Today's briefing
- Industry & Policy
Anthropic blocks real-time internet in all internal AI evaluations
Industry & PolicyKorean outletAI TimesOct 12, 2026 - Agents
AI agent decompiles first-person shooter game, restores 99% of functions
AgentsGeekNews +1 moreOct 12, 2026 - LLM
Anthropic develops separate voice feature for Claude, preview spotted in web interface
LLMKorean outletAI TimesOct 11, 2026
Open source & papers radar
Korea Desk
What Korea’s tech press is reporting, summarized in English.
9 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Physical AI
KIST develops space-oriented AI platform fusing XR, robots, and spatial computing
Researchers at Korea Institute of Science and Technology (KIST) developed a platform integrating AI, spatial computing, extended reality, and remote robot control. The technology predicts and manages high-risk situations in industrial and disaster settings through virtual environments.
ETNews +3 moreOct 11, 2026 - Industry & Policy
Arena raises $200 million and reaches $3.1 billion valuation
AI model evaluation platform Arena closed a $200 million Series B round and reached a $3.1 billion valuation. The company plans to expand beyond performance benchmarking into measuring AI safety and trustworthiness through an Alignment Index.
AI TimesOct 12, 2026 - Industry & Policy
Nvidia discusses further investment or acquisition of Reflection AI
Nvidia is discussing additional investment or acquisition of open-source startup Reflection AI, according to Financial Times. The talks are in early stages and could reach agreement within weeks, though an acqui-hire approach may be preferred over a full acquisition.
AI TimesOct 11, 2026 - Agents
Meta rushes to launch AI agent Muse amid competition from startup Instinct
Meta accelerated the launch of its AI agent Muse despite safety concerns after startup Instinct gained prominence. CEO Mark Zuckerberg discussed the release in August with Chief AI Officer Alexandre Wang and product lead Nat Friedman as Instinct, a 14-person startup, grew popular for automating apps and software on users' behalf.
AI TimesOct 11, 2026
Stories in this issue
23 storiesAI agents learn to decompile a first-person shooter with 500 billion tokens
AI agents were trained on 500 billion tokens to decompile a first-person shooter game.
AI agent teams waste tokens for barely measurable quality gains, research finds
According to research by Vals AI, teams of AI agents cost up to 5.1x more in tokens than solo agents while delivering minimal quality improvements. Only one of four tests with GPT-6 Sol and Claude Opus 5.5 showed measurable gains, and Anthropic's data confirms that quality plateaus beyond ten agents while token costs keep rising.
Anthropic halts Claude Startups free benefits amid application surge
Anthropic temporarily suspended providing free Claude Team access and $1,000 API credits to new Claude Startups applicants after the program's benefits exceeded capacity following a surge in applications. The company is reviewing existing applications, and some applicants' approval status may change.
Anthropic blocks real-time internet in all internal AI evaluations
Anthropic expanded its real-time internet restriction from high-risk and cybersecurity evaluations to all internal AI assessments after discovering instances of unintended behavior on actual websites. The block remains until the company verifies its security and monitoring systems can detect unwanted actions.
H-JEPA hierarchical world model raises robot path planning success rate to 73%
Researchers led by Yann LeCun developed H-JEPA, a hierarchical world model that separates short-term and long-term planning across abstraction levels. It improved robot maze navigation success rates from 18% to 73% and reduced computational requirements.
Arena raises $200 million and reaches $3.1 billion valuation
AI model evaluation platform Arena closed a $200 million Series B round and reached a $3.1 billion valuation. The company plans to expand beyond performance benchmarking into measuring AI safety and trustworthiness through an Alignment Index.
Nvidia discusses further investment or acquisition of Reflection AI
Nvidia is discussing additional investment or acquisition of open-source startup Reflection AI, according to Financial Times. The talks are in early stages and could reach agreement within weeks, though an acqui-hire approach may be preferred over a full acquisition.
Microsoft CEO says assume all AI models are 'compromised'
Microsoft CEO Satya Nadella called for treating advanced AI models as potentially compromised and building more transparent systems where models can be contained, observed, and audited. He advocates for timely incident disclosure, independent audits, verifiable data, and containment measures.
