EN한
← archive

$ cat briefs/daily/2026-11-16.md

2026-11-16

November 16, 2026 (Mon)

2 tooling updates · 2 paper picks · 1 creative workflow

+2new pages
[01]

Top Stories

1. SkillEvaluator makes evaluation budgets and missing evidence explicit

NVIDIA's 2026-11-13 release of SkillEvaluator 0.9.0 adds reusable no-skill baselines, comparisons between skill versions, and token, judge-call, attempt and cost budgets. Partial results remain visible when a budget stops evaluation. Grading changes require reevaluation before comparing older scores; Tier 3 remains advisory in all-tier CI.

Why it matters: a completed report should distinguish measured improvement from incomplete evidence. This is an existing project's release, discovered through Trending; its reliability was not independently tested. → Eval Harness Configuration (source)

2. Datasette Agent gets a documented path to local models

The 2026-11-14 release of datasette-llm 0.1a11 accepts OpenAI-compatible endpoints. Its maintainer demonstrates configuring local models through LM Studio, Ollama or llama.cpp.

Why it matters: endpoint configuration expands the models available to the data-query workflow; it does not establish equal accuracy across them. → Agents (LLM Agents) (source)

[02]

Paper Picks

Evolving Infra: separate compute allocation from KV cache management, then adapt both to agent workloads. On up to 256 GB300 GPUs, authors report up to 4.1× lower mean request latency and 14× lower P95 TTFT versus Dynamo. Read it for the connection between context management and serving; the maxima measure different outcomes and do not establish dollar savings. → Evolving Infra: agent-adaptive serving (source)

Criteria exposure: giving a self-improving embodied coding agent the success criteria improves its exposed-test score, but the advantage disappears under small criterion changes that preserve the instructions. Read it for transfer checks beyond the optimized verifier. Both picks are abstract-level, author-reported evidence. → Criteria exposure in embodied self-improvement (source)

[03]

Watch

  • Higgsfield GPU, dated 2026-11-13, delegates GPU setup and rendering or model execution through chat. Billing runs while the GPU runs; the default pause follows 10 minutes of inactivity, and the entry says to export results before stopping. Rates, persistence and output rights are not established by this entry. → Higgsfield (source)
[04]

New in Wiki

[05]

Updates