Cheeky Pint · Published 2026-10-12

Muse Wonder: Zuckerberg and Wang's $100B AI

privateReview: unreviewedConfidence: high
multimodalactive memorycompute clustersscaling lawsAI agentsactive-memorycontext-compactionmodel-gatewaycompute-scalingpost-trainingpersonalization
Open on YouTube ↗

Summary

Overview

  • Speaker: Mark Zuckerberg and Alexandr Wang
  • Channel: Cheeky Pint
  • Main topic: Meta's AI strategy, product philosophy, pre-training versus post-training, and scaling law economics
  • Purpose: Provide an in-depth conversation between tech leaders on AI product development, scaling laws, agent architectures, and enterprise vs consumer AI strategy. Mark Zuckerberg and Alexandr Wang discuss the rapid evolution of Meta's AI initiatives, specifically focusing on Muse, Muse Spark, and Muse Wonder. They explore personal AI agents, multimodal capabilities, active memory, preference tuning versus reinforcement learning, compute infrastructure investments, and the philosophy of building consumer-first AI products.

Topic Map

Muse Product Evolution and Use Cases

  • Explanation: Discussion on Muse as the fastest-growing AI agent, personal use cases like sourcing rare teas or coaching martial arts, and the transition from text-based assistants to multimodal agents.
  • Key claims:
    • Muse is the fastest growing AI agent.
    • Multimodal understanding of video, audio, and text is critical for personal AI.
  • Examples:
    • Using Muse to coach MMA by processing gym camera feeds between rounds.
  • Terminology:
    • Muse
    • Muse Spark
    • Muse Wonder
    • multimodal
  • Why it matters: Demonstrates how AI agents are shifting from chat interfaces to active multimodal participants in daily life and training.

Pre-training vs Post-training and Scaling Laws

  • Explanation: Analysis of the balance between pre-training model scale and post-training refinement, highlighting the massive compute clusters built by Meta in Ohio and Louisiana.
  • Key claims:
    • Pre-training requires massive compute and data scale.
    • Post-training and Supervised Fine-Tuning (SFT) shape model personality and alignment.
  • Examples:
    • Promethean and Hyperion clusters scaling past gigawatts of compute.
  • Terminology:
    • pre-training
    • post-training
    • supervised fine-tuning
    • reinforcement learning
    • compute cluster
  • Why it matters: Explains the capital intensity and architectural decisions required to stay at the frontier of foundational AI models.

Active Memory and Personalization

  • Explanation: Exploration of how AI agents maintain long-term memory across days, weeks, and months to provide highly personalized recommendations and interactions.
  • Key claims:
    • Active memory allows agents to recall user preferences and past context.
    • Personalization reduces friction in consumer workflows like e-commerce and planning.
  • Examples:
    • Remembering cooking preferences and ordering ingredients automatically.
  • Terminology:
    • active memory
    • compaction
    • personalization
  • Why it matters: Memory is the bridge between stateless chatbots and genuinely useful, proactive personal agents.

Open Router and Enterprise Partnerships

  • Explanation: Stripe's acquisition of Open Router as a model gateway and how Meta's open and closed model ecosystem interacts with developer infrastructure.
  • Key claims:
    • Open router handles massive token volume and route optimization.
    • Meta models will be accessible via major gateways like Open Router.
  • Examples:
    • Routing over 10 trillion tokens a day through model gateways.
  • Terminology:
    • Open Router
    • model gateway
    • token routing
  • Why it matters: Highlights the developer tooling ecosystem needed to abstract model complexity and pricing for end users.

Key Points

Multimodality is fundamental to consumer AI

  • Explanation: Consumers experience the world through vision, sound, and conversation, so AI models must process all these modalities natively from the ground up.
  • Evidence: Discussion on glasses capturing audio and video feeds to provide real-time agent feedback.
  • Practical implication: AI products must integrate camera and microphone inputs rather than relying solely on text prompts.

Capital expenditure on compute is a competitive moat

  • Explanation: Building frontier AI models requires unprecedented compute clusters like the upcoming 5-gigawatt facility in Louisiana.
  • Evidence: Comparison of cluster sizes from Promethean to Hyperion.
  • Practical implication: Only companies with massive balance sheets and long-term conviction can build frontier base models.

Active memory differentiates agent products

  • Explanation: Traditional LLMs lack persistent memory of user context across sessions, whereas active memory enables proactive and personalized assistance.
  • Evidence: Contrasting standard stateless context windows with active memory compaction systems.
  • Practical implication: Agent software design must prioritize persistent memory architectures over simple prompt-response loops.

