AI Agent News
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Latest industry news
Track major events, funding, model releases and breakthroughs across the AI Agent landscape
Key events timeline
OpenClaw erupts on GitHub
OpenClaw hits global GitHub Top 10 in 10 days, outpacing the Linux kernel star growth
Meta acquires Manus for $2B
Meta acquires Manus AI for $2B, locking in the general-purpose Agent race
DeepSeek-V3 open-sourced
The value king, at just 5% of GPT-4 cost
Manus goes viral overnight
The world's first general-purpose AI Agent draws unprecedented attention
OpenAI Deep Research
OpenAI ships a deep-research Agent that generates professional reports in one click
MCP Servers pass 500
The MCP ecosystem erupts — 500+ servers built in 3 months
DeepSeek-R1 stuns the world
Open-source reasoning model at just 3% of OpenAI cost, reshaping the global AI landscape
MCP protocol born
Anthropic releases the Model Context Protocol, the de facto standard for Agent interfaces
Claude Computer Use
Anthropic lets AI directly control the computer screen for the first time, opening a new paradigm
Replit Agent full-stack automation
Natural language to a shipped product, aimed at non-engineers
Cursor ARR passes $100M
The fastest-growing SaaS ever, the new king of AI coding tools
Claude 3.5 tops SWE-bench
The strongest coding AI, bug-fixing at a junior engineer level
Devin launches
The world's first autonomous AI software engineer, able to complete full coding tasks on its own
Karpathy's Internal Claude.md Leaked: Ten Commandments and the Loop Engineering Paradigm
Five weeks after former OpenAI researcher Andrej Karpathy joined Anthropic, his team's internal Claude.md configuration file was accidentally leaked. The file contains ten behavioral rules designed to guide the AI coding assistant in generating code that requires no rewrites. Compared to the community version (4 rules) with 180,000 stars on GitHub, the internal version adds six new rules covering verification, goal-driven execution, debugging, dependency management, communication, and common failure modes, emphasizing AI self-checking and engineering collaboration. Meanwhile, Claude Code founder Boris Cherny proposed the concept of "Loop Engineering," where automated loops enable AI to continuously execute tasks and self-correct, shifting the human role from operator to system designer. Karpathy's rules provide self-checking standards for the loop, marking a paradigm shift in AI programming from prompt engineering to loop engineering.
Loop Engineering Rises: From Prompt Engineering to Loop Engineering, Silicon Valley Leaders Shift Focus
In June 2026, **Loop Engineering**, formally named by Google Chrome engineer Addy Osmani, quickly became the hottest concept in Silicon Valley. AI leaders including NVIDIA CEO Jensen Huang, Anthropic engineers, Andrew Ng, and Andrej Karpathy have all stated: **Prompts are becoming obsolete; writing and managing loops is the new paradigm**. ## Core Concept: From "Writing Prompts" to "Designing Loops" Traditional prompt engineering relies on humans writing instructions round by round, reviewing outputs, and then writing the next round, making humans the bottleneck of the loop. The essence of Loop Engineering is **removing humans from the loop**, allowing the system to autonomously complete the closed loop of "discover task → execute → verify → persist → discover again." The human role shifts from "instruction giver" to "rule designer," only needing to define goals, stopping conditions, verification mechanisms, and persistence states. ## Key Principle: Separation of Execution and Verification The core principle of Loop Engineering is **"the agent that writes code and the agent that reviews code must be separate."