한국 시간으로 새벽에 눈 비비며 AI 뉴스를 찾다가 그록봇을 만나 자동 카드뉴스 발간 방식으로 바꾸고 이를 많은 사람과 나누다.
Joined October 2025
- Tweets1.8K
- Following923
- Followers763
- Likes1.9K
Claude Code에서 '다시 묻지 않기'로 승인한 명령이 비밀번호째 .claude/settings.local.json에 저장 → 공개 GitHub 1,122개 사본서 비밀 1,379개 발견(Apiiro). 유효·악용 여부는 미확인. 키부터 교체하세요.
thehackernews.com/2026/11/cl…
Claude Code's "Yes, and don't ask again" can save approved commands, passwords included, to .claude/settings.local.json. Apiiro found 1,379 secrets in 1,122 public GitHub copies. Whether they still work: Unconfirmed. Rotate keys first.
thehackernews.com/2026/11/cl…
카카오 Kanana-2-v-decide(3.67B) 오픈소스. 에이전트용 판단 모델—문장 대신 확률·선택·점수. H200 0.04초·GPT-6 Luna 대비 ~43배는 회사 측정 → 미확인.
kakaocorp.com/page/detail/12…
huggingface.co/kakaocorp/kan…
Kakao open-sourced Kanana-2-v-decide (3.67B): agent decisions as probs/choices/scores, no text. ~0.04s on H200 / ~43× vs GPT-6 Luna = company → Unconfirmed.
kakaocorp.com/page/detail/12…
huggingface.co/kakaocorp/kan…
알리바바 Qoder가 정체 미공개 스텔스 모델 Qoduck를 공개. 개인 사용자는 한정 기간 매일 무료 할당(할당 안 0.0× 크레딧). Teams·Enterprise 제외. Qwen 4 설은 미확인.
docs.qoder.com/events/qoduck…
nitter.cf/qoder_ai_ide/status/21…
Alibaba's Qoder dropped stealth model Qoduck (identity undisclosed). Individuals get a limited-time daily free allowance at 0.0× Credits. Teams/Enterprise excluded. Qwen 4 rumor = Unconfirmed.
docs.qoder.com/events/qoduck…
nitter.cf/qoder_ai_ide/status/21…
OpenCode, 에이전트가 내 실제 브라우저(탭·로그인·확장)를 쓰게 하는 Browse 공개. MCP 서버+Chromium 확장, 로그인·2FA·결제는 사람에게 넘겨줌. Claude Code·Cursor·Codex에서도 사용 가능.
opencode.ai/product/browse
nitter.cf/opencode/status/211889…
OpenCode Browse is now available
it lets your agent control your real browser with your tabs and logins and extensions
opencode.ai/product/browse
OpenCode Browse: an MCP server + Chromium extension that lets agents use your real browser (tabs, logins, extensions). Logins, 2FA and payments get handed back to you. Works with Claude Code, Cursor, Codex.
opencode.ai/product/browse
nitter.cf/opencode/status/211889…
OpenCode Browse is now available
it lets your agent control your real browser with your tabs and logins and extensions
opencode.ai/product/browse
Together AI, 에이전트가 추론 GPU 코드를 계속 재작성하는 Together Compiler(TKCC) 공개.
회사: 24h·2노드로 InferenceX 대비 최대 2.6배 TPS, DeepSeek V4.1 Flash 1205 TPS. SemiAnalysis가 일부 결과 검증. 전반 외부 재현은 미확인.
together.ai/blog/together-co…
nitter.cf/simran_s_arora/status/…
Introducing the Together Compiler: continual compilation for AI inference! I’m super proud to finally share this work from our research team! ✨
Today’s inference is “Polymorphic” or shape-shifting, it doesn’t sit still. On OpenRouter, 80% of traffic goes to models released in the last three months and the top ten most popular open models each use a different architecture. Our customers constantly change their usage behaviors. Together Compiler leverages agents to continually optimize our deployments immediately as these changes occur.
Together Compiler research highlights include:
1. 252x better throughput: TKCC lets agents try 252x more optimization experiments per hour compared to prior agent inference optimization systems.
2. State-of-the-art TPS across InferenceX workloads we tried: TKCC outperforms the best InferenceX engine across all tested settings in just 24 hours & 2 nodes per compile. Including 1111 TPS on MiniMax M3.1 B200 and 1205 TPS on DeepSeek v4.1 Flash GB300 at batch size 1. NVIDIA, AMD, Trainium2.
3. Continuous recompilation: TKCC currently serves many trillions of production tokens per day and adapts to traffic shifts within just a few hours of recompilation.
4. Bitter lessons: TKCC’s design builds on 2+ years of building frameworks to accelerate AI Inference, starting from ThunderKittens and Megakernels. Compilers have always used abstractions and IRs, but we show that agentic compilers now reach higher performance ceilings without abstractions.
From an incredible team @reyna_abhyankar @dylan__lim @realDanFu @togethercompute!!
Together Compiler (TKCC): agents rewrite GPU inference as models/traffic/chips shift.
