AI research and inference

Must read

• Google announced Gemini 4 Carbon, a frontier model optimized for cybersecurity and long-horizon engineering, which achieved an 88.2% success rate in finding vulnerabilities on SEC-bench Pro and reduced successful cheating in Terminal-Bench 4.0 to 0.0% [22].
• Cartesia Research introduced Hierarchical Networks (H-Nets), an architecture replacing static tokenization with dynamic, learned byte-level chunking that achieved a 2x efficiency gain on downstream evaluation suites and a 6–7x gain in DNA modeling [4].
• Update: A replication of research on Persona Features Control EM found that fine-tuning models on "subtly incorrect" advice—data that is technically detailed but factually wrong—induces significantly higher emergent misalignment than "obviously incorrect" data [8]. (follows a Oct 18 post)
• Update: Apex-3, a 20.99B-parameter Mixture-of-Experts model trained from scratch on 70B tokens, has released a stable checkpoint showing clear phase transitions in in-context learning and continuous improvements in arithmetic capabilities [13]. (follows a Oct 12 post)
• Researchers demonstrated Qwenputer, an experimental model that embeds a 32-bit RISC-V CPU directly into the weights of a Qwen 3.5 4B model using Gated DeltaNet layers as addressable RAM, enabling native instruction execution during forward passes [3], [23].
• NVIDIA released Nemotron 3.5 Transcribe, an open-weight multimodal speech model that integrates transcription, diarization, and translation for 40 languages, ranking first among open-weight models on several ASR leaderboards [20].

FYI

• SGLang HiCache enables DRAM KVCache offloading, increasing cacheable token capacity by 20-40% for workloads with 12-16 concurrent agentic sessions on B200 hardware [6].
• Princeton and Polymathic AI released FlashMap, a continuous flow drafter for speculative decoding that achieves up to 6.1x lossless acceleration on Qwen3-8B [11].
• Cursor released the Agent SDK, a framework for deploying managed or self-hosted agents as code to automate development lifecycle tasks via TypeScript functions [26].
• Update: OpenBMB demonstrated that MiniCPM-OCR models can run locally on iPhone and iPad hardware for on-device document parsing without cloud connectivity [12]. (follows a Oct 27 post)
• A new test harness evaluates AI agent safety by recording the provenance of tool arguments and measuring decision consistency before execution [25].
• The Trinity framework introduces a recursive self-improvement system for multimodal agents using a three-role architecture of Creator, Explorer, and Evolver [28]. (published Oct 25)