Signals: local AI in the wild.
The shift to open-weight and on-device AI is not our opinion alone. Here is where practitioners, labs, and builders are working it out in public. We link to the source, we do not reproduce it. Follow a link to read it where it was published.
Pinned
Recent
- Running Qwen 3.8 Flash Next on a $1000 Local AI Build with Strata
- Building AI for Reliable Execution: Lessons From Industrial Robotics
- ThursdAI - Oct 8 - OpenAI drops 722 math papers, Haiku 5.5 hits 10 cents & more
- ttok 1.0
- I expect rapid progress but not towards general superintelligence
- Step 5 Preview (Tested): This MODEL BLEW MY BRAINS AWAY!
- Reflection AI Announces Beam 501B Open-Weight MoE Model
- I Gave Two AI Supercomputers a Real Job
- Chatting with Alex Ziskind about the Future with AI
- Google Launches EmbeddingGemma 2 Open Model
- We're using GLM-5.3 Flash instead of frontier models on a massive production codebase
- Mistral Is BACK - Mistral Large 4 First Test (Le Chonk!)
- DeepSeek Harness Just Got a Web Agent
- Ollama v0.40.0
- The Cyber Risk Discourse is Broken
- llm-openai-decisions 0.1a0
- Quoting Felix Rieseberg
- Reflection AI Is About to Release a US Open-Weight Model to Take On DeepSeek and Qwen
- You can now try Aleph Alpha's Kolibri 78B for free online here.
- Qwen3.8 27B addition in words
- Strata takes the promise of "MoE models just need a total amount of VRAM+RAM" and makes it a reality
- Strata is seriously impressive, running Qwen 3.8 Flash Next on hermes at 512k context.
- ThursdAI - Oct 1 - OpenAI joins the assistant race, CoreWeave drops serverless GPUs & more
- Ollama v0.35.1
- Phone-Sized Open Models Match Opus Coding Benchmarks
View the full archive (99 items)
Picked by an on-device, open-weight AI pipeline running on our own hardware. An item lands here when a trusted source carries it, or when several independent sources are echoing the same story. One link per story, never a reproduction, and we prune by hand.