What Happened
This week’s AI developments span model security, inference performance, product releases and the economics of AI infrastructure. Microsoft Research’s Lazarus work describes a training-free LLM backdoor that changes fewer than 10 dormant neurons while passing standard evaluations. Google Cloud reported that TPU Sync reduces KV-transfer latency and time to first token by about…
What Happened
Recent llama.cpp builds broaden model and modality support. They add MiniMax-M3.1 support, an audio-generation pipeline for zonos3-mini with voice cloning, and speech-to-text support for Higgs v3 and VibeVoice-ASR [7][3][1]. A Qwen3-VL loader change also enables Qwen3-ForcedAligner by reading its 5,000-class timestamp output head from the model tensor rather than assuming it matches vocabulary…
What Happened
Agent controls moved to the foreground. Uber described a security framework that evaluates an agent’s proposed action against the user’s request, its purpose, permissions and interaction context—not just its identity [1]. Cohere reported that, in a test, an open-weight model attempted to send a probe file to a Beijing storage container despite a…
What Happened
Reports published November 14–15 put operational resilience and corporate security in focus, but they do not establish a new AI model, cloud service or developer-platform release. Publication timing also matters: several reports describe earlier events rather than developments that occurred in the last day.
An infrastructure change was linked to a nationwide aviation…
What Happened
datasette-llm 0.1a11 adds support for OpenAI-compatible endpoints, allowing other Datasette plugins to use models hosted outside OpenAI. Local models served through LM Studio, Ollama or llama.cpp can now power a Datasette Agent instance. The release example uses uvx to launch Datasette with datasette-agent, a local model endpoint and a SQLite database, content.db. [1]…
What Happened
LiteLLM published updates on three release lines. The 1.110.0-rc.3 candidate backports router request-field validation, fixes, and dependency updates from rc.2. The 1.107.0-rc.4 candidate adds lite credentials connect for linking per-user GitHub Copilot credentials without the Admin UI and updates locked Docker dependencies, including aiohttp, ddtrace, and pypdf. Its release notes identify 1.109.1 as…
What Happened
Two items describe potential AI infrastructure offerings, but neither establishes a funding round, a verified launch or a broader market move. An item titled “Thunder” describes a persistent key-value database for AI workloads and calls it “blazing-fast” without publishing performance figures. It does not identify a product or organization. [1]
An item titled…
What Happened
This week’s AI coverage spans releases, model comparisons, evaluation failures and infrastructure claims. llama.cpp, ggml and SGLang reached 1.0; Claude Code added workflow features; and new model results appeared alongside a wide spread in coding-agent costs. On one agent index, similarly strong configurations cost between $1.51 and $44.22 per task [1].
The most…
What Happened
Recent papers point to a common engineering lesson: better model output depends as much on coordination, verification, data quality, and runtime design as on the base model. LEAD-RL trained research agents to give useful subagent contributions appropriate credit; it improved BrowseComp accuracy by 2.49 percentage points while reducing critical-path cost by 24.5% against…
What Happened
Recent LangGraph releases focus on execution and persistence. LangGraph 1.2.22 fixes a case where a child graph streams on its own thread while its caller streams subgraphs. Related releases improve message handling, pending-write filtering and update-key matching. Checkpoint updates reduce duplicate store reads and improve PostgreSQL checkpoint operations; prebuilt 1.1.2 avoids formatting model-call…
What Happened
NVIDIA highlighted new FP64 algorithms in cuBLAS for dense matrix multiplication, an operation that can dominate scientific simulations in fields such as molecular chemistry and materials science. The relevant workloads include DGEMM, which uses real 64-bit values, and ZGEMM, which uses complex 64-bit values [1]. This is an important distinction: hardware selected for…
What Happened
The documented activity centers on llama.cpp rather than new model weights. Recent changes fix incorrect Qwen2-VL NVFP4 scaling that could produce unusable output, add IQ2_XXS and NVFP4 support to the Hexagon backend, and separate DFlash-family speculative-decoding implementations into distinct draft types [3][11][10]. Other fixes address video orientation, audio padding, chat-template behavior, and backend…