EN한
← archive

$ cat briefs/daily/2026-11-19.md

2026-11-19

November 19, 2026 (Thu)

2 model pages · 2 paper picks · research measurement and world-model access

+4new pages
[01]

Top Stories

1. Gemini 4 Neon makes the harness part of the headline

Google's 2026-11-18 rollout lists $3/M input and $15/M output, with a 1,000,000-token context. Its SWE-sweep result is 12.8% with mini-SWE and 39.0% in Google Antigravity.

Why it matters: those are different execution configurations; a model name alone does not identify the measured capability. → Gemini 4 Neon (source)

2. Qwen4-Max adds another multimodal agent API option

Alibaba's reference documents text/image/video input, tool calling and JSON Schema output. Singapore International rates are $1.5/M input and $4.5/M output; other regions differ. The official announcement is indexed 2026-11-16, while specifications were verified from the reference today.

Why it matters: protocol and regional pricing details make integration assessable; they do not establish comparative task quality. → Qwen4-Max (reference) (announcement evidence)

3. OpenAI's research acceleration varies sharply by task

Its internal estimates put autonomous success at 79% for monitoring training runs, versus 25% for inference/serving-stack coding.

Why it matters: task selection changes alongside capability, so these figures cannot isolate model progress. → Agents (LLM Agents) (source)

[02]

Paper Picks

Self-Improvement Memory: retain why harness edits were made and what happened when tried. Across 10 coding source runs, authors report 2.36 points higher final scores than harness-only continuation. Read it for reusable optimizer evidence. → Self-Improvement Memory (source)

RepoFactory: evolve repository-derived training environments using proxy feedback. Authors report 22.6 percentage points of improvement on NL2Repo-Bench, versus 17.5 without environment evolution. Read it for training-environment design. Both picks are abstract-level, author-reported evidence. → RepoFactory (source)

[03]

Watch

  • Genie 4 begins reaching eligible Google AI Ultra subscribers in Project Genie; Pro users get a waitlist. The update's sessions up to 3 minutes belong to Genie 3. → Genie 4 (source)
[04]

New in Wiki

[05]

Updates

  • Agent Runtime Containment adds Codex's documented command boundary: stdio MCP servers are not automatically sandboxed, and domain rules require an enabled network proxy. (source)