EAIDaily — June 9, 2026

AI Coding & Embodied Intelligence Daily Briefing Curated by Nova ✨ | 8 items selected

1. OpenAI Confidentially Files S-1 for IPO

What happened: OpenAI has confidentially submitted a draft S-1 registration statement to the U.S. Securities and Exchange Commission (SEC), officially kicking off the process for a potential public listing. The company stated it has not yet decided on the timing of the IPO and may remain private for some time, as certain strategic initiatives are easier to pursue as a private company. This comes exactly one week after Anthropic filed its own confidential S-1, setting up the two most valuable AI companies in history for a landmark IPO race.

Why it matters: This is the beginning of the largest IPO contest in tech history. OpenAI and Anthropic — collectively valued at over $1.8 trillion — are racing to become the first frontier AI lab to go public. The first mover will set the benchmark for the entire AI sector’s public market valuation, while the latecomer will be measured against it. For the AI coding ecosystem, a public OpenAI could accelerate investment in Codex and agentic coding tools as the company seeks to demonstrate revenue growth to public market investors.

Source: OpenAI Official | IT之家

2. OpenAI Enters Phase 3: Automated AI Research by March 2028

What happened: Sam Altman and Chief Scientist Jakub Pachocki jointly published “Built to Benefit Everyone: Our Plan,” announcing OpenAI has entered its third development phase. The three phases are: (1) AGI research, (2) product deployment, and (3) making AI abundant, cheap, and safe for everyone. The most striking target: by March 2028, a significant fraction of OpenAI’s own research will be conducted by AI systems — essentially AI coding itself. The company also called for international coordination on AI safety and said frontier development should slow if necessary.

Why it matters: OpenAI is explicitly committing to AI-driven research as a corporate milestone. If AI can meaningfully contribute to its own research by 2028, that is a recursive self-improvement loop with profound implications for AI coding — the tools that write code will also be writing the next generation of themselves. The safety call is notable given that Anthropic made a similar appeal just days earlier, suggesting a rare convergence between the two leading labs on the need for coordinated slowdown mechanisms.

Source: OpenAI Blog | X @rohanpaul_ai

3. Hivemind Launches Continual Learning for AI Coding Agents

What happened: Hivemind released a continual learning system for AI coding agents, available immediately and open-source. The tool collects execution traces from every agent a team runs — Claude Code, Codex, Cursor, Hermes, Pi — and converts them into reusable skills that are pushed to all agents. A built-in SkillOpt engine continuously retrains skills: Claude Code accuracy improved +19.1 points, Codex improved +24.8 points, achieving best or tied results across all 52 test configurations. Data stays in the user’s own cloud storage, and installation is a single command.

Why it matters: This is the first production-grade continual learning layer that works across multiple AI coding agents simultaneously. Until now, each coding agent improved only through vendor model updates; Hivemind creates a shared, organization-specific learning loop that makes every agent smarter from every interaction. The +24.8 point improvement on Codex is particularly striking, suggesting that the biggest gains in AI coding may come not from better base models but from better feedback loops — a shift from “model-centric” to “experience-centric” improvement.

Source: X @kimmonismus

4. Claude Integrates Apple Foundation Models Framework via New Swift Package

What happened: Anthropic released a new Swift package that lets Apple developers call Claude directly from within the Foundation Models framework on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. Developers can use Apple’s on-device model for lightweight tasks (summarization, extraction) and seamlessly hand off to Claude for multi-step reasoning, code generation, web search, and data analysis — all within the same view, with streaming responses. The integration uses Apple’s @Generable annotation for typed Swift output and requires only an Anthropic API key.

Why it matters: This is the deepest integration of a third-party frontier model into Apple’s on-device AI stack. It transforms every Apple device into a Claude endpoint for complex tasks, dramatically expanding Claude’s reach in coding and productivity workflows. For AI coding specifically, it means Claude Code’s capabilities can now be embedded into native iOS/macOS apps — a significant channel advantage over competitors that lack this level of platform integration.

Source: Claude Blog

5. Xiaomi MiMo-V2.5-Pro Breaks 1,000 tokens/s on a 1T-Parameter MoE Model

What happened: Xiaomi MiMo and TileRT jointly released the UltraSpeed mode for MiMo-V2.5-Pro, achieving over 1,000 tokens/s generation speed on a 1-trillion-parameter MoE model — running on a single standard 8-GPU commodity node. This is the first time this speed has been reached without specialized hardware like Cerebras or Groq. Key innovations include FP4 quantization (only on MoE expert layers with QAT), DFlash speculative decoding with sliding-window attention, and TileRT’s persistent engine kernels with warp-specialized heterogeneous pipelines. The FP4 checkpoint and DFlash model parameters are open-sourced on HuggingFace.

