Independent projects and a research desk: what we build, what we watch and what may matter next.
01
Build to understand.
An illustrated stroll with Thomas and Thor, following Marseille’s light and weather. From the Old Port to the Calanques, stopping at Les Goudes and waving to Mamie Pauline. The game is in French.
Fifty-six texts in French, accompanied by selected paintings and images. French poets, the Chinaski notebook and the À contre-vent and Au bord de la nuit collections, to read and contemplate.
02
A few developments that materially change our view.
NVIDIA is adding a 64 GB DGX Spark configuration, due October 23, while keeping the local-agent positioning and the ability to cluster two systems.
Our takeThe decision is no longer simply cloud versus local, but how much memory, bandwidth and software ecosystem the actual workload needs.
NVIDIAOn 1 October, ANYbotics announced Shift, an evolution of Data Navigator. The platform connects robot missions, inspection data and maintenance systems. DCS/SCADA signal integration is part of the 1.0 release announced for October.
Our takeThe useful test is traceability from asset to reading, anomaly and maintenance decision. Third-party robot support and agent orchestration remain announced directions, rather than evidence of capabilities available in operation today.
ANYboticsBrittany Ferries, Adrena and Bureau Veritas report energy savings of up to 6% after thousands of commercial voyages and nearly two years of onboard monitoring aboard Galicia. The system is now deployed across nine vessels.
Our takeStrong signal: value comes from a restrained stack — operational data, a physical model, domain expertise and controlled recommendations — rather than a generic AI layer.
Brittany FerriesIEEE Robotics & Automation Society highlights that dexterous manipulation is still constrained by the lack of high-quality tactile data even as vision-language-action models advance quickly.
Our takeWorth watching: value may shift from models alone toward sensors, datasets and interfaces that can measure contact, force and deformation reliably.
IEEE Robotics & Automation SocietyDatapace reports a pilot agent-memory benchmark: ten questions, one seed, on free models. Tested memory systems use fewer tokens without beating file search on this limited sample.
Our takeBefore adding complex memory, measure retrieval, freshness, cost and failure modes against a trivial baseline. The most sophisticated architecture is not automatically the best one.
Datapace — agent-memory-benchmark03
Topics tracked over time. Personal, commercial or confidential monitoring stays off the public site.
0 in this collection
Orchestration, tool use, memory, permissions and workflow automation.
1 in this collection
Benchmarks, observability, guardrails, provenance and recovery from failure.
0 in this collection
Open models, local RAG, cloud-independent deployment and data control.
1 in this collection
Field deployments, digital twins, performance and decision support.
2 in this collection
Robotics, world models, perception, touch and interaction with the physical world.
1 in this collection
Unified memory, accelerators, local workstations and onboard compute.
04
Editorial assessments as of 4 Oct 2026. Arrows indicate our attention, not measured growth.
Tool and permission control is becoming an architectural layer in its own right.
Serious document and vision workloads are becoming viable without cloud dependency.
Sensors, data and hardware durability remain the bottleneck.
Value is most credible when the model is tested against real operations over time.
Benchmarks are becoming more useful than isolated demos.
Bandwidth and software now matter as much as raw memory capacity.
05
Ideas that emerge when several signals start telling the same story.
An agent architecture only earns its complexity if it beats a simpler option on a metric that actually matters: accuracy, cost, latency, freshness or control. A well-indexed file can still be a better starting point than sophisticated memory.
Local systems are improving fast enough to make private processing credible for more workloads. But the right decision remains workload-specific: model size, bandwidth, latency, confidentiality, maintenance and total cost.