PJFP.com

Pursuit of Joy, Fulfillment, and Purpose

Tag: Richard Sutton

  • SubQ 1.1 Small Explained: How Subquadratic Sparse Attention Hits 98% Retrieval at 12 Million Tokens With 64.5x Less Compute Than Dense Attention

    Subquadratic, a frontier AI research and infrastructure company, has released the model card and technical report for SubQ 1.1 Small, a long-context language model built on a new attention mechanism the company calls Subquadratic Sparse Attention (SSA). The headline claim is unusual in two directions at once: the model retains 98% single-fact retrieval accuracy at 12 million tokens, roughly twelve times the length it was primarily trained on, while cutting attention compute by 64.5x against dense attention at a 1 million token context. The deeper argument in the report is not really about a single model at all. It is about what happens to the entire retrieval-and-orchestration stack once reasoning over a complete artifact stops being prohibitively expensive.

    TLDR

    SubQ 1.1 Small is a small long-context model that replaces the dense attention of an existing open-weight frontier model with Subquadratic Sparse Attention, a learned, content-dependent sparse attention mechanism that scales linearly in compute and memory rather than quadratically. On retrieval it posts 99.12% on NVIDIA’s 13-task RULER suite at 128K tokens and 100% needle-in-a-haystack accuracy at 1M and 2M tokens, holding at 98% out to 6M and 12M tokens while attending to only 0.13% of token pairs. It keeps competitive general ability, scoring 85.4% on GPQA Diamond and 89.7% pass@4 on LiveCodeBench v6, and reaches 13% on the long-horizon AutomationBench Finance agentic benchmark, close to Opus 4.8 and GPT-5.5 and well ahead of mid and small tiers. The efficiency story is a scaling win rather than a constant-factor one: 64.5x fewer attention FLOPs than dense attention at 1M tokens and 56x faster than FlashAttention-2 on a single attention layer. The report frames cheap long-context compute as a research accelerator that let the team run more than one hundred million-token experiments and find a training recipe (long-context continued pretraining is the strongest lever) rather than guess at one, positions SSA against FlashAttention, DeepSeek’s Lightning Indexer line, state space models like Mamba, and hybrids, invokes Sutton’s Bitter Lesson to argue that RAG, chunking, and agentic scaffolding are partly workarounds for context scarcity, and was independently verified by Appen. Deployment is starting with design partners now, with a 2M to 12M token lineup planned by year end.

    Thoughts

    The most interesting move in this report is the framing, not the benchmark. Subquadratic plants its flag on Richard Sutton’s Bitter Lesson and argues that much of the modern AI stack, the retrieval pipelines, the chunkers, the re-rankers, the agentic orchestration, is scaffolding built around a single computational constraint: dense attention costs grow with the square of context length. If that constraint relaxes, a lot of hand-engineered machinery that exists to feed a model the right fragments at the right moment starts to look like the task-specific pipelines that learned representations eventually displaced. That is a genuinely provocative thesis, and it is the right lens for reading the rest of the document. The company is not selling a longer context window as a feature. It is betting that whole-artifact reasoning is a different shape of capability than retrieval over fragments, and that fragmentation destroys the cross-references a contract or a codebase actually depends on before the model ever sees them.

    The part of the paper most teams will undervalue is the claim that the real payoff of efficient attention is not cheaper inference but cheaper experimentation. A dense long-context training campaign is expensive enough that most groups get a handful of attempts and are forced to guess at the recipe. Subquadratic says SSA let them run more than a hundred experiments across six model generations with per-step iteration under a minute at million-token context, which is how they discovered that long-context continued pretraining, not clever post-training, was the dominant lever. If that holds, algorithmic efficiency becomes a first-class scaling variable alongside parameters and data, because capability becomes responsive to iteration velocity rather than raw compute alone. It reframes efficiency from a deployment line item into a research multiplier, and that is a more durable advantage than any single benchmark number.

    The generalization result deserves scrutiny precisely because it is so clean. A model trained overwhelmingly at 1M tokens, with a sliver at 2M and nothing beyond, holds 98% retrieval at 12M. The proposed explanation is that SSA routes attention by content relevance rather than fixed positional pattern, so there may simply be no obvious length boundary once the routing behavior is learned. That is plausible and the report is careful to say the 12M result emerged rather than being designed for. But single-needle NIAH is a deliberately clean probe with one target and a binary answer. The far harder RULER suite is only reported at 128K, the longest standardized length in the original benchmark, so the multi-hop, aggregation, and distractor-heavy capability that whole-artifact reasoning actually requires has public numbers at 128K, not at 12M. The honest read is that precise retrieval generalizes spectacularly and composite reasoning at extreme length is still an open question the report does not over-claim on.

