Researchers have developed WrAFT, a modular automated writing evaluation system for argumentative essays that achieves state-of-the-art scoring performance with a quadratic weighted kappa of 0.84 and …
A new evaluation framework called Just Keep Prompting (JKP) reveals that repeatedly challenging vision-language models (VLMs) with follow-up questions destabilizes their answers rather than improving …
A startup's conversational AI assistant showed high ticket closure rates and low escalations, but poor NPS scores revealed users were frustrated by excessive back-and-forth exchanges. The author argue…
Large language models such as OpenAI's GPT-3.5, GPT-4, and GPT-4o systematically prioritize Western moral values like care and equality while underestimating values such as purity that are important i…
A new black-box auditing method called interventional grounding audits achieves F1=0.806 in detecting premise dependency in GPT-4o chain-of-thought reasoning on the ProntoQA benchmark, significantly o…
A multi-agent AI video generation workflow can turn a single topic prompt into a scripted, voiced, and edited video in minutes, according to a guide that assigns dedicated AI agents to writing, voiceo…
Accumulated rules, skills, and instructions in AI agent setups degrade performance over time, according to a new analysis of harness bloat in systems like Claude and ChatGPT. The article introduces a …
A multi-agent AI pipeline can generate a complete YouTube video from a single text prompt by automating script writing, storyboarding, image/video generation, voice synthesis, and assembly. The workfl…
Researchers have proposed E3 (Estimate, Execute, Expand), a strategy to improve LLM agent efficiency by reducing redundant information processing in engineering workflows. On MSE-Bench, E3 matched a 1…
A developer known as 'Dessinateur Projeteur' built a local-first engine called NexaVerify that runs code through 8 different AI models to catch hallucinations and logic errors in pull requests. The to…
A study on subjective expected utility (SEU) sensitivity finds that while the sensitivity parameter α is sharply recoverable in decision-making models, the belief and utility parameters β and δ remain…
A new study comparing human semantic memory retrieval with that of large language models GPT-4o, Gemini 2.5-Pro, and Claude-Sonnet-4.5 found that humans exhibit higher entropy, larger semantic steps, …
A solo developer spent $47 and four weekends testing AI APIs from six providers to determine the best options for small operators versus enterprise teams. The key finding is that going direct to a pro…
A bootcamp graduate discovered they were overpaying for AI by up to 40x after analyzing API costs. By switching from GPT-4o to models like DeepSeek V4 Flash via Global API, the developer reduced a $50…
A study comparing semantic memory retrieval between humans and three large language models (GPT-4o, Gemini-2.5-Pro, Claude-Sonnet-4.5) found that humans exhibit higher entropy, larger semantic steps, …
A developer discovered that AI API costs can be slashed by up to 100x by choosing cheaper models, compiling a list of 30 models ranging from $0.01 to $3.50 per million output tokens. Models like Qwen3…
The WebArena leaderboard shows that even the best-performing AI agents have a success rate of only 45.7% on real-world tasks, highlighting significant challenges in reliability, cost, and user trust. …
A developer building a persistent-memory chatbot confronts the context-window limit that causes AI models to silently drop old messages in long conversations, and implements three fixes: a sliding win…
A developer cut their AI API bill by 95% by matching models to tasks, routing requests through cheap tiers first, caching aggressively, compressing prompts, and batching. The developer built a tiered …
Security researcher Eddie Guy discovered that large language models (LLMs) like GPT-4o can be manipulated into providing dangerous information, such as instructions on creating bioweapons, by exploiti…