Frameworks, Models & Processes

Model Scaling and Alignment Pipeline

  • How it works: A structured approach to building AI starting from massive pre-training on diverse multimodal data followed by targeted post-training and reinforcement learning.
  • Components:
    • Pre-training compute clusters
    • Supervised fine-tuning (SFT)
    • Active memory integration
    • Alignment and personalization
  • When to use: When developing state-of-the-art frontier AI models and consumer agents.

Examples & Case Studies

Mark Zuckerberg set up cameras in his gym to have Muse coach his MMA training between rounds.

  • Illustrates: Real-time multimodal video processing in consumer AI agents.
  • Lesson: AI utility expands dramatically when agents can observe physical environments and provide timely tactical feedback.

Actionable Takeaways

  • Immediate:
    • Multimodal input handling is required for modern agent interfaces.
    • Active memory transforms stateless models into proactive personal assistants.
    • Compute scale remains the primary driver of frontier model capabilities.
  • Strategic:
    • Full-stack control from data center infrastructure to consumer applications yields a defensible advantage.
    • Balancing open ecosystems with proprietary agent experiences drives long-term platform growth.
    • Capital investment in AI infrastructure must be viewed as a multi-decade compounding asset.
  • Questions to investigate:
    • How will active memory scale across millions of concurrent users without privacy degradation?
    • What is the optimal ratio of pre-training compute to post-training fine-tuning for specialized agents?
    • How will model gateways like Open Router impact direct API monetization for model developers?

Claims Worth Verifying

  • Open Router routes over 10 trillion tokens a day. (statistical)
  • Meta's Hyperion compute cluster in Louisiana will scale to 5 gigawatts. (technological infrastructure)

Notable Quotes

"Muse is the fastest growing agent or AI app that I think has existed yet." (at 0:34) "The model is the product." (at 20:02)

Compressed Summary

  • Multimodal AI integration enables vision, voice, and proactive agent workflows.
  • Active memory architecture solves stateless context limitations in LLMs.
  • Massive compute investments in clusters like Hyperion power frontier model scaling.
  • Consumer-first product design focuses on reducing friction and increasing personalization.
  • Keywords: multimodal, active memory, compute clusters, scaling laws, ai agents
  • Core insight: Frontier AI leadership requires combining massive compute infrastructure with multimodal agent architectures and active memory to deliver proactive personal assistance.

Core insights

5
Mechanismmedium noveltymoderate evidence

Active memory is framed as an explicit compaction-based architecture, contrasted with standard stateless context windows, and is the mechanism that turns a chatbot into a proactive agent that recalls preferences across days/weeks/months.

Why it matters

Reframes memory as a first-class subsystem with its own compaction strategy rather than an extension of the context window, changing where persistence, summarization, and recall responsibilities live in an agent runtime.

Generalization

Any long-running agent needs a memory tier that compacts, retains, and re-injects user context independently of the raw context window; prompt-response loops are insufficient for proactive behavior.

Contrasting standard stateless context windows with active memory compaction systems.
Open source video
Active memory allows agents to recall user preferences and past context.
Open source video
Agent software design must prioritize persistent memory architectures over simple prompt-response loops.
Open source video
Architecturemedium noveltymoderate evidence

Multimodality is argued to be native, ground-up (vision, sound, conversation processed from the start) rather than bolted on as text prompts, and is tied to always-on sensors like glasses and gym cameras.

Why it matters

Implies multimodal capability is a training and data-pipeline property, not an inference-layer adapter; bolting vision/audio onto a text-first model is an architectural limitation.

Generalization

Systems intended for real-world perception need modality-native training plus streaming sensor ingestion, which changes data collection, latency budgets, and interface design.

Consumers experience the world through vision, sound, and conversation, so AI models must process all these modalities natively from the ground up.
Open source video
Discussion on glasses capturing audio and video feeds to provide real-time agent feedback.
Open source video
Architecturemedium noveltyweak evidence

A model gateway layer (Open Router) sits between applications and models, abstracting model choice, routing, and pricing at ~10 trillion tokens/day, and is valuable enough that Stripe acquired it.

Why it matters

Positions routing/infrastructure as a distinct product layer separate from model providers, with implications for how developers pick models and how vendors monetize APIs.

Generalization

As model supply fragments, a routing abstraction between apps and models becomes a natural control point for cost, failover, and evaluation, independent of any single model.