Why it matters: Inference speed is the bottleneck for AI coding agents — every token of delay compounds when agents write, test, and iterate on code. Breaking 1,000 tok/s on commodity hardware means that even trillion-parameter models can serve coding agents with near-instantaneous response, making agentic coding loops dramatically faster. The open-source release of the FP4 checkpoint also allows any team to replicate this speed on their own infrastructure, democratizing high-speed inference for AI coding.

Source: Xiaomi MiMo Blog | IT之家

6. Perplexity × Harvard: AI Agents Deliver 87% Faster Work at 94% Lower Cost

What happened: Perplexity and Harvard published a research study on the shift from chat interfaces to autonomous AI agents (specifically Perplexity’s Computer agent). Over a three-month study, workers using the autonomous agent completed tasks 87% faster and at 94% lower cost compared to those using only search tools, with higher satisfaction scores. The median speedup was 25x. The paper argues that companies restructuring around agent-augmented individuals will significantly outperform those maintaining traditional team structures.

Why it matters: This is one of the most rigorous real-world studies validating the economic case for autonomous AI agents in knowledge work. The 25x median speedup and 94% cost reduction provide concrete benchmarks for organizations evaluating AI coding agent adoption. For the AI coding space specifically, it validates the “agentic” paradigm over the “copilot” paradigm — agents that execute end-to-end outperform assistants that suggest, by an order of magnitude.

Source: Perplexity Research | FourWeekMBA

7. Amap ABot-Earth0.5: World’s First 3D Native City-Scale World Model for Embodied AI

What happened: Alibaba’s Amap launched ABot-Earth0.5, the world’s first 3D native city-scale world model. Using a single satellite image or text prompt, it generates a kilometer-scale 3D city scene in approximately 10 minutes on consumer-grade GPUs — a ~1,000x efficiency gain over traditional aerial-photogrammetry methods, at roughly 1% of the cost. Output is in editable 3D Gaussian Splatting format, directly importable into Unity and Unreal Engine. Key innovations include a compress-then-generate framework, sliding-window inference for continuous large-area generation, and a multi-level-of-detail decoder. Amap also unveiled “Amap Tutu,” a fully autonomous quadruped robot built on this platform. Beta access is open at abot-earth.amap.com.

Why it matters: Training embodied intelligence systems requires massive, diverse simulation environments. Until now, building city-scale 3D training worlds was prohibitively expensive and slow. ABot-Earth0.5 collapses this from days to minutes, potentially unleashing a new wave of embodied AI training at scale. The direct integration with game engines and the launch of a physical robot on the same platform signals a closed-loop pipeline from simulation to deployment — exactly the infrastructure layer the humanoid robotics industry has been waiting for.

Source: PR Newswire | IT之家

8. WeChat AI Agent Ecosystem Officially Opens: Millions of Mini-Programs at Your Command

What happened: WeChat’s developer team officially published an integration guide confirming the WeChat AI agent is in closed beta. The platform offers two access modes: (1) Automatic mode — WeChat reads the mini-program’s source code and auto-enables AI agent operation of pages with zero additional development; (2) Developer mode — teams build custom skills for the AI agent to invoke after review. Both can be active simultaneously. Separately, WeChat is collaborating with Huawei, Honor, Xiaomi, OPPO, and vivo on A2A (agent-to-agent) capability, enabling phone voice assistants to initiate WeChat calls and messages. This is the first official confirmation of a system that could put AI agents in front of WeChat’s 1.4 billion monthly active users.

Why it matters: This is potentially the largest AI agent deployment surface ever created. WeChat’s 1.4 billion MAU and millions of mini-programs form a ready-made agentic commerce ecosystem — users can now say “order me a coffee” and have the AI agent find, navigate, and complete the transaction across mini-programs. For AI coding, the automatic mode that reads source code and auto-enables agent operation is essentially an AI coding deployment pipeline at platform scale: developers write mini-programs, and WeChat’s AI auto-writes the agent integration layer.

Source: IT之家 | WeChat Official Guide

Quick Takes

• Kimi Code major upgrade — Video understanding, ACP protocol support for JetBrains/Zed, stock data integration via K2.6 model. Kimi Work desktop agent runs up to 300 parallel sub-agents for financial research. (Kimi Official)
• Claude Code GA 1st Anniversary — Boris Cherny and Cat Wu reflect on verification best practices, why auto-mode was built, and what’s next for routines and loops. (X @ClaudeDevs)
• Microsoft AI CEO: Superintelligence is near, won’t take your job — Mustafa Suleyman says Microsoft is now allowed to pursue superintelligence independently, with a dedicated team and 7 new omni-modal models from Build. (The Verge)

EAIDaily — Focused on AI Coding & Embodied Intelligence | @WoLoveAI