Technologie · KI in der FabrikNVIDIAs FOX fragt Fabriken, wo die Befugnisse der KI enden sollenVerfasst 8. November 2026 um 13:29 UTCAnsehenTranskript0:07NVIDIA's factory blueprint lets plants decide which AI decisions need human0:11approval.0:11Those choices govern software designed to investigate faulty products and coordinate0:15equipment.0:16A recommendation to an engineer and permission to move a circuit board carry0:19different responsibilities.0:20Where does the software's authority end?0:22And what do the advertised savings actually measure?0:25Nvidia's reference example starts with an optical inspection camera flagging a0:29circuit board.0:29Software takes another look, then coordinates an investigation.0:32The robot handling demonstration uses simulation, so the sequence illustrates a0:37proposed workflow rather than documenting an autonomous production line.0:43Fox stands for Factory Operations Blueprint.0:46It's an early access design that developers adapt to a plant.0:49NVIDIA's November4 technical description sets out three jobs: connect existing data0:53and tools, configure specialist agents, and define operator review.0:57A manager agent sits above manufacturing execution software, business planning1:01systems, and machine databases, bringing their different information into one1:04investigation.1:05The camera alert leads to a visual recheck.1:08The reference sequence includes a simulated transfer toward rework, followed by1:11machine data analysis.1:13A root cause agent assembles a PDF report with evidence and a recommended corrective1:17action.1:18The plant decides which steps need approval.1:20The visual agent uses NVIDIA Cosmos to compare the suspect component with the1:24factory handbook without retraining the model.1:26Updating that handbook could help with product changes, though the savings still1:30needs testing.1:31Kumo Anomaly examines machine readings, including reflow, oven temperatures.1:36An unusual temperature pattern can suggest a cause, but an engineer still needs to1:40judge whether the proposed response fits.1:42A plant could require approval before every rework dispatch.1:45Another could allow board movement while reserving process changes for a person.1:49Those are examples of policies a plant could configure.1:52Each implementation needs testing to show how its permissions behave under operating1:55conditions.1:56The manager can use a language model such as NVIDIA Nematron.1:59NemoClaw supplies its agent framework, and OpenShell isolates execution.2:03The design combines task interpretation, persistent memory, and access policies.2:08Plant teams still have to connect their systems and evaluate the result.2:11Configurable approval rules alone can't guarantee that every consequential action2:15will receive human review.2:16Siemens Electronics Factory in Erlang in Germany supplies the main numerical claims2:20surrounding Fox.2:21Frequent product and order changes make planning and inspection important there.2:24But the figures describe separate activities, and each needs to stay attached to the2:28job it measures.2:29The financial pressure reaches beyond inspection.2:32Siemens research estimates annual unplanned downtime costs of two hundred2:35fifty-three million dollars for an average large plant in the sector surveyed.2:39That estimate describes those surveyed sectors with no promised Fox saving attached.2:43The practical problem is time spent gathering evidence across systems, finding2:47specialists, and agreeing on a response while an operational issue continues.2:51Siemens says agent based production planning can enable cost savings of up to twenty2:55percent.2:56That's a potential upper bound.2:57Siemens hasn't specified the cost basis or comparison period behind it, so factories3:02can't treat that figure as an expected return on their own implementation.3:05In the planning example.3:07A rush order triggers simulations through Siemens Intelligence Center X.3:11Nemotron powered Fox agents weigh costs and existing orders alongside supplies and3:15production line setups.3:16Siemens is integrating those agents into Intelligence Center X.3:19The twenty percent claim belongs to that planning work.3:22Nobody has published a benchmark showing how accurately the complete defect3:26investigation workflow diagnoses faults, or how much time it saves.3:29NVIDIA also claims a forty percent reduction in inspection calibration time at3:33Erlangen, with the baseline and method unclear.3:36Separately, Siemens and NVIDIA describe up to forty percent faster time to market3:40using product development and preparation software.3:42Those percentages measure different work.3:44The time too, market work combines Siemens product, Lysophale software, and process3:49preparation X with Cosmos and Omniverse.3:51Synthetic defect images and virtual circuit boards help prepare inspection before a3:55physical prototype exists.3:57That preparation claim can't validate the calibration figure or the complete Fox4:01investigation.4:02NVIDIA's separate AI factory operations agent handles DGX Superpuid infrastructure,4:06investigating computing, network hardware, and scheduling problems.4:09Fox supports a fully local design using an NVIDIA DGX station.4:13Keeping production records, images, prompts, and results inside the factory requires4:17every model and service to stay there too, including Cosmos and Kumo Anomaly.4:21Running only the manager locally leaves that job unfinished.4:24Factory software teams and systems integrators can request early access.4:28NVIDIA offers its agent toolkit and video search and summarization blueprint to help4:32extend specialist agents.4:33Plants control the approval policy, and real world testing must show whether the4:38combined investigation earns their trust.4:40Fox remains in early access with no announced general availability date.Artikel lesenChronologie der GeschichteHeyDex
