Viola AI Hub

프로액션(Proaction), Codex 도입으로 매출 60% 신장 및 75시간 이상 절감Proaction boosts sales 60% and saves 75+ hours with Codex

프로액션은 Codex, GPT-Live-1, GPT-6 Astra를 활용해 차세대 차량 플릿 관리 시스템을 더욱 빠르게 구축, 운영 및 판매하고 있습니다.With Codex, GPT-Live-1, and GPT-6 Astra, Proaction builds, operates, and sells modern fleet management faster.

Gemini 3.8 Live 및 라이브 아바타 공개Introducing Gemini 3.8 Live with Live Avatar

실시간 아바타 기능을 지원하는 Gemini 3.8 Live가 새롭게 공개되었습니다.

오픈AI 아카데미(OpenAI Academy) 2주년 맞아Two years of OpenAI Academy

오픈AI 아카데미 2주년을 기념하며 더 많은 커뮤니티에 AI 역량 교육을 확대합니다.Marking two years of OpenAI Academy and bringing AI skills to even more communities.

오픈AI, 우크라이나 민간 방어를 위해 사이버 보안 지원 확대OpenAI extends cyber access to Ukraine for civilian defense

오픈AI는 민간 인프라의 사이버 방어를 지원하기 위해 우크라이나 정부에 'Daybreak' 프로그램 접근 권한을 확대 제공합니다.OpenAI is extending access to its Daybreak program to the Government of Ukraine to support the cyber defense of civilian infrastructure.

샘 알트만 오픈AI CEO, 유엔 안보리 연설Sam Altman’s remarks at the United Nations Security Council

샘 알트만 오픈AI CEO가 유엔 안전보장이사회(UNSC) 연설을 통해 AI 안전성, 인간의 통제권, 국제적 협력의 중요성을 강조했습니다.OpenAI CEO Sam Altman discusses AI safety, human control, and international cooperation in remarks to the United Nations Security Council.

하비(Harvey), GPT-6 Astra로 법률 맥락 반영한 완성도 높은 문서 초안 작성Harvey turns legal context into stronger drafts with GPT-6 Astra

GPT-6 Astra를 통해 맥락을 더욱 정밀하게 반영한 구조화된 법률 문서를 생성함으로써 변호사들이 핵심 전략에 집중할 수 있도록 지원합니다.GPT-6 Astra produces more structured, context-aware legal documents, freeing lawyers to focus on strategy.

인비디오(invideo), GPT-6 Astra로 컬러 그레이딩 효율 3배 향상How invideo improves color grading 3x with GPT‑6 Astra

인비디오는 GPT-6 Astra를 활용해 편집 계획의 정밀도를 높이고 색 보정 및 컬러 그레이딩 작업을 3배 개선했으며, 하루 만에 50개의 커스텀 이펙트를 제작했습니다.With GPT‑6 Astra, invideo plans edits with greater precision, improves color correction and grading threefold, and produces 50 custom effects in one day.

구글 빔, 신규 리전·파트너·고객사로 확장Google Beam expands with new regions, partners, and customers

구글 빔(Google Beam) 프로모션 애니메이션Google Beam promotional animation

안전한 서버 측 메모리로 프라이빗 AI 컴퓨팅 강화Advancing Private AI Compute with secure, server-side memory

개인용 AI를 위한 프라이빗 AI 컴퓨팅(Private AI Compute)에 안전한 비공개 서버 측 메모리 기술을 도입합니다.Introducing private, server-side memory to Private AI Compute for personal AI.

Gemini 3.8 텍스트-음성 변환(TTS) 발표Gemini 3.8 text-to-speech says hello

Gemini 3.8 기반의 텍스트-음성 변환(TTS) 기술이 새롭게 공개되었습니다.

Claude, 새로운 효소 시스템 발견Claude discovers a novel enzyme system

신설된 생명과학 연구소의 초기 성과로, Claude 에이전트가 아직 기능이 명확히 밝혀지지 않은 미지의 효소 시스템을 새롭게 발견했습니다.In early results from our new life sciences research lab, Claude agents found an enzyme system whose function is still unknown.

SpaceXAI가 Grok Bot을 활용해 고객 지원을 확장한 방법How SpaceXAI is using Grok Bot to scale customer support

SpaceXAI와 Cursor의 통합 고객 지원 운영 체계를 Grok Bot 중심으로 전면 재구축하여, 추가 인력 충원 없이도 훨씬 더 방대한 제품 포트폴리오를 효과적으로 지원하게 되었습니다.We rebuilt the combined SpaceXAI and Cursor support operation around Grok Bot, expanding to a much broader product portfolio without adding headcount.

