Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Anthropic's Claude Opus 4.5 and Zhipu AI's GLM-5.2 are frontier reasoning models competing on cost, context length, and coding benchmarks. GLM-5.2 offers a 1M-token context window and leads by 20.3 points on SWE-bench Pr…
A developer explains how to manage cost and latency when integrating large language models into products. Key strategies include constraining output tokens, caching requests, routing to cheaper models, and streaming resp…
The Washington Post tested six major AI chatbots on political questions and found systematic differences in ideological slant. ChatGPT and DeepSeek leaned left in most responses, while Gemini offered balanced answers 93%…
Reddit COO Jen Wong targets 1 billion users as the platform, once an internet oddity, expands its appeal and turns profitable for the first time since its 2024 IPO. With over 120 million daily users and 94% of revenue fr…
Technical writers are not being replaced by AI, as the human elements of the job—such as chasing subject-matter experts and understanding product tensions—remain irreplaceable. The role is evolving toward editing, orches…
A glitch in Google's Gemini AI exposed its internal reasoning process, revealing a raw thought stream that included translating Japanese subtitles for a Doraemon episode. The incident highlights transparency and safety c…
OpenAI's Codex surpassed 3 million weekly active users, a fivefold increase since January 2026, driven by new multi-agent execution and computer control capabilities. Enterprise clients including Goldman Sachs and DoorDa…
Meta has revived Facebook Creator Studio as a standalone AI companion app, integrating an AI chatbot that provides performance insights, engagement recommendations, and comment drafting. The app is currently in testing w…
Google integrated 'Computer Use' into Gemini 3.5 Flash, enabling the AI model to operate computers, browsers, and mobile devices autonomously. The model achieved a score of 78.4 on the OSWorld benchmark, matching GPT-5.5…
An engineer who built production AI pipelines at scale describes the '80/20 rule' where the first 20% of an AI feature takes 20% of the time, but the last 20% takes 80% due to real-world data surprises, cost overruns, an…
OpenAI, Meta, and Google are using food-related codenames for AI projects, including OpenAI's 'Jalapeño' chip and 'Garlic' LLM, Meta's 'Avocado' model, and Google's 'Nano Banana' image generator. The trend highlights a p…
A team at an unnamed company built a customer support agent on LangChain that hallucinated a wrong return policy in a multi-step process, logging success while being confidently wrong. This incident highlights the observ…
A developer with 25 years of experience reports that AI coding agents became genuinely productive around late 2025 with models like GPT-5.3 and Claude Sonnet 4.5. The developer's role shifted to designing environments wh…
OpenAI quietly upgraded all free ChatGPT users to GPT-5.5 Instant, reducing hallucinations by 52.5% and improving multi-turn reasoning, now serving 900 million weekly users. The move narrows the gap between free and paid…
A developer built a knowledge base in 2019 and spent seven years iterating on it, only to find that adding the Model Context Protocol (MCP) simplified the system dramatically. By replacing complex embedding and vector se…
A developer replaced 2.5 hours of daily busywork with a $0 AI agent setup running on a Mac Mini M4. The system uses local LLMs (Ollama with Qwen models), Python scripts, and cron jobs to automate email triage, calendar c…
Wan-Streamer v0.1 introduces a streaming Transformer that processes audio, video, and text as a single interaction loop, reducing latency and enabling full-duplex real-time AI assistants. The system achieves approximatel…
A developer describes setting up T3 Code with Claude Code and the open-source LLM gateway Lynkr to create a layered stack for coding agents. The architecture separates UX, coding behavior, and model traffic control, addr…
A developer dogfooded a human-controlled AI PR workflow by updating the LogFabric website using Cursor and Claude. The workflow ensured AI-assisted changes were scoped, reviewed, and merged only after human approval, dem…
Correctover has built a detection framework to catch silent model swaps in production LLM systems, where providers return responses from different models than requested without notice. The framework operates across six d…