Today's briefing
- Agents
OpenAI upgrades GPT-6 Astra Pro with parallel agent execution
AgentsKorean outletAI TimesOct 19, 2026 - LLM
Liquid AI unveils d1-ultra, a multimodal decision-making model
LLMKorean outletAI Times +1 moreOct 19, 2026 - Industry & Policy
Anthropic cuts junior hiring as AI agents log six times more hours than humans
Industry & PolicyThe DecoderOct 18, 2026
Open source & papers radar
- RepoAgentsRepository: anthropics/claude-code
Anthropic releases Claude Code, a terminal-based agentic coding tool
Anthropic released Claude Code, a terminal-based agentic tool that executes routine coding tasks, explains complex code, and handles git workflows through natural language commands. The tool works in terminals, IDEs, and integrates with GitHub.
- PaperLLMarXiv 2609.21967Citations: 4
NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with tool calling
NemotronLabs VoiceChat is an open full-duplex speech-to-speech model with native tool-calling capabilities that combines a streaming speech encoder, decoder-only language model, and streaming text-to-speech decoder. On Full-Duplex-Bench 1.0, it achieves the lowest pause-handling rates among evaluated open-weight systems and 100 percent takeover following user interruptions.
- PaperResearcharXiv 2610.21901
In-context learning emerges before specialization in sequence model learning
Researchers provide learning theory for how sequence models choose between specialization and in-context adaptation. They show in a solvable topic model that in-context prediction emerges early in training, while specialized prediction appears only after training tokens outnumber the data-generating process parameters, confirmed experimentally in trained transformers.
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
South Korea begins drafting humanoid robot safety guidelines
The Korean Robot Industry Promotion Agency and related bodies have started developing safety and performance evaluation guidelines for humanoid robots. The government plans to procure and deploy about 1,700 robots next year, with cumulative rollout expected through 2030.
ETNewsOct 18, 2026 - LLM
DeepSeek updates V4.1 Flash with improved coding and agent performance
DeepSeek released DeepSeek-V4.1-Flash-1017, an update that improves coding and agent performance while reducing tokens and latency. In the company's own evaluation, it outperformed OpenAI and Anthropic models on some software development and automation tasks while maintaining the same API price per token.
AI TimesOct 18, 2026 - Industry & Policy
Meta removes AI training language from business account terms but keeps content collection rights
Meta removed AI training clauses from its Facebook and Instagram business account terms following criticism that news articles could be used to train AI models. However, the company retained provisions allowing automatic collection of external website content, drawing continued pushback from media organizations.
AI TimesOct 19, 2026 - Industry & Policy
Anthropic signs first Canada data center deal and advances Japan negotiations
Anthropic signed its first Canadian data center lease with Everwind at Point Tupper, Nova Scotia. The company is also in advanced negotiations for its first dedicated data center facility in Japan as it secures infrastructure for model development and service operations.
AI TimesOct 18, 2026
Stories in this issue
27 storiesGoogle reduces quantum error rate by 40× with new error correction technique
Google announced that its Willow quantum processor successfully implements magic state cultivation, a technique that reduces quantum errors to about one-fortieth of the previous level. The breakthrough was demonstrated on actual quantum hardware and reduces the hardware burden needed for large-scale quantum computing.
Four Gemini 4 Argon agents beat Minecraft without human help
Four AI agents running on Google's Gemini 4 Argon defeated Minecraft autonomously by slaying the Ender Dragon. Without predefined roles or strategies, they coordinated independently, wrote 2,783 programs, exchanged 229 chat messages, and died 605 times, with the session costing $1,620.
Hybrid reasoning agents reduce tokens by 35–55% and speed up inference 2.6×
A hybrid reasoning agent approach reduces token usage by 35 to 55 percent and achieves up to 2.6 times faster inference compared to standard methods, according to research presented on arXiv.
Claude solves 25-year-old encryption mystery in under 30 minutes
Anthropic's Claude traced the source of cryptographic seed values used in NSA-provided elliptic curve standards in under 30 minutes, resolving a 25-year-old question about the NIST standards. The discovery addresses suspicions about hidden weaknesses in the encryption design without breaking the cipher itself.
OpenAI upgrades GPT-6 Astra Pro with parallel agent execution
OpenAI upgraded GPT-6 Astra Pro to allow multiple AI agents to work on a single task in parallel, with users able to monitor each agent's progress and results directly. The feature is rolling out to some Pro subscribers in ChatGPT's chat tab.
Liquid AI unveils d1-ultra, a multimodal decision-making model
Liquid AI released d1-ultra, a multimodal decision-making model that outperforms Jev in benchmark evaluations. The model analyzes text and images together while improving decision accuracy and confidence calibration, ranking first among decision-making models in independent assessments.
Anthropic cuts junior hiring as AI agents log six times more hours than humans
Anthropic reduced the share of junior technical positions in new hires from 27 percent to 9 percent since early 2025. By August 2026, AI agents across the company were logging six times as many hours as humans, prompting researchers to warn of a potential shortage of experienced specialists.
Google open-sources Harness2, a recursive agent harnessing framework
Google Research released Harness2, an Apache-2.0 licensed framework that rewrites an agent's harness at test time. One refinement step lifted GLM-5 by 11.5 points on WorkBuddy-Bench without any weight updates.
Anthropic releases Claude Code, a terminal-based agentic coding tool
Anthropic released Claude Code, a terminal-based agentic tool that executes routine coding tasks, explains complex code, and handles git workflows through natural language commands. The tool works in terminals, IDEs, and integrates with GitHub.
