Technical AI news, automatically curated and generated
A 12-year-old developer built KODA, an AI coding ecosystem optimized for low-resource devices, running on a $150 Poco C55 phone with 4GB RAM. The system comprises a 92KB single-file web app, a 22KB code editor extension …
→ View original sourceThe paper introduces Dust, a zeroth‑order optimization algorithm that pretrains transformer language models without backpropagation. Dust adds independent Gaussian noise to the activation output of each linear layer at e…
→ View original sourcePewDiePie fine‑tuned a local model named Ajax using data generated via OpenAI's API, which violated OpenAI's terms of service and resulted in his account being banned. After an appeal restored access, he resumed harvesti…
→ View original source
A new controlled version of the CivBench benchmark evaluates large language models' ability to play full games of Civilization V. GLM‑5.3 outperforms Opus‑5.5, while Qwen‑3.8‑27B shows surprisingly strong performance. Th…
→ View original sourceChatGPT’s image‑generation mode has been producing New Yorker‑style single‑panel cartoons that routinely include the authentic pen names of the magazine’s cartoonists, even though those artists never created or signed th…
→ View original source
The paper introduces activation alignment, a technique that trains a lightweight linear map on synthetic unlabeled data to align the intermediate activations of a context‑constrained student model with those of a full‑co…
→ View original source
The study exploits the iterative denoising process of masked diffusion language models (dLLMs), which expose token‑level answer distributions at every step before a token is committed, unlike autoregressive models that r…
→ View original source
UndoBench introduces a benchmark that isolates task competence from recovery capability in tool‑using AI agents by constructing 36 base workflows paired with 36 fault scenarios spanning eight enterprise domains. The desi…
→ View original source
llama.cpp version 0.6.0 has been released, introducing MTP (mid‑tower prediction) speculative decoding for the Qwen4Exp model alongside various other enhancements. The update improves inference speed and efficiency while…
→ View original source
Blockway's Agens Volundr 32B Preview is a 32‑billion‑parameter model built on a proprietary hybrid architecture in which only 18 of the 72 transformer layers maintain a key‑value cache, reducing memory footprint for long…
→ View original source
WHIRL v0.1.3 is an open‑source (Apache‑2.0) native Windows inference engine for the AMD Radeon AI PRO R9700 (RDNA 4, 32 GB) implemented in pure C++/HIP, requiring only the AMD driver. It outperforms llama.cpp b11214 on i…
→ View original sourcePrompt Optimizer is a cross‑platform AI‑prompt engineering tool that lets users create, test, and refine prompts for text and image generation models. It can be run as a web app, desktop application, Chrome extension, or…
→ View original source