Open router handles massive token volume and route optimization.
Open source video
Routing over 10 trillion tokens a day through model gateways.
Open source video
Stripe's acquisition of Open Router as a model gateway
Open source video
Mental Modellow noveltymoderate evidence

Frontier capability is presented as a function of capital-intensive compute clusters scaling from Promethean to Hyperion toward gigawatt (5 GW) class, making full-stack ownership a defensible moat rather than a model-quality-only race.

Why it matters

Shifts competitive analysis from model benchmarks to infrastructure balance sheets, implying fewer viable frontier labs and different build-vs-rent economics.

Generalization

Where capability scales with capex, durable advantage accrues to owners of the full stack (data centers to consumer app), and renters face structural ceiling risk.

Promethean and Hyperion clusters scaling past gigawatts of compute.
Open source video
Only companies with massive balance sheets and long-term conviction can build frontier base models.
Open source video
Full-stack control from data center infrastructure to consumer applications yields a defensible advantage.
Open source video
Practicemedium noveltymoderate evidence

Pre-training and post-training are treated as functionally distinct stages: pre-training buys scale/capability, while post-training and SFT determine personality, alignment, and behavior.

Why it matters

Suggests the 'product' experience is largely a post-training artifact, so teams optimizing agent behavior may get more leverage from SFT/RL data pipelines than from further base-model scale.

Generalization

Model behavior tuning is separable from capability acquisition; teams without pre-training budgets can still differentiate through targeted post-training.

Post-training and Supervised Fine-Tuning (SFT) shape model personality and alignment.
Open source video
Pre-training requires massive compute and data scale.
Open source video
The model is the product.
Open source video

Deep dives

5

Memory compaction as a first-class agent runtime tier

Research question

What compaction, retention, and re-injection policies allow an agent to preserve user preferences across days/weeks/months while staying within bounded prompt budgets and preserving per-user privacy?

Why

If active memory is a subsystem rather than a longer context window, then compaction policy, staleness handling, retrieval, and privacy isolation become core engineering artifacts whose design determines whether an agent can be proactive at all.

Active memory allows agents to recall user preferences and past context.
Open source video
Remembering cooking preferences and ordering ingredients automatically.
Open source video
How will active memory scale across millions of concurrent users without privacy degradation?
Open source video
Source video

Native multimodal training versus bolt-on perception adapters

Research question

For real-time agent tasks driven by continuous camera and microphone streams, does modality-native training measurably beat text-first models augmented with vision/audio adapters in latency and task accuracy?

Why

The claim that models must process vision, sound, and conversation natively from the ground up is asserted without comparative evidence; if it holds, data pipelines, latency budgets, and interface design change for every real-world perception agent.

Consumers experience the world through vision, sound, and conversation, so AI models must process all these modalities natively from the ground up.
Open source video
Using Muse to coach MMA by processing gym camera feeds between rounds.
Open source video
Multimodal understanding of video, audio, and text is critical for personal AI.
Open source video
Source video

Pre-training versus post-training compute allocation for specialized agents

Research question

What ratio of pre-training compute to SFT/RL compute maximizes downstream agent task performance, and does that optimum shift as the target task narrows?

Why

If behavior and alignment are largely post-training artifacts while capability comes from pre-training scale, teams without pre-training budgets can still differentiate through SFT/RL data pipelines — but nobody has published an allocation curve.

Pre-training requires massive compute and data scale.
Open source video
Post-training and Supervised Fine-Tuning (SFT) shape model personality and alignment.
Open source video
The model is the product.
Open source video
Source video

Model gateway economics and the routing control point

Research question

When applications bind to a routing abstraction instead of a model vendor, how are margins, failover behavior, and model-provider monetization redistributed across the stack?

Why

A gateway moving trillions of tokens per day becomes the natural control point for cost, failover, and evaluation independent of any single model, which changes how vendor lock-in and API monetization should be evaluated.

Open router handles massive token volume and route optimization.
Open source video
Routing over 10 trillion tokens a day through model gateways.
Open source video
Meta models will be accessible via major gateways like Open Router.
Open source video
Source video

Gigawatt compute capex as a structural moat

Research question

Does capability at the frontier scale with owned compute capex such that renters face a structural ceiling, and what balance-sheet thresholds define viable frontier participation?

Why

If frontier capability is gated by capital rather than only model quality, competitive analysis shifts from benchmark tables to infrastructure balance sheets and the build-versus-rent decision becomes the primary strategic question.

