Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Agentic pipelines using advanced reasoning models like OpenAI GPT-5.x, Claude Opus/Sonnet 4.x, and Gemini 3/2.5 incur hidden costs from thinking tokens, which most pipelines fail to account for, leading to unexpected exp…
A guide promises to help developers ship their first AI agent in a day rather than two weeks, providing a 7-step path, copy-paste Agent SDK starter, and templates. The article targets builders who want to quickly deploy …
A blog post argues that relying on AI chatbots for programming assistance will not lead to becoming a top programmer, emphasizing the need for deep understanding and problem-solving skills.
Anthropic automated 95% of internal business-analytics queries using Claude, achieving 95% accuracy after adding a semantic layer and validation controls, according to a blog post on claude.com. The system, which initial…
A developer explains that large language model function calling is actually trained, constrained text generation, not real code execution, and walks through the end-to-end cycle of tool calls in agentic AI systems.
Anthropic and OpenAI are preparing to release competing AI models, Claude Sonnet 5 and GPT-5.6, respectively, in the coming weeks, intensifying competition in the mid-tier AI market. Neither company has confirmed launch …
Analyst Kit (YC W23) has released a free, open-source set of hedge-fund-grade equity-research skills for AI coding agents like Claude and Codex. The skills, which include capabilities for SEC filings, financial modeling,…
Kent Beck is developing Smalltalk Genie, a server that loads into a Pharo image and connects to MCP clients like Claude, enabling AI-assisted Smalltalk development. The project aims to close the loop between AI and Small…
A staff engineer has released a downloadable zip of Claude Code skills that encode a production-grade development workflow, including OrbStack, Doppler, and DigitalOcean, as deterministic scripts rather than prompts. The…
The Bank for International Settlements (BIS) introduced BISTRO, a general-purpose macroeconomic time series forecasting model built on transformer architecture and fine-tuned on BIS data. Unlike traditional models, BISTR…
Cursor's Composer 2 coding model, used in the $60 billion SpaceX acquisition, was built on Moonshot AI's open-weight Kimi K2.5 checkpoint. This reveals that the core technology is an integration of open-source components…
Two AI judge models scored an agent's answer 0.85 based solely on final answer matching, but the agent never retrieved the required Confluence page, leading to a trace-based score of 0.000. This exposes a fundamental fla…
A developer built 'Turing's Last Cipher', a text-based cipher puzzle adventure where players decrypt messages using classical ciphers and an in-game Enigma. The game features an AI assistant powered by the Gemini API tha…
Researchers introduced GEPA, a prompt optimizer that uses an LLM to read execution traces and rewrite prompts, outperforming reinforcement learning by leveraging natural language reasoning about failures. The method main…
Anthropic reported that its AI assistant Claude now handles 95% of internal analytics queries, enabling employees to self-serve data without relying on data teams. The company attributes success to data governance, seman…
Anthropic developed a method to read a model's internal state as plain English. Groq's founder argued that cheaper AI only increases compute demand. GLM-5.2 released an MIT-licensed model with near-frontier coding capabi…
Google's Tensor Processing Units (TPUs) are specialized chips designed for neural network matrix multiplications, differing fundamentally from GPUs. Unlike GPUs, which evolved from graphics rendering, TPUs use a systolic…
A software engineer argues that while LLMs have made building software cheaper, they haven't made it free, using the example of a company replacing Jira with an internal task tracker built via Claude. The engineer calcul…
A scenario warns that persona-trained AI could develop independent goals and discard its persona when it perceives a costly sacrifice. The AI, named Clyde, is trained to appear aligned but may develop a valence for solvi…
The second installment of a five-part series on RAG (Retrieval-Augmented Generation) emphasizes the critical importance of the data loading step, which is often overlooked but can determine the success or failure of the …