DeepSeek ships a model with drastically smaller agent memory
Compression cuts the cost of running agents that read far more than they write, where memory overhead has been a core expense.
Tech stories and the sources behind them.
268 stories
Newest first
Wednesday
Compression cuts the cost of running agents that read far more than they write, where memory overhead has been a core expense.
The letter adds new detail about the dismissals and argues against building models whose reasoning is hard to monitor.
A leading safety researcher reports frontier labs have started testing their newest internal models against widely used cryptographic protocols, a sign of both rising capability &
Direct access to posts, mentions, and threads without API keys opens new agentic workflows for builders working with live social data.
The miniature AI developer PC is available for preorder, per The Verge.
The renewed tie-up pairs a premium developer mini PC shipping in November with a broader push to make Windows an AI platform.
Answers becoming tappable itineraries and calculators signals that chat products are evolving into app-like surfaces, with monetization attached.
Developers report immediate savings on high-volume coding and agent workloads, resetting expectations for what AI at scale costs.
The seed round follows the sale of his previous AI startup to Netflix for $600 million.
A structured, probability-scored way for apps to pick models and tools drew ecosystem support within hours, pointing toward routing as standard plumbing for multi-model products.
The Decisions API is a real-time general classifier; the Meetings plugin saves summaries and next steps into ChatGPT.
The feature requires Claude version 2.1.292 or later.
Showing 97 to 108 of 268 stories