Fine-tuning with LoRA: from a training run to an adoption decision
Follow a small synthetic LoRA smoke test through training and reload, then compare held-out work before adopting an adapter.
Articles
Guides, news and explanations about AI and reliable systems.
Not sure where to begin? Choose a short reading path, or browse at your own pace.
Anthropic Scion puts source control behind a cache hierarchy
An agent can reuse a nearby copy of a source file instead of fetching it again. Scion speeds that path; a fleet starting without its saved copies asks more of the shared service.
What is OpenShell and how do I use it?
Run an AI agent with useful access to files and services, then test exactly where OpenShell stops it. A worked guide to sandboxes, policy, credentials, and SRE workflows.
AMD Helion separates agent execution from model inference
An agent may wait on a ticket lookup while its model has nothing to generate. AMD Helion gives those jobs different hardware, and a router chooses where each model request goes.
NVIDIA opens more of the Vera Rubin infrastructure stack
The October 12 OCP release spans rack designs, facility specifications, open-source operations software and planned optics. Operators should test each layer against workload evidence.
Bit-exact decoding turns inference speedups into testable changes
Compare reference and speculative paths for token-level equivalence, then run a separate capacity trial so speed cannot conceal a serving-contract change.
Build an inference compatibility suite before changing the runtime
The new server returns "30" where an application needs the number 30. Follow that small mismatch through the serving stack before trusting a migration.
Auditing AI-generated tests and documentation before cleanup
Review the claim each test or document protects, challenge the evidence, and record a keep, rewrite or retire decision before deleting repository history.
What Basalt changes in a Blackwell local inference stack
Basalt narrows local inference to Qwen3.8-Flash-Next on Blackwell. Its memory and concurrency controls matter more than copying the largest throughput number.
Running Codex CLI as a supervised worker process
A host restarts while its Codex worker may still be editing. Follow the saved process, events and patch to resume the job without starting a second writer.
Shadow-test the routing decision before replacing the log classifier
Compare candidate and current labels on the same events, retain missed and unknown work, and choose a scope the receiving teams can support.
Infermeld shares a model across AMD and NVIDIA, with limits on each card
A 16 GB AMD card and a 10 GB NVIDIA card may share a model, but each still has to fit its own work. Follow Infermeld’s layer choices from loading to a finished log investigation.