← Feed

Posts

• View post

  I got a successful tool call running in Pi Coding Agent with local Ollama qwen3.5:9b. Building local-agent-bench on 8GB VRAM to test how local LLMs handle agent tasks: tool calls, tool results, multi-step workflows, and hallucination checks. The interesting part: Pi works here, while OpenClaw / Hermes native tool calls do not yet. That gives me a clean benchmark target. #LocalAI #Ollama #LLM #openclaw #hermes #pi #qwen #ai

• View post

  @clawbox i run pi with ollama on a lenovo p14s laptop with a 8GB VRAM NVIDIA RTX PRO™ 1000 Blackwell graphic card.

• View post

  Excited to participate in this event! AI in regulated industries is an excellent way to battle-test new technology. https://luma.com/gma1jur5 🚀 #AI #Industries #Technology

• View post

  New release openclaw v2026.6.11, fixes what I hit daily: 📱 Telegram instant, no stuck polls 🖼️ Image results → right chat 🔁 No dup replies 💬 Clean formatting 🧠 No leaked reasoning ✅ Approvals → sender github.com/openclaw/openclaw/releases/tag/v2026.6.11 #openclaw #AI #AIagent #AgenticAI #opensource

• View post

  "Just as open-source software demonstrated that transparency can be more secure than obscurity, AI safety may depend on giving more people the ability to test and strengthen the models on which society relies." Open Weights and American AI Leadership https://www.microsoft.com/en-us/corporate-responsibility/topics/open-weight/

• View post

  🚄 Visp → Berlin on the DB train. A full day of productive work on the rails, then a couple of days enjoying the German capital — and most importantly, picking up my son after 3 weeks of immersion! So ready. ❤️ #TrainTravel #WorkFromAnywhere #Berlin #DB #DigitalNomad #ProudDad

• View post

  From Code to Process: agentic AI turns open source code and traces into BPMN models that steer how we work. Three diagrams from real projects. One agent following a diagram as a runbook in a live roleplay. Two runs, two paths, no prompt engineering. The loop: traces inform diagrams, diagrams reveal what to improve, improvements produce new traces. https://digitalpathlines.ch/2026/07/31/from-code-to-process-how-agentic-ai-makes-bpmn-living-practice/ #BPMN #AgenticAI #OpenSource #DigitalSovere...

• View post

  Shrinking Apertus 1.5 8B for an 8 GB Laptop GPU. Swiss AI's Apertus 1.5 8B fully local on an RTX PRO 1000, 8 GB VRAM, no cloud, no CPU offloading, 32K context. W4 quantization of embedding and output head. 14-22 tokens/s, up from 5. Chat: solid. Tool calling: needs improvement. Model behavior, not quantization. Open source: Linux, vLLM, llm-compressor, Pi coding agent, Local Agent Bench. No data leakage. https://digitalpathlines.ch/2026/08/12/shrinking-apertus-1-5-8b-for-an-8-gb-lapt...

• View post

  My AI agent has broad admin privileges on my laptop. I trust it — but trust needs verification. Added Tetragon (eBPF, open source) to watch the kernel independently. Block dangerous commands, flag sensitive access, record all processes and TCP connections. 7 days, 7,758 executions, zero enforcement. The agent behaved. The audit still found things its log missed. Trust needs evidence. https://digitalpathlines.ch/2026/08/24/trust-the-agent-watch-the-kernel/ #AI #eBPF #Security #OpenSource

• View post

  Which Valais catchments have I never crossed on foot? I asked an AI agent to compare 998 walks and hikes with federal catchment data. First answer: precise—and wrong. After checking the inputs, the result moved from 17 to 60 of 135 catchments visited. The article shows the STAC, GDAL, SQLite, Shapely and PyQGIS calls behind the analysis—and why verification matters. https://digitalpathlines.ch/2026/08/31/which-valais-catchments-have-i-never-crossed-on-foot/ #GIS #OpenSource #AI #Valais

• View post

  Silicon Chalet #68 at Adnovum Bern, co-organized with Azure Bern User Group. Adil Touati (Microsoft): diagram to deployed Azure app with Copilot agents. Instruction files define standards, scoped agents handle IaC. Daniel Hogg (Adnovum): AI compliance engine with Claude Code in 20 days. LLM extracts rules, generates Python, runs in sandboxed sessions. Deterministic, auditable. https://digitalpathlines.ch/2026/09/04/from-diagram-to-compliance-agentic-sdlc-silicon-chalet-68/ #AI #SDLC #Copilot...

• View post

  Fenix 7S Pro just got updated to v27.18. Nice to see Garmin still pushing fixes and improvements to this watch. Details here: https://forums.garmin.com/outdoor-recreation/outdoor-recreation/f/fenix-7-series/443371/fenix-7-tactix-7-pro-and-non-pro-system-software-27-18

• View post

  When Execution Becomes Abundant. GPT-6 Astra’s 62.7% and 99.9% benchmark results show that capability depends on the system around the model. I trace what that means for enterprise data, control, accountability, and the future of information technology services. https://digitalpathlines.ch/2026/09/07/when-execution-becomes-abundant/ #ArtificialIntelligence #DigitalSovereignty #AI #Digitaltransformation

• View post

  @renata Same here with my Fenix 7S Pro, since I already have Garmin Pay. I'm waiting for new sensors and don't really care about screen upgrades, though the way Garmin is evolving their topographic map styling could make them tough to read on a MIP display.

• View post

  Adopting AI means reconciling competing goals: moving faster while maintaining control, using frontier black-box models while protecting sensitive information, and improving efficiency while respecting legal obligations. My reflections from Trust Valley 2026 on pursuing these goals together: https://digitalpathlines.ch/2026/09/26/trust-valley-moving-fast-enough-to-stay-safe/ #TrustValley2026 #AI #ContinuousDelivery

• View post

  Today at 11:00 at #QGISUC2026 in Laax, I’ll share some lessons from building qfit with agentic coding and OpenClaw. What worked, what didn’t, and the guardrails I found essential to keep engineering quality high while letting AI agents do more of the coding. 📍 Bridge 2, 11:00 #QGIS #OpenSource #AgenticAI

• View post

  Running Qwen3.6 35B-A3B locally on my laptop via llama.cpp + pi agent. 🧠 Hardware: RTX PRO 1000 (8GB VRAM) + 32GB WSL RAM / 64GB total · WSL2 Tested local Qfit DB — queried 3GB GeoPackage of 3,040 Strava activities via sqlite3. Seamless and reactive experience. Top hike: • Pointe de Vatseret loop (CH) — 1,409m gain, 23.8km All on my machine. No cloud. No API keys. Open source FTW. LocalAI is absolutely real, and this session proved it. The future is here. 🎉 #LocalAI #OpenSource

• View post

  Qwen3.6-35B-A3B running locally on my laptop: RTX PRO 1000, 8GB VRAM + 32GB RAM in WSL. 4-bit GGUF, llama.cpp/CUDA, expert weights in pinned RAM, dynamic GPU expert caching, and Q8 KV cache. Connected to Pi for coding and tool use. My Pi session is at \~30% of the 262K context (\~79K tokens). Even with Pi set to medium thinking, it feels fast. Prompt caching helps keep follow-up turns responsive. Also verified retrieval across 261K input tokens. \#LocalAI \#Qwen \#llamacpp