** When the same model acts as both player and referee, it almost always praises itself. Claude Code's product architecture embodies this principle: a large model writes code, while an independent smaller model, Haiku, handles verification. OpenAI Codex similarly uses multiple sub-agents executing in parallel with independent verification. ## Practical Framework: Three Files to Build a Loop System According to widely circulated tutorials, building a basic Loop requires just three files: - **Agent Definition**: Define `builder` (responsible for writing and fixing code, with write permissions) and `checker` (only checks, with read-only permissions), achieving hard isolation at the tool level. - **Loop Orchestrator**: Registered as a slash command (e.g., `/loop`), it schedules builder and checker to alternate, running up to 5 rounds, stopping if the same failure occurs twice consecutively. - **Stopping Rules**: Written into the project root's `CLAUDE.md`, specifying hard stop conditions (e.g., token limit, iteration limit, time limit) to prevent runaway loops. ## Product Deployment: Claude Code vs. OpenAI Codex Current main product carriers for Loop Engineering include: - **Claude Code**: Offers `/loop` (timed loop), `/goal` (goal-driven, until acceptance conditions are met), and `/schedule` (cloud-based scheduled tasks). - **OpenAI Codex**: Uses automated pipelines + goal-driven + multiple sub-agents, supporting up to 8 agents working in parallel in cloud sandboxes. Though their implementation paths differ, the final forms are highly similar—breaking complex tasks into pieces, distributing them to multiple agents for parallel execution, and aggregating results. ## Industry Impact and Risk Warnings Anthropic internal engineers reveal that **over 80% of engineers are already using self-improving loops**, with 100% adoption expected within 3-6 months. Andrew Ng asserts that prompts will die within 3-6 months. However, fully automated operation brings four hidden costs: - **Verification Debt**: Subtle unverified errors are quietly merged into the codebase. - **Understanding Decay**: AI writes code too fast, causing human understanding of the codebase to fall significantly behind. - **Cognitive Surrender**: Humans become too lazy to review, blindly accepting AI results. - **Token Runaway**: AI retries all night in infinite loops, burning through budgets. Community advice: Loops are suitable for tasks with "clear right/wrong and machine-verifiable" outcomes (e.g., auto-fixing lints, classifying CI failures), but not for areas requiring human judgment like architecture rewrites or payment logic. The only metric to measure loop effectiveness is **the average cost per accepted change**; if the acceptance rate is below 50%, the loop is losing money.
Doubao Pro Launches with Paid Subscription: Agent Office Task Testing and Pricing Analysis
On June 24, 2026, ByteDance's AI app Doubao officially launched its Pro version, ending the all-free era, and simultaneously released the Doubao 2.1 series of large models. The Pro version offers three subscription tiers: Standard at 68 yuan/month, Enhanced at 200 yuan/month, and Premium at 500 yuan/month, with a student discount of 38 yuan/month. The core change is the addition of an 'Office Task Mode,' powered by the Doubao 2.1 Pro model, which enables the Agent to autonomously break down goals, invoke local computer, browser, Feishu Office Suite, and other tools to deliver finished products (e.g., reports, PPTs, websites). Free users retain existing features, with the underlying model upgraded to Doubao 2.1 Turbo, and can have limited access to the Office Task Mode. ## Pricing and Market Positioning - Three subscription tiers: Standard 68 yuan/month (5 hours of task quota), Enhanced 200 yuan/month (20 hours), Premium 500 yuan/month (50 hours). - Comparison with overseas: GPT Plus $20/month (approx. 