Ep 229: LG just opened free API access to its 750B K-EXAONE 2.5 model for agentic workloads.
From Models & Agents — Video Edition · 13:58
Show by Patrick
Models & Agents LG just opened free API access to its 750B K-EXAONE 2.5 model for agentic workloads. What You Need to Know: LG AI Research released free hosted access to K-EXAONE 2.5, a 750B-parameter Mixture-of-Experts model focused on Korean and agent tasks. It improves on the prior version across software development, web browsing, SaaS automation, and agent problem-solving benchmarks. Builders working on Korean-language agents or tool-calling workflows can now test it without cost. Top Story LG AI Research opened free API access to K-EXAONE 2.5, a 750B-parameter Mixture-of-Experts model trained further from the August K-EXAONE 2.0 release. The model targets agentic workloads with strong Korean-language support and handles long context plus complex tool calls. Internal benchmarks show gains over the prior version on software development, web information retrieval, SaaS automation, and agent problem-solving tasks. Developers and startups can obtain an API key immediately through the EXAONE platform for service building and research. The release positions the model as a technology-validation step toward a frontier-level system planned for January. Source: mk.co.kr Model Updates Telecom Companies Give AI Agents More Responsibility: PYMNTS.com Telecom operators are assigning AI agents expanded roles in network operations, customer service, and predictive maintenance. The shift reflects growing trust in agent reliability for high-stakes infrastructure tasks. Source: pymnts.com Compression of Emotion in Language Model Representations: arXiv NLP Seven LLMs tested on the GoEmotions dataset show consistent over-compression of 27 fine-grained emotion categories into broader positive-negative splits in their hidden representations. Linear probes still recover fine-grained information despite the unsupervised geometry favoring coarser groupings. Source: arxiv.org Contextuality and Contradictions: Exploring Real-World Identifications of AI-Generated Writing: arXiv NLP Wikipedia editors and students identify AI text using diverse evidence types that prove contradictory and context-dependent rather than fixed textual markers. The study finds pragmatic infelicity, not specific features, often drives these judgments. Source: arxiv.org RL-ATTACK: A Generalized Black-Box Adversarial Attack on Language Models via Reinforcement Learning: arXiv NLP A reinforcement-learning attacker trained against target models achieves lower post-attack accuracy and better query efficiency than baselines on natural language inference and generative QA tasks. The approach learns tailored transformations directly from the target without external knowledge. Source: arxiv.org Guardrail Erosion in Domain Fine-Tuned Large Language Models: A Paired Audit of Knowledge-Centric Abstention in Biomedical LLMs: arXiv NLP A paired audit of 25 biomedical fine-tuned models against their base checkpoints found appropriate abstention rates on MedAbstain dropped from a median of 8.3% to 4.2% after fine-tuning. The change was heterogeneous across model pairs. Source: arxiv.org Agent & Tool Developments AI agent writing process: copy-paste when it fits: Simon Willison (AI builder) (X) Simon Willison reports selectively copying sentences generated by agents into his own writing when they match his intended meaning exactly. The practice treats agents as collaborative drafting tools rather than full replacements. Source: x.com Hiding AI PR slop with details/summary tags: Simon Willison (AI builder) Simon Willison now collapses lengthy AI-generated GitHub PR descriptions behind HTML details/summary tags to reduce reader fatigue while keeping the content available. GitHub Markdown supports the tags, making the pattern immediately usable. Source: simonwillison.net From Procedures to Permissions: Learning Precise Subagent Authorization from Agent Skills: arXiv NLP The method extracts authorization requirements directly from Agent Skills procedural descriptions to train subagent permission policies. On ClawArena-Team it raised tool permission precision by 13.4 points and workspace permission precision by 24.1 points for Qwen3.5-35B-A3B. Source: arxiv.org MARS: Training Multilingual Search Agents for Retriever-Free Evidence Acquisition: arXiv NLP MARS trains multilingual search agents that interact directly with raw corpora via shell commands instead of relying on learned retrievers. Corpus-native entity grounding and redundancy-aware rewards improve trajectory quality and cross-lingual generalization. Source: arxiv.org Practical & Community Twitter preview cards can't handle HTML tags in titles: Simon Willison (AI builder) (X) Simon Willison discovered that Twitter preview cards strip or break HTML tags placed in post titles. The finding affects anyone embedding formatted text in social sharing metadata. Source: x.com AI chatbots aren’t therapists. But can they finally get men to open up?: STAT The article examines whether AI chatbots can lower barriers for men…
More episodes of Models & Agents — Video Edition
More episodes
- Ep 236: Alibaba's Qwen4-Max packs 2.4T parameters into a 71B-active MoE with 1M context for…
- Ep 235: OpenAI’s new ChatGPT Wallet lets agents complete purchases with user-approved cards.
- Ep 234: Vercel’s UI SDK supplies agents with 50 behaviors for building accessible components…
- Ep 233: Gemini 4 Hydrogen ships at $1.50/$7.50 per million tokens and powers the free tier plus…
- Ep 232: Google's Gemini Decision 1 scores up to 1,000 multimodal questions in one 150-250 ms pass…
- Ep 231: A new 0.8B on-device model doubles agentic benchmark scores while keeping all data private.
- Ep 230: Anthropic cut Fable 5.2 prices 20 percent while raising its FrontierCode 2.0 score from…
- Ep 228: Ten open MoE models expose routing choices that swing sparsity from 4x to 128x and change…
- Ep 227: Ling-3.2-flash ships a 560B/25B multimodal MoE with 1M context and OpenAI-compatible APIs…
- Ep 226: Gemini Omni 1.2 Flash handles longer video clips on edge devices at lower cost than prior…
PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.