Frontier 학습의 안전 제동, GLM-5.3의 오픈 웨이트 경쟁, 에이전트 실행 도구 확장

As models become more capable, the risks associated with developing and testing them internally also grow. We temporarily paused reinforcement learning (RL) training on our latest models intended for deployment for two weeks while we hardened and red-teamed our research environments and expanded monitoring coverage. Our largest planned frontier RL run remains on hold while smaller-scale training and evaluations validate these safeguards and establish more evidence of alignment.

We temporarily slowed some frontier training to strengthen security and monitoring. Our largest planned frontier RL run remains on hold while smaller-scale training and evaluations help us test safeguards and gather more evidence of alignment. I expect confidence in safety to increasingly set the pace of AI development. We urgently need tools for labs and countries to coordinate on this, which is why I signed Pacing the Frontier. In the meantime, we’re taking practical steps ourselves - and will continue to share what we learn as our approach evolves. https:// openai.com/index/pacing-m odel-development-cyber-capabilities/ … https:// pacingthefrontier.com

GLM-5.3 achieves 60 on the Artificial Analysis Intelligence Index, on par with Kimi K3 and up 7 points from GLM-5.2. Once the weights are released it will be tied as the leading open weights model @Zai_org has just launched GLM-5.3, which ties Kimi K3 (60) for the most intelligent open weights models on the Artificial Analysis Intelligence Index (assuming weights are released as per Z AI’s guidance). Between Kimi K3 and GLM-5.3, the open weights frontier is closer than ever to the proprietary frontier. GLM-5.3 keeps GLM-5.2's size at 753B total parameters and 40B active parameters. The Z AI team has shared that GLM-5.3’s weights are expected to be released within the next week. Key Takeaways: ➤ GLM-5.3 sees the largest gain in agentic capabilities, placing it amongst frontier models. The model's Elo rating in GDPval-AA v2, our real-world agentic knowledge work evaluation, rises from 1524 to 1770, a 246-point jump. That places GLM-5.3 second among all models, behind only Claude Opus 5 (1855), and surpassing the previous open weights leader on this evaluation, Kimi K3 (1668), by more than 100 points. ➤ GLM-5.3 is less token efficient than its predecessor. Across the Artificial Analysis Intelligence Index v4.1, GLM-5.3 uses roughly 18,700 output tokens per task, up about 20% from GLM-5.2 (15,700) and 27% more than Kimi K3 (14,700). This would have implications for cost of deployment. ➤ At $0.68 per Intelligence Index task, GLM-5.3 costs 1.5x GLM-5.2's $0.44, but is still cheaper than other models in the same intelligence tier. GLM-5.3 is 19% cheaper per task compared to Kimi K3 ($0.84) and 45% cheaper than GPT-5.6 Sol ($1.23). GLM-5.3’s increase in cost is partially driven by a 20% token usage increase compared to its predecessor. ➤ GLM-5.3 makes an improvement in real-world knowledge accuracy, scoring 14 on AA-Omniscience up from 4 for GLM-5.2. GLM-5.3 is now the second-best open weights model on AA-Omniscience, behind only Kimi K3 (20). A more detailed breakdown of where improvements are made reflects a genuine gain in accuracy as opposed simply more cautious abstention, as accuracy rate (24% to 34%) and attempt rate (46% to 55%) both increased. However, GLM-5.3’s hallucination rate regressed up slightly, from 26% to 30%. Additional model details: ➤ Size: 753B total parameters, 40B active (MoE), unchanged from GLM-5.2 ➤ Context window: 1M tokens ➤ Pricing: $1.40 per 1M input tokens and $4.40 per 1M output tokens. An 81% cache hit discount is applied ($0.26 per 1M cached input tokens) ➤ License: MIT ➤ Accessibility: GLM-5.3 is currently accessibly through Z AI’s first party API. The Z AI team has shared they expect to release the weights shortly Check out the full analysis of GLM-5.3 on Artificial Analysis: https:// artificialanalysis.ai

GLM-5.3 API is now live. - Built for coding, defensive cybersecurity, and long-horizon agentic tasks - Priced the same as GLM-5.2 - Available via the official API and partner model gateways Get started: https:// docs.z.ai/guides/llm/glm -5.3 …

Elon Musk

@elonmusk

Improvements to Grok @Bot

We temporarily slowed some frontier training to strengthen security and monitoring. Our largest planned frontier RL run remains on hold while smaller-scale training and evaluations help us test safeguards and gather more evidence of alignment. I expect confidence in safety to increasingly set the pace of AI development. We urgently need tools for labs and countries to coordinate on this, which is why I signed Pacing the Frontier. In the meantime, we’re taking practical steps ourselves - and will continue to share what we learn as our approach evolves. https:// openai.com/index/pacing-m odel-development-cyber-capabilities/ … https:// pacingthefrontier.com

Claude

@claudeai

Claude can now send emails in Gmail and manage files in Google Drive. Ask Claude to reply to a thread, and it drafts and sends the response. You control when it needs your approval. Connect Gmail or Google Drive from the connectors menu to try. Available on all paid plans.

Claude Code 2.1.235 has been released. 19 CLI changes Highlights: • Optional spellcheck underlines typos in prompts via local aspell/hunspell/ispell, improving prompt accuracy • SendMessage now rejects messages too large for cross-session delivery up front, preventing lost output • Dialogs show exact grant text and add 'withhold - don't ask again' when hidden, preventing accidental grants Full details are in thread ↓

Really interesting paper. I recommend it to anyone interested in training agents using existing harnesses. (bookmark it) ClawGym II runs RL through OpenClaw and Claude Code as opaque boxes. A serving proxy sits at the model boundary and captures every call the harness makes, then those calls get organized into prefix trees so PPO and GRPO can optimize over the recovered multi-turn structure. Qwen3-30A3B gains 9.98 points of Pass@1 through OpenClaw and 14.81 through Claude Code, stable across 200 to 400 optimization steps. Mix-harness training pushes further. One model gets optimized jointly by heterogeneous harnesses, which points at policies that generalize across execution systems instead of overfitting to a single one. Paper: https:// arxiv.org/abs/2608.16798 Track more trending AI papers in our academy: https:// academy.dair.ai

New Tuned Evaluators from @LangChain Versioned judges you just 'turn on' in a tracing project The 'Perceived Error' Judge flags conversations where the agent probably messed up (proxy for user feedback!) They used a specially trained judge model & cut eval cost by ~82%

Tesla

@Tesla

BAMF