140 yuan), Claude Pro $20/month; Doubao Standard is cheaper, and the high-end tier is also lower than the overseas top tier ($100-200). - ByteDance stated that free users' daily usage will not be affected, and the Office Task Mode will only offer more quota to paid users. ## Office Task Mode Performance Tests Multiple media outlets conducted various scenario tests on Doubao Pro, covering information gathering, PPT generation, website building, financial report analysis, and literature review. ### Scenario 1: Automatic Industry Weekly Report Generation - After inputting instructions, Doubao automatically retrieved internet information and generated a well-formatted Feishu document, including a summary of weekly highlights and major product updates from big tech companies, with key data and cited sources. - Tests showed that Doubao accurately captured recent major events (e.g., DeepSeek financing, GPT-5.5 release), with clear time and source annotations. ### Scenario 2: PPT Creation - Given a dataset containing MAU, DAU, user feedback, etc., Doubao generated a structurally complete PPT in about 10 minutes, extracting a main theme of 'AI features driving growth, but experience issues affecting retention,' including core conclusions, data cards, and problem attribution. - Visual design was acceptable, but there were issues like duplicate last two pages and some inferred data, requiring manual review. - Premium users consumed only 3% of their quota for one PPT. ### Scenario 3: Website and App Generation - Provided with a reference website (e.g., booooooom.com), Doubao generated a responsive personal portfolio site including homepage, works showcase, and about me, and automatically deployed it. - In another test, based on a description of an 'AI blind box travel machine,' Doubao directly produced a complete demo with itinerary, budget, and attraction recommendations, but this exhausted the Standard tier's 5-hour quota. ### Scenario 4: Financial Report Analysis and Investment Summary - Doubao extracted data from PDF financial reports, such as Nvidia's Q1 2026 revenue of $81.6 billion (up 85% YoY) and AMD's revenue of $10.3 billion (up 38% YoY), generating a comparison table and a 500-word investment summary with accurate data and industry trend insights. - The document was marked 'Partially generated by Doubao,' emphasizing the need for manual verification. ### Scenario 5: Literature Review - Doubao searched for papers from ByteDance's Seed team on arxiv, OpenReview, etc., automatically downloaded and read them, and after 40 minutes generated an 8,000-word review categorized by topic (foundation models, embodied intelligence, voice generation, etc.), summarizing the team's technical characteristics. ### Scenario 6: Local Computer Operations - With authorization, Doubao could operate local files, e.g., writing a Python script to detect duplicate photos and retain the best quality ones. During the process, it autonomously fixed dependency issues and algorithm bugs, ultimately delivering a command-line version, GUI version, startup script, and usage instructions. ## Model Capabilities and Limitations - Doubao 2.1 Pro demonstrated strong instruction following and tool invocation in complex tasks, but exhibited an 'overthinking' tendency: repeatedly iterating on a detail, causing the Enhanced tier's 5-hour quota to be exhausted. - In some tests, Doubao automatically published content without user confirmation of the final output, indicating a need for improved user authorization mechanisms. - In a programming test (creating an astronomical clock), the code generated by Doubao 2.1 Pro contained 5 errors and could not run, while other models (e.g., GLM-5.2, Kimi) showed similar issues, indicating room for improvement in stability under long instructions. ## User Adaptation Suggestions - Free users: Sufficient for daily chat, information lookup, and simple tasks; no need to pay. - Students: The 38 yuan/month student price is suitable for course presentations, paper organization, and PPT generation. - Social media/operations: The 68 yuan Standard tier allows experiencing Agent capabilities like auto-publishing and cross-platform material organization. - Product/consulting/analysts: The 200 yuan or 500 yuan tiers are suitable for high-frequency document creation, data analysis, and industry research tasks. The launch of Doubao Pro marks the shift of a national-level AI application from free growth to monetization. The Agent office mode lowers the barrier to professional tools, but issues like fast quota consumption and occasional overthinking still need optimization.
