Today's briefing
- Agents
Google DeepMind's AlphaProof Code verifies Signal's cryptographic core in four weeks instead of years
AgentsThe DecoderNov 12, 2026 - Industry & Policy
OpenAI's quarterly revenue reportedly crosses $10 billion for the first time
Industry & PolicyTechCrunch +1 moreNov 12, 2026 - Research
Google DeepMind formally verifies Signal's encryption code with AlphaProof Code
ResearchKorean outletAI TimesNov 11, 2026
Open source & papers radar
- RepoOpen Source & ToolsRepository: openai/codex-securityStars: 11,633New stars today: 20Language: TypeScript
OpenAI releases Codex Security CLI and SDK for vulnerability scanning and remediation
OpenAI released Codex Security, a CLI tool and TypeScript SDK for finding, validating, and fixing security vulnerabilities. The tool is available as an npm package.
- PaperResearcharXiv 2611.14320Upvotes: 51
WikiTrace: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
The authors identify memory-induced degradation in LLM agents—failures on tasks the agent would solve without memory, often from following incorrect lessons—and propose WikiTrace, an evidence-linked memory system that links each lesson to its source trajectory. Tested on five agentic benchmarks, WikiTrace consistently outperformed existing memory approaches and reduced degradation.
- PaperPhysical AIarXiv 2611.12885Upvotes: 63
RoboScript: World In-Context Learning for agentic robot manipulation
The authors propose World In-Context Learning (WICL), where a frozen model learns manipulation tasks by rehearsing demonstrations in a simulator and writing feedback into its context. Their RoboScript system achieved 83.4% mean success across all 50 RoboTwin 2.0 tasks with one demonstration per task and no gradient updates, exceeding trained policies by over 20 points and transferring to a physical dual-arm robot.
- PaperResearcharXiv 2611.11995Upvotes: 31
Infrathon: Benchmarking Language Agents for Real-World AI Infrastructure Optimization
The authors introduce Infrathon, a benchmark for evaluating frontier agents on long-horizon AI infrastructure optimization tasks. Testing six models across 150 runs with 24-hour budgets showed that agents discovered solutions outperforming human baselines—Claude Opus 5.5 raised KV-cache compression from 6.82% to 19.86%—and that shorter evaluations can mask optimization gains and model ranking changes.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
5 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Agents
Meta expands Muse AI agent to Apple Watch and Wear OS smartwatches
Meta extended its Muse AI agent to Apple Watch and Wear OS smartwatches after launching on smartphones and tablets in September. Users can request tasks by voice or text and check progress on their smartwatch.
AI TimesNov 11, 2026 - Industry & Policy
TSMC to invest up to $1 billion in OpenAI and Anthropic
TSMC approved investments of up to $500 million each in Anthropic and OpenAI as the two AI companies develop their own chips. The moves follow OpenAI and Anthropic's efforts to reduce their dependence on Nvidia and other external semiconductor suppliers.
AI TimesNov 11, 2026 - Industry & Policy
OpenAI acquires Odessia, a travel AI agent startup
OpenAI acquired Odessia, a travel-focused AI agent startup, and brought aboard Francis Davidson, founder of Sonder, a rental platform similar to Airbnb. The acquisition was not disclosed in terms but is characterized as an acqui-hire focused on personnel.
AI TimesNov 11, 2026 - Physical AI
UNIST develops stretchable diode to reduce e-skin signal errors in robots
Researchers at UNIST's Department of Chemistry developed a stretchable organic diode that blocks unnecessary current flow between pressure sensors in electronic skin for robots, reducing signal errors for precise tactile-based manipulation.
iRobotNewsNov 11, 2026
Stories in this issue
27 storiesGoogle launches Gemini Decision 1, a decision model that scores up to 1,000 questions in a single pass
Google DeepMind released the Gemini Decisions API, which evaluates up to 1,000 closed-ended questions in a single pass across text, images, audio, and video. Gemini Decision 1 returns results in 150 to 250 milliseconds at about five times lower cost than Gemini 3.5 Flash-Lite, Google says.