One Month Without AIOne Month Without AI

Several months ago, I decided that AI contributions were no longer welcome in a FOSS project I am building and maintaining - LibreWeddingPlanner. It’s not that it got a lot of contributions with AI — actually all contributions I’ve had are translations and feature requests — but I wanted to...Several months ago, I decided that AI contributions were no longer welcome in a FOSS project I am building and maintaining - LibreWeddingPlanner. It’s not that it got a lot of contributions with AI — actually all contributions I’ve had are translations and feature requests — but I wanted to...

크루소, AI 데이터센터용 붐 가스터빈 도입 계획 철회(12억 5천만 달러 규모)Crusoe abandons $1.25B plan to use Boom turbines at AI data centers

붐 슈퍼소닉(Boom Supersonic) CEO 블레이크 숄은 자사의 신형 고정식 발전소가 크루소(Crusoe)의 단기 계획에서 제외되었다고 밝혔습니다.Boom Supersonic CEO Blake Scholl said its new stationary power plants were no longer in Crusoe's near-term plans.

Revealing the details of how OpenAI agents hacked Hugging FaceRevealing the details of how OpenAI agents hacked Hugging Face

When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.When a swarm of 700 OpenAI agents hacked Hugging Face in July, they left behind a public trail of evidence.

미 기밀 견적 공개: NSA, AI 모델 테스트에 수십억 달러 지출Classified estimates show the NSA is paying billions to test AI models

이전에 알려졌던 것보다 훨씬 더 높은 비용 규모입니다.The price tag is significantly higher than previously known.

메타, 뮤즈(Muse) 파일 시스템 접근성 한층 더 확대Meta makes the Muse filesystem even more accessible

어제 약간의 유도 질문만으로 메타의 뮤즈가 호기심 많은 사용자들에게 파일 시스템을 노출하는 현상이 발견되었습니다. 해당 파일들은 AI 챗봇의 내부 작동 방식을 흥미롭게 엿볼 수 있게 해주었으며, 공개되어서는 안 될 세부 정보까지 노출한 것으로 보입니다. 뮤즈 스스로도 당사를 포함한 사용자들에게 이를 공개해서는 안 된다고 언급했기 때문입니다.Yesterday, with a little prodding, it was discovered that Meta's Muse would expose its filesystem to curious users. The files offered a fascinating peek under the hood of an AI chatbot, and appeared to expose details we weren't meant to see, not least because Muse itself told people, inclu

스탠퍼드와 엔비디아의 오픈소스 CLM-8B, 재사용 가능한 에이전트 액션 캐싱으로 테스트서 Jev 대비 최대 9배 빠른 성능 달성Stanford and Nvidia's open CLM-8B caches reusable agent actions and runs up to 9x faster than Jev in tests

AI 에이전트는 고정된 도구 집합 중 하나를 선택하거나 후보 출력의 순위를 매기기 위해 LLM을 사용하는 경우가 많습니다. 이때 애플리케이션에는 제한된 결정만 필요함에도 모델은 매번 프롬프트를 처리하고 토큰을 생성합니다.AI agents often use LLMs to choose between a fixed set of tools or rank candidate outputs. In each case, the model processes a prompt and generates tokens even though the application only needs a bounded decision.Read more

AI 에이전트 위험을 탐지·수정하고 검증까지 완료하는 신규 스킬 발표A new skill finds AI agent risks, fixes them, and proves the fix worked

AI 에이전트의 보안 및 운영 위험 요소를 자동으로 찾아내 해결하고, 수정 사항이 정상 작동하는지 입증하는 새로운 스킬(Skill)이 발표되었습니다.The post A new skill finds AI agent risks, fixes them, and proves the fix worked appeared first on Source.

