Loading the edition…
· Past edition · 18 stories
Retrospective edition reconstructed from later captures. Paper versions may include later updates; coverage labels describe the evidence used.
Infrastructure & deployment
GPT‑6 caching reuses eligible shared prompt prefixes for 30 minutes, with new tools to monitor hit rates, diagnose misses, and shape what gets cached.
Why it matters: For persistent agents, keeping reusable context stable and inspecting misses can help reduce repeated processing and latency. The dashboard and diagnostics give teams ways to measure those changes rather than relying on guesswork.
Infrastructure & deployment
Open each story in the reader. Summaries are shown when available; some sources provide an abstract only. With an account you can keep your own notes.
Greedy decoding can produce different outputs in BF16 and FP16 on identical hardware; the key vulnerability is the LM-head decision, not simply accumulated body error.
Why it matters: Precision choice can change generated outputs even when the model, prompt, decoding algorithm, and hardware are unchanged. Selective LM-head recomputation can reduce divergence in tested low-batch settings, but it is not a universal determinism guarantee.
Infrastructure & deployment
CliffCompaction reduces context cost by trimming or dropping original content, not by rewriting it or repeatedly compacting earlier compactions.
Why it matters: The reported cost savings could make additional test-time rollouts more affordable: the authors say CliffCompaction adds over 10 percentage points on Terminal-Bench for less than the cost of two full-context runs.
Infrastructure & deployment
Hugging Face describes running llama.cpp quantized models through Transformers.
Why it matters: Interoperability can make quantized model formats easier to use.
Infrastructure & deployment
The vLLM team describes a plugin that combines its scheduling and cache management with MLX execution on Apple Silicon.
Why it matters: The project addresses concurrent local-model serving.