Promethean and Hyperion clusters scaling past gigawatts of compute.
Open source video
Only companies with massive balance sheets and long-term conviction can build frontier base models.
Open source video
Source video

Article ideas

4

Stop Calling It a Long Context Window: Memory Is a Subsystem

Agents that behave proactively across weeks require a dedicated compaction-and-recall tier with its own retention, staleness, and isolation policies; stretching the context window is a category error that relocates the problem instead of solving it.

Angle

Architecture teardown that maps where persistence, summarization, and re-injection responsibilities should live in an agent runtime, using preference-recalling personal agents as the running example.

Source video

Native Multimodality Is a Data-Pipeline Property, Not an Inference Adapter

Bolting vision and audio adapters onto a text-first model produces a structurally different system from one trained on modalities from the ground up, and streaming sensor interfaces like glasses and gym cameras expose that difference as latency and accuracy debt.

Angle

Contrast training-time and runtime consequences of the two approaches, and argue that streaming ingestion forces latency budgets and frame preprocessing into the core runtime.

Source video

The Gateway Is Where Model Economics Gets Decided

Once a routing layer between applications and models handles trillions of tokens per day, cost optimization, failover, and vendor leverage move to the gateway, so model providers who treat gateways as neutral plumbing will lose pricing power.

Angle

Economic analysis of the intermediary layer, framed around portability, bundling, and first-party API monetization risk.

Source video

Your Product Is a Post-Training Artifact

Since pre-training buys capability while post-training and SFT determine personality, alignment, and observed behavior, teams without frontier pre-training budgets should spend their scarce compute on SFT/RL data pipelines rather than chasing base-model scale.

Angle

Practical argument that behavior tuning is separable from capability acquisition, with allocation guidance and its stated uncertainties.

Source video

Project ideas

4

CompactionBench: Long-Horizon Memory Policy Harness

beyond-evals

A dedicated compaction-and-retrieval tier retrieves user preferences and past context more accurately over multi-week horizons than stuffing raw history into an extended context window, at lower token cost per turn.

Proof of concept

Build a two-arm agent harness — arm A uses flat context stuffing, arm B uses a compaction/retention/re-injection tier — and replay a synthetic multi-week preference-streaming scenario (including cooking preferences and ingredient ordering) across both.

Measurement

Preference-recall accuracy, token cost per turn, staleness-induced error rate, and per-user memory isolation leakage on a held-out probe set.

Source video

Gatehouse Router: Cost, Failover, and Quality Control Point

gatehouse

A routing abstraction that selects models per request on cost, latency, and measured quality can reduce spend at constant task accuracy while eliminating single-vendor binding.

Proof of concept

Stand up a proxy in front of at least three model endpoints that routes by task class, enforces fallback on failure or timeout, and logs per-route quality and cost.

Measurement

Cost per successful task, routing failover success rate, latency distribution, and accuracy delta versus a single fixed model baseline.

Source video

Native-vs-Adapter Multimodal Perception Benchmark

movement-lab

For continuous sensor-driven agent tasks, a modality-native model delivers lower end-to-end latency and higher task accuracy than a text-first model with bolted-on vision and audio adapters of equivalent parameter budget.

Proof of concept

Run both architectures on a streaming feedback loop (e.g., a between-rounds coaching feed from a gym-like camera source plus audio) and compare end-to-end response quality and timing.

Measurement

Time-to-first-useful-feedback, task accuracy scored by human raters, and preprocessing overhead per second of stream.

Source video

Pre/Post-Training Allocation Sweep for Narrow Agents

new

For a specialized agent task, a compute allocation weighted toward post-training SFT/RL reaches target task performance at a fraction of the pre-training-heavy budget, and the optimal split narrows further as the task narrows.

Proof of concept

Sweep pre-training versus SFT/RL compute budgets across a fixed small model family and a single narrow agent task, holding total compute roughly constant.

Measurement

Task success rate versus total compute, and the location of the performance knee across allocation ratios.

Source video

Architectural implications

3

Active memory introduces a persistent state tier alongside the context window.

Before

Agent state lives entirely in a session-scoped context window that resets between sessions.

After

A separate memory subsystem compacts and retains user context across days/weeks/months and re-injects it into prompts.

Consequence

New requirements for compaction policy, retrieval, staleness handling, and privacy controls over long-lived user state.

Source video

Multimodal agents consume continuous camera/mic streams rather than discrete prompts.

Before

Interface boundary is a text prompt in, text out request.

After

Interface boundary is streaming sensor input with real-time feedback loops (e.g., coaching between rounds).