    What lends the report credibility is how much counter-evidence it volunteers. It walks through MiniMax abandoning its hybrid M1 architecture and returning to full attention for M2 after efficient variants showed multi-hop reasoning deficits at scale. It admits that earlier SubQ checkpoints improved retrieval while regressing on knowledge benchmarks, forcing dedicated capability-balancing work. It describes catching a case where the MRCR benchmark moved up while the model felt worse in real workflow spot-checks, and switching its development signal to RULER as a result. That last point is a quietly important methodological argument: benchmark score and deployment behavior diverged enough to change checkpoint selection, which is a warning every team shipping long-context models should internalize. A vendor confident enough to show where its own metrics misled it is more trustworthy than one that only shows the wins.

    A few caveats keep the enthusiasm grounded. AutomationBench Finance at 13% is genuinely strong relative to peers, but it is a low absolute score across the board, including for GPT-5.5 at 18% and Opus 4.8 at 16%, so this is early evidence of agentic transfer rather than proof of a finished agent. The efficiency comparisons isolate a single attention layer rather than full end-to-end model throughput, which is the right way to expose the scaling shape but not the same as a wall-clock serving benchmark. The model is built from an unnamed donor open-weight frontier model, so some of its general-knowledge and coding strength is inherited rather than created here. And the most aggressive claims about the future, a 2M to 12M lineup and much higher sparsity, are roadmap, not released artifacts. None of that undercuts the core result. It just means the right posture is to treat SubQ 1.1 Small as a strong proof of concept for an architecture that, if it scales as advertised, could quietly remove a layer of the AI stack that everyone currently takes for granted.

