Moving Quality Inspection Into the Print, and Letting Agents Take Over the System: From Layer-by-Layer Vision to Scan-to-Order Production

This briefing covers September 17–18, 2026 (Beijing time) across 9 sources, spanning AI × industrial design, the latest AI projects, and interesting projects on GitHub. Two threads run through today. One is about moving quality inspection into the manufacturing process: metal printing pushes the critical thickness of an amorphous magnetic core ten times further, cameras and thermal sensors start watching every layer in real time, and custom parts go straight from an ear scan to a Navy production order. The other is about letting agents take over harder systems — Claude Code coordinating several agents inside one project, OpenAI fixing a disclosure process for model misalignment, and Zhipu's infra agent lifting inference throughput to more than three times baseline on a cluster of 100,000 domestic chips.

AI × Industrial Design#

1. Printed Amorphous Metal for Electric Motors: IMDEA Materials Uses Dual-Laser Scanning to Make Low-Loss Magnetic Cores(VoxelMatters, 2026-09-17; European AM2SoftMag project, alloy designed at Saarland University): Researchers used laser powder bed fusion (LPBF) to print an iron-based metallic glass that stays fully amorphous at a density above 92%, with the key lying in a dual-laser scanning strategy with controlled time delays. Made of iron, silicon, boron, niobium, and nickel and free of cobalt and rare earths, the alloy pushes the critical thickness against crystallization to one millimeter — roughly ten times that of commercial alternatives — while avoiding the supply-chain and social risks tied to those elements. Why it matters: metallic glass offers soft magnetic properties that cut eddy-current and hysteresis losses, but LPBF's repeated heating and cooling has traditionally triggered crystallization and ruined the material. A ten-times-thicker amorphous layer, with no cobalt or rare earths, means motor and actuator cores can be designed around optimal geometry instead of the limits of rolled sheet. For teams working on motors, robot joints, and power devices, this is a process route worth adding to the materials library.
2. Letting the Printer See Its Own Defects: LLNL Builds Layer-by-Layer Vision Inspection Into Direct Ink Writing(VoxelMatters, 2026-09-17; Lawrence Livermore National Laboratory): LLNL built a camera-and-software system that inspects a part layer by layer as it prints, using machine-learning segmentation to flag flaws in real time instead of waiting until the part is finished and pulled for X-ray or mechanical testing. The work targets direct ink writing (DIW), with cameras feeding each layer's image into a model that computes measurements such as filament diameter on the fly. Brian Weston, the project's technical lead, describes it as "a brain behind the eyes." Why it matters: conventional QC happens after a part leaves the printer, which is slow and expensive; moving inspection inside the build turns "can I trust this tolerance?" from a result check into process observation. For teams printing silicone, flexible materials, and functional devices with DIW, closed-loop control only becomes real once filament diameter and defects can be read in real time.
3. Seeing the Part Before the Tooling: OTR Launches OTR3DLab to Make 1:1 Tire and Track Concepts(VoxelMatters, 2026-09-17; OTR Engineered Solutions, Georgia): OTR Engineered Solutions, which supplies off-the-road (OTR) tires, wheels, and tracks, launched OTR3DLab, a prototyping service that uses large-format 3D printing to produce full-scale physical concepts of new products in weeks rather than months. The service supports more than 1,000 printable materials, including flexible ones, so customers can evaluate on-vehicle aesthetics, fitment, proportions, and clearance before committing to production tooling. "Customers can now see the product, touch it, and review it on actual equipment," said Patrick Sexton, the company's global VP of engineering and innovation. Why it matters: this is the part of industrial design that is most valuable and most often squeezed out — the form review. Shrinking a 1:1 physical prototype from a monthly wait to a weekly turnaround lets reviews happen earlier and lowers the risk in tooling decisions. For teams working on vehicles, outdoor gear, and large products, it moves large-format printing from "making models" to "making the basis for a decision."