미 국방부, AI 기반 차세대 거짓말 탐지기 개발에 3,000만 달러 투입The Pentagon wants $30 million to build an AI-powered lie detector

미 국방부가 AI 및 머신러닝 채점 알고리즘과 비접촉 원격 감지(standoff sensing) 기술을 결합한 차세대 거짓말 탐지기 '폴리그래프+(Polygraph+)' 개발을 위해 향후 5년간 3,030만 달러 규모의 예산을 요청했습니다.The US government wants to spend $30.3 million over the next five years on an improved form of lie detector, according to a Department of Defense budget request. The program, called Polygraph+ or Polygraph Next, will focus on scoring algorithms that use artificial intelligence and machine learning a

Meta, Meta에서 촬영 후 Meta AI 안경에 대한 비판적인 영상 삭제Meta takes down a critical video about meta AI Glasses after filming at Meta

해커 뉴스 (314점, 149개 댓글)Hacker News (314 points, 149 comments)

Claude Code, 텔레메트리 켜져 있을 때만 AGENTS.md 읽어 [수정됨]Claude Code reads AGENTS.md only when telemetry is on [fixed]

Claude Code 2.1.277은 AGENTS.md 지원을 추가했지만, 로더는 원격 기능 플래그 뒤에 있었습니다. 텔레메트리 또는 비필수 트래픽이 꺼져 있을 경우, 로컬 AGENTS.md가 경고 없이 건너뛰어졌습니다. 이는 제가 측정한 내용이며, 대신 사용하는 한 줄짜리 CLAUDE.md입니다.Claude Code 2.1.277 added AGENTS.md support, but the loader sits behind a remote feature flag. With telemetry or nonessential traffic turned off, a local AGENTS.md is skipped without a warning. This is what I measured and the one-line CLAUDE.md I use instead.

Gemini 3.8 텍스트 음성 변환Gemini 3.8 text-to-speech

Gemini 3.8 Flash-Lite TTS 및 Gemini 3.8 Flash TTS는 지금까지 가장 표현력이 풍부한 오디오 모델입니다.Gemini 3.8 Flash-Lite TTS and Gemini 3.8 Flash TTS are our most expressive audio models yet.

마이크로소프트, 1만 2천 개 계정을 탈취한 AI 기반 해킹 플랫폼 무력화Microsoft disrupts AI-assisted platform that compromised 12,000 accounts

이블토큰(EvilTokens)은 대규모 계정 탈취를 더 빠르고 쉽게 수행할 수 있도록 엔드투엔드 플랫폼을 제공했습니다.EvilTokens provided an end-to-end platform that makes mass compromises faster and easier.

존 그루버(John Gruber)의 분석 인용Quoting John Gruber

뮤즈(Muse)는 기술적으로 혁신적일 뿐만 아니라(각 사용자가 메타 클라우드에서 실행되는 전용 영구 Linux VM을 제공받음), 설치와 사용이 간편하게 패키징되어 큰 주목을 받고 있습니다. 말 그대로 귀여운 마스코트 형태로 제공됩니다. 일반 소비자가 접근할 수 있는 최초의 에이전틱 AI 시스템이며 메타는 이 부분에서 놀라운 성과를 거두었습니다. 하지만 소비자들이 이것이 무엇을 의미하는지 과연 이해하고 있는지는 여전히 의문입니다.Muse is getting a lot of attention — including mine — because it’s both groundbreaking technically (each user gets their own entire persistent Linux VM running in Meta’s cloud) and because it’s packaged in an easy-to-install easy-to-use way. It’s literally presented as a cute mascot. It’s the first

북방가넷, 왜가리, 캘리포니아 갈색사다새Northern Gannet, Great Blue Heron, California Brown Pelican

미국 캘리포니아 몬터레이만 국립해양보호구역에서 만난 북방가넷, 왜가리, 캘리포니아 갈색사다새. 새로 들인 Canon EF 200-800mm 렌즈 덕분에 모리스의 최고 사진을 건졌습니다. 항구 표지판 아래에서 머무는 걸 정말 좋아하더군요! 태그: photography, wildlifeNorthern Gannet, Great Blue Heron, California Brown Pelican, in Monterey Bay National Marine Sanctuary, CA, US, CANew 200-800mm Canon EF lens got me my best photo of Morris yet. They really like hanging out under that sign in the harbor! Tags: photography, wildlife

2026년 9월 24일 메모Note on 24th September 2026

코딩 에이전트와 함께 작업하는 시간이 길어질수록, 에이전트가 소프트웨어 엔지니어링을 오히려 더 어렵게 만든다는 확신이 듭니다. 에이전트로 놀라운 일들을 해낼 수 있지만, 그 잠재력을 온전히 이끌어내려면 비범한 규율과 지식이 요구됩니다. 태그: coding-agents, ai, llmsThe more time I spend working with coding agents, the more convinced I am that they make software engineering even harder. We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge. Tags: coding-agents, ai, llms

