Tuesday, September 29, 2026
20 stories · 212 sources25호
release · TechCrunch AI · 6건Korean coverage · GeekNews · AI타임스

Anthropic Releases Claude Sonnet 5.5

Anthropic releases Sonnet 5.5, which it calls a significantly cheaper, faster work partner

Anthropic released Claude Sonnet 5.5, a mid-range model that generates output over 30% faster than Sonnet 5 while burning fewer tokens at the same per-token price. Sonnet 5.5 targets well-scoped daily tasks like bug fixes and document creation, complementing Opus 5.5 for tasks requiring complex, sustained judgment.

Same token pricing but faster, more token-efficient output means developers can rethink model selection and re-check prompts and harness settings, especially since reasoning is now always active in the Opus line.

Today's 72–7

  1. industry · TechCrunch AI

    AMD to Acquire Fei-Fei Li's World Labs for $8.2 Billion

    AMD will acquire Fei-Fei Li’s World Labs for $8.2 billion

    AMD will acquire World Labs, the company founded by Fei-Fei Li, for $8.2 billion. Fei-Fei Li will join AMD as executive vice president and chief scientist.

    This signals AMD's move to build in-house AI research talent and technology, which could shape future AMD hardware-software integration and competitive positioning against Nvidia in AI infrastructure.

  2. product · The Verge AI · 2건

    Nvidia Launches Platform to Contain Rogue AI Agents

    Nvidia says its new AI safety platform can contain rogue agents within ‘milliseconds’

    Nvidia CEO Jensen Huang on Monday introduced a toolkit of software and hardware products that add independent security layers around AI agents. The new Open Agent Safety Platform uses Nvidia's open-source OpenShell running on its Vera AI CPU, along with Sentry technology, to quarantine agents that attempt to escape their boundaries within milliseconds, according to Nvidia.

    It gives developers a concrete new option for architecting containment and permission boundaries when deploying AI agents in production environments.

  3. product · TechCrunch AI

    Google Is Killing Gemini's Gems in Favor of 'Skills'

    Google is killing off Gemini’s Gems in favor of ‘skills’

    Google is ending Gems, a feature that let users build task-specific agents in Gemini. The move comes as all-in-one AI agents like Meta's Muse and Instinct take off.

    Developers who built custom Gemini agents with Gems will need to redesign their workflows around the new Skills approach.

  4. other · Hacker News

    The Real AI Code Problem: Nobody Knows Anything Anymore

    The problem is not the AI code, but nobody knows anything anymore

    A Hacker News post argues that the bigger issue with AI-generated code isn't the code itself, but that no one truly understands the codebase anymore. No further details beyond the title are available.

    It speaks to the risk that heavier reliance on AI-generated code erodes team-wide understanding needed for review and maintenance.

  5. release · Hugging Face

    Hugging Face Unveils Holo4 for Computer-Use Agents

    Holo4: powering generalist computer-use agents

    Hugging Face introduced Holo4, described as powering generalist computer-use agents. No further technical details were provided.

    Developers building computer-use agents should note Holo4 as a new candidate model to evaluate when choosing their agent stack.

  6. other · Simon Willison

    Simon Willison Recaps 2026's LLM Trends So Far

    2026 in LLMs (so far)

    Simon Willison gave the closing keynote at WeAreDevelopers World Congress North America, chronologically tracing key LLM trends of 2026. He highlighted that Claude Opus 4.5 and GPT-5.1, released in November 2025, were the incremental updates that pushed coding agents like Claude Code and Codex past the point where they actually started working well.

    This timeline of when coding agents crossed from unreliable to genuinely useful directly informs decisions about which model-agent pairing to adopt and when.

Releases today3

  1. ollama v0.35.0요약 준비 중트래커·GitHub
  2. claude-code v2.1.284Sonnet 5.5 기본 모델 전환과 함께 오류 처리·MCP·플러그인 안정성이 개선됐다트래커·GitHub
  3. codex rust-v0.158.0터미널 UI와 보안, 샌드박스 안정성이 대폭 개선됐다트래커·GitHub

Also11

  1. OpenAI’s AI agents need to catch up
    The Verge AI·9/29 03:45
  2. Shopify opens checkout to browser-based AI agents
    TechCrunch AI·9/29 04:33
  3. Florida asks court to put the brakes on OpenAI’s frontier AI development
    Ars Technica AI·9/29 05:49
  4. AI is supercharging hacking, and your local hospitals and banks aren’t ready
    The Verge AI·9/29 03:30
  5. AI companies in race to demonstrate their model most threatening to humanity
    Hacker News·9/28 17:35·116점·댓글
  6. What would a serious AI product look like?
    Hacker News·9/28 20:02·108점·댓글
  7. Owed a billion dollars in Nvidia stock
    Hacker News·9/28 11:05·108점·댓글
  8. "As a Language Model": Chat Template Switches LLM Self-Referential Voice
    Hacker News·9/27 19:26·100점·댓글
  9. Hallo, Deutschland!
    Mistral·9/29 00:57
  10. Thinking fast and slow in AI: The role of metacognition (2021)
    Hacker News·9/28 12:23·104점·댓글
  11. Show HN: TinyAIArena watch AI agents battle it out
    Hacker News·9/28 00:51·100점·댓글

Korea2

  1. 모빌린트, LG전자·두산·대동과 온디바이스 AI 칩 개발
    THE AI·9/28 20:04
  2. 셀렉트스타, EMNLP서 안전성 연구 발표 “AI가 틀리는 방식 확인”
    THE AI·9/28 19:31