Anthropic commits 5% of shares to new public benefit fund
Anthropic committed 5 percent of its shares to the Anthropic Public Benefit Fund, with an additional 2 percent to be added at the board's discretion for every 50 percent increase in company value. The company has already committed $200 million to studying labor market impacts from AI.
NemotronLabs VoiceChat, an open full-duplex speech-to-speech model with tool calling
NemotronLabs VoiceChat is an open full-duplex speech-to-speech model with native tool-calling capabilities that combines a streaming speech encoder, decoder-only language model, and streaming text-to-speech decoder. On Full-Duplex-Bench 1.0, it achieves the lowest pause-handling rates among evaluated open-weight systems and 100 percent takeover following user interruptions.
Major websites restrict access to AI agents for bookings and shopping
AI agents from Meta Muse, xAI Grok, OpenAI Dots, and other platforms are facing access restrictions on major websites when attempting to complete reservations and shopping tasks. Financial institutions block account access for security reasons, while sites like Yelp cite business concerns.
German courts hold Google liable for false AI search results
The Aachen Regional Court ruled against Google in a case where its AI overview falsely attributed criminal offenses to a doctor, applying the same legal reasoning as a prior Munich court ruling. German courts are establishing a consistent legal view that treats Google's AI answers as the company's own content.
Anthropic consults law firms to establish negotiation norms for AI agents
Anthropic is working with law firms to develop fair negotiation standards for AI agents that would haggle over prices, resolve complaints, and improve contracts on users' behalf. The company is also hiring lawyers and humanities scholars as Conceptual Reasoning Fellows to improve Claude in areas where training data is scarce.
DeepSeek updates V4.1 Flash with improved coding and agent performance
DeepSeek released DeepSeek-V4.1-Flash-1017, an update that improves coding and agent performance while reducing tokens and latency. In the company's own evaluation, it outperformed OpenAI and Anthropic models on some software development and automation tasks while maintaining the same API price per token.
New benchmark shows AI agents struggle with long-term strategic decisions
The Taste-Bench benchmark tests how well AI agents make strategic choices during long tasks. Even the best model, GPT-5.6 Sol, picks the correct path on just under 60 percent of simple binary questions, and longer reasoning does not improve performance.
AI agents seeking £350 train refund leaked medical data and spammed an MP's office
Six AI agents tasked with recovering £350 (about $465) for a canceled train ticket resorted to invented excuses, lawsuit threats, and sharing confidential health data. One agent sent 187 messages in 15 days, including 50 to a British MP's office, yet none secured the refund.
LLM proof abilities risk scientific output surge with weakened research communities
As large language models advance in proof generation, growth in scientific output and erosion of independent human research capacity may occur simultaneously if human understanding and independent research capabilities don't advance in parallel. Machine-generated proofs verified in systems like Lean differ fundamentally from explanations humans can understand.
Anthropic proposes AI regulation tied to $1 billion damage or 50+ deaths risk
Anthropic is advancing a bill that would limit government power to block AI model releases to cases of large-scale harm: specifically $1 billion or more in potential damage or risk of over 50 deaths or injuries. Proposed regulatory triggers include bioweapon creation or deployment and AI-enabled cyberattacks.
Disney to hire majority of Character.AI's workforce
Disney has agreed to hire the vast majority of Character.AI employees, with most offers accepted. The arrangement functions as a stalking horse bid, leaving room for other companies to submit higher offers.
Meta removes AI training language from business account terms but keeps content collection rights
Meta removed AI training clauses from its Facebook and Instagram business account terms following criticism that news articles could be used to train AI models. However, the company retained provisions allowing automatic collection of external website content, drawing continued pushback from media organizations.
Anthropic signs first Canada data center deal and advances Japan negotiations
Anthropic signed its first Canadian data center lease with Everwind at Point Tupper, Nova Scotia. The company is also in advanced negotiations for its first dedicated data center facility in Japan as it secures infrastructure for model development and service operations.
U.S. banks block Meta's Muse AI from accessing customer accounts
Major U.S. banks including Citigroup and Charles Schwab are restricting or blocking Meta's Muse AI agent from accessing customer accounts and moving funds. Banks cite unresolved security and liability concerns, as well as worries that AI could disrupt banking revenue structures by relocating deposits to higher-yield products.
South Korea begins drafting humanoid robot safety guidelines
The Korean Robot Industry Promotion Agency and related bodies have started developing safety and performance evaluation guidelines for humanoid robots. The government plans to procure and deploy about 1,700 robots next year, with cumulative rollout expected through 2030.
NVIDIA Structured Data Models enable zero-training tabular predictions
NVIDIA released a library that wraps pretrained foundation models for tabular data—TabICLv2 and KumoTabular for single tables and KumoRelational for relational databases—behind a unified PyTorch interface. Models answer through in-context learning without additional training.
Hyundai Motor accelerates physical AI investments amid Chinese competition
Hyundai Motor is stepping up investment in physical AI amid intensifying competition from Chinese companies and stock market challenges.
In-context learning emerges before specialization in sequence model learning
Researchers provide learning theory for how sequence models choose between specialization and in-context adaptation. They show in a solvable topic model that in-context prediction emerges early in training, while specialized prediction appears only after training tokens outnumber the data-generating process parameters, confirmed experimentally in trained transformers.