AI summary · this digest is compiled by AI, not yet reviewed by Jason
The key trend today is the maturation of agent infrastructure, spanning low-cost decisions, passwordless delegated login, and context-retrieval evaluation as AI moves from chat toward controllable real-world workflows.
Claude Mods proposes an agent interface designed around humans rather than machines and tools. Instead of treating the harness as an invisible execution layer, it organizes tasks around user attention, decisions, and workflows. The concept suggests that the Skills ecosystem may evolve beyond prompts into complete human-agent collaboration interfaces.
OpenAI has launched the Decisions API for fast, focused, and low-cost decisions inside applications. It targets tasks such as classification, routing, and moderation that do not require long-form generation. Developers can use it to separate lightweight judgments from general-purpose model calls, reducing latency and inference costs.
1Password has introduced a way for AI agents to log into personal accounts on a user’s behalf without exposing the underlying password. The design addresses credential leakage as agents begin performing real-world actions, creating important infrastructure for safer automation across payments, productivity, and consumer services.
Canva has released its first in-house image generation model and made it available to all users. The company emphasizes improved customization, controllability, and lower costs. For designers and indie developers, image generation is becoming a native platform capability rather than an external add-on, shortening the path from concept to finished content.
Uber is rolling out ChatGPT Work to more than 30,000 employees for research, data analysis, presentation creation, and complex workflows. The deployment shows enterprise AI adoption moving beyond isolated chat use cases toward general-purpose agent platforms embedded in departmental processes and measured through productivity gains.
Latent Space reports that DeepSeek DeepInfer can process roughly 155 trillion tokens per cluster per day, generating about $22 million in daily revenue at an estimated 80% margin. If accurate, it demonstrates how open-source inference infrastructure can turn extreme serving scale into substantial cash flow.
A study using a deterministic geochemistry simulator found that agents can compute the correct result yet answer incorrectly when long tool outputs are moved out of context and retrieved on demand. For developers, the lesson is that context compression must be evaluated for information recovery and answer accuracy, not only token savings.
FrontierCS 2 introduces 48 open-ended computer science research tasks in which agents work inside existing repositories and reconstruct deleted paper contributions. Grounding evaluation in the authors’ released implementations makes it more realistic than single-program benchmarks and helps teams assess sustained research and iteration capabilities.
Investors: Rally Ventures、Skyward Group
AI原生嵌入式保险平台，把小额保单直接嵌入结账流程；27万美元种子轮显示保险分销正借助AI走向场景化和自动化。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free