· 4 min read
AI Ideas
Open WebUI and Ollama: test an internal AI workspace before buying a GPU
A scoped internal AI pilot: choose tasks, map data flow, measure quality and latency, then decide whether server investment makes sense.
Read moreIDEAS, DECISIONS & LESSONS
Journal
Notes on AI, architecture and building useful products.
24 articles
· 5 min read
AI Ideas
A document assistant with Dify: start with a reviewable pilot
A practical proposed pilot for internal document lookup: sources, human review, evaluation, self-hosting and license checks.
Read more
· 4 min read
ai
MiMo-V2.6-Pro Is MIT-Licensed and 1.02T Params — I Ran the Self-Hosting Math
Xiaomi's new open-weight model tops Terminal-Bench 2.1 at 89.9% and jumps from 19.0 to 71.9 on DeepSWE. Before you get excited about MIT licensing, here's what 1.02T total params vs. 42B active params actually costs you in hardware.
Read more
· 5 min read
ai
Datasette's Security Patch Shows What an AI-Assisted Audit Actually Looks Like
Datasette 1.0a39 and 0.65.4, released September 11, 2026, fix permission bugs found through a week-long audit using Claude Fable 5.1, GPT-5.6 Sol, and GPT-6 Astra. A technical lead's breakdown of the two-person-plus-three-models workflow, and why it's a better template than 'just ask the AI to find bugs.'
Read more
· 4 min read
ai
Alibaba Cloud and Cambricon Just Got PyTorch Board Seats — Why That Matters More Than the Press Release Suggests
The PyTorch Foundation added Alibaba Cloud and Cambricon as Platinum members (with governing board seats) and Ant Group as Gold at PyTorch Conference China. A Tech Lead's read on what non-CUDA silicon getting a direct vote in PyTorch's roadmap means for anyone betting on the framework's neutrality.
Read more
· 5 min read
ai
Qwen3.8-Flash-Next Beats Opus 4.6 on Agent Benchmarks With 6B Active Params — I Ran the Numbers
Alibaba's Qwen3.8-Flash-Next, an early preview of Qwen4's architecture, scores 62.5 on SWE-bench Pro against Claude Opus 4.6 Max's 53.4 — while activating just 6B of its 180B parameters. Here's what the routing architecture actually buys you, and where the benchmark story gets more complicated.
Read more
· 5 min read
ai
GLM-5.3-Flash: The First Open-Weight Model I'd Actually Trust With a 1M-Token Context
Z.ai's MIT-licensed GLM-5.3-Flash pairs a 320B-A18B MoE with a hybrid linear/sparse attention scheme built specifically to make million-token context usable, not just possible. A technical lead's read on IndexPool, the self-hosting math, and where it fits next to closed frontier models.
Read more
· 4 min read
ai
Nvidia Is Reportedly Buying Hugging Face for $12.9B — Here's Your Vendor Lock-In Checklist
If your team's model pipeline touches the Hugging Face Hub, Inference Endpoints, or Transformers, a hardware company owning your model registry changes the risk calculus. A practical audit for what to check before this deal closes.
Read more
· 5 min read
agentic coding
GLM-5.3: What It Means When an Open-Weight Model Catches Up on Coding and Security
Z.ai shipped GLM-5.3 without retraining the base model and still doubled coding and cybersecurity benchmark scores. A technical lead's read on why post-training gains matter more than model size, and what it changes about build-vs-buy on agent models.
Read more