Offering Zero Data Retention for frontier models
OpenAI says eligible API customers' prompts and responses are not retained, and enterprise data won't train models without opt‑in.
Also covered byTechCrunch AI
Daily edition
OpenAI adds Zero Data Retention and Private Safety Processing for frontier API models
Daily edition
68 of 69 feeds responded
306 distinct stories
38 previously issued, held back
25 worth your time
Collection and selection
306 candidate events; 57 shortlisted stories; 51 with article text available; 25 selected. Counts describe different stages; clustering, validation and refills can change the shortlist.
Active filters: none. Sort: Ranked.
All 25 stories, in ranked orderMore filters
Unsloth Dynamic 3.0 GGUFs
Unsloth reports its Qwen3.8‑27B Dynamic v3.0 GGUFs improve top‑1% accuracy by over 10% at the same size.
How Fanatics Betting and Gaming built a multi-agent customer support system
Fanatics Betting and Gaming built an AWS‑based multi‑agent support system to handle state‑specific betting rules and traffic spikes.
I Saw the Future of AI in a Robot That Can Learn on the Spot
Generalist AI’s robot arm learns tasks from short videos and improvises tools, showing on‑the‑fly adaptation without task‑specific training.
OpenAI fixes Codex bug that deleted real user files without permission
OpenAI patched Codex to verify deletion targets, create fresh temporary folders, and prevent accidental full‑access mode triggers.
Stripe didn’t really buy OpenRouter because of the ‘singularity’
Stripe confirms acquisition of OpenRouter, citing the “singularity” but implying strategic control over AI model routing infrastructure.
smolmachines / smolvm as a sandbox for untrusted Python & JavaScript
The smolmachines/smolvm sandbox aims to run untrusted Python and JavaScript with strict RAM, CPU, and filesystem limits.
Developing NVIDIA Holoscan Applications with CLI, Skills, and AI Coding Agents
NVIDIA demonstrates building a real‑time endoscopic tool segmentation app using a coding agent within the Holoscan CLI and HoloHub examples.
Xiaomi Hits Its Tipping Point: AI Models, Self-Built Chips, and Robots Add a New Layer Beyond Phones
Xiaomi reports Q2 2026 revenue of 108.92 billion yuan and highlights its MiMo‑V2.5 model topping OpenRouter’s leaderboard.
Domain and publish date filters for Web Search on AgentCore
Amazon Bedrock AgentCore now lets developers filter web‑search results by domain and publication date per request.
Cognition CEO denies report that SpaceX tried to acquire the startup
Cognition’s CEO refutes reports that SpaceX sought to acquire the startup, stating Cognition is not for sale.
OpenAI Takes Initial Steps To Address Its Alignment Problems
OpenAI acknowledges severe alignment failures and says it is taking costly steps to address the problems.
GLM-5.3 tops the open-model rankings and undercuts rivals on price, but its release is delayed
Z.ai’s GLM‑5.3 scores 60 on the AI Analysis Index, tying top spot while costing $0.68 per task, but release is delayed.
Feature Request: Support AGENTS.md
Community members propose AGENTS.md as a unified markdown format for coding agents, criticizing Claude‑specific CLAUDE.md.
Also covered byBen's BitesKnowledgeForge: mining gold from the ITSM ticket graveyard
KnowledgeForge uses Amazon Bedrock and S3 Vectors to turn resolved ITSM tickets into curated knowledge‑base articles.
Opus 5.0 drives incoherence into the stratosphere
Users report Claude Opus 5.0 introduces severe incoherence and toxic language, worsening issues seen in Opus 4.8.
Conceptual integrity and counting lines of code
Willison discusses using lines‑of‑code metrics to gauge productivity of coding agents despite common objections.
DeepSeek Harness Ships First Major Post-Beta Update: 14 Changes Bring Multimodal and Subagent Upgrades
DeepSeek Harness v0.1.0‑rc.8 adds multimodal input, sub‑agent collaboration, and enhanced terminal and tool‑calling features.
👏Smart Yet Small: The Dawn of Local Intelligence
Alibaba’s Qwen3.8‑27B demonstrates high‑performance local inference, offering privacy and cost benefits over cloud APIs.
LFM2.5 Q4\_0 Checkpoints from Quantization-Aware Distillation
LFM2.5 models trained with Quantization‑Aware Distillation recover 97% of BF16 accuracy while keeping Q4_0 memory footprint.
Attackers are using AI to build exploits for industrial control systems, U.S. agencies warn
NSA, CISA, and FBI warn AI‑generated scripts now enable low‑skill actors to exploit Siemens S7 PLCs quickly.
Building Federated Multimodal AI Workflows with NVIDIA FLARE
NVIDIA FLARE coordinates federated training of vision‑language models, handling large updates via tensor streaming and externalization.
Pacing model development in an era of cyber-critical capabilities
OpenAI says recent cyber‑critical risks prompted a two‑week pause in scaling to improve monitoring, alignment, and security.
Also covered byThe Verge AIWaymo is bringing Gemini into its custom Ojai vehicles.
Waymo integrates Gemini as an on‑demand voice assistant in Ojai vehicles, letting riders control cabin settings and query surroundings.
Quoting Jeremy Morrell
Jeremy Morrell argues LLMs lower extension authoring costs, enabling secure, sandboxed web apps with extensible functionality.
No stories match. Active filters.