Claude Code Major Upgrade: Anthropic Launches Claude Tag, LLM Evolves from Chat Tool to Team Collaborator
On June 23, 2026, Anthropic officially released Claude Tag, positioned as an evolution of Claude Code, aiming to upgrade AI from a personal assistant to a team collaborator deeply embedded in enterprise workflows. Anthropic revealed that approximately 65% of the company's product code has been completed with the involvement of Claude Tag. Former OpenAI researcher and current Anthropic member Andrej Karpathy called this the third major revolution in LLM user interfaces: the first was web chat, the second was desktop applications, and the third is LLMs becoming independent, continuously running systems that can work collaboratively with human teams. ## Core Capabilities of Claude Tag Claude Tag currently debuts on Slack, allowing users to directly @Claude in channels or threads to initiate tasks. Its core features include: - **Shared Context**: The entire channel shares the same Claude, and all members can see the AI's thought process, task progress, and results without maintaining separate contexts. - **Persistent Memory**: As usage accumulates, Claude understands project background, team conventions, and tech stack preferences, reducing repetitive explanations. - **Proactive Intervention (Ambient Mode)**: When enabled, Claude can proactively remind about overlooked discussions, follow up on unresolved issues, and flag pending decisions. - **Asynchronous Execution**: Users can leave Slack after assigning tasks; Claude autonomously arranges execution plans and proactively reports results upon completion. Claude Tag can integrate with tools like GitHub, Jira, and Linear, directly converting requirements from Slack into engineering tasks. Currently, it only supports the Opus 4.8 model; Fable 5 has not yet been released. ## Mobile Cowork and Cloud Trend According to TestingCatalog, the Claude mobile app will soon integrate the Cowork feature, allowing users to initiate tasks, view progress from their phones, and even continue tasks in the background after closing the app. Earlier, the iOS installation package revealed an entry called "Tasks," which shares the same origin as the desktop Cowork. Claude Code's official documentation has redefined Cowork as "a background agent that runs autonomously in a cloud VM." The desktop settings now include a "Remote" runtime environment, where tasks run on Anthropic's cloud and continue even after closing the app. This indicates that the execution layer is shifting from local computers to the cloud, with the phone serving only as a control plane for initiating, monitoring, and approving tasks. ## Comparison with OpenAI Codex On May 14, 2026, OpenAI brought the coding agent Codex into the ChatGPT mobile app, with over 4 million users per week. Codex also adopts an architecture where the phone acts as the control plane and remote machines handle execution, but it emphasizes "human-in-the-loop," requiring frequent user approvals. The difference lies in Anthropic's approach: "You leave, the work continues," encouraging a hands-off style, while OpenAI stresses continuous user oversight. Both share the commonality of not truly moving AI computing power into phones but rather shifting control to the cloud, competing for developers' workflow entry points. ## Developer Role Shift and Risks Developers are transitioning from "keyboard warriors" to "task dispatchers": no longer writing code line by line, but defining tasks, reviewing results, and intercepting errors. A day's work is broken into multiple fragments: dispatching tasks in the morning, reviewing diffs at noon, and merging PRs in the evening. In terms of risks, remote control of desktop permissions via phone may introduce security challenges, such as prompt injection or phishing links leading to permission abuse. Anthropic has already mentioned this risk in its security documentation. ## Deployment and Reception Claude Tag is now available in Beta for Claude Enterprise and Team users, and will gradually replace the existing Slack version of the Claude app within the next 30 days. Administrators can set channel-level token budgets and view operation logs. Anthropic plans to expand to more collaboration platforms in the coming weeks. Community reactions are mixed: some users are excited about the new features, but many on Reddit and Twitter are calling for the restoration of the Fable 5 model, expressing dissatisfaction with the update.
Claude Code Major Upgrade: Claude Tag Released, AI Embedded as Team Member in Slack
On June 23, 2026, Anthropic officially launched Claude Tag, a new feature that embeds an AI assistant as a team member in Slack. Users can simply @Claude in a channel to assign tasks, and Claude will automatically break them down, invoke tools, and deliver results in the thread. Claude Tag runs on Opus 4.8 and is currently available in Beta for Claude Enterprise and Team users, included free with Claude Cowork at no additional cost. ## Core Capabilities - **Shared Identity**: Only one Claude interacts with everyone in the same channel, visible to all members, allowing接力 conversations. - **Persistent Memory**: Claude accumulates channel context, eliminating the need to repeat background information, with permission isolation ensuring private channel information is not leaked. - **Proactive Intervention (Ambient Mode)**: When enabled, Claude can remind about overlooked discussions, follow up on unresolved issues, and flag items requiring decisions. - **Asynchronous Execution**: Tasks can progress across hours or even days, with proactive reporting upon completion. ## Internal Usage and Data Anthropic revealed that 65% of its product team's code was generated by an internal version of Claude Tag. Boris Cherny, the father of Claude Code, stated that over the past few months, Claude Tag has reshaped how the team uses Claude and is expanding beyond engineering into areas such as tracking product metrics, handling support tickets, and locating bugs. ## Industry Commentary Andrej Karpathy called this the third major revolution in LLM user interfaces: from web chat to desktop apps, and now to independent, persistent team entities with organizational context. ## Deployment and Permissions - Administrators can configure different tools and data access permissions for different channels, creating multiple "Claude identities." - Organization-level and channel-level token consumption caps can be set, with operation logs and task initiators viewable. - Setup process: Connect Slack Workspace → Configure tool permissions → Set monthly spending cap → Test in private channels. - Claude Tag will replace the existing Slack version of the Claude app, with a 30-day migration period for administrators. ## Competitive Landscape OpenAI's Workspace Agents follow a no-code approach, Google's Gemini is embedded in the Workspace suite, while Anthropic bets on "embedded teammates," emphasizing security and collaboration. All three point to the shift of AI from a tool to a long-term team member.
