Jason Says
OpenAI acquiring Ona is today's clearest signal: equipping Codex with persistent cloud environments transforms AI agents from one-shot assistants into long-term collaborators — and that makes the entire agent-skills ecosystem exponentially more valuable.
OpenAI is acquiring Ona to equip Codex with secure, persistent cloud environments. This move upgrades Codex from a code-completion tool into a stateful, long-running agent platform capable of handling complex enterprise workflows autonomously — a significant step toward production-grade AI agents.
Addy Osmani open-sourced agent-skills, packaging senior engineering workflows (DEFINE→PLAN→BUILD→VERIFY→REVIEW→SHIP) into production-grade skills for AI coding agents. The goal: make agents consistently follow engineering best practices across every phase, not just wing it.
pm-skills marketplace launches with 68 PM skills and 42 chained workflows covering the full product lifecycle — from discovery to strategy, launch, and growth. Built for Claude Code and Cowork, this signals the agent-skills ecosystem rapidly expanding beyond engineering into product management.
Abacus AI's Bindu Reddy reports Claude Fable as a split personality: terrifyingly capable in agent mode with a strong bias for action, yet surprisingly poor at basic chat compared to GPT-5.5 and Grok. Meanwhile, credible rumors suggest GPT-5.6 may launch significantly cheaper.
ModSleuth introduces an automated framework to audit the recursive dependency chains in modern LLMs — tracking which upstream models generated training data, filtered corpora, or guided decisions. Think of it as an SBOM for AI models. Critical for compliance, IP risk, and understanding why your model behaves the way it does.
Computer-use agents waste massive tokens encoding near-identical consecutive screenshots. ReVision exploits temporal visual redundancy to compress visual tokens, finally enabling agents to meaningfully leverage interaction history. First method to show real performance gains from history in computer-use tasks — key for developers building long-horizon automation.
Long-context inference faces two walls: exploding KV cache memory and O(T²) sparse selection cost. SparDA introduces a fourth 'Forecast' projection to decouple sparse selection from attention computation, tackling both bottlenecks simultaneously. Directly relevant for developers deploying 128K+ context models who need lower inference costs.
Astrophysicist Chi-kwan Chan uses Codex to build black hole simulation software, studying extreme physics and testing general relativity. This case demonstrates AI coding agents delivering real value in highly specialized scientific computing — far beyond typical CRUD apps, accelerating research at the frontier of human knowledge.
Investors: Base10 Partners
专注物流、薪资、建筑等实体经济自动化的 AI 投资基金，关闭两支共 8.5 亿美元基金，押注 AI 向传统行业深度渗透的趋势。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free