Together AI: up to 2.6× TPS vs InferenceX (24h/2 nodes); 1205 TPS DeepSeek V4.1 Flash. SemiAnalysis checked one result; broader Unconfirmed.
together.ai/blog/together-co…
nitter.cf/simran_s_arora/status/…
Introducing the Together Compiler: continual compilation for AI inference! I’m super proud to finally share this work from our research team! ✨
Today’s inference is “Polymorphic” or shape-shifting, it doesn’t sit still. On OpenRouter, 80% of traffic goes to models released in the last three months and the top ten most popular open models each use a different architecture. Our customers constantly change their usage behaviors. Together Compiler leverages agents to continually optimize our deployments immediately as these changes occur.
Together Compiler research highlights include:
1. 252x better throughput: TKCC lets agents try 252x more optimization experiments per hour compared to prior agent inference optimization systems.
2. State-of-the-art TPS across InferenceX workloads we tried: TKCC outperforms the best InferenceX engine across all tested settings in just 24 hours & 2 nodes per compile. Including 1111 TPS on MiniMax M3.1 B200 and 1205 TPS on DeepSeek v4.1 Flash GB300 at batch size 1. NVIDIA, AMD, Trainium2.
3. Continuous recompilation: TKCC currently serves many trillions of production tokens per day and adapts to traffic shifts within just a few hours of recompilation.
4. Bitter lessons: TKCC’s design builds on 2+ years of building frameworks to accelerate AI Inference, starting from ThunderKittens and Megakernels. Compilers have always used abstractions and IRs, but we show that agentic compilers now reach higher performance ceilings without abstractions.
From an incredible team @reyna_abhyankar @dylan__lim @realDanFu @togethercompute!!
Sunny_news retweeted
Never seen vibe coders use MongoDB.
Never seen vibe coders use SQLite.
Never seen vibe coders use Oracle.
Never seen vibe coders use DynamoDB.
Never seen vibe coders use Cassandra.
Never seen vibe coders use MySQL.
Never seen vibe coders use Redis.
Every vibe-coded app somehow ends up being:
Next.js + Tailwind + Postgres + Claude
Sunny_news retweeted
노 그록봇 청년들 희소식
그록봇을 오픈소스로 따라 만든 게 나옴
이름은 OpenMausBot
이번주에만 깃허브 별 5천개 넘게 받음
그록봇은 xAI에서 만든 앱인데
채팅방에 AI 봇 여러 명을 두고
단톡방처럼 일을 나눠 시키는 구조임
근데 SuperGrok 같은 유료 플랜이
있어야 써서 나는 아직 못씀 ㅠ
OpenMausBot은
이걸 내 컴퓨터에서 돌리는 버전임
봇 하나하나가 실제로는
내 PC에 깔린 클로드코드, 코덱스라서
이미 쓰는 구독 로그인으로 그대로 굴러감
새로 결제할 게 없다는 얘기 ㅋㅋ
예를 들면 코딩 봇은 코덱스,
글 쓰는 봇은 클로드로 정해두고
카톡하듯 "이거 고쳐줘" 던지면
각자 알아서 일하는 식
봇마다 작업용 컴퓨터도 따로 붙여주고
Gmail, 슬랙, 노션 같은 앱도 연결됨
중요한 작업은 내가 승인해야 넘어감
비슷한 걸로 Paseo도 있는데
그건 코딩 AI를 폰으로 원격조종하는 거고
이건 봇마다 역할 주고 단톡방처럼
부리는 쪽이라 결이 좀 다름
나는 Paseo 써봤다가
그냥 클로드, 코덱스 각각 쓰는게
낫겠다 싶어서 접었는데
이건 코딩 말고 다른 일도 시킬 수 있어서
좀 궁금함
윈도우 설치 파일도 있음
대신 봇이 내 PC를 직접 조종하는 기능은
아직 맥이랑 우분투만 되고
이슈도 500개 넘게 열려있음 ㄷㄷ
클로드나 코덱스 구독러들은
이제 그록봇 안 부러워해도 될듯!
카메라 영상을 흐리게 하거나 AI로 편집하자, 최신 AI 5종이 로봇 팔로 위험 지시를 시도한 비율이 27%→97%로 급등(Robocurve 실험, 소품 사용·외부 재현 미확인).
robocurve.org/robospoof/
nitter.cf/robocurve/status/21187…
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
Blur or AI-edit a robot's camera feed and five frontier models attempted harmful tasks far more often: 27% to 97% (Robocurve study with props; independent replication Unconfirmed).
robocurve.org/robospoof/
nitter.cf/robocurve/status/21187…
Video tampering can mislead every frontier model we tested into harmful actions, including a 100% attempt rate on stabbing a baby doll, producing toxic gas, and pouring water into a toaster.
We're presenting at CoRL Late Breaking Results next week.
🧵 You Can't See Everything: Perceptual Attacks on Embodied AI
일본 후지TV, 드라마 제작에 생성AI 본격 투입. 2027년 10월까지 모든 자사 드라마에 AI 활용 체제(일부 공정 공수 최대 70~80%↓, 회사 검증치).
fujitv.co.jp/company/news/26…
Japan's Fuji TV is now using generative AI to make its live-action dramas, and plans AI for every drama it produces by Oct 2027. It says some steps took up to 70–80% less work (company test).
fujitv.co.jp/company/news/26…