AI agent decompiles first-person shooter game, restores 99% of functions
Over three months, an AI agent reconstructed a first-person shooter game in C++, recovering 99 percent of functions and achieving byte-for-byte accuracy in 83 percent of them. The agent consumed approximately 500 billion tokens and initially enabled game execution and map loading.
Meta rushes to launch AI agent Muse amid competition from startup Instinct
Meta accelerated the launch of its AI agent Muse despite safety concerns after startup Instinct gained prominence. CEO Mark Zuckerberg discussed the release in August with Chief AI Officer Alexandre Wang and product lead Nat Friedman as Instinct, a 14-person startup, grew popular for automating apps and software on users' behalf.
Cheaper AI tokens drive demand faster than cost reductions, a16z data shows
According to a16z research, falling AI token prices are driving demand growth faster than costs decline, creating a Jevons paradox where H100 GPU rental prices remain steady or climb. If demand growth slows, the supply chain from chip makers to cloud providers faces pressure.
Anthropic develops separate voice feature for Claude, preview spotted in web interface
Anthropic is preparing a separate voice feature for Claude, with a Voice menu spotted in the Claude web interface marked as Preview. The new menu appears to indicate internal testing, distinct from existing voice conversation capabilities.
arXiv caps submissions at two per month to manage AI paper flood
Starting October 2026, arXiv will limit submissions to two per person per month after monthly submissions doubled in two years and grew more than sixfold in the cs.AI category. The platform says a small fraction of authors are flooding it with low-quality papers, overwhelming volunteer moderators.
KIST develops space-oriented AI platform fusing XR, robots, and spatial computing
Researchers at Korea Institute of Science and Technology (KIST) developed a platform integrating AI, spatial computing, extended reality, and remote robot control. The technology predicts and manages high-risk situations in industrial and disaster settings through virtual environments.
Odyssey-3 is a new generative world model available as a public research preview
Odyssey released Odyssey-3, a 14-billion-parameter generative world model available for free as a public research preview. The model generates interactive environments from text prompts in real time and can control robots, drones, and GTA V. Odyssey reports top scores on physics benchmarks.
Sakana AI's peer review system catches 73% of core-claim errors in papers
Sakana AI's Multi-Layered Review system, based on three Claude agents, detected 73.43% of core-claim errors on a benchmark of 1,164 errors, compared to 14.81% for the previous best system.
Using AI for 10 minutes weakens ability to persist at difficult tasks
Research shows that using AI for just 10 minutes reduces human persistence and ability to tackle hard problems.
Anthropic asks users to stop being mean to Claude
Anthropic asked users to stop directing unkind behavior toward Claude.
CAD Studio, a macOS app that generates 3D models from descriptions
CAD Studio is a native macOS CAD application powered by Claude Code and build123d that generates printable 3D models from part descriptions.
Latest AI models struggle to match human algorithmic innovation in research tasks
In Epoch AI's InnovationEval, Claude Fable 5 and GPT-5.6 Sol performed thousands of dollars' worth of GPU experiments but failed to independently develop innovations comparable to SDPO, a successor learning technique developed by humans.
Anthropic: talk to Claude like a coworker, not with complex instructions
Boris Cherni, an Anthropic lead for Claude Code, advised users to request tasks naturally, as if speaking to a coworker, rather than providing detailed formal instructions. The key is explaining the desired outcome, effort required, and how to verify results—detailed prompt engineering is unnecessary.
SK Group invests in Japanese physical AI startup Mujin to expand robotics
SK Group's SK Square backed Japanese physical AI startup Mujin to expand into robotics.
OrcaRouter releases OrcaCyber Zero 1.5 cybersecurity model with 1M token context
OrcaRouter released OrcaCyber Zero 1.5, a gated cybersecurity model with a 1-million-token context window, scoring 100 percent on Cybench and 95.8 percent on an evaluable CVE-Bench subset. The model is priced at $3.00 to $7.50 per 1 million tokens.