A study of 535 real scanned receipts from the ICDAR-2019 SROIE dataset found that metamorphic testing — checking whether an LLM returns consistent totals when the same receipt is presented in differen…
Group Relative Policy Optimization (GRPO) eliminates the value model used in Proximal Policy Optimization (PPO), reducing active memory footprint from 42 GB to 28 GB for a 7B parameter model in FP16 b…
DictaFlow's development testing shows Parakeet offers near-instant local dictation on iPhone and Mac with a 483 MB download footprint, while Whisper provides better vocabulary, accent handling, and la…
Apple filed a federal lawsuit against OpenAI on June 14, 2026, alleging trade secret misappropriation after former engineer Chang Liu kept his work MacBook and credentials post-employment, downloaded …
GPT-4 has 1.8 trillion total parameters but activates only about 36 billion (2%) per token due to its Mixture-of-Experts (MoE) architecture, where a routing function selects k experts for each token. …
Mira Murati's Thinking Machines Lab released Inkling, a 975B-parameter open-weight model, on July 15. Within hours, Artificial Analysis scored Inkling at 41 on its Intelligence Index, making it the to…
A full-stack developer who knew nothing about AI security learned prompt injection by playing Lakera's free game Gandalf, which has logged nearly 9 million interactions from over 200,000 people since …
In the tenth installment of a series on building AI agents in Rust, the author explores driving a local model directly via stdin/stdout instead of HTTP, using Salvatore Sanfilippo's DwarfStar inferenc…
A 2026 career-series post breaks down the hierarchy of AI, Machine Learning, Deep Learning, Generative AI, and Agentic AI, explaining that each is a more specialized subset of the previous. The author…
A new engineering playbook details three production techniques to cut unnecessary LLM token costs after request optimization: adaptive retrieval that varies context size by query complexity, reranking…
88% of enterprise AI projects never reach production, and the primary cause is not the model, data, or infrastructure but the cost of running agentic loops, which scales quadratically (O(N²)) with the…
A team built a multi-agent system with an orchestrator, three sub-agents, and a memory layer to classify customer support tickets, but the task could have been handled by a single prompt with a well-w…
A technical guide demonstrates the limitations of base language models like GPT-2 for domain-specific tasks and introduces fine-tuning as a solution, showing that GPT-2 produces inconsistent outputs w…
Visual prompting is the deliberate design of visual context to steer a model's attention, constrain its hypothesis space, improve grounding, and shape downstream reasoning, representing a change in th…
A new analysis reveals that while AI token prices dropped roughly 80% between 2025 and 2026, nearly 60% of companies running AI agents in production exceeded their budgets by 30–50%, according to rese…
A guide to building AI products maps the core structure into five nodes, starting with the base LLM and moving through wrapping layers that add tools, memory, and RAG to overcome the model's limited k…
Anthropic's Claude Cowork platform has five core components—skills, connectors, plugins, projects, and artifacts—but most users only leverage one, according to a company guide. Skills are how-to files…
In part 9 of a series on building AI agents in Rust, the author introduces sandboxed code execution, allowing the agent to run arbitrary shell or Python code on the user's machine. The post emphasizes…
Boris Cherny, who leads Claude Code at Anthropic, and Peter Steinberger, creator of the open-source agent OpenClaw, have stopped manually prompting AI agents and now design systems that automate promp…
A new LLM security paper from Prof Luyi Xing's group shows how an attacker can distort an AI agent's view of the world by leaving carefully formatted content in a product review, GitHub comment, email…