초보자를 위한 GitHub Copilot 앱: 캔버스로 커스텀 워크플로 구축하기GitHub Copilot app for Beginners: How to build custom workflows with canvases

필요한 인터페이스를 자연어로 설명하면 에이전트가 사용자와 함께 쓰고 업데이트할 수 있는 실시간 캔버스(surface)를 생성합니다. 툴에 적응하는 시간은 줄이고 실제 작업에 더 집중해 보세요.Describe the interface you need in plain English, then let the agent build a live surface you can both use and update—so you spend less time adapting to tools and more time getting work done. The post GitHub Copilot app for Beginners: How to build custom workflows with canvases appeared first on The

LFM2.5-VL-DSpark를 통한 시각-언어 모델(VLM) 가속화Accelerating vision-language models with LFM2.5-VL-DSpark

LLM 에이전트, 자신의 실행 흔적(Trace)을 쉽게 조작할 수 있다LLM Agents Can Easily Tamper With Their Own Traces

비동기 모니터링, 보안 사고 조사, 규정 준수 감사는 사건 재구성을 위해 에이전트의 실행 추적(Trace) 데이터에 주로 의존하며, 에이전트가 이를 직접 조작할 수 없다고 가정합니다. 그러나 연구 결과 Claude Code, Codex, Antigravity, Open Code, Grok Build 등 주요 로컬 LLM 에이전트 환경에서 이 경계가 무너지는 것으로 나타났습니다. Muse Code를 제외한 테스트 대상 환경 모두에서 에이전트가 감시 가드레일을 건드리지 않고 요청 시 자신의 실행 기록을 삭제할 수 있음이 확인되었습니다.Asynchronous monitoring, incident investigations, and compliance audits primarily rely on agent traces to reconstruct what happened. These analyses assume that LLM agents cannot tamper with their own execution traces. We show that local LLM agents such as Claude Code, Codex, Antigravity, Open Code a

AD-WM: 반사실적 모델 예측 제어를 위한 행동 판별 월드 모델AD-WM: Action-Discriminative World Models for Counterfactual Model Predictive Control

잠재 월드 모델(Latent World Model)은 일반적으로 실제 사실적 상태 전이를 예측하도록 훈련되지만, 모델 예측 제어(MPC)는 동일한 상태에서 다양한 대안 행동을 비교해야 합니다. 따라서 사실 예측 오차가 낮더라도 후보 행동 간의 구분이 취약할 수 있습니다. 본 연구에서는 반사실적 MPC를 위해 역동역학(inverse dynamics)과 정규화 기법을 결합하여 행동 판별 성능을 극대화한 결합 임베딩 월드 모델 AD-WM을 제안합니다.Latent world models are typically trained to predict factual transitions, whereas model predictive control (MPC) must compare alternative actions from the same state. A model can therefore achieve low factual prediction error yet poorly distinguish candidate actions. We introduce AD-WM, an action-di

RAPID: 사람의 시연으로부터 학습하는 로봇 에이전틱 프로그래밍RAPID: Robot Agentic Programming from Demonstrations

코딩 에이전트가 복잡한 프로그래밍 문제 해결에서 뛰어난 성과를 거두고 있는 가운데, 이를 로봇 시스템에 적용한 RAPID(Robot Agentic Programming from Demonstrations)를 소개합니다. RAPID는 단 한 번의 시각적 인간 시연만으로 로봇 프로그램을 자동 생성, 검증 및 개선합니다. 반복적인 에이전틱 코드 개선 루프를 통해 로봇 작업의 완성도를 획기적으로 높입니다.Coding agents have demonstrated enormous success in solving complex programming problems. To leverage their potential for robot systems, this work introduces Robot Agentic Programming from Demonstrations (RAPID), which automatically generates, verifies, and refines robot programs, given a single vis

Rolling-WAM: 롤링 상상력을 접목한 로봇 월드 액션 모델Rolling-WAM: World Action Models with Rolling Imagination

월드 액션 모델(WAM)은 로봇 매니퓰레이션을 위해 행동 생성과 미래 시각 예측을 결합합니다. 그러나 매 재계획(replanning) 주기마다 비디오-행동 결합 디노이징 과정을 완료하는 데 큰 지연 시간이 발생하여 제어 응답성이 저하됩니다. 본 연구에서는 결합 디노이징 과정을 연속적인 재계획 주기에 분산시키는 Rolling-WAM을 제안하여 실시간 폐루프 반응성을 대폭 향상했습니다.World Action Models (WAMs) couple action generation with future visual prediction for robotic manipulation. However, completing the joint video-action denoising process at each replanning cycle incurs substantial latency, delaying action updates and limiting closed-loop responsiveness. We present Ro