Google announces Gemini Decisions API for fast, calibrated multimodal decisions
Google announced the Gemini Decisions API, powered by its Gemini Decision model, for fast and calibrated multimodal decision-making.
Google's 48 AI agents discover unmapped Maya pyramid in NASA Lidar data in 72 hours
Google deployed 48 AI agents that analyzed NASA Lidar data and discovered an unmapped Maya pyramid in 72 hours.
OpenAI releases Codex Security CLI and SDK for vulnerability scanning and remediation
OpenAI released Codex Security, a CLI tool and TypeScript SDK for finding, validating, and fixing security vulnerabilities. The tool is available as an npm package.
Google DeepMind releases AlphaGenome 2 for cell-type-specific variant effect prediction
Google DeepMind released AlphaGenome 2, a conditional DNA sequence model that predicts regulatory activity and variant effects per cell type. It is available via the AlphaGenome API for non-commercial research.
Google DeepMind's AlphaProof Code verifies Signal's cryptographic core in four weeks instead of years
Google DeepMind's AlphaProof Code mathematically verified Signal's cryptographic core code in four weeks, compared to the years the process typically takes. AI agents powered by Gemini Carbon wrote more than six million lines of proof code and flagged about one percent for human review. Deepmind plans to extend the method to critical infrastructure like power grids and medical devices.
OpenAI's quarterly revenue reportedly crosses $10 billion for the first time
OpenAI's quarterly revenue reportedly reached $10 billion for the first time, representing an almost 50% jump from $6.7 billion in the second quarter.
WikiTrace: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
The authors identify memory-induced degradation in LLM agents—failures on tasks the agent would solve without memory, often from following incorrect lessons—and propose WikiTrace, an evidence-linked memory system that links each lesson to its source trajectory. Tested on five agentic benchmarks, WikiTrace consistently outperformed existing memory approaches and reduced degradation.
RoboScript: World In-Context Learning for agentic robot manipulation
The authors propose World In-Context Learning (WICL), where a frozen model learns manipulation tasks by rehearsing demonstrations in a simulator and writing feedback into its context. Their RoboScript system achieved 83.4% mean success across all 50 RoboTwin 2.0 tasks with one demonstration per task and no gradient updates, exceeding trained policies by over 20 points and transferring to a physical dual-arm robot.
Infrathon: Benchmarking Language Agents for Real-World AI Infrastructure Optimization
The authors introduce Infrathon, a benchmark for evaluating frontier agents on long-horizon AI infrastructure optimization tasks. Testing six models across 150 runs with 24-hour budgets showed that agents discovered solutions outperforming human baselines—Claude Opus 5.5 raised KV-cache compression from 6.82% to 19.86%—and that shorter evaluations can mask optimization gains and model ranking changes.
WaveGen: End-to-End Text-to-Waveform Generation via Spectral Trajectory Learning
The authors propose WaveGen, a fully end-to-end diffusion-based text-to-waveform framework that jointly learns spectral and waveform trajectories without relying on intermediate representations, external alignment modules, or self-supervised embeddings. Scaling from 0.1B to 0.6B parameters, WaveGen achieved state-of-the-art results on Emilia, reaching WER of 1.47 and SIM of 0.75 on the Seed-en benchmark.
Google's Tabletop generates custom collaborative apps from text prompts
Google launched Tabletop, an experimental AI tool that generates customized collaborative hubs to help coordinate with family and friends. Users can describe interactive tools—such as savings trackers, chore charts, event plans, and more—and the tool creates them, with a visual style similar to workplace management apps like Asana.
Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days
Anthropic claims it used its in-house model Mythos to port Claude to AMD GPUs in just three days, with almost no human intervention. Anthropic now runs Claude on four chip families and aims to cut costs and reduce vendor dependencies.
Meta expands Muse AI agent to Apple Watch and Wear OS smartwatches
Meta extended its Muse AI agent to Apple Watch and Wear OS smartwatches after launching on smartphones and tablets in September. Users can request tasks by voice or text and check progress on their smartwatch.