    Key Takeaways

    • SubQ 1.1 Small is a long-context language model from Subquadratic AI, built on a new attention mechanism called Subquadratic Sparse Attention (SSA), released June 16, 2026 alongside a model card and technical report.
    • SSA is a learned, content-dependent sparse attention mechanism that scales linearly in both compute and memory with sequence length, rather than quadratically like dense attention.
    • The central result is context-length generalization: the model was trained primarily at 1M tokens, with some training at 2M and none beyond, yet retrieval held far past the training window.
    • Needle-in-a-haystack accuracy is 100% at 1M and 2M tokens and 98% at both 6M and 12M tokens, roughly twelve times the primary training length.
    • At 12M tokens the model attends to only 0.13% of token pairs, close to a 1,000x reduction in attention relationships, while still retrieving accurately.
    • On NVIDIA’s 13-task RULER benchmark at 128K tokens, SubQ 1.1 Small scores 99.12%, with the remaining errors concentrated in aggregation-style tasks rather than retrieval.
    • RULER tests beyond single-fact lookup: single-key and multi-key retrieval, common-word and frequent-word extraction, and multi-hop variable tracing across positions.
    • At 1M tokens, SSA requires 64.5x fewer attention FLOPs than dense attention (3.9 PFLOP versus 252 PFLOP per attention layer).
    • On a single attention layer, SSA runs 56x faster than FlashAttention-2 at 1M tokens (966 ms versus 54,164 ms on an H100), reaching parity near 16K tokens and pulling away as context grows.
    • The efficiency gain is a scaling-law win, not a constant-factor speedup: the advantage over dense attention grows as context length increases.
    • On general knowledge, SubQ 1.1 Small scores 85.4% on GPQA Diamond (pass@1), below GPT-5.5 (93.2) and Opus 4.8 (92), near Sonnet 4.6 and GPT-5.4-mini (87.5), and above GPT-5.4-nano (81.7) and Haiku 4.5 (67.2).
    • On coding, it reaches 89.7% pass@4 on LiveCodeBench v6, close to the absolute frontier (GPT-5.5 92, Opus 4.8 92.2) and ahead of the smaller tiers.
    • On AutomationBench Finance, a long-horizon agentic benchmark, it scores 13%, close to Opus 4.8 (16%) and GPT-5.5 (18%) and ahead of Sonnet 4.6 (8%), Haiku 4.5 (3%), and GPT-5.4-mini (0%). Absolute scores are low across all models.
    • The model was not trained from scratch. The team converted an existing open-weight frontier model by replacing dense attention with SSA, then built long-context ability through staged context extension and continued pretraining.
    • Context was extended in stages (262K, 512K, 1M, 2M) using YaRN positional scaling, with long-context continued pretraining performed between extension stages on naturally long data: books, long documents, and repository-scale code.
    • Roughly one trillion tokens of continued pretraining were performed, most of it at the 1M-token stage.
    • Long-context continued pretraining was the most consistent predictor of long-context retrieval gains across the experiments, more so than post-training tweaks.
    • The team ran more than one hundred long-context experiments across six major model generations, which the report argues is only possible because SSA made million-token iteration cheap (under a minute per step).
    • Capability balance was a recurring challenge: gains in long-context retrieval often regressed short-context knowledge and reasoning unless training was explicitly managed for both.
    • Benchmark scores and real deployment behavior diverged. The MRCR benchmark moved up while qualitative workflow spot-checks got worse, so the team switched its primary development signal to RULER.
    • The report frames RAG, chunking, summarization, and agentic orchestration as scaffolding built around context scarcity, drawing an analogy to Sutton’s Bitter Lesson, where hand-engineered mechanisms get displaced by larger-scale learning.
    • SSA is positioned against FlashAttention (a memory optimization that does not change quadratic compute), fixed-pattern sparse attention, DeepSeek’s learned sparse line, state space models, and hybrid architectures.
    • DeepSeek’s Lightning Indexer (used in DSA and CSA) is the closest published comparison. Its quadratic scoring overtakes the sparse attention it feeds around 52,000 tokens, reaching roughly 16x the attention cost at 1M and 190x at 12M.
    • State space models like Mamba achieve linear cost through a compressed fixed-size state, but that compression is lossy and weakens exact retrieval, which is why production efficient models are usually hybrids with some dense attention layers retained.
    • MiniMax is cited as a cautionary case: it moved from a hybrid M1 to a full-attention M2 after hybrids showed multi-hop reasoning deficits at scale and less mature supporting infrastructure.
    • The benchmark results were independently verified by Appen, a third-party evaluation firm.
    • The named use cases are financial analysis and due diligence, legal and contract work, and software engineering (architecture-level reasoning, cross-file refactoring, dependency tracing, planning, review, and long-horizon memory).
    • Sparsity settings were deliberately conservative, tuned for maximum context length rather than maximum sparsity. Limited experiments at 4x the sparsity reported positive early results.
    • The training infrastructure used a memory-scaling ladder: single node, intra-node sequence parallelism, CPU offload, multi-node sequence parallelism, nested offloading, and Ring Attention for the longest contexts.
    • Beyond about 8M tokens, BF16 numerical underflow and stability became practical constraints on evaluation.
    • The technical report is authored by Saul Ramirez, Alex Whedon, Ashmal Vayani, and Phong Vo of Subquadratic AI.
    • Deployment is starting with a first cohort of design partners, with broader rollout through the quarter and a general model lineup ranging from 2M to 12M tokens by the end of the year.
    • The company’s framing line is “Efficiency is intelligence,” and its broader thesis is that the point is not bigger context windows for their own sake but reasoning directly over complete artifacts with less surrounding scaffolding.

    Detailed Summary

    The problem: whole-artifact reasoning and context scarcity

    The report opens by naming a class of tasks it calls whole-artifact reasoning: problems whose structure requires reasoning across a complete artifact rather than over isolated fragments. A legal agreement may define a term on page 2, qualify it on page 12, carve out an exception on page 46, and amend it in a schedule. A function may be defined in one file, called from forty others, and constrained by invariants encoded in the architecture rather than in comments. A financial review may require connecting filings, earnings reports, contracts, and internal records. In each case the difficulty is not locating a passage, it is reasoning over relationships distributed throughout a large artifact. Most production systems do not do this directly. They rely on retrieval pipelines, chunking, summaries, and agentic workflows that partition information and reconstruct fragments at inference time, because dense attention scales quadratically with context length and makes direct reasoning over large artifacts expensive. Subquadratic argues that much of the modern AI stack is therefore designed to manage context scarcity rather than reason over complete artifacts, and it connects this to Sutton’s Bitter Lesson: sophisticated hand-engineered mechanisms historically get displaced once larger-scale learning becomes practical.