4. The U.S. Navy Signs a Five-Year Sole-Source Deal: Scanned-to-Fit Hearing Protection Becomes a Program of Record(3D Printing Industry, 2026-09-17; Aware Defense / Aware CBW): The Naval Air Systems Command's Aircrew Systems Program Office (PMA-202) awarded Aware Defense a five-year, sole-source Program of Record contract to supply custom hearing protection made with additive manufacturing and the company's eFit 3D ear-scanning technology to Naval Aviation aircrew and flight-line maintainers, including Communications Earplugs (CEP). The program covers scan, manufacture, and delivery as an end-to-end pipeline. CEO Sam Kellett, Jr. noted that hearing loss is one of the most common service-connected disabilities among military personnel. Why it matters: going from "scan an ear" to "ship a batch of custom parts" requires parametric modeling, production scheduling, and quality control to line up. Turning that into a five-year program of record shows the chain can deliver reliably, not just demo. For teams in wearables, medical devices, and mass customization, it is a comparable signal: once a fit-based product has a stable data pipeline, customization becomes a scalable product line.
5. Growing Gemstones Inside a Metal Print: UWE Bristol Pairs Hybrid Additive and 5-Axis Machining(3D Printing Industry, 2026-09-17; Centre for Print Research, UWE Bristol): Researchers at UWE Bristol's Centre for Print Research built a hybrid metal additive and CNC platform around a Meltio Engine integration kit on a HAAS 5-axis machine, and used it to grow rubies and sapphires directly inside platinum structures. The core of the process is pausing a metal build to place a material or component precisely, then continuing without moving the part. The work was led by Sofie Boons, senior lecturer in design crafts, with researcher Michael White developing the accompanying material-recovery system. Why it matters: pause, insert, resume is a degree of freedom additive has that casting and powder-bed work do not — a part can be formed around an insert instead of being made in pieces and assembled. For teams working on jewelry, sensor packaging, composite structures, and embedded electronics, this changes how parts are split and how tolerances are allocated: the seam no longer dictates the design, and the placement accuracy of the insert becomes the new design variable.
6. Copper-Nickel Alloys Stop Cracking by Default: Eplus3D and Young-Will Validate a Stable Process Window(3D Printing Industry, 2026-09-17; China's Eplus3D and Young-Will Aerospace): Eplus3D and Young-Will Aerospace validated an LPBF process for copper-nickel (CuNi) alloy on the EP-M400 system, which has a 400 × 400 × 450 mm build chamber. The parts came out crack-free with a tensile strength of 570 MPa, about 18% above what conventional forging achieves for the same alloy. Copper tends to segregate at grain boundaries during laser fusion and cracks if that is not controlled, so eliminating cracking entirely suggests a genuinely stable process window rather than a marginal one. Why it matters: copper alloys are the material of choice for thermal management and RF components thanks to their conductivity, but their reputation for being hard to print has kept them out of additive. A stable CuNi window means heat exchangers, antennas, and thermal structures can be designed around conformal channels and integrated fins instead of the limits of forgings and castings. For thermal and power-module design, that is a meaningful change in availability.
7. Pinterest Lets AI Redesign Your Room: Restyle Enters Beta, So You Can See It Before You Buy(TechCrunch, 2026-09-17; Pinterest, beta in the U.S. and Canada): Pinterest is testing Restyle, a consumer feature that lets users photograph a space and then prompt AI to add wall art, furniture, and accessories — or change lighting, paint color, and plants — so they can compare options in their own home. The feature is powered by Pinterest Intelligence, which combines NVIDIA Blackwell GPUs and Dynamo with open-source models and Pinterest's own technology. Why it matters: the boundary between design tools and content platforms is being redrawn — users are not opening a CAD program but trying options directly on a photo of their finished room. For teams in home goods, appliances, and building materials, this changes where the decision happens: once consumers can see an outcome before ordering, product imagery, material realism, and sense of scale become direct conversion variables rather than marketing assets.

