Meta Muse 与 AMD 联手押注消费级 agent

Mike @MikeLongTerm原推发布于 2026.10.13

在 X 上打开原推

$AMD & $META are ready to J-curve beyond $1,000 IMO | WINNING Consumer & Enterprise Agentic SI together🧵 Not Financial Advice! DYOR! @AMD and @Meta @Muse are positioned to jointly win the consumer, small business, and enterprise agentic AI race. Meta’s Muse agent now powered by the new Muse Wonder frontier models gives every user a dedicated Secure VM (an isolated cloud computer with browser, terminal, files, and persistent memory). These VMs run on high density AMD EPYC server CPUs (Turin and next gen Venice class chips), while GPUs handle model inference. OpenAI’s Dots follow a similar pattern and also run on EPYC. This architecture is the key, where agents don’t just generate tokens, they orchestrate multi step workflows, use tools, recover from errors, and keep working for hours or days. That persistent runtime layer is CPU bound, and AMD’s core dense EPYC processors are purpose built for packing large numbers of agent sandboxes efficiently. Muse Wonder and Muse Wonder Full are Meta’s first true frontier models from Superintelligence Labs. They upgrade the earlier Muse Spark family with stronger Active Memory (retaining and retrieving user context across days, weeks, and months) and Active Focus (sustained effort, alternative approaches, and error recovery on hard tasks). Meta’s benchmarks show particular strength in personal and professional agentic work, legal document tasks (Harvey LAB), finance analysis, long horizon coding (SWE-Marathon), business automation, and an internal Personal Superintelligence Index; areas that matter far more for everyday users and small teams than pure leaderboard scores. Wonder will roll into the Muse app (subscribers first) and is already available via the Meta Model API and Muse Code. Pricing and access reflect the frontier positioning: Muse Wonder on the Meta Model API is roughly $4 / $20 per million input/output tokens (comparable to high end Claude/GPT pricing); Wonder Full is reserved for Muse Code subscribers (higher tier plans in the ~$50–$200 range). Context is 1M tokens; modalities include text, image, video, and audio input. While a basic single agent Muse session can operate efficiently inside one user’s Secure VM on roughly 2 vCPUs, delegating work to multiple sub-agents multiplies the compute footprint. Muse Code documentation shows concurrent child agent limits that are explicitly CPU derived (defaults around 8 simultaneous agents including the root, configurable higher, with queuing beyond that and support for up to 1,000 child tasks over a workflow’s lifetime). Parallel tool calls, isolated worktrees, independent sub tasks, and long running concurrent execution all increase CPU core demand, memory bandwidth, and scheduling overhead inside the sandbox. At consumer and small business scale this raises the average cores per active user; at enterprise or high volume multi agent deployments it drives the need for denser high core count EPYC flagship hosts and larger rack scale systems (such as AMD’s Helion designs with up to 32,768 cores in a CPU only configuration). The result is a steeper curve of CPU infrastructure growth more sandboxes, more cores per rack, and higher overall utilization to keep multi agent workflows responsive and non-blocking rather than serialized or throttled. This further amplifies the long term demand signal for AMD to innovate higher density server CPUs as Meta’s Muse ecosystem (and similar agent platforms) expands to 100-1,000 agents capabilities in the coming quarters/years! 1. Massive Fuel for AMD CPU Growth (Next 3–5 Years) This is significant. Pure chatbot inference is heavily GPU bound. Persistent personal/professional agents change the mix: each user needs a dedicated (or densely packed) virtual machine with CPU cores, memory, and storage for long running workflows, browser use, code execution, file management, and background tasks, while GPUs handle the model inference. Muse (and OpenAI Dots) already run on AMD EPYC high core density server CPUs. Reports from the initial Muse launch showed AMD EPYC Turin hosts, with agents themselves reporting the hardware. High core counts (128+ class) allow packing many user VMs per physical server. As Muse scales already millions of downloads shortly after launch, with daily actives in the millions and Zuckerberg’s stated ambition toward broader personal superintelligence, the CPU demand multiplies potentially 50-100x. Analysts have highlighted agents as a CPU demand driver; AMD’s data center CPU business has been strong, and Meta is a major existing EPYC customer. Projections around the agent wave make this a multi year tailwind. The 3–5 year horizon aligns with continued growth in personal agents, more complex multi agent workflows, higher utilization of the VMs, and broader geographic/rollout expansion. AMD GPUs remain critical for training and inference, but the agent runtime layer is a distinct, large CPU opportunity that Wonder’s capabilities will help expand by making the agents more useful and sticky. 