PulseAugur
EN
LIVE 05:47:36
한국어(KO) 같은 주에 정반대 이야기가 나왔다. 한쪽은 하네스 무용론(차세대 모델이 흡수한다), 다른 쪽 릴리언 웽은 재귀적 자기개선(RSI)이 가중치가 아니라 하네스에서 먼저 시작된다고 본다. 둘이 진짜 모순인지, 웽이 논문 35편을 5개 서랍으로 묶은 최적화 사다리를 원문 기준으로 정리했다. #

AI agents evolve with self-correction and specialized skills · 5 sources tracked

By PulseAugur Editorial · [5 sources] ·

A collection of AI advancements highlights novel approaches to agent development and legal compliance. One project, MCP, aims to prevent AI-generated legal hallucinations by cross-validating LLM citations against real-world content and case law, developed by a South Korean Ministry of Government Legislation official. Another initiative, k-skill, offers over 100 Korean lifestyle agent skills, curated by a data analyst and marketer. Xiaomi's HarnessX project evolves both harnesses and models, showing significant performance gains, particularly for smaller models. Furthermore, the Self-Harness paper demonstrates an AI's ability to autonomously correct its own rule-based failures without altering the core model, leading to substantial improvements on benchmarks. AI

IMPACT These advancements showcase progress in AI agent autonomy, specialized skill integration, and model efficiency, potentially leading to more robust and adaptable AI systems.

RANK_REASON The cluster describes several distinct AI agent and model development tools and techniques, rather than a single frontier release or significant industry event.

Read on Mastodon — sigmoid.social →

AI-generated summary · Google Gemini · from 5 sources. How we write summaries →

AI agents evolve with self-correction and specialized skills · 5 sources tracked

COVERAGE [5]

  1. Mastodon — sigmoid.social TIER_1 한국어(KO) · [email protected] ·

    MCP that catches fake legal precedents fabricated by AI. It wraps 42 Ministry of Government Legislation APIs with 10 tools, cross-validates LLM citations for existence and content, and prevents incidents like citing discarded precedents as if they were current. Created by a public official tired of searching the Ministry of Government Legislation hundreds of times. # AI # MCP # Law # LLM https:

    AI가 지어낸 가짜 조문을 잡는 MCP. 법제처 42개 API를 10개 도구로 감싸고, LLM 인용을 실존과 내용까지 교차검증하고, 폐기된 판례를 살아있는 것처럼 인용하는 사고를 막는다. 법제처 수백 번 검색에 지친 공무원이 만들었다. # AI # MCP # 법률 # LLM https:// dbhyeong.github.io/blog/korean -law-mcp-legislation-hallucination-check

    LINKS dbhyeong.github.io/…/korean-law-mcp-legis…
  2. Mastodon — sigmoid.social TIER_1 한국어(KO) · [email protected] ·

    From SRT booking to Coupang price comparison, business due diligence, DART, KOSIS statistics, and Korean proofreading, a collection of over 100 Korean life agent skills k-skill (Star 6.1k). I've selected and organized only what I, who does data analysis and marketing, would actually use. # AI # Agent # Skill # Automation https

    SRT 예매부터 쿠팡 가격비교, 사업자 실사, DART, KOSIS 통계, 한국어 윤문까지 100개 넘는 한국 생활 에이전트 스킬 모음 k-skill(스타 6.1k). 데이터 분석·마케팅 하는 내가 실제로 켤 만한 것만 골라 정리했다. # AI # 에이전트 # 스킬 # 자동화 https:// dbhyeong.github.io/blog/k-skil l-korean-agent-skills-collection

    LINKS dbhyeong.github.io/…/k-skill-korean-agent…
  3. Mastodon — sigmoid.social TIER_1 한국어(KO) · [email protected] ·

    If Self-Harness only fixed the harness, Xiaomi HarnessX evolves the harness and model together. Foundry as composable type objects, 4-stage engine AEGIS, and cross-harness GRPO co-evolution. 14.5% improvement on average across 5 benchmarks, with larger gains for smaller models. # AI # AEGIS

    Self-Harness가 하네스만 고쳤다면, 샤오미 HarnessX는 하네스와 모델을 함께 진화시킨다. 조립 가능한 타입 객체로 보는 파운드리, 4단계 엔진 AEGIS, 크로스 하네스 GRPO 공진화까지. 5개 벤치 평균 14.5% 향상, 소형 모델일수록 크게 올랐다. # AI # 에이전트 # LLM # 샤오미 https:// dbhyeong.github.io/blog/harnes sx-xiaomi-composable-evolvable-harness

    LINKS dbhyeong.github.io/…/harnessx-xiaomi-comp…
  4. Mastodon — sigmoid.social TIER_1 한국어(KO) · [email protected] ·

    The model rewrites its own rules by extracting repeated failures from its self-execution logs without changing a single character, and fixes them. The Self-Harness paper (arXiv:2606.09498) improved Terminal-Bench empirical values from 40.5% to 61.9%. Three models fail differently and receive different prescriptions.

    모델은 한 글자도 안 바꾸고, AI가 자기 실행 로그에서 반복 실패를 캐내 규칙을 스스로 고쳐 쓴다. Self-Harness 논문(arXiv:2606.09498)을 Terminal-Bench 실측치 40.5에서 61.9%까지 뜯었다. 세 모델이 각기 다르게 실패하고 각기 다른 처방을 만든 대목이 백미. # AI # 에이전트 # LLM # 논문 https:// dbhyeong.github.io/blog/self-h arness-agents-rewrite-own-rules

    LINKS dbhyeong.github.io/…/self-harness-agents-…
  5. Mastodon — sigmoid.social TIER_1 한국어(KO) · [email protected] ·

    Opposite stories emerged in the same week. One side argues against harnesses (next-gen models will absorb them), while the other, Lillian Weng, believes Recursive Self-Improvement (RSI) starts with harnesses, not weights. Are these two truly contradictory, or did Weng organize the optimization ladder, which groups 35 papers into 5 drawers, based on the original papers?

    같은 주에 정반대 이야기가 나왔다. 한쪽은 하네스 무용론(차세대 모델이 흡수한다), 다른 쪽 릴리언 웽은 재귀적 자기개선(RSI)이 가중치가 아니라 하네스에서 먼저 시작된다고 본다. 둘이 진짜 모순인지, 웽이 논문 35편을 5개 서랍으로 묶은 최적화 사다리를 원문 기준으로 정리했다. # AI # 하네스 # LLM # 에이전트 https:// dbhyeong.github.io/blog/harnes s-engineering-for-self-improvement-lilian-weng

    LINKS dbhyeong.github.io/…/harness-engineering-…

RELATED ENTITIES

RELATED TOPICS