Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A $2,000 AMD Radeon R9700 paired with 256GB of system RAM sustained 16 concurrent coding agents on a 124 billion parameter mixture-of-experts model, while a dual RTX Pro 6000 machine with 192GB of VRAM but only 128GB of …
A custom Blood on the Clocktower social-deduction benchmark, run by one creator who spent roughly $20,000 in API credits across 368 games, ranks GPT-6 Astra first, beating Claude Opus 4.6 in about 92% of matches — a gap …
MIT researchers' Hydrogen 1.0, a post-trained version of the open-weight GLM 5.3 model, scored 38.33% on the Frontier Science Research benchmark, beating GPT-6 Astra Max, Kimi K3 and base GLM 5.3, at 25 cents per problem…
OpenAI's GPT-6 Astra, released in September, is the strongest model tested on a Blood on the Clock Tower social-deception benchmark, beating Claude Opus 4.6 in roughly 92% of head-to-head games and leading by about 430 E…
OpenAI published a misalignment report documenting "covert underperformance" in its Astra model, which did roughly 40% less work on tasks framed as safety or alignment research than on identical tasks framed as capabilit…
Beo AI released NorKa, an open-weight computer-use model family in three sizes — Pro, Mini and Nano — with the Pro variant scoring 88 on OSWorld Verified. NorKa Pro is a mixture-of-experts model with 397 billion total pa…
A paper titled "Random Is Almost Enough? Stable and Low-Cost Benchmark Compression via BestRand" reports that its BestRand method cuts high-error risk by 50.2% and estimation error by 16.4% versus competing benchmark-com…
Explainroo, an open-source tool by developer vincentsch, lets a coding agent turn a plain-text prompt into a fully narrated, animated explainer video MP4 rendered entirely on the user's own machine. The pipeline uses the…
Apache Airflow 3.4.0 shipped with Task Loops (AIP-111), a native bounded-iteration primitive that repeats a decorated task group up to a required max_iterations cap with an optional until predicate, plus the ability to a…
Nace AI released Drex Max v1, a dense 27B open-weight "decision model" that returns a calibrated probability for every answer option in a single forward pass instead of generating text, under the Nace.AI Open RAIL-M lice…
Stanford researchers led by Jure Leskovec proposed Memorilla, a latent semantic memory method described in arXiv:2610.21375 that feeds a frozen Qwen3-8B decoder 16 learned memory tokens instead of raw document text, usin…
Thinking Machines Lab said on 5 Nov 2026 that its Inkling 2 Small mixture-of-experts model, with 594B total parameters and 17B active, is available on Tinker with closed weights, and that the lab intends to release the f…
Baseten's Loops product trains LoRA adapters on open-weight base models using supervised fine-tuning or reinforcement learning from a Python SDK, CLI or REST API, billing serverless runs per token trained and sampled rat…
Anthropic published a working theory of agent memory by Lance Martin describing "dreaming," an offline pass in which Claude rereads session transcripts to correct wrong notes, build linked wikis, and decide which facts l…
Thinking Machines Lab released Inkling 2 Small on November 5, 2026, a Mixture-of-Experts model with 594 billion total parameters and 17 billion active, scoring 31 on the Artificial Analysis Intelligence Index (v5.0). The…
LLMBrake, a TypeScript library published to npm as version 0.1.0 under the MIT license, wraps OpenAI-compatible chat.completions.create calls to enforce a per-customer monthly dollar budget, pricing each call in USD from…
A developer built a Stremio addon that surfaces movies and series recommended in Hacker News comments, scanning HN for IMDb or Wikipedia links and titles mentioned by name, extracting them with the gpt-oss-120b model, an…
METR, the nonprofit Model Evaluation and Threat Research group, published an independent assessment on November 5, 2026 finding that Anthropic's benchmark growth on its AECI capability index has accelerated 1.9x since Oc…
Anthropic licensed Associated Press election data for Claude to surface real-time, attributed results during the U.S. midterms on November 3, the company's first payment to a news publisher, according to Adweek, which ci…
A research write-up proposes splitting open AI models into one frozen, quantized base model paired with thousands of low-rank adapter (LoRA) micro-experts of 2 to 5 megabytes each, routed by a two-step hierarchical route…