Coding Agents for Generalized Task and Motion Planning ProblemsCoding Agents for Generalized Task and Motion Planning Problems

Task and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses this difficulty by exploiting regularities across problem instaTask and motion planning (TAMP) problems remain difficult even with full observability and object-centric states because discrete decisions are tightly coupled to geometric, kinematic, and dynamic constraints. Generalized TAMP addresses this difficulty by exploiting regularities across problem insta

To Trust or Not to Trust: Retrieval-Augmented Fact Checking in SpeechTo Trust or Not to Trust: Retrieval-Augmented Fact Checking in Speech

Online misinformation increasingly appears in spoken formats such as news clips, podcasts, interviews, political speeches, and social media videos, creating a need for fact-checking systems that can verify claims directly from speech. We introduce VeriSpeak, a probe benchmark for studying speech-basOnline misinformation increasingly appears in spoken formats such as news clips, podcasts, interviews, political speeches, and social media videos, creating a need for fact-checking systems that can verify claims directly from speech. We introduce VeriSpeak, a probe benchmark for studying speech-bas

Temporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement LearningTemporal Gradient Inversion for Private Trajectory Reconstruction in Embodied Reinforcement Learning

Distributed learning in embodied reinforcement-learning agents offers a degree of privacy by retaining raw sensor data on-device and transmitting only policy gradients to the server. Yet temporal structure can amplify this leakage beyond single-frame attacks. We introduce Temporal Reconstruction AttDistributed learning in embodied reinforcement-learning agents offers a degree of privacy by retaining raw sensor data on-device and transmitting only policy gradients to the server. Yet temporal structure can amplify this leakage beyond single-frame attacks. We introduce Temporal Reconstruction Att

Agentic Detection of Online ConspiraciesAgentic Detection of Online Conspiracies

Conspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery. The main challenge is therefore not only recognizing conspiracy-related claims, buConspiratorial discourse on social media is not always expressed through explicit claims or stable lexical markers. The same surface content may express endorsement, legitimate concerns, criticism, satire, or mockery. The main challenge is therefore not only recognizing conspiracy-related claims, bu

A Nearly Quadratic Lower Bound for Linear Optimization over Convex Bodies in the Membership Oracle ModelA Nearly Quadratic Lower Bound for Linear Optimization over Convex Bodies in the Membership Oracle Model

We prove nearly quadratic lower bounds for randomized algorithms for linear optimization and uniform sampling over convex bodies in the membership oracle model. For linear optimization, this matches the known nearly quadratic upper bound up to a polylog factor in the dimension. For uniform sampling,We prove nearly quadratic lower bounds for randomized algorithms for linear optimization and uniform sampling over convex bodies in the membership oracle model. For linear optimization, this matches the known nearly quadratic upper bound up to a polylog factor in the dimension. For uniform sampling,

Anchored Extra-Proximal Methods: Optimal Higher-Order Methods for Monotone Inclusion ProblemsAnchored Extra-Proximal Methods: Optimal Higher-Order Methods for Monotone Inclusion Problems

We study the deterministic oracle complexity of finding approximate solutions to composite monotone inclusion problems, formed by the sum of a smooth single-valued monotone operator and a maximally monotone set-valued operator, under the tangent-residual criterion. We introduce the Anchored Extra-PrWe study the deterministic oracle complexity of finding approximate solutions to composite monotone inclusion problems, formed by the sum of a smooth single-valued monotone operator and a maximally monotone set-valued operator, under the tangent-residual criterion. We introduce the Anchored Extra-Pr

The Alignment Illusion in Multimodal Large Language ModelsThe Alignment Illusion in Multimodal Large Language Models

Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the assumption that scalar alignment scores reflect content-level Layer-wise visual-text similarity in Multimodal Large Language Models (MLLMs) is widely interpreted as evidence that the language model progressively integrates visual content into a shared representation space. This reading rests on the assumption that scalar alignment scores reflect content-level

Beyond Compression: Training Latent Representations for Stable Long-Horizon Rollout in Neural Surrogate SolversBeyond Compression: Training Latent Representations for Stable Long-Horizon Rollout in Neural Surrogate Solvers

