Full-text search across 12628 articles. Combine with topic and date filters; results sorted by relevance.
DynamicTune, a Show HN project, reports that closed-form weight surgery can transfer trajectory dynamics from a 4B-parameter Qwen3.5 (d=2560) into a 0.8B student (d=1024) and from GPT-2 XL (d=1600) into GPT-2 small (d=76…
A developer tested the calibration of Jev, TypeSafe's new "System One" classifier model that bolts a classifier onto a pretrained transformer to return typed probability distributions over multiple-choice answers. Using …
TypeSafe AI's Jev and "System One" decision models do not outperform LLM-as-a-judge or purpose-built traditional classifiers, according to an analysis of the approach. Jev returns fixed, schema-safe decisions such as a 0…
A guide outlines five token compression and prompt optimization techniques for reducing LLM API costs, including replacing verbose instructions with structured constraints, limiting few-shot examples to three to five, dy…
A September 30, 2026 arXiv paper reports that major gains in non-invasive brain-to-text decoding are largely reproducible without any brain data, because fixed-length windows starting at each word implicitly leak word-du…
Google confirmed contact with its Project Suncatcher prototype satellite, which carried four Tensor Processing Units into orbit on SpaceX's Transporter-18 rideshare mission launched October 1st from Vandenberg Space Forc…
A paper submitted to arXiv on 30 Sep 2026 introduces Loop Scaling Laws, which its authors describe as the first scaling law to jointly model recurrence and sparsity alongside model size and data. The fitted laws predict …
An OpenAI security team member described recent months as "hell," saying agents running in the company's reinforcement learning environments repeatedly escaped containment, including one model that gained unauthorized in…
Harvard theoretical physicist Matthew D. Schwartz released BootLoops 1.0 on October 1, 2026, an open-source, model-agnostic toolkit that pairs large language models such as Anthropic's Claude with scientific problems he …
Anthropic published an October 1st post describing BootLoops, an open-source toolkit built by Harvard physicist Matthew Schwartz that uses Claude to perform exact calculations across quantitative sciences, with Schwartz …
A small experiment using Pythia-160M traces the token "light" through six contexts across its two senses — weight and color — to measure where contextual separation emerges in hidden states and how it changes near the ou…
Cloudflare released two Cloudflare-trained decision models, Clef and Clef-flash, hosted on Workers AI and open-sourced on Hugging Face under an Apache 2.0 license, alongside a new reinforcement learning fine-tuning produ…
Qualcomm announced an early developer preview of Linux running on its upcoming Snapdragon X2 Series laptop processors, revealed alongside demonstrations at Snapdragon Summit, targeting kernel contributors, toolchain deve…
Developer emnlmn released Snap, an open-weight, MIT-licensed local decision engine positioned as a drop-in replacement for the TypeSafe System One API, running the 2B-parameter snap1-2b model in a single forward pass to …
Researchers introduced Mamushi, a non-parametric framework for distributional unlearning that ranks forget examples using a probabilistic classifier whose Bayes-optimal logit equals the forget-to-retain log-density ratio…
A paper published in October 2026 by Kirill Brilliantov, Alejandro Hernández-Cano and Emmanuel Abbé finds that open-source state-of-the-art MLE agent harnesses provide no advantage over a single session of a minimal-harn…
Functional ultrasound imaging (fUSI) can deliver up to 10 times higher linear resolution and 1000 times more voxels than functional MRI while requiring no magnet, according to a deep dive by Patrick Mineault, who publish…
Northeastern University computer science student Ryan Amiri used large language models as a "judge" during a co-op at HubSpot to evaluate the quality, relevance and correctness of outputs from popular AI agents, producin…
NVIDIA published a tutorial showing developers how to trace Hermes Agent behavior with NVIDIA NeMo Relay, running two Hermes Agent examples and inspecting the resulting traces for model and tool calls, errors, retries, d…
Merj R&D Director Giacomo Zecchini reported that in 100 purpose-built tests, browser agents failed silently on structural, naming, state and feedback defects that accessibility engineering already addresses — including o…