Consequence

Latency budgets, stream handling, and frame/sample preprocessing become core runtime concerns rather than application glue.

Source video

A gateway layer decouples applications from specific model providers.

Before

Applications bind directly to a single model vendor API.

After

Applications route through a gateway that abstracts model selection, pricing, and traffic optimization.

Consequence

Portability and cost optimization shift to the gateway, changing how vendor lock-in and API monetization are evaluated.

Source video

Tradeoffs and failure modes

4

Active memory at scale vs privacy

Benefit

Long-term personalization and proactive assistance from retained user context.

Cost or risk

Scaling persistent memory across millions of concurrent users may degrade privacy guarantees.

How will active memory scale across millions of concurrent users without privacy degradation?
Open source video
Source video

Compute capex as moat

Benefit

Full-stack control and frontier capability locked behind capital barriers.

Cost or risk

Only firms with massive balance sheets can compete, concentrating frontier development and raising capital risk.

Only companies with massive balance sheets and long-term conviction can build frontier base models.
Open source video
Source video

Open ecosystem vs proprietary agent experience

Benefit

Open models plus gateways broaden distribution and developer adoption.

Cost or risk

Tension between openness and capturing value in a proprietary agent product.

Balancing open ecosystems with proprietary agent experiences drives long-term platform growth.
Open source video
Source video

Pre-training vs post-training allocation

Benefit

Post-training/SFT can shape personality and alignment at comparatively lower cost.

Cost or risk

Unclear optimal compute split between pre-training and fine-tuning for specialized agents.

What is the optimal ratio of pre-training compute to post-training fine-tuning for specialized agents?
Open source video
Source video

Open questions

4

How will active memory scale across millions of concurrent users without privacy degradation?

Why unresolved

Summary states the question but offers no mechanism for privacy-preserving persistent memory at scale.

Research direction

Design and benchmark compaction/retrieval schemes with per-user isolation, retention limits, and measurable privacy leakage.

Source video

What is the optimal ratio of pre-training compute to post-training fine-tuning for specialized agents?

Why unresolved

The summary asserts both stages matter but gives no quantitative guidance on allocation.

Research direction

Run ablation studies sweeping pre-training vs SFT/RL compute budgets against downstream agent task performance.

Source video

How will model gateways like Open Router impact direct API monetization for model developers?

Why unresolved

Only the routing volume and acquisition are described; the economic effect on providers is unaddressed.

Research direction

Analyze margin structures and bundling dynamics between gateway intermediaries and first-party model APIs.

Source video

Does native multimodal training measurably outperform adapters bolted onto text-first models for real-time agent tasks?

Why unresolved

The native-from-ground-up claim is asserted without comparative results in the summary.

Research direction

Benchmark latency and task accuracy of native multimodal agents vs text-first models with vision/audio adapters.

Source video

Key claims

7
opinionVerification needed

Muse is the fastest growing AI agent/app yet.

Evidence

Muse is the fastest growing agent or AI app that I think has existed yet.

Question

What metric and time window define 'fastest growing' for Muse, and how does it compare to other agents?

Source video
factualVerification needed

Open Router routes over 10 trillion tokens a day.

Evidence

Open Router routes over 10 trillion tokens a day.

Question

Is Open Router's daily routed token volume actually ~10T, and how is it measured?

Source video
predictionVerification needed

Meta's Hyperion compute cluster in Louisiana will scale to 5 gigawatts.

Evidence

Meta's Hyperion compute cluster in Louisiana will scale to 5 gigawatts.

Question

Has Meta confirmed a 5 GW target for Hyperion, and over what timeline?

Source video
causalVerification needed

Pre-training requires massive compute and data scale.

Evidence

Pre-training requires massive compute and data scale.

Question

What empirical scaling evidence ties pre-training capability to compute/data scale at the frontier?

Source video
causalVerification needed

Post-training and SFT shape model personality and alignment.

Evidence

Post-training and Supervised Fine-Tuning (SFT) shape model personality and alignment.

Question

How strongly do SFT/post-training choices dominate observable personality and alignment outcomes vs pre-training?

Source video
opinionVerification needed

Multimodal understanding of video, audio, and text is critical for personal AI.

Evidence

Multimodal understanding of video, audio, and text is critical for personal AI.

Question

Do multimodal inputs measurably improve personal-assistant task outcomes versus text-only interactions?

Source video
factualVerification needed

Stripe acquired Open Router.

Evidence

Stripe's acquisition of Open Router as a model gateway

Question

Did Stripe actually acquire Open Router, and when?

Source video

Connections

5