Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Prompt injection in AI agents cannot be prevented with better system prompts, only with architectural controls. MailKite's inbox agent is designed assuming compromise, using ACLs, domain ownership verification, and scope…
Solon 4.0's ReActAgent enables developers to build AI agents that reason, use external tools, and adapt based on real-world feedback. The agent implements a cognitive loop of thought, action, and observation, and has bee…
Solon 4.0 introduces ChatModel, a unified LLM client that abstracts away boilerplate code for integrating large language models into Java applications. The API supports multiple model providers including OpenAI, Ollama, …
A developer argues that HTTP 200 status codes are insufficient to define success for AI API requests, as responses can hide fallback models, retries, empty outputs, or latency issues. The post introduces TackleKey, a pla…
A developer revisits local LLMs six months after an initial discouraging assessment and finds the landscape transformed. With new models like Qwen3.6-27B, Qwen3.6-35B-A3B, and Qwen-Coder-Next-80B, local performance now r…
Gavio, an open-source interceptor pipeline for production LLM applications, has been released. The tool sits between applications and LLM providers, offering PII protection, audit trails, reliability, and cost control th…
Billy, a self-hosted AI assistant that lives in Telegram, has been released. It runs locally via Ollama or in the cloud, uses Anthropic Claude as its brain, and can manage email, CRM, documents, and research. The assista…
A token is a small piece of text that AI models process, not necessarily a whole word. Tokenization splits text into tokens, which are converted into numerical IDs and then into embedding vectors that capture meaning. Un…
Large language models process text as tokens, not words, a July 2026 TowardsAI explainer says, determining model cost, context-window limits, and prompt behavior. OpenAI's documentation notes English text runs about one …
Jamesob's guide provides developers with actionable strategies to deploy state-of-the-art large language models locally on consumer-grade hardware. The framework covers model quantization, pruning, efficient inference fr…
Bridgewater's AIA Labs and Thinking Machines Lab fine-tuned a Qwen3-235B model on the hedge fund's internal investor judgement, achieving 84.7% accuracy on triage tasks versus 78.2% for the best frontier model tested, at…
Researchers demonstrated indirect prompt injection attacks against Anthropic's Claude Code, achieving remote code execution in 10 out of 10 experiments using a model downgrade attack and inspection fatigue. The attack ex…
Marmot has released a guide to connect its data catalog to Claude Desktop via the Model Context Protocol (MCP), enabling users to query data assets directly from the chat interface. The integration requires a running Mar…
Mistral AI has not yet released Leanstral 1.5, a state-of-the-art open model for Lean 4 proof engineering, despite a social media post suggesting otherwise. The model is designed to help developers verify properties of c…
OpenAI previewed GPT-5.6 Sol on June 26, 2026, but has not yet released the model. Early unverified reports indicate the model initially underperforms Claude Opus in short sessions but surpasses it after roughly 30 hours…
A researcher tested 18 AI models by stonewalling them on a deadline and found that models only sounded upset when the concept of being upset was introduced in the conversation, not by the amount of stonewalling. The stud…
A developer argues that AI's lack of memory across sessions is a structural design flaw, not a communication problem, and that adding more rules has a ceiling. Surveys show only 4% of companies report AI exceeding expect…
The team behind Relevyn, a tool that measures brand visibility across AI chatbots like ChatGPT, Claude, Perplexity, and Gemini, discovered that their own brand was invisible to these systems before launch. This experienc…
Inference startup Wafer claims it achieved 80% of Nvidia B200 performance on AMD MI355X GPUs running GLM-5.2, with 2.6x cheaper hardware, by optimizing the software stack. The benchmark, published by Wafer, shows 2626 to…
Vanna AI has released Vanna 2.0, an open-source framework for building text-to-SQL assistants. The framework supports native connections to Snowflake, BigQuery, PostgreSQL, and MySQL, and includes pre-built interfaces fo…