    What SSA is and the three requirements it targets

    Subquadratic Sparse Attention is a content-dependent sparse attention mechanism designed to satisfy three requirements at once, a combination the report argues prior approaches never achieved in a practical long-context system. First, dense-attention-level retrieval and reasoning quality, which requires routing that is content-dependent (determined by the tokens themselves) rather than driven by a fixed positional pattern. Second, subquadratic scaling, where selection, retrieval, and attention are each linear in sequence length so the mechanism is linear end to end, not only within the attention read. Third, full-context training with standard autoregressive generation, so the model can optimize over the entire context during training while keeping efficient token-by-token decoding at inference. The internal mechanism by which SSA achieves this is held back as outside the scope of the report, which focuses instead on the requirements and the experimental program that followed.

    Where SSA sits among prior approaches

    The background section is effectively a taxonomy of long-context modeling. FlashAttention is treated not as a competitor but as the standard dense-attention baseline: it solved the memory problem by never materializing the full attention matrix, but it left the quadratic compute cost untouched, so doubling context still quadruples attention computation. Fixed-pattern sparse attention (sliding-window, strided, as in Longformer, BigBird, and the sliding window in Gemma) scales well but sacrifices content-dependent routing and tends to fail on retrieval benchmarks like RULER. Compression methods like Multi-head Latent Attention reduce KV-cache memory at inference but do not change the quadratic prefill cost. Learned sparse attention, exemplified by DeepSeek’s Native Sparse Attention and its Lightning Indexer, learns where to route but pays a quadratic cost in the indexer itself. State space models and linear attention (Mamba, Mamba-2 and Mamba-3, RetNet, RWKV, gated delta networks) achieve linear cost through a compressed fixed-size state, but that compression is lossy and weak on exact retrieval. Hybrids (Jamba, Kimi Linear, Qwen3 Next, Nemotron) keep a few dense layers to preserve retrieval, which means the quadratic component still dominates at long context. System-level workarounds (RAG, agentic frameworks, recursive language models) move retrieval outside the model entirely. The report’s stated open problem is to combine subquadratic scaling end to end with content-dependent retrieval, arbitrary-position access, and practical ultra-long-context training in one system, which it claims no widely deployed architecture provides and which SSA targets.

    Training: conversion, staged context extension, and continued pretraining

    Rather than training from scratch, the team converted an existing open-weight frontier model that supported a 262K-token context by replacing its dense attention with SSA. They then extended the context window in stages (262K to 512K to 1M to 2M) using YaRN to rescale positional representations, performing long-context continued pretraining between extension stages rather than jumping straight to the final length. The training mixture emphasized naturally long data such as books, long documents, and repository-scale code, packed to the target length with document separators and without masking cross-document attention boundaries. Most continued-pretraining tokens were trained at the 1M-token stage, with roughly one trillion tokens total. Post-training played a separate role: shaping how the long-context capability was expressed while preserving reasoning, coding, and instruction following. The team explored sample-level loss aggregation to keep a few extremely long examples from dominating gradient updates, and staged the post-training corpus across synthetic retrieval tasks, long-context reasoning, coding, educational material, and general instruction following, alternating capability-building phases with recovery phases.

    Results: retrieval, knowledge, coding, and agentic tasks

    On retrieval, SubQ 1.1 Small scores 99.12% on the 13-task RULER average at 128K, with errors concentrated in aggregation-style tasks like common-word and frequent-word extraction. On needle-in-a-haystack, evaluated on 50 held-out UUID samples per length, it scores 100% at 1M and 2M (within the training window) and 98% at 6M and 12M (held out), attending to only 0.13% of token pairs at 12M. On knowledge, GPQA Diamond pass@1 is 85.4%, landing between the small and mid frontier tiers and confirming that long-context optimization need not sacrifice reasoning, a result the report credits to its capability-balancing stages after earlier checkpoints showed retrieval gains coming at the cost of knowledge. On coding, LiveCodeBench v6 pass@4 is 89.7%, and the report notes coding data played a dual role, also improving non-code long-context retrieval because code is dense with the cross-position dependencies that train general routing. On long-horizon agentic work, AutomationBench Finance is 13%, where agents must discover the right endpoints among roughly 500 across 47 applications, make interdependent API calls, follow layered business rules, and ignore seeded distractors, graded on binary end-state correctness with no partial credit.

