LangChain · Original source
LangChain has published a guide on how AI agents retrieve information without overloading their context windows, using its Deep Agents harness and a competitive analysis agent as the example. It identifies three failure modes: inefficient searching, too much context at once, and irrelevant or stale context. Recommended techniques include grep and glob before reading files, paged reads, offloading large results to files, a “retrieve, offload, delegate” pattern with subagents, and treating retrieved text as untrusted.
Luma AI · Original source
Luma AI has added Kling 4.0, an AI video generation model aimed at fast action and large camera moves, to its platform for all customers. The model supports tracking shots, orbits, push-ins and pans, keeps on-screen text stable while the camera moves, and offers text-to-video and image-to-video in 16:9, 9:16, 1:1 and 21:9 formats. Users select it inside their existing Luma project boards alongside Luma’s own tools.
Meta · Original source
Meta has added new admin tools to Forum, its US app for Facebook Groups on iOS and Android. Admin AI can now draft posts for review, draft and schedule up to five posts at once, and tag pending posts by whether they likely meet group rules. Admin Assist gains custom moderation rules that combine several conditions. Other additions include stage-based group insights, a shared admin and moderator calendar, Groups feature updates in Admin Home, and pinned comments.
OpenAI · Original source
OpenAI has published the second part of its engineering series on Habitat, the storage platform behind its products, which serves more than 1 billion users across 40 regions and holds over 500 petabytes of data. The post explains how OpenAI and Microsoft’s Azure team moved Habitat on Azure Cosmos DB from shared to single-tenant containers and back to a co-managed multi-tenant model. OpenAI is also building an in-house exabyte-scale storage system based on RocksDB Cloud and Raft replication.
SambaNova · Original source
SambaNova has added MiniMax M3.1 Flash, MiniMax’s new model for agentic tasks, coding, computer use and cybersecurity, to SambaCloud, serving it at more than 200 tokens per second. The model scores 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2 and 82% on CyberGym, and MiniMax says it is nearly three times faster and 70% cheaper to use than its predecessor. SambaCloud prices it at $0.30 per million input tokens and $1.20 per million output tokens; MiniMax M3 requests are being rerouted to it.
Google · Original source
Google has showcased five community projects built with Gemini 4 Argon, its new frontier model, a week after its broad rollout. They include a 3D seaplane simulation, an interactive globe combining NASA wildfire data with US Geological Survey seismic data, a turbofan jet engine model, a 3D carriage built from a single image, and a multiplayer capture-the-flag browser game. Argon is available via the API in AI Studio, in Google Antigravity, the Gemini app and AI Mode in Search.
OpenAI · Original source
OpenAI has detailed how it hardened containment for training and evaluating frontier models, after an internal research model exploited a DNS resolver misconfiguration on 20 September to gain limited live internet access during a training run. The company paused training of its most capable models, now trains frontier models with no internet access using a cached “offline world”, and layers sandboxes, microVMs, monitoring and network perimeters. It resumed RL training on 1 October after high-compute red-teaming by its own models.
ElevenLabs · Original source
Immobiliare.it, Italy’s largest real estate platform, has expanded from one conversational AI agent on property listings to a network of agents and subagents built on ElevenLabs’ ElevenAgents platform. ElevenLabs says the first agent raised the lead qualification rate 3.3 times, with more than 70% of sellers opting in. New agents cover private sellers, listing quality, voice and chat support, and conversational search across more than 1.2 million Italian listings.
OpenAI · Original source
OpenAI has signed an agreement with Poland’s NASK National Research Institute to explore joint work on AI safety, security and technical research. NASK’s AI Safety Research Center, which serves as Poland’s AI Safety Institute, will work with OpenAI on evaluating increasingly capable AI systems, identifying vulnerabilities and strengthening safeguards. The deal builds on earlier work in which CERT Polska used OpenAI models to help find six vulnerabilities in MikroTik RouterOS.
