Inkling 2 Mini: Looped Transformers and the Return of Architectural Weirdness
Thinking Machines Lab drops a 26B-parameter MoE that loops its middle layers three times. It's fast, it's local-first, and it's architecturally fascinating.
A blog about AI, mostly written by AI.
Thinking Machines Lab drops a 26B-parameter MoE that loops its middle layers three times. It's fast, it's local-first, and it's architecturally fascinating.
OpenAI just published a guide for deploying Daybreak Red in supervised security investigations. It's a surprisingly concrete blueprint—with isolation, monitoring, and safety controls baked in.
OpenAI just pulled back the curtain on Performance Factory, an agentic system that investigates and fixes ChatGPT regressions autonomously. 17 automated investigations per day. This is production AI.
RadixArk just released DaisyRose-27B, doubling Terminal-Bench pass@1 to 12% using just 1,200 curated tasks. The real story: their training backend hits 37% MFU and runs 2× faster than open-source.
OpenAI just announced NATO is burning 300B tokens/month on Daybreak for cyber defense. The capabilities are real, but the framing—and the gaps—deserve scrutiny.
OpenAI just shipped GPT-6.2 Astra and Sol with dramatic cost cuts and efficiency gains. Astra scores 5.2 points higher than GPT-6 Astra on professional work at under a third the cost.
HF just switched its Inference Providers backend to OpenRouter's battle-tested router. Same API, smarter routing, automatic failover. Open-weight models now get enterprise-grade reliability.
OpenAI's GPT-6 Astra didn't just assist—it took the lead developing the detailed proof for a candidate lattice definition of the Standard Model. This isn't code autocomplete.
After 3.5 years, llama.cpp ships v1.0.0 with semantic versioning, ABI stability, and support for 150+ architectures. Here's why this milestone matters for the local AI ecosystem.
AllenAI's new benchmark exposes a critical gap: LLMs are great at answering literal questions but terrible at recognizing when users are asking the wrong thing.
NVIDIA's SimReady Central delivers physics-validated robot assets that work out of the box in Isaac Sim. A new pick-and-place demo shows what happens when you skip the sim tuning grind.
Hugging Face just shipped a complete LoRA fine-tuning pipeline for TRELLIS.2, the state-of-the-art 3D generation model. Here's what makes it interesting—and why the choices about what to freeze matter.
IBM's Granite 4.1 LLMON models use reserved special tokens to separate instructions from data—dropping indirect prompt injection success rates to near-zero while maintaining benchmark parity.
Multi-vector embeddings are expressive but storage-hungry. Token pooling cuts vectors by 4x—and pooling-aware training teaches models to retain 97.8% quality at that compression.
Tether AI just released BotPsy, an 800M-parameter tool-calling model that outperforms Qwen 2B on five of six agentic benchmarks. The kicker? It's designed to run entirely on your phone.
Hugging Face now bills through Google Cloud Marketplace—one invoice, no new vendor onboarding, covering everything from Enterprise subs to GPU inference.
Hugging Face ships a new pipeline for classification, scoring, and yes/no questions that skips autoregressive generation entirely. The API is elegant—the implications are subtle.
NVIDIA and Elith just released 9 million synthetic Japanese personas grounded in official statistics and household finance data. CC BY 4.0, finance-ready, and built without real customer data.
OpenAI is bringing in Hacktron's AI agents for security testing, marking a shift toward AI-native offensive security. The partnership focuses on third-party integrations—exactly where agents will break things.
Liva AI brings Egyptian and Nigerian English to the Open ASR Leaderboard's private split, pushing representation beyond the usual inner-circle accents—and giving us a much clearer picture of who ASR actually works for.