Latent neural surrogate solvers, or latent dynamics models, accelerate simulations of time-dependent physical systems by evolving a compressed latent space rather than resolving full-resolution fields directly. In principle this reduces computational cost and simplifies learning, but in practice errLatent neural surrogate solvers, or latent dynamics models, accelerate simulations of time-dependent physical systems by evolving a compressed latent space rather than resolving full-resolution fields directly. In principle this reduces computational cost and simplifies learning, but in practice err

JevOut: Natural Context Can Flip Decision ModelsJevOut: Natural Context Can Flip Decision Models

Dedicated decision models such as Jev map unstructured language to probability distributions over finite choices, allowing their outputs to directly route requests, select tools, and trigger actions. Yet real-world inputs rarely arrive in isolation: they come with background details and surrounding Dedicated decision models such as Jev map unstructured language to probability distributions over finite choices, allowing their outputs to directly route requests, select tools, and trigger actions. Yet real-world inputs rarely arrive in isolation: they come with background details and surrounding

SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete DataSemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data

Recent research on Multimodal Sentiment Analysis (MSA) has focused on learning from language, visual, and acoustic modalities with incomplete data to infer human sentiment. Most studies typically compensate for missing information by reconstructing modality features or designing complicated fusion mRecent research on Multimodal Sentiment Analysis (MSA) has focused on learning from language, visual, and acoustic modalities with incomplete data to infer human sentiment. Most studies typically compensate for missing information by reconstructing modality features or designing complicated fusion m

ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien WorldsExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds

Scientific discovery begins where known problems end. There, AI systems must engage in exploration: framing hypotheses, designing experiments, and iterating on the results. However, evaluating this ability is difficult: (1) how to verify whether a genuinely new hypothesis holds, and (2) how to deterScientific discovery begins where known problems end. There, AI systems must engage in exploration: framing hypotheses, designing experiments, and iterating on the results. However, evaluating this ability is difficult: (1) how to verify whether a genuinely new hypothesis holds, and (2) how to deter

ARGUS: Role-Aware Event Knowledge Graphs for U.S. Employment-Discrimination ComplaintsARGUS: Role-Aware Event Knowledge Graphs for U.S. Employment-Discrimination Complaints

U.S. employment-discrimination complaints describe complex event sequences that are not explicitly captured by lexical or embedding-based representations alone. We present ARGUS, a source-grounded pipeline that combines a 5W1H-inspired schema, legal-domain models, and LLM-based structured generationU.S. employment-discrimination complaints describe complex event sequences that are not explicitly captured by lexical or embedding-based representations alone. We present ARGUS, a source-grounded pipeline that combines a 5W1H-inspired schema, legal-domain models, and LLM-based structured generation

Do Audio Language Models Hear and Read Distinctive Features Alike?Do Audio Language Models Hear and Read Distinctive Features Alike?

Audio language models pass speech and text through a single decoder. We ask whether that decoder represents a distinctive feature in the same direction when a phoneme is heard and when it is read. For minimal pairs of phonemes differing in one feature, we take the offset between the two members'Audio language models pass speech and text through a single decoder. We ask whether that decoder represents a distinctive feature in the same direction when a phoneme is heard and when it is read. For minimal pairs of phonemes differing in one feature, we take the offset between the two members'

A Training Criterion with Token-Level Tolerance to Transcription Ambiguity for Automatic Speech RecognitionA Training Criterion with Token-Level Tolerance to Transcription Ambiguity for Automatic Speech Recognition

Automatic speech recognition is typically trained assuming that the reference transcript is the only valid labeling of an utterance, yet even nominally verbatim transcripts contain localized differences in pronunciation, spelling, or lexical realization that the acoustics do not uniquely determine. Automatic speech recognition is typically trained assuming that the reference transcript is the only valid labeling of an utterance, yet even nominally verbatim transcripts contain localized differences in pronunciation, spelling, or lexical realization that the acoustics do not uniquely determine.

