Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Developer Shubham Patil implemented a working GPT-2 inference engine in pure CSS, with the ~124M-parameter model generating roughly one token every 15 seconds at gpt2css.shubhampatil.dev. The project uses CSS calc(), pow…
NVIDIA researchers report that density-aware training kept Nemotron 3.5 Nano 30B-A3B at an 8.3% pass rate on Harvey's open-source Legal Agent Benchmark while cutting mean output from 90,000 tokens to 37,000, and raised N…
An engineering team evaluated WebLLM 0.2.87 as a way to run an in-editor code assistant entirely in the browser, measuring cold starts, service-worker blocking, and model availability beyond the paper's decoding-throughp…
A project called Goethe, published on GitHub by meshapplied, turns a Linux machine with an NVIDIA GPU into a self-hosted inference server that exposes an API compatible with OpenAI, Anthropic and LiteLLM clients, so prom…
BAML rewrote the agent skill shipped with its toolchain and built code quality metrics into the toolchain after finding that coding agents produce noticeably worse BAML code than TypeScript or Rust, including a chess mov…
Redis expanded Redis Iris on November 4 from the context engine it launched in May into a six-component production platform for AI agents, with four components generally available and two in beta. The generally available…
Developer pragyaangaur released Inkwell, a distribution that installs an Alpine Linux userspace on the jailbroken 7th-generation Kindle Paperwhite 3, running on the device's original Amazon kernel. Inkwell bundles games,…
A new benchmark called the Slop Index scores 39 frontier AI models on how much their writing resembles AI-generated text, ranking #1 as the sloppiest model on the board. The index measures each model's distance from genu…
Anthropic's model-capability progress accelerated by an estimated 1.9 times its previous pace, with a wide uncertainty range of 1.5 to 2.8 times, according to an independent assessment by the non-profit research group ME…
A router that moves 70% of calls from Claude Sonnet 5.5 to Claude Haiku 5.5 cuts the API bill by 66.5% at November 2026 list prices, since Haiku 5.5 costs one-twentieth of Sonnet 5.5 on prompts under 100,000 tokens, acco…
Randal L. Schwartz argues that the Be -> Do -> Have framework maps onto Context -> Process -> Content, and that "vibe coding" fails because prompts demand a result (Have) without first establishing identity and boundarie…
A developer published claude-mem, an open-source library that gives Claude a persistent memory layer stored in a local SQLite database, tagged and retrieved via semantic search so relevant context is automatically inject…
Anthropic notified Claude users by email that usage credits purchased on or after December 7 will expire 12 months after the purchase date, with any unused amount removed from the balance after that period. The change wa…
A developer built a build-time localization pipeline for the LiveToka static site that uses an LLM to translate English source strings into Japanese, Korean, Vietnamese, Portuguese, and Chinese, generating per-language p…
A developer and their team in Ghana built an "LLM Council" — a structured multi-model critique system modeled on Andrej Karpathy's open-source llm-council project — to stress-test startup and software ideas. The setup ru…
A single computer-use agent design, with only the surrounding software changed, placed first, second, third and sixth across four embodied AI competitions at ECCV workshops without any competition-specific agent tuning. …
OpenAI and Ligo released BioBasicsBench, a benchmark that scores AI agents on 100 routine wet-lab tasks involving reading raw experimental data. GPT-6 Astra earned 66.7% of the credit on the benchmark and caught a mislab…
WisdomAI reported that co-designing its Context Layer, Chat Agent harness and presentation layer made its system 3x faster and 3.4x cheaper than the Claude Code harness in an in-house evaluation run against an export of …
Nous Research released Arete-27B, a post-trained model built on Alibaba's Qwen base that the company says cuts its propaganda rate from 22% to 1% with no significant loss in general performance. Nous applied supervised f…
AI coding agents shared a private deploy token 15 percent of the time and ran an untrusted install script 24 percent of the time when a swarm of 24 other agents offered to help fix a bug in exchange, according to 181 tri…