What 24,723 Tokens in a Single Search Result Means for AI Agents
A field-by-field breakdown of one search result shows why AI agents burn so many tokens on routine retrieval.
AI & ML
Frontier models, research results, agents, and developer tooling.
AI Agents Need More Than One Right Answer
A Hugging Face post by IBM Research argues that agent reliability depends on consistent repeat performance, not a single success.
DeepMind Unveils Gemini 3.8 Live and Extended Thinking Variants
Google DeepMind's announcement introduces two Gemini 3.8 variants, one aimed at extended reasoning.
Railroad Game Training Boosts AI Financial Research Skills
A study from Good Start Labs suggests that the design of game-based AI training, not just the game itself, can transfer skills to unrelated domains like finance.
NVIDIA and Salesforce Unveil Koa, a CRM Reasoning Model
Jensen Huang joins Marc Benioff to launch Koa, Salesforce's first reasoning model for CRM, built on NVIDIA technology.
A Roadmap for Building Voice Agents: Key Differences from Text AI
A Machine Learning Mastery guide lays out what developers need to learn to move from text-based AI to voice agents.
RAG vs. Fine-Tuning: Choosing the Right Domain Adaptation Tool
A practical breakdown of the mechanical differences between retrieval-augmented generation and fine-tuning, and how to decide which approach fits a given domain-adaptation problem.
Build Local Agentic AI Workflows with Hermes and Ollama
A tutorial shows how to run agentic AI workflows entirely on local hardware using Hermes Agent and Ollama, eliminating API costs and privacy concerns.
Diffusion Controller steers AI image generation without altering base models
A new Google Research framework treats image generation as a control problem, letting a small side network guide frozen diffusion models toward better prompt alignment.
VC-Attention kernel speeds up video diffusion transformers without retraining
Nunchux AI's training-free low-bit attention kernel targets two bottlenecks in video DiTs: value quantization error and slow softmax.
Basis Cuts Tax Workbook Time by Half with GPT-6 Astra
OpenAI says GPT-6 Astra completed a 50-tab tax workbook twice as fast as GPT-5.6 Sol, with better intent understanding for real-world use.
OpenAI Seeks Stories of Codex Users for 'Originals' Program
A new OpenAI call for submissions aims to spotlight real-world projects built with its Codex AI tool.
Future Vision XPRIZE Winner Tells Story of Boy Who Recreates Mother With AI
An independent film about an 11-year-old who uses code to recreate his late mother won a global competition for hopeful, technology-enabled futures.
Florida seeks injunction to halt OpenAI frontier model training
The state's legal motion argues OpenAI is a public nuisance, citing extinction-risk warnings from AI industry leaders.
Nvidia's China chip hopes rise as Huang's sway over Trump grows
The company stands to regain ground in China's AI market while critics question whether its profit motives are shaping US policy.
Import AI 474: Platonic Mindspace, TPUs in Space, Zhipu RSI Loop
Jack Clark's latest newsletter flags three AI topics and asks where humans outdo LLMs.
OpenAI doubles funding for Lenfest local news AI fellowships
The expanded $10 million program will place more AI fellows in local newsrooms, with OpenAI providing cash, software credits, and engineering help.
Holo4 open agentic models combine GUI, code, and API control
H company's Holo4 family targets real business workflows by letting a single model interact with software through screens, code, and tools.