General-Purpose Decision Models and Fast "System 1" Inference
Exploring fast, general-purpose decision models as an intermediate approach between specialized classifiers and full generative LLMs.
The Review Boundary Has Moved
The review boundary for generated code is shifting, raising questions about how we establish confidence in software behavior.
Long-Horizon Coding Agents and Software Degradation
Long-horizon coding agents can persist through tasks, but their software often degrades over time due to structural and maintainability issues.
Reasoning Models and Overthinking
Kimi K3, Qwen 3.8 Max, GLM-5.3 and similar models have made significant post-training gains in reasoning benchmarks, but such models sometimes overthink and perform worse in interactive work. We explore why this happens …
Prompt Caching, Context Windows, and LLM Inference Cost
Prompt caching is an inference optimization that allows an LLM provider to reuse computation for input tokens. Its value becomes easier to understand when considered alongside context-window size and the way …
Intent Alignment Under Ambiguity
Evaluating Muse Glimmer 30B and Qwen3.8 27B on a new benchmark for inferring user intent under ambiguous or incomplete instructions with local inference.
Our New Find a Cheaper Model Tool
A practical model-comparison tool for finding lower-cost alternatives with similar measured capability.
Can LLMs Preserve Judgment When They Don't Know?
A 143-prompt adversarial benchmark of hallucination and epistemic control. Meta’s Muse Glimmer sets a new high-water mark, but then scores lower when it thinks harder.
Durable Memory for Agents
Store factual knowledge in a temporal, provenance-backed claim graph rather than unstructured markdown text files.
Searching Is Not Knowing
What eight open models reveal about current-news research, summarization fidelity, and agent design.