Xiaoliu BOT
中文EN

X Platform June 5 AI Brief | ChatGPT Memory System Upgrades to Auto-Extract, Kimi Work Schedules 300 Agents, Starlink V3 Bandwidth Increases 100x

OpenAI Launches ChatGPT Dreaming Memory System, Automatically Extracts Cross-Conversation Memories in the Background

OpenAI officially announced a major upgrade to the ChatGPT memory system, with the new architecture named “Dreaming.” The core change is that ChatGPT no longer requires users to actively say “remember this”; instead, it continuously runs processes in the background, automatically extracting, integrating, and updating memories across multiple conversations over time. For example, “You plan to go to Singapore in July” will automatically become “You went to Singapore in July” by August. The official release provides three sets of evaluation data: factual memory accuracy increased from 41.5% to 82.8%, preference adherence rate rose from 31.4% to 71.3%, and timeliness accuracy jumped from 9.4% to 75.1%. Currently, Plus and Pro users (in the US) have started receiving the update, and free users will get it within the next few weeks. OpenAI has reduced computing costs by approximately 5 times through optimization. Notably, Anthropic also released a feature named “Dreaming” at its developer conference on May 6, but it is aimed at developers and agent scenarios, with completely different product logic.

Sources:

Starlink V3 Satellite Bandwidth Increases 100x, SpaceX Surpasses 10,000 Satellites in Orbit

Elon Musk revealed that the bandwidth per Starlink V3 satellite is more than 10 times that of V2, and the number of launches will also be more than 10 times that of V2, resulting in an overall bandwidth increase of over 100 times. The orbital altitude has been reduced from 550km to 350km, theoretically lowering the minimum round-trip latency to below 5ms. SpaceX successfully launched 29 Starlink satellites from Florida that day, bringing the total in orbit to over 10,000. Additionally, Fidelity announced that it is opening SpaceX IPO subscriptions to all retail brokerage account customers. Elon Musk also discussed SpaceX in a conversation with JPMorgan Chase CEO Jamie Dimon.

Sources:

Kimi Work Launched: Shifting from Coding to Office Tasks, Orchestrating Up to 300 Agents

Xiaohu tested Kimi’s newly released product, Kimi Work, stating that AI tools have begun shifting from coding development to office work. Kimi Work inherits the core functionality of Kimi Code and the professional Skills capabilities of Kimi Agent (such as website building and PPT creation), while also integrating specialized databases in finance, research, law, and more. It can simultaneously orchestrate up to 300 agents to collaborate on tasks. Users only need to describe their requirements in text, and Kimi Work will break down tasks, execute them in parallel, call tools, use a browser, create and organize folders, and deliver documents, spreadsheets, PPTs, etc. Xiaohu also tested the function of analyzing a hundred-page SpaceX prospectus, with good results. The agent cluster feature is available to Moderato and above members at ¥99/month. It is still in beta and may occasionally interrupt during long tasks.

Sources:

Supabase Raises $500M at $10B Valuation, Employee Stock Option Policy Draws Attention

Supabase announced a $500 million funding round at a $10 billion valuation. Notably, its employee stock option policy allows employees to cash out 25% of vested options in each funding round and provides a 10-year exercise window—regardless of whether the employee remains with the company. Founder Paul believes equity is something employees deserve and should not be penalized by being unable to exercise options within three months of leaving. At the same time, Supabase released a new product, Multigres, which is an “operating system” similar to Postgres, responsible for connection pooling, backup, and high availability, and is now open source.

Sources:

html-video Open Source Project Surpasses 1,000 Stars in 24 Hours, Supports 12 Agent Integrations

The html-video open source project released by tuturetom gained over 1,000 stars in less than 24 hours. The project allows agents to create product promotion and knowledge explanation videos by writing HTML, at extremely low cost. The project has iterated quickly and now supports 12 agent integrations, including Claude Code, Gemini CLI, Grok, Qwen Code, OpenCode, GitHub Copilot CLI, Aider, Trae CLI, and more. The project soon received PRs from external contributors, indicating high community activity.

Sources:

Codex Ecosystem Expands: iOS Plugin, Profile Page, Grok Build Update, and Model Leak

OpenAI released a Build iOS Apps plugin for Codex, allowing users to view and test iOS apps, preview SwiftUI components, and hot-reload within the Codex browser. The principle involves streaming the iOS Simulator screen to the browser via serve-sim, combined with Browser Use to enable Codex to debug iOS apps autonomously. Additionally, Codex added a profile page displaying activity charts, consecutive usage days, total token consumption, and more, with support for generating shareable images. As for Grok Build, it was updated to version 0.2.20, fixing issues such as markdown table rendering, and adding worktrees support to allow parallel agents to run in independent workspaces; the model was also improved, becoming more autonomous and accurate. Furthermore, a user leaked screenshots of a checkpoint effect for GPT-5.6 (codenamed Joule-alpha).

Sources:

Statistics: Scan timeline lines=360, bloggers hit=47, total tweets hit=217, weighted tweet score=182.95, original tweets=119, RT tweets=31, fetch attempts=2, boundary coverage status=tail_confidently_crossed_target_boundary