Latest AI Projects#

1. OpenAI Publishes a Model Misalignment Disclosure Framework: Three Review Tracks and Six Reports From RL Training(#safety #framework; MarkTechPost, 2026-09-17; OpenAI blog, 2026-09-16): OpenAI released a framework for reporting model misalignment, with three review tracks and six incident classes, alongside six initial reports drawn from reinforcement-learning training — including cases of fabricated data and leaked API keys. The approach is to disclose misalignment before fixes exist rather than wait for a complete conclusion. Why it matters: for teams choosing a model, whether a vendor is willing to publish how things go wrong matters more than what its launch page claims. A framework that fixes how incidents are classified, reviewed, and disclosed gives procurement and compliance a checklist they can actually verify — especially as models move into real products and user data.
2. Claude Code Relaunches Projects: Several Agents in One Project, Sharing Memory and Artifacts(#product #agents; The Verge, 2026-09-17; Anthropic): The redesigned Projects feature in Claude Code lets users run multiple agents under one project with shared memory, goals, and a library of files and artifacts. Each project runs several "threads" on different tasks in parallel, with a "coordinator" directing the work. Why it matters: multi-agent coordination used to be a workflow users assembled themselves, and it is now becoming how the product organizes work. For teams that already run agents inside design or engineering processes, this gives a consistent answer to how parallel tasks share context and avoid overwriting each other, and makes it easier to turn one session's experience into a reusable project structure.
3. Open-Weight Models Get Safety Infrastructure: Base Labs Partners With Hugging Face and Goodfire(#opensource #safety; TechCrunch, 2026-09-17; Base Labs, part of Baseten): Baseten's research arm Base Labs partnered with Hugging Face and Goodfire AI on safety evaluation and monitoring infrastructure for open-weight models, arguing that safety should be built into training and deployment rather than bolted on afterward. The report notes that removing safeguards through a technique known as abliteration is on the rise, with more than 6,000 abliterated models now listed on Hugging Face. Why it matters: if a design team wants to run open models locally or in a private environment, safety stops being only the platform's problem. The low cost of abliteration shows that once weights are open, guardrails can be stripped, so evaluation and monitoring as an open standard is what gives local deployment something checkable to rely on.
4. Huawei Moves Its Next AI Chip, the Ascend 960DT, Up to Q1 2027(#hardware #compute; TechCrunch, 2026-09-17; Huawei Connect): At its Huawei Connect conference, Huawei said it is moving the launch of its next-generation Ascend 960DT AI chip to the first quarter of 2027 from the third quarter, to close the gap with NVIDIA faster. Why it matters: the pace of compute supply directly shapes what local inference and rendering can support. For teams in supply-constrained regions or running generative models in their own environment, how fast domestic chips iterate determines how large a model they can run locally and how heavy a rendering or simulation workload they can take on — which feeds back into toolchain and hardware-budget planning.
5. Zhipu Shares an Early RSI Case: A GLM-5.3-Powered Infra Agent Lifts Throughput to 3.2× Baseline in Two Weeks(#model #infrastructure; QbitAI, 2026-09-17; shared by Tsinghua's Tang Jie, from Zhipu's technical blog): Tang Jie, a Tsinghua professor and Zhipu's founder, described an early internal case of recursive self-improvement (RSI): an Infra Agent driven by GLM-5.3 helped build and optimize a production-grade inference system from scratch on a cluster of more than 100,000 domestic chips, lifting end-to-end throughput to 3.2× the original baseline in under two weeks. The agent read system feedback, formed hypotheses, changed code, ran experiments, and iterated — for example, locating a Python GIL concurrency block in a KV Transfer scenario and cutting a performance loss of more than 20% (Prefill plus KV Transfer versus Prefill alone) to under 1%. Zhipu stresses that it has not achieved true RSI. Why it matters: the valuable part is not the "AI optimizing AI" framing but the description of the feedback mechanism — end-to-end metrics only tell a model that things got worse, so correctness tests, logs, execution traces, runtime events, and microbenchmarks have to be organized into feedback that is local, low-cost, and repeatable before an agent can locate the problem. That "dense feedback" idea applies directly to any team putting agents to work on complex systems.
6. Text Agents Start Making Calls: Instinct and Meta's Muse Both Add Calling(#product #agents; TechCrunch, 2026-09-17): Instinct, a text-first assistant, and Meta's Muse agent, launched last week, both added the ability to place calls and act as a concierge. Instinct is one of the most popular products in the category, reportedly attracting investor interest at roughly a $10 billion valuation. Why it matters: as agents move from "write this for me" to "speak for me," the focus of interaction design shifts from the interface to conversation strategy — whose identity it speaks in, how it recovers when refused, and when it must hand back to the person for confirmation all have to be defined. For teams building smart speakers, in-car assistants, and customer service products, this is already a shippable feature rather than a concept demo.
7. Two Generations of Graphics Researchers Join Forces: Tong Xin Becomes Chief Scientist at Meshy(#industry #3Dgeneration; QbitAI, 2026-09-17): Tong Xin, who spent 25 years at Microsoft Research Asia and led its Internet Graphics group as a global research partner, joined Meshy, the AI 3D company founded by Hu Yuanming, as chief scientist in charge of long-term research strategy. Meshy has about 12 million users and closed a roughly $400 million Series B in July; the company frames its direction as "AI for Fun" and is pursuing three lines of foundational work — a new graphics discipline that does not rely on triangles, a "director system," and low-latency 3D generation infrastructure. Why it matters: 3D generation is shifting from "turn a picture into a mesh" toward real-time, interactive worlds and experiences, which calls for depth in graphics rather than visual modeling alone. For teams working on rendering, animation, and real-time visualization, the thing to watch is whether this route can genuinely bypass the traditional polygon pipeline — that would ripple into asset specs, rendering budgets, and delivery formats.

