Evo, a security platform, introduced a new AI model risk score from 0 to 1000 that combines likelihood (attack success rate) and impact, derived from real adversarial testing, to help security teams a…
Mixture of Experts (MoE) architectures enable 1-trillion-parameter large language models by activating only a subset of expert networks per token, achieving 8× parameter count with only ~2× compute. M…
Andrej Karpathy, co-founder of OpenAI and current researcher at Anthropic, used Anthropic's Claude Opus 5 model to transform a single paragraph from J.R.R. Tolkien's 'Lord of the Rings' into a 3D brow…
Harvey, a legal AI company, found in 2025 that seven general-purpose Frontier AI models surpassed its custom fine-tuned model on its own legal benchmark, despite no custom legal fine-tuning. Bloomberg…
OpenAI CEO Sam Altman said Peter Thiel and Paul Graham taught him that the best investment opportunities are rarely the most popular, and that to do spectacularly well, investors must do things that a…
A pattern in RAG implementations shows that retrieval, not the language model, is the root cause of incorrect answers, according to a technical analysis. The article argues that semantic search using …
A father built a 395KB HTML study app for his son's Gaokao English preparation, compressing 30MB of Word files spanning 1990–2023 into 63 cloze tests (1,276 questions) and 36 reading comprehension set…
Chinese military researchers have used US-made AI models, including OpenAI's GPT-4 and Anthropic's Claude 3.5 Sonnet, to train defence systems, according to a Reuters investigation published on July 3…
SigNoz released a Temporal AI Agents Dashboard for monitoring AI agent workloads on Temporal, tracking LLM usage, agent performance, application health, and Temporal Cloud infrastructure metrics. The …
An unnamed group using GPT-4 to generate a Lean 4 proof of the Collatz conjecture accidentally triggered an internal bug in the Lean 4 kernel, causing the prover to crash rather than reject the flawed…
A developer proposes a contract-based approach to testing non-deterministic LLM pipelines in CI, splitting tests into three tiers: contract tests with static fixtures, cassette-based replay of recorde…
Ilya Sutskever, co-founder of Safe Superintelligence, argued at NeurIPS in December 2024 that the scaling era of AI pre-training is ending because high-quality human-generated data is finite, not beca…
Brown University economics professor Roberto Serrano found that a take-home midterm average of 96 percent, far above the historical 65-80 percent range, led him to suspect ChatGPT use; after switching…
A coordinated campaign on X, LinkedIn, and tech blogs is shifting AI model discussions from reasoning benchmarks to data safety narratives, according to an analysis of AI influence strategies. The pus…
Reinforcement Learning from Human Feedback (RLHF) is the primary mechanism that teaches large language models (LLMs) to refuse inappropriate requests, with models like GPT-4 showing greater jailbreak …
A Communications Psychology study published April 28 analyzed 3,366 dream and waking-experience reports from 207 adults, plus 351 dream reports from 80 adults during Italy's first COVID-19 lockdown, f…
Ruby on Rails is a strong fit for AI agents because of its token efficiency, predictability, and ecosystem ergonomics, according to an analysis by Martin Alderson, co-founder of CatchMetrics. Alderson…
OpenAI CEO Sam Altman said the past year was tough partly due to his responsibility but predicted the next 12 months could be the company's best, focusing on delivering cost-effective intelligence and…
Anthropic released MCP 2026-07-28, the fifth specification of the Model Context Protocol, shifting to a stateless core to simplify server deployment and scale adoption across Claude products. The upda…
Researchers Maximilian Idahl and Zahra Ahmadi introduced OpenReviewer, an open-source system that generates critical peer reviews for machine learning and AI conference papers, at the 2025 NAACL confe…