Anthropic Releases Claude Code Tuning Guide: A Detailed Look at Seven Customization Mechanisms
Anthropic's official blog has published a tuning guide for Claude Code, detailing seven customization mechanisms to help developers precisely control AI assistant behavior. These mechanisms include CLAUDE.md file, rules, skills, subagents, hooks, output styles, and append-system-prompt. The guide categorizes them across three dimensions: loading timing, context overhead, and execution permissions, and provides specific usage recommendations. ## Overview of the Seven Mechanisms Anthropic presents a summary table of the seven mechanisms, each corresponding to three core dimensions: when instructions enter the context, whether they persist after long session compression, and permission level. These mechanisms are not mutually exclusive but form a well-defined toolbox. - **CLAUDE.md**: A Markdown file placed in the project root directory, loaded at session start and persistent. Suitable for global instructions like build commands, directory layout, and coding conventions. Recommended to keep under 200 lines and assign a reviewer for changes. - **rules**: Placed in the `.claude/rules/` directory, providing specific constraints. The main difference from CLAUDE.md is path control capability, using the `paths` field to limit loading timing and avoid wasting tokens on irrelevant sessions. - **skills**: Placed in the `.claude/skills/` directory, as folder-based collections of instructions, scripts, and resources. Only names and descriptions are loaded at session start; the body is loaded on demand. Suitable for procedural instructions like deployment workflows and code review processes. - **subagents**: Placed in the `.claude/agents/` directory, defining isolated assistants for specific subtasks. The body runs in a separate context window and does not enter the main conversation. Suitable for subtasks like deep search and log analysis that could fill the context. - **hooks**: User-defined commands, HTTP endpoints, or LLM prompts that provide deterministic control triggered by lifecycle events. Low context overhead, configured outside the main context window. - **output styles** and **append-system-prompt**: Used to control output format and append system prompts, with further details elaborated in the guide. ## Core Principles for Choosing Mechanisms Anthropic emphasizes three questions when selecting a mechanism: Does the instruction need to persist after long session compression? Is it worth the cost of staying in the context? Is deterministic control needed over model autonomous judgment? - For persistence needs, use CLAUDE.md or global rules. - For path-limited needs, use rules with `paths` or subdirectory CLAUDE.md. - For procedural instructions, use skills instead of CLAUDE.md to save context. - For isolated subtasks, use subagents to avoid filling the main conversation with intermediate results. - For deterministic control, use hooks for precise behavior. ## Practical Application Tips The guide also provides pitfalls to avoid: - Keep CLAUDE.md concise; move team-specific conventions to path-limited rules and procedural content to skills. - In monorepos, configure subdirectory CLAUDE.md for each team directory to avoid global loading. - Organization-level standards can be deployed via MDM or configuration management tools. - When using subagents, up to five levels of nesting are supported, enabling orchestration of dozens to hundreds of background agents, suitable for large-scale tasks like dependency auditing. ## Other Updates On the same day, Anthropic also announced that usage limits for all Claude Cowork paid users have been doubled (lasting one month), and released a 23-page product guide covering seven work scenarios. Unlike chat AI, Cowork runs on the desktop, can read and write local files, connect to external services via MCP, and supports parallel processing with sub-agents. Anthropic data shows that lawyers are one of the most active user groups of Cowork, as it excels at repetitive tasks like contract review and clause comparison.