    Efficiency and the DeepSeek comparison

    Efficiency is measured on one attention layer against a dense baseline on the same backbone. Per-forward-pass attention FLOPs scale from a 2.1x reduction at 32K to 8x at 128K, 31.5x at 512K, and 64.5x at 1M tokens (3.9 PFLOP for SSA versus 252 PFLOP for dense). Measured against FlashAttention-2 in isolation, SSA reaches parity near 16K tokens and pulls away to 56x at 1M, where it runs in 966 ms versus 54,164 ms on an H100. The report devotes a discussion section to DeepSeek’s sparse attention line as the closest published comparison. DeepSeek’s Lightning Indexer is a learned selector, but it is a full-attention distilled transformer, so it scales quadratically: in a V3.2-style configuration the indexer is cheaper than the sparse attention it feeds only below about 52,000 tokens, then overtakes it, reaching roughly 16x the attention cost at 1M tokens and 190x at 12M. SSA targets that same selection role with a selector the report says is dramatically cheaper and linear throughout, and notes SSA could conceptually replace the selector over either uncompressed or compressed representations.

    Efficiency as a research accelerator and the evaluation lessons

    A recurring theme is that the most valuable effect of cheap long-context compute was on the research loop, not just inference. Where a dense campaign would allow a handful of attempts, SSA enabled more than a hundred experiments across six model generations with per-step iteration under a minute at million-token context. That throughput is what surfaced the finding that long-context continued pretraining is the strongest lever, and it leads the authors to argue that algorithmic efficiency should be treated as a first-class scaling variable alongside model and dataset size. The report is unusually candid about evaluation pitfalls. It describes how the MRCR benchmark diverged from deployment behavior, with MRCR-optimized checkpoints often feeling worse on repository-scale code reasoning, multi-document synthesis, and contract analysis, which pushed the team to rely on RULER and a fixed set of qualitative workflow spot-checks as development signals. It also cites MiniMax returning from a hybrid M1 to a full-attention M2 as evidence that reducing asymptotic cost is not sufficient on its own if retrieval quality, reasoning at scale, and system maturity are not preserved at the same time.

    Implications, availability, and what comes next

    The report’s deployment argument is that the most important enterprise implication of long-context models is not larger windows but the ability to reason directly over complete or more-complete artifacts, moving retrieval, re-ranking, and orchestration logic into the model where the task is naturally whole-artifact rather than naturally decomposable. It is careful not to declare retrieval obsolete: for corpora larger than any plausible context window, fast-changing knowledge, and genuinely multi-stage workflows, RAG and orchestration remain the right tools. The narrower claim is that the class of scaffolding that exists only to compensate for context limits gets smaller as efficient long-context models extend the reachable window. The benchmark results were independently verified by Appen. Subquadratic is deploying SubQ 1.1 Small with a first cohort of design partners now, with broader rollout through the quarter and a general lineup spanning 2M to 12M tokens planned by the end of the year, and it flags much higher sparsity as future work.

    Notable Quotes

    “Much of the modern AI stack is therefore designed to manage context scarcity rather than reason over complete artifacts directly.”

    SubQ-1.1-Small Technical Report, framing retrieval and orchestration as workarounds for an architectural limit

    “The hybrid has moved the line, but not changed its shape.”

    SubQ-1.1-Small Technical Report, on why hybrid models keep their quadratic component at long context

    “A routing mechanism intended to make long context affordable becomes the dominant long-context cost, reintroducing quadratic scaling after providing scalar compute savings.”

    SubQ-1.1-Small Technical Report, on DeepSeek’s Lightning Indexer overtaking the attention it feeds

    “If the cost of long-context experiments is too high, teams are forced to guess at the recipe. If the cost falls far enough, they can search for it.”

    SubQ-1.1-Small Technical Report, on efficient attention as a research accelerator

    “Fragmentation systematically destroys those relationships before the model ever sees them.”

    SubQ-1.1-Small Technical Report, on why chunking hurts whole-artifact reasoning

    “Holding the whole artifact in context changes the shape of the task rather than only the speed of it.”

    SubQ-1.1-Small Technical Report, on the difference between bigger windows and direct reasoning

    “The value of SSA is therefore not only that it makes long-context inference cheaper. It makes long-context experimentation cheaper.”

