Tech AI News - 테카이

2026.10.30 AI News Briefing — Kakao Open-Sources Embeddings, OpenAI Reshapes Pricing Plans

· 4 min read

This post was translated from the Korean original by AI.한국어 원문 읽기 →

Kakao open-sourced embedding models for personalized search, while Google DeepMind unveiled 'Genie 4,' a world model that generates sound in real time. OpenAI is cutting ChatGPT Pro usage limits in half, and Anthropic revealed in its IPO filing why it paused model training for over a month.


1. Kakao open-sources two personalized AI embedding models, 'Kanana-2'

  • Source: Asia Today | 2026.10.29
  • Key details: Kakao released two embedding models on Hugging Face: the 3-billion-parameter 'Kanana-2-3b-embedding' and the 1.3-billion-parameter 'Kanana-2-1.3b-embedding.' Embedding models convert the meaning of sentences or past conversations into numerical vectors to calculate similarity, forming the foundation for personalized AI search and recommendations. The models were presented as usable for queries that reference earlier conversations, such as "that quiet restaurant I mentioned before."
  • Why it matters: This reads as a push to grow the developer ecosystem by open-sourcing the embedding infrastructure behind search and recommendations, rather than the large language model itself.

2. Google DeepMind unveils 'Genie 4,' a world model that generates sound in real time

  • Source: OfficeChai | 2026.10.27
  • Key details: Google DeepMind unveiled 'Genie 4' as a research preview — a world model (AI that simulates real-world physics and space to generate virtual environments) that can create explorable virtual environments from text or a reference image. Scenes change based on user actions, and ambient sound, music, and dialogue are generated in real time alongside them; Google said it's the company's first world model to generate audio together with visuals. Physics simulation and human rendering have also improved, but the generated worlds are still at an early stage, staying consistent for only a few minutes, and are currently available only to a small group of trusted testers.
  • Why it matters: It shows the 'world model' race — training robots and AI agents in virtual environments instead of the real world — now extending into audio as well.

3. OpenAI halves ChatGPT Pro usage limits starting October 30

  • Source: OfficeChai | 2026.10.29
  • Key details: The $200-a-month ChatGPT Pro plan's usage — including ChatGPT Work and Codex — drops from 20x the Plus plan to 10x, and the weekly message limit for GPT Pro chat falls from 200 to 100. Pricing stays the same, and existing subscribers could keep the old limits through October 29. At the same time, OpenAI introduced a new $500-a-month 'Pro 500' plan offering 25x the Plus plan's limits.
  • Why it matters: It's an example of inference-cost pressure reshaping subscription plan structures, highlighting the profitability strain on premium AI subscription services.

4. Anthropic's IPO filing discloses it paused pretraining of its next model for over a month

  • Source: Let's Data Science | 2026.10.26
  • Key details: Anthropic filed an S-1 with the US Securities and Exchange Commission (SEC) to pursue a Nasdaq listing under the ticker 'ANT.' The filing disclosed that pretraining of its next frontier model was paused for over a month to strengthen security, including preventing model weight leaks. The 94-page risk factors section named cyberattacks and the risk of misuse for bioweapon design, among others, and the US Federal Trade Commission (FTC) had previously sent a consumer-protection-related inquiry.
  • Why it matters: This is a case of an AI company disclosing model security issues directly to investors as a filing risk ahead of an IPO, formally signaling to the market that model security is tied directly to business continuity.

5. Ministry of Science and ICT reports on three AI megaprojects — data centers, physical AI, and K-semiconductors — for H2

  • Source: Daily Ian | 2026.10.16
  • Key details: The Ministry of Science and ICT reported its second-half plans, saying it will focus on three megaprojects: AI data centers, physical AI (AI that interacts with the physical environment, such as robots and autonomous vehicles), and K-semiconductors. It also set a goal of launching an AI service usable by every citizen within the year to accelerate realizing an 'AI-based society.'
  • Why it matters: The government's direction on AI infrastructure and semiconductor investment serves as a benchmark for gauging where private investment and policy benefits will flow.

Today's insight

  • Competition over foundational infrastructure — embeddings and world models — stands out as much as competition over model performance. Kakao open-sourcing its embeddings and Google previewing its world model can both be read as attempts to claim the groundwork for next-generation agents and search.
  • It's also notable that AI companies heading toward an IPO are formally listing security and safety issues as risk factors. The tension between growth and safety is being passed directly on to investors.

References

  • #AI News
  • #Daily
  • #Kakao
  • #Google DeepMind
  • #OpenAI
  • #Anthropic