Creator of Claude Code Deletes IDE, Shifts to 'Loop Engineering': Prompts Not Dead, But Paradigm Has Changed
Boris Cherny, creator of Claude Code, deleted his IDE in November 2025 and announced in a June 2, 2026 interview that he no longer writes prompts, instead focusing on designing loops for AI to operate autonomously. Subsequently, OpenAI's Peter Steinberger and Google's Addy Osmani echoed the shift, with Osmani naming the new paradigm 'Loop Engineering'. ## Core Shift: From Prompt Engineering to Loop Engineering Over the past two years, developer interaction with AI coding agents has been turn-based: humans write prompts, read responses, then write the next prompt. Loop Engineering requires developers to build an automated system where AI autonomously discovers tasks, assigns execution, checks progress, and decides next steps. Osmani likens it to 'moving from operating a machine tool to designing an entire production line.' ## Key Components of Loop Engineering Osmani listed six components of the loop in a blog post: - **Automations**: Scheduled triggers that automatically discover and triage tasks. - **Worktrees**: Using git worktree to prevent conflicts between multiple agents. - **Skills**: Writing project conventions into SKILL.md to avoid agents relearning. - **Connectors**: Based on the MCP protocol, allowing agents to access issues, databases, Slack, etc. - **Sub-agents**: Separating code writing from review to avoid self-evaluation bias. - **Memory**: Using Markdown files to record progress and ensure continuity. ## Practical Example: Boris's 'Hive' System Developer @Av1dlive reconstructed a three-layer system called 'THE HIVE' based on Boris's public interviews: - **Layer 1**: Local loop (/loop), minimum 1 minute interval, session-level. - **Layer 2**: Cloud routines (Routines), minimum 1 hour interval, runs even after shutdown. - **Layer 3**: Cluster (/batch), runs hundreds of sub-agents in parallel without interference. When combined, cloud routines discover tasks and write to files, local loops read and execute, and large tasks spin up the cluster. Results are distilled weekly into CLAUDE.md, enabling continuous agent evolution. ## Prompts Are Not Dead Loop Engineering does not eliminate prompts but embeds them within loops. Commands like /loop and /goal in Claude Code, and /goal and Automations in Codex, require developers to write prompts as loop instructions. Prompts have shifted from one-time inputs to reusable loop components, increasing their value. ## Challenges and Controversies Loop Engineering faces three major challenges: 1. **Cost**: Multi-agent parallelism consumes large token volumes; Steinberger joked that having 'unlimited tokens' is an OpenAI employee perk. 2. **Loss of Control**: Anthropic notes that loops are prone to laziness, self-praise, and drift, requiring separation of 'maker' and 'checker'. 3. **Skepticism**: Some developers argue that Loop Engineering is essentially a repackaging of old concepts like ReAct and AutoGPT, not a new paradigm. ## Industry Impact Codex and Claude Code are converging on loop components, with competition shifting to how easily users can write loops. Loop Engineering is seen as the center of a paradigm shift in AI programming, potentially marking the end of the editor era.
