models-as-research-infrastructure/ https://alaintural.com/writing/open-models-as-research-infrastructure/ Mon, 29 Jul 2024 09:00:00 +0000 Models and evaluation Open model comparisons focus on inference cost and benchmark gaps. Access changes what kind of experiment a team can run in the first place. https://alaintural.com/writing/the-model-is-becoming-a-runtime/ https://alaintural.com/writing/the-model-is-becoming-a-runtime/ Mon, 24 Jun 2024 09:00:00 +0000 Agents and runtimes Tool calling and structured output turn inference into a loop. The model starts behaving less like a text generator and more like a control layer. https://alaintural.com/writing/multimodality-changes-the-architecture/ https://alaintural.com/writing/multimodality-changes-the-architecture/ Thu, 16 May 2024 09:00:00 +0000 Models and evaluation GPT-4o is described as a better interface. The more interesting change is a single model trained across text, vision and audio rather than a pipeline. https://alaintural.com/writing/benchmark-scores-are-not-capability/ https://alaintural.com/writing/benchmark-scores-are-not-capability/ Mon, 11 Mar 2024 09:00:00 +0000 Models and evaluation Benchmark numbers measure a narrow slice under one setup. Contamination, prompt format and evaluation method sit between the score and the work. https://alaintural.com/writing/long-context-is-not-memory/ https://alaintural.com/writing/long-context-is-not-memory/ Tue, 20 Feb 2024 09:00:00 +0000 Models and evaluation A million-token context window enlarges working memory. It does not replace the write, select and retrieve layer that makes memory a system.