    SubQ-1.1-Small Technical Report, conclusion

    Read the full SubQ 1.1 Small technical report and model card here.

    Related Reading

    • Subquadratic (subq.ai) the company behind SubQ 1.1 Small and the Subquadratic Sparse Attention architecture, where you can join the waitlist.
    • The Bitter Lesson by Richard Sutton the short essay whose argument the report leans on, that hand-engineered mechanisms lose to general methods that scale with computation.
    • Attention Is All You Need the original Transformer paper that introduced the dense attention whose quadratic cost SSA is built to remove.
    • RULER (arXiv) NVIDIA’s long-context benchmark that the report uses as its primary retrieval signal, and that fixed-pattern sparse methods historically struggle with.
    • Retrieval-augmented generation (Wikipedia) background on the RAG approach that the report frames as scaffolding around context scarcity rather than a permanent fixture.

  • Gavin Baker on Orbital Compute, TSMC, Frontier AI Models, Anthropic’s Vertical Take Off, and the Coming Wafer Shortage

    Gavin Baker, founder and CIO of Atreides Management, returns to Patrick O’Shaughnessy’s Invest Like the Best for his sixth appearance. He calls the current AI moment the most extraordinary moment in the history of capitalism, walks through what Anthropic’s vertical takeoff in revenue actually means, lays out why orbital compute is closer than skeptics believe, dissects the TSMC bottleneck that may be the only thing standing between today’s market and a full-on AI bubble, and rates every hyperscaler on how they have positioned for a world where frontier model providers may stop selling API access altogether.

    TLDW

    Anthropic added eleven billion dollars of ARR in a single month, which is roughly the combined business of Palantir, Snowflake, and Databricks built over a decade. That is the setup. From there Gavin Baker covers the March and April selloff, the contrarian read that a closed Strait of Hormuz was actually bullish for American manufacturing competitiveness, why Anthropic and OpenAI multiples may be misleadingly cheap on an unconstrained run rate basis, why Elon Musk’s discipline on SpaceX valuation created a superpower of permanent access to capital, the practical engineering case for orbital compute as racks in space rather than Pentagon sized space stations, why TSMC’s capacity discipline is the single most important variable in whether the AI cycle becomes a bubble, what Terafab in Texas changes, why the Pareto frontier of AI models has flipped from Google dominance to Anthropic and OpenAI dominance in nine months, the shift from all you can eat AI subscriptions to usage based pricing and what that means for revenue scaling, Richard Sutton’s bitter lesson as the largest risk to the AI trade, why frontier tokens still capture an overwhelming share of economic value, the role of continual learning as the third great open question, why most new chip startups should not try to build a better GPU, why Cerebras did something different and hard, why disaggregated inference may extend GPU useful lives to ten or fifteen years and rescue the private credit industry, why being in the token path is the new venture filter, the new prisoner’s dilemma around releasing frontier models via API, an honest rating of Google, Meta, Amazon, and Microsoft, why personal safety is becoming a real AI era risk, and why he remains an AI optimist maximalist who believes this could be the next Pax Americana.