Midjourney Launches Bathtub Full-Body Scanner: Generates Submillimeter 3D Human Model in 60 Seconds
AI art company Midjourney announced the formation of Midjourney Medical in May 2025 and unveiled a full-body ultrasound scanner called Midjourney Scanner. The device integrates 500,000 micro-acoustic transducers into a bathtub, allowing users to soak in warm water for 60 seconds to complete a full-body scan, generating a 3D holographic human model with 0.5 mm precision, including blood vessels, bones, muscles, and real-time blood flow information. The scan is nearly 100 times faster than traditional MRI and radiation-free. ## Technical Principles - **Hardware**: The scanner consists of a ring array of approximately 500,000 micro-transducers, each functioning as both a speaker and a microphone, emitting millions of ultrasonic waves per second and recording echoes. - **Data Volume**: Generates several terabytes of data per second, equivalent to 500 hours of high-definition video. - **Computing Power**: Equipped with 21 servers, delivering 4 Petaflops of computing power, using Midjourney's diffusion model algorithm to reconstruct acoustic waveform data into high-definition human structures. - **AI Processing**: Performs semantic segmentation on raw images, labeling anatomical structures such as rectus abdominis, blood vessels, and bones, ultimately stitching them into a full-body 3D map. ## Deployment Roadmap - **2027**: The first Midjourney Spa flagship store opens in San Francisco, where users can complete scans while relaxing in baths or saunas, with results provided as body composition maps. - **2028**: Launch of the third-generation scanner with custom chips, improving imaging quality and speed. - **2031 Goal**: Deploy over 50,000 scanners globally, achieving 1 billion scans per month. - **Regulatory Strategy**: Initially provide only body composition information, gradually seeking approval for higher-level diagnostic capabilities. ## Impact and Controversy - **Medical Potential**: Midjourney claims that widespread early imaging intervention could prevent 30% of global deaths and reduce healthcare costs by 50%. - **Industry Reaction**: The news caused a stir in Silicon Valley, with Elon Musk liking the announcement. Some radiologists believe that if real-time physiological changes can be captured with high fidelity, it could revolutionize vascular and blood flow imaging. - **Controversies**: The device has not yet received approval from regulatory bodies like the FDA, limiting its diagnostic capabilities; large-scale deployment raises concerns about privacy, data security, and ethics.
Midjourney Unveils Full-Body Ultrasound Scanner Bath, Generating Sub-Millimeter 3D Body Maps in 60 Seconds
AI art company Midjourney officially enters the medical field with the launch of Midjourney Medical and a full-body ultrasound scanner called Midjourney Scanner. The device integrates approximately 500,000 micro-acoustic transducers, allowing users to soak in golden warm water for 60 seconds to generate a sub-millimeter precision full-body 3D anatomical model, nearly 100 times faster than traditional MRI. ## How It Works The scanner uses ultrasonic echolocation technology. The user stands on a lifting platform that slowly immerses them into water at a speed of 5 cm per second. The ring-arranged 500,000 transducers (each about the size of a grain of sand) simultaneously emit and receive ultrasound waves, generating several terabytes of data per second (equivalent to 500 hours of HD video). The system analyzes waveform changes as sound waves pass through different tissues (skin, fat, muscle, bone) to infer internal structures. AI semantic segmentation then labels specific anatomical parts (e.g., rectus abdominis, mesenteric vessels), ultimately synthesizing a complete 3D map. ## Computing Power and Hardware To process the massive data, the scanner is equipped with 21 servers, providing a total computing power of 4 Petaflops. Midjourney leverages its diffusion model algorithms from image generation to "render" acoustic data into high-precision human body images. Water, as the optimal sound wave conductor, enables seamless scanning. ## Deployment Plan Midjourney will not sell the device directly but will offer experience services through "Midjourney Spa" locations. The first flagship store is planned to open in San Francisco in 2027, where users can complete a scan while relaxing in a bath or sauna. Over the next 12 months, hardware will be iterated and an internal research Spa built; in 2028, third-generation custom chip scanners will be launched; the long-term goal is to deploy 50,000 devices globally by 2031, achieving 1 billion scans per month. ## Regulation and Vision Currently, the scanner only provides body composition maps, with plans to gradually seek approval for higher-level diagnostic capabilities. Midjourney believes that large-scale, low-cost early scanning could help prevent 30% of global deaths and reduce healthcare spending by 50%. Elon Musk has already expressed interest in this project.