2. Upgraded Version of Regular Muse / Muse Spark + Continued VM Usage Muse Spark was the deliberate first rung: smaller, faster, capable enough for early agent work but not frontier. Wonder is the larger, more advanced class that “brings us to the frontier,” with better post training gains on stronger base models, stronger out of distribution generalization, and the Active Memory / Active Focus features. It will continue powering (and upgrading) the same Muse agent architecture that launched not long ago. The VM model stays central. Meta’s security write-ups describe each user’s Muse Secure VM as an isolated Linux environment with enough CPU, memory, and storage for real work (code compilation, custom tools, concurrent sub agents, persistent files and memory). Typical reported specs have been ~2 vCPUs and ~8 GB RAM per user on high-density AMD EPYC Turin hosts (9D25-class chips), with inference on separate GPU infrastructure. The design prioritizes isolation, credential protection via Sentinel, and the ability for the agent to act as if it has “a computer under the desk.” Wonder’s improvements in persistence, memory, focus, coding, and tool use make that VM more capable rather than replacing it.⁠ 3. Small Businesses and Personal Agentic Usefulness This is where the combination of frontier agent models + dedicated persistent VM shines for non enterprise users: Personal use: ~Navigate multi step tasks that unfold over time (insurance claims, understanding a lease or medical information, personal financial analysis, calendar/message coordination, shopping or booking with confirmation gates). ~The agent remembers relevant details across sessions and keeps working toward the goal (with check-ins) on its own cloud computer, even when the user’s device is off. ~Lower barrier than pure professional coding agents because of the consumer facing Muse app, free tier with paid upgrades, and deep integration points. Small businesses: ~Frontier coding for building or iterating apps, games, internal tools, or websites (strong SWE Marathon / long project results help here). ~Professional agent work in law (document drafting/review from source materials), finance (analysis, workflows), health-related queries, and business automation. ~A solo founder or small team can offload sustained, multi-step work (research + drafting + tool use + verification) to an agent that maintains context and operates inside a dedicated environment, without standing up their own infrastructure or large staff. ~Muse Code provides a developer facing path with the same models for heavier software workflows. Conclusion: For @finkd , the stakes of the agentic superintelligence race are existential to Meta’s long term vision. As he wrote in his letter on personal superintelligence: “I am extremely optimistic that superintelligence will help humanity accelerate our pace of progress. But perhaps even more important is that superintelligence has the potential to begin a new era of personal empowerment where people will have greater agency to improve the world in the directions they choose.” He has repeatedly framed Meta’s goal as bringing “personal superintelligence to everyone,” with Muse as the vehicle an always available agent that can “understand their goals and can just work on their behalf 24/7.” Muse Wonder’s Active Memory, Active Focus, multi-agent delegation, and Secure VM architecture are the concrete tools to deliver that vision at consumer, small business, and enterprise scale. Falling behind would leave Meta’s massive distribution and product surface dependent on others’ models and infrastructure. For Dr. @LisaSu , the same race is equally decisive for AMD. She has described demand as “very, very high for the next several years,” noting that even after capacity expansions “the demand is even higher than our supply,” and that the company will “substantially increase our supply in 2027.” On the specific impact of agents, Su has been explicit: agentic AI is “increasing the need for server CPU compute” for orchestration, data movement, and parallel execution, making CPU demand “largely additive” to GPU demand. She has observed the traditional CPU to GPU ratio shifting from 1:4 or 1:8 toward 1-4-8 CPU : 1 GPU, and potentially beyond, with “lots and lots of agents” driving more CPUs racks than GPUs. Meta’s Muse (@OpenAI Dots, $MSFT, $AMZN, $GOOGL, @AnthropicAI , $AAPL and similar smaller platforms) running persistent, multi-agent sandboxes on high density EPYC processors turns that thesis into volume. Winning the agentic layer with the right CPUs and rack scale systems (including designs like Helion) locks in years of data center growth; losing it cedes the fastest expanding part of the compute stack. Not Financial Advice! DYOR!

这条是引用着另一条写的：x.com/finkd/status/2109689305073614918