Only Mon, September 21, 2026 · All dates →
jevals: replacing LLM judges with typed Jev decisions for agent evalsjevals is an open-source tool that replaces LLM judges with typed Jev decision models for agent evals and guardrails. All evals for a trace are sent as one HTTP request costing a few thousandths of a cent and returning in a few hundred milliseconds, suitable for every trace or inside the agent loop. It supports Jev, Kev, Laya, and regular chat LLMs.
Google Releases Open Agentic Orchestrator AXGoogle introduced AX, an open agentic orchestrator that runs declarative YAML-defined tasks at scale, sandboxing each agent with workspace setup and network fencing. It offers primitives to create, suspend, resume, and delete tasks, targeting agent workloads that accumulate state and call model APIs and tool servers.
AI's Wobbly House of Cards Puts Markets at RiskA Bloomberg column warns that trillions in AI-driven investment gains and market value rest on the assumption of continued rapid growth, and that a slowdown could trigger hard-to-unwind shocks for Wall Street and the wider U.S. economy.
Would a $1k Qwen3.8-27B Dedicated Chip Have a Market?A LocalLLaMA discussion asks whether users would buy a consumer-oriented Taalas chip running a fixed Qwen3.8-27B at 7,000 TPS for $1,000. The poster argues such chips resemble game discs: the model is fixed but runs far faster than GPUs and won't become worthless when newer models launch.
Autolith: a self-rewriting Common Lisp agentAutolith is a self-modifying terminal programming agent running in a single Common Lisp image that can inspect, edit, break and recover its own implementation at runtime, maintaining states such as MCP servers, memories, tools and LSP servers, with file search powered by the Rust library fff and support for Linux, macOS, BSD and Windows.
Swiss-South African lab open-sources 27B creative-writing model Hemmingway-1A small Swiss-South African lab open-sourced Hemmingway-1, a 27B creative-writing fine-tune of Qwen3.8-27B, under Apache-2.0 with bf16 weights (~54.7 GB) on HuggingFace and vLLM compatibility. It scores 1330 on EQ-Bench 4 and beats tested frontier models on the lab's internal blind writing benchmarks for communication and human-likeness, while math, code, and factual recall are unchanged from the base model, making it a deliberate specialist.
Qwen 27B runs autonomously for 3 weeks on one 3090A user ran a quantized Qwen 27B locally on a single RTX 3090 for about 21 days with the goal of having the agent build its own CUDA inference engine. The run consumed roughly 230M tokens, spent about 83 hours in 699 compactions, and reached about half of llama.cpp's prefill speed while producing working kernels and benchmarks.
Swedish Startup's Drone Uses Jetson Orin Nano to Pick TargetsA Swedish startup's attack drones run a small AI model on Nvidia's Jetson Orin Nano to independently select and bomb targets, requiring no human input and zero external communications.
Guardian: mathematicians still vital, but AI firms won't see itA Guardian editorial criticises OpenAI's claim that its AI agents solved the Navier-Stokes problem, arguing LLM results rely on prior human mathematical work without adequate credit or compensation. Mathematician Tristan Buckmaster raised concerns about OpenAI's use of his Codex data; OpenAI denied directly accessing it but could not rule out the data helping improve its model.
Bessent calls meeting with China's He Lifeng successfulUS Treasury Secretary Bessent met Chinese Vice Premier He Lifeng ahead of a Trump-Xi summit and called the meeting "successful." The report did not disclose specific details or outcomes of the talks.
Competition for small neural networks playing strategy gamesA competition for small neural networks that play strategy games has drawn discussion on Hacker News. It focuses on training compact models to handle strategy gameplay under constrained compute, with the discussion linked in the HN thread.
High schooler open-sources C++ tensor library and autograd projectA third-year high school student built a simple C++ tensor library and autograd implementation over a few weeks to learn machine learning basics, open-sourced it on GitHub, and is asking the ML and C++ communities for advice and constructive criticism.
Decontamination reports can't fix benchmark contamination, evaluator must own the testA write-up argues lab-run decontamination reports cannot fix benchmark contamination: labs check themselves, corpora contain copyrighted works that can't be disclosed, and paraphrase, forum, GitHub, or synthetic-derived data evade matching. It proposes evaluators own the test, submissions never see labels, evaluation runs offline and is rebuilt from a named commit, and where possible test data is generated after submissions freeze, with only reproduced results counting.
The A.I. Daily