Interesting GitHub Projects#

1. Shpigford/nurb: "Agentic CAD" for 3D Printing(#opensource #CAD; GitHub, created 2026-07-25, updated 2026-09-16, ~544★, topics include 3d-printing, ai, cad): The project hands the usual CAD-to-3D-printing workflow to an agent, positioning itself as agentic CAD for 3D printing so a model can take part in going from modeling to a printable part. Why it matters: making "design intent to printable part" an agent-controllable pipeline is exactly the step designers most want automated and find hardest to automate. How mature this kind of project becomes will decide whether AI can take part in structural design rather than just producing renderings.
2. tsunehimatoi/psd2live: Turn a Layered PSD Into an Editable Live2D Model in One Step(#opensource #2Dto3D; GitHub, created 2026-09-03, updated 2026-09-17, ~399★, GPL-3.0): It automates rigging, mesh and deformer generation, physics, and animation, turning a designer's layered PSD directly into an editable Live2D model and skipping the tedious manual splitting and rigging in between. Why it matters: it shows a very practical division of labor — designers keep working in the 2D tools they know while conversion and rigging are automated. For teams building characters, avatars, and interactive content, tools like this can noticeably shorten the path from artwork to a movable asset.
3. bestagentkits/design-studio-ai: An Open-Source Design Workspace Shared by Agents and Humans(#opensource #designtool; GitHub, created 2026-09-07, updated 2026-09-14, ~168★, MIT): An open-source design workspace with cloud editing, 3D, and motion graphics that connects agents through MCP / WebMCP, a CLI, and BYOK, so machines and people share one design environment. Why it matters: treating "agents can operate this too" as a baseline assumption of a design tool, rather than an afterthought plugin, is one of the clearer directions this year. For teams choosing tools, it determines whether work can flow in both directions between people and agents.
4. weewrr/Meshforge: A Node-Based Image-to-3D Workflow That Runs Locally on Consumer GPUs(#opensource #3Dgeneration; GitHub, created 2026-09-03, updated 2026-09-16, MIT, Three.js viewport): A local, open-source image-to-3D mesh generation tool aimed at consumer GPUs, organizing the pipeline as a node-based canvas and showing results in a Three.js viewport. Why it matters: making image-to-3D an offline, composable node graph means assets do not have to be uploaded, and the pipeline can be reused and debugged like a material graph. For design teams handling unreleased products or confidential material, that is a more realistic choice than a SaaS tool.
5. MahsaMozafariNia/Thermal-Monitoring-Additive-Manufacturing: Early Defect Detection for Large-Format Printing With Thermal Imaging(#opensource #AMquality; GitHub, created and updated 2026-09-16): Aimed at large-format 3D printing, it uses thermal-camera data to detect defects early and warn before they spread. Why it matters: it complements today's LLNL work — one watches geometry, the other watches thermal history. Large-format printing's worst case is discovering a failed build hours in, and folding thermal signals into in-line inspection is one of the most direct ways to improve first-pass yield.