Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Google released Data Cloud connectors in Gemini for business in preview, letting business users and analysts query live enterprise data in natural language across BigQuery, Spanner, AlloyDB, Cloud SQL and Looker. The con…
OpenAI Group PBC released GPT-6 Cyber, a version of its GPT-6 Astra model trained for authorized vulnerability research, which went out of scope in 0.07% of graded runs without deployment safety instructions versus 5.51%…
In a five-run test of ten plain-English data questions against a fictional three-store coffee chain's January–March 2026 sales workbook, only one locally runnable open model came close to OpenAI's cheapest GPT-6 tier, GP…
A repeatable eval — collecting example inputs, defining what a good answer looks like, and scoring outputs consistently — lets teams compare two prompts or two models with numbers rather than impressions, according to gu…
Ali Ahmed built a research-to-LinkedIn automation pipeline that uses deterministic code, not the LLM, to decide what counts as true: every STATED claim must carry a verbatim quote found in the source paper via normalized…
A University of Washington research team's multi-agent system, VibeSys, generated serving engines for two hybrid Qwen3.8 models on AWS Trainium2, raising Qwen3.8-27B throughput from 195 to 861 tok/s on a single Trainium2…
Ali Ahmed documented the specific configuration needed to stream LLM Server-Sent Events through a Next.js route handler without buffering, drawn from building Turbo Alan Refiner, a multi-pass LLM text-refinement app. The…
Ali Ahmed merged his first contribution to LiteLLM, the open-source LLM gateway, with PR #38591, which fixed a bug where content retrieved by an agent's tool loop was re-compressed into the same reference, causing agents…
A developer's comparison of leading models found that hallucination rates do not track parameter count: DeepSeek V4 Pro, with 1.6 trillion parameters, posted a 94% hallucination rate on the AA-Omniscience benchmark, whil…
A Rutgers University study published in Nature Machine Intelligence found that five reasoning LLMs — OpenAI's o3-mini, DeepSeek's R1, Anthropic's Claude 3.7 Sonnet, OpenAI's gpt-oss-20b and Alibaba Cloud's Qwen-3 8B — co…
OpenRouter released Auto Code, a router built specifically for coding agents and harnesses that uses the same routing engine as the Auto Router and treats every request as part of an agent session. Auto Code selects mode…
Cloudflare cut the cost of its AI code-review pipeline from $2.43 to $0.77 per merge request, a 68 percent reduction, by relocating a prompt-cache breakpoint, removing unnecessary model calls, and stepping down models an…
A developer published a hands-on guide to Claude Code Subagents, showing how to define specialized AI assistants as markdown files with YAML frontmatter that run in isolated context windows with their own toolsets and sy…
A developer published a decision guide comparing three ways to build a Claude-powered agent: a hand-written tool loop using client.messages.create(), the beta client.beta.messages.toolRunner() helper in @anthropic-ai/sdk…
Developer Paolo Viviani released Cerea, an Apache 2.0-licensed, self-hostable chat and agent frontend for LLMs that works with any OpenAI-compatible inference provider and is designed to pair with his Pystino model gatew…
GPT-6 Astra, released in September, ranks as the best model at the social deception game Blood on the Clocktower, leading Claude Opus 4.6 by about 430 Elo points and winning roughly 92% of games against it, according to …
OpenAI announced GPT-6 Cyber on Thursday, a security-focused model it says scores 48.5% on the ExploitGym vulnerability-exploitation benchmark, up from 32.1% for GPT-5.6 Cyber and above the 42.4% posted by its most advan…
Hugging Face released a Mixture-of-Experts training handbook aimed at engineers who have already trained a dense model and want to train an MoE next, according to author Nouamane, who worked on the Ultra-Scale and Smol T…
CCORE released Qwen3.5-4B-GLB-3D-Coder, a 4-billion-parameter model built on Alibaba's open-source Qwen3.5-4B under the Apache 2.0 license, which turns natural-language prompts into standalone HTML and JavaScript Three.j…
A developer built "Tin tức hôm nay có gì?", a TypeScript/Bun service that ingests roughly 95 Vietnamese RSS feeds, embeds articles with jina-embeddings-v3 (falling back to local LaBSE), and clusters them with HDBSCAN to …