TSMC to invest up to $1 billion in OpenAI and Anthropic
TSMC approved investments of up to $500 million each in Anthropic and OpenAI as the two AI companies develop their own chips. The moves follow OpenAI and Anthropic's efforts to reduce their dependence on Nvidia and other external semiconductor suppliers.
Google DeepMind formally verifies Signal's encryption code with AlphaProof Code
Google DeepMind used AlphaProof Code, a multi-agent AI system, to mathematically prove that Signal's core encryption library meets key security requirements. The move extends AI applications from cryptographic analysis to real-world software security validation.
OpenAI acquires Odessia, a travel AI agent startup
OpenAI acquired Odessia, a travel-focused AI agent startup, and brought aboard Francis Davidson, founder of Sonder, a rental platform similar to Airbnb. The acquisition was not disclosed in terms but is characterized as an acqui-hire focused on personnel.
Fish Audio releases Live 3, a full-duplex voice model reacting in 186 ms
Fish Audio released Live 3, a full-duplex voice model for voice agents that listens while speaking, distinguishes filler sounds from real interruptions, and responds in as little as 186 ms. English is supported at launch.
Tree-Ising repairs language models with bit-level discrete optimization
Tree-Ising casts model repair as sparse discrete optimization over BF16 weights, using sign flips and exponent shifts. On two layers, it achieves 54.6% on MATH-500, outperforming the agent's edit (51.7%), while requiring only four forward-pass requests and zero LLM tokens at about $0.90 in GPU time.
APEX achieves higher instruction fulfillment in few-step streaming video generation
The authors extend APEX to streaming video using condition shifting to estimate the velocity field of the generated distribution. On 180-second videos, APEX achieves higher instruction fulfillment (3.856 versus 3.831) and visual aesthetics (0.641898 versus 0.632606) compared with S2-Avatar, with human raters showing +30.0% net preference on instruction following.
ASpec, a high-throughput training framework for speculative decoding models
ASpec is a high-throughput training framework designed for speculative decoding models to accelerate LLM inference.
MMDM measures decision-making models across 16 task categories with 5,448 questions
The authors introduce MMDM, a benchmark for evaluating System-1 decision models. It covers 16 task categories with 5,448 high-quality questions, option spaces ranging from 2 to 1,024, and inputs up to 64,817 tokens. Testing 28 models reveals that many perform well on common knowledge but struggle with interdependent options, constrained decisions, and multi-select tasks.
FOMO-Dex applies force-driven corrections for contact-rich robotic manipulation
Force-Modulated Dexterity (FOMO-Dex) is a unified compliance formulation for contact-rich manipulation that applies force-driven motion corrections along selected directions. In card picking, paper cutting, screwdriving, and nut-and-bolt assembly, it reduces excessive contact forces compared with stiff control, and policies trained on FOMO-Dex demonstrations achieve higher success rates than those trained on raw operator commands.
MERL improves VLA post-training with trust-calibrated imagination
Model-Evolving Reinforcement Learning (MERL) couples policy optimization with an online-updated simulator comprising a world model and reward proxy. Simulator reliability controls the imagined-data ratio and rollout horizon. On perturbed LIBERO-PRO tasks, MERL reaches 79.7% average success, improving high-threshold learning speed by 10.9% over matched model-free control.
Horkos launches NX70 Duo AM combining metal additive manufacturing and precision machining
Horkos and Meltio 3D launched the NX70 Duo AM, a hybrid machining center combining metal additive manufacturing and high-precision cutting. The twin-head non-bed system integrates additive and subtractive manufacturing in a single machine.
Memories.ai provides standardized egocentric video collection and annotation for robot training data
Memories.ai offers standardized egocentric video collection and AI-powered annotation to build training data infrastructure for robotics.
UNIST develops stretchable diode to reduce e-skin signal errors in robots
Researchers at UNIST's Department of Chemistry developed a stretchable organic diode that blocks unnecessary current flow between pressure sensors in electronic skin for robots, reducing signal errors for precise tactile-based manipulation.