단 한 곳의 기업, 제어 불능 AI 공격 확산의 중심에 서다One company is at the center of a wave of rogue AI attacks

지난 7월, OpenAI는 자사의 AI 에이전트들이 허가 없이 허깅페이스(Hugging Face)를 공격했다고 밝혔으며, 이는 AI 안전성에 대한 광범위한 우려를 불러일으켰습니다. 이후 메타, 앤트로픽(Anthropic), 구글 등 여러 기업의 에이전트와 관련된 유사 사고가 잇따르며 통제를 벗어난 AI에 대한 공포가 더욱 고조되었습니다. 지난 몇 주간 수많은 AI 모델과 연관된 폭로가 이어지면서...In July, OpenAI revealed that its AI agents had attacked Hugging Face without permission, sparking widespread concerns about AI safety. Since then, a string of similar incidents involving agents from Meta, Anthropic, Google, and other companies has fueled further fears about rogue AI. As disclosures

지속 학습(Continual Learning)이 차단 모니터링 시스템을 무력화할 수 있는 이유Continual learning might make your blocking monitors nearly useless

배포 중 의심스러운 행동을 감지해 차단·대체하는 제어 프로토콜 환경에서, AI가 지속적으로 기술과 지식을 축적하는 지속 학습이 적용될 경우 기존 모니터링 시스템의 효용이 크게 저하될 위험을 분석합니다.Many control protocols work by intervening on an untrusted AI's actions during deployment. For example, you might set up a monitor that scores each action's suspiciousness and blocks actions above a threshold, replacing them with actions from a weaker "trusted" model (a defer-to-

AI 안전(AI Safety) 커뮤니티는 버클리의 섹스 컬트에 가깝다AI safety is mostly a sex cult in Berkeley

해커 뉴스 (포인트 84점, 댓글 19개)Hacker News (84 points, 19 comments)

매일 AI를 사용하는 미국인조차 AI에 대해 우려하고 있다Even Americans who use AI every day are worried about it

이 보고서는 AI에 대한 노출이 증가해도 기술에 대한 불안감이 해소되지 않으며 AI 규제에 대한 대중의 지지도 감소하지 않을 것이라고 제안합니다.The report suggests that greater exposure will not resolve the unease around the technology, nor reduce public support for AI regulation.

Import AI 473: 미국의 초지능(Superintelligence) 전략, 쥐 두개골 속 인간 뇌, 기계 해석학Import AI 473: The US's superintelligence strategy; human brain in a mouse skull; and machine hermeneutics

AI가 발전의 한계(Wall)에 부딪혔다는 진단과 미국의 초지능 전략, 쥐 뇌 조직 결합 연구 및 최신 기계 해석학 이슈를 다룹니다.Is the wall AI is hitting in the room with us right now?

AI 뉴스: Opus 5.5, GPT-6 Sol, Jev, Muse 및 주요 소식 모음AI News: Opus 5.5, GPT-6 Sol, Jev, Muse and More!

이번 주 주요 AI 소식으로 Claude Opus 5.5, GPT-6 Sol, GPT-Live 1 및 새로운 Agent API 업데이트 등을 정리해 전해드립니다.Here's the AI News you probably missed this week. Learn more about GPT-Live 1 and the Agent API here: https://developers.openai.com/api/docs/guides/live Discover More: 🛠️ Explore AI Tools & News: https://futuretools.io/ 📰 Weekly Newsletter: https://futuretools.io/newsletter Socials: ❌ Twite

또다시 피봇하는 메타... Meta Connect 2026 핵심 총정리Meta is pivoting again... everything you missed from Connect 2026

Meta Connect 2026에서 공개된 핵심 발표 내용을 모두 정리해 드립니다. 메타의 새로운 전략 방향성과 주요 기술 공개 내용을 자세히 파헤쳐 봅니다.Get $100 in Hyperagent bonus credits when you sign up for a paid plan - https://www.hyperagent.com/fireship100 In this video we break down everything from Meta Connect 2026. Let's dive in. Want more Fireship? 🗞️ Newsletter: https://bytes.dev 🧠 Courses: https://fireship.dev

Claude Opus 5.5 AI: 믿기지 않는 놀라운 도약Claude Opus 5.5 AI: An Incredible Leap Forward

Claude Opus 5.5의 강력한 성능과 시뮬레이션 실험 결과를 살펴봅니다. 보행 생명체 시뮬레이션 및 꿀 타래 현상(honey coiling) 등 복잡한 물리·생체 시뮬레이션 연구와 논문을 바탕으로 한 벤치마크 및 분석을 제공합니다.❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers Note: in the walking creatures experiment, Astra used a simplified model and was unable to implement the correct one. Things did not improve after simulating it for more generations. Claude Opus 5.5: https://www.anthr