    Key Takeaways

    • Anthropic added eleven billion dollars of ARR in one month, more than the combined businesses of Palantir, Snowflake, and Databricks built across a decade. There is no precedent for this in the history of capitalism.
    • The SaaS and cloud revolution created between five and ten trillion dollars of value over twenty years. AI is replaying that compression on a timeline measured in months.
    • The March selloff was a drawdown driven by disagreement with price action, not invalidated thesis. That is the kind of drawdown an investor can lean into.
    • Deep Seek Monday in January 2025 was a similar setup. By the day of the selloff, AWS Asia GPU prices had already doubled, GPU availability had fallen, and it was obvious reasoning models would be vastly more compute hungry at inference. The market priced the opposite.
    • The Strait of Hormuz closing was actually positive for America. US natural gas (the primary input into US electricity, which feeds AI) fell twenty percent on Bloomberg while Asian and European natural gas doubled or tripled. American manufacturing competitiveness improved overnight.
    • The US is now the world’s largest producer and exporter of oil and gas. The economy is dramatically less energy intensive than in the 1970s. The shortage trauma comparison does not hold.
    • Tech as a sector traded as cheaply versus the rest of the market in early April as at any point in the last ten years, into the single most bullish moment for AI fundamentals on record.
    • Anthropic is dramatically more capital efficient than OpenAI, having burned roughly eighty percent less to reach a similar revenue scale. They have very different structural returns on invested capital.
    • Anthropic at roughly nine hundred billion for fifty billion of ARR (growing a thousand percent) is striking. Adjusted for compute constraint, the unconstrained run rate could be one hundred fifty to two hundred billion, putting the implied multiple closer to five times.
    • Claude Opus generates roughly seventy percent fewer tokens for the same question than previously, with token quantity tied to answer quality. Subscribers on flat-fee plans are getting a lobotomized model.
    • Elon Musk’s superpower is twenty years of making investors money. He never pushes valuation. SpaceX compounded low thirty percent per year for a decade because Musk treats fair pricing as a sacred covenant.
    • Capitalism will solve the watts shortage. The current bottleneck has shifted from chips and energy to zoning and political approval. Many capex decisions are paused until after the US midterms.
    • The watts shortage probably begins to alleviate in 2027 and 2028. Orbital compute solves it longer term.
    • Orbital compute is not Pentagon sized data centers in space. It is racks in space. A Blackwell rack is three thousand pounds, eight feet tall, four feet deep, three feet wide. SpaceX has shown a satellite roughly that size.
    • The satellites operate in sun synchronous orbit so solar wings (around five hundred feet per side) always face the sun and the radiator on the dark side always points to deep space.
    • Starlink V3 satellites already run at around twenty kilowatts. A Blackwell rack runs at one hundred kilowatts. SpaceX engineers express genuine confidence they have already solved cooling and radiator design at these scales.
    • Racks in space are connected with lasers traveling through vacuum, the same lasers already on every Starlink. SpaceX operates the world’s largest satellite fleet and, via xAI Colossus, the world’s largest data center on Earth.
    • Inference will move to orbit. Training will stay on Earth for a long time. Terrestrial data centers remain valuable for the rest of an investor’s career.
    • The wafer bottleneck is structural and political. TSMC is essentially Taiwan’s GDP, water, and electricity. The leaders see themselves as inheritors of Morris Chang’s sacred legacy and they do not behave like a Western public company.
    • Jensen Huang has never had a contract with TSMC. The relationship is run on handshakes and the assumption that things will be fair over time.
    • If TSMC did everything Jensen wanted, Nvidia could be selling two to three trillion dollars of GPUs in 2026 and 2027. TSMC’s discipline is the single largest factor preventing a true AI bubble.
    • Historically, foundational technologies always get a bubble. Railroads, canals, the internet. The current AI buildout is overwhelmingly funded out of operating cash flow, GPUs are running at one hundred percent utilization, and that is fundamentally different from the year 2000 fiber overbuild.
    • If one of Intel or Samsung Foundry catches up at the leading node, the other will follow, and TSMC’s discipline collapses. Watch TSMC capacity decisions to predict a bubble.
    • Terafab, the SpaceX and Tesla joint venture to build the world’s largest fab in America, has a partnership with Intel that grants access to fifty years of institutional foundry knowledge. The A teams at ASML, KLA, Lam Research, and Applied Materials will follow Elon’s reputation in hardware engineering.
    • The hiring playbook for Terafab includes building Taiwan Town, Japan Town, and Korea Town next to the fab. Recruit the engineers and import their families, their restaurants, and their staff.
    • Frontier tokens still capture an overwhelming share of all economic value created at the model layer. This is surprising and is one of the three big open questions for AI investing.
    • The Pareto frontier of intelligence versus cost has flipped. Nine months ago Google’s TPU dominated every point on the frontier. Today Anthropic and OpenAI dominate, with Grok 4.3 on the frontier and Gemini 3.1 hanging on.
    • Google’s conservative TPU V8 design (partly an attempt to reduce dependence on Broadcom and Nvidia) is the leading explanation for the loss of per token cost leadership.
    • AI pricing is shifting from all you can eat to usage based, mirroring the cellular and long distance industries. Cellular stopped being a great growth industry when it went all you can eat. AI just made the opposite move.
    • OpenAI and Anthropic together could exceed two hundred billion in ARR this year if compute keeps coming online and frontier token pricing holds.
    • The two hundred fifty dollar a month consumer AI plan is no longer enough to evaluate frontier capability. Enterprise plans with usage based billing are required because rate limits are now severe.
    • The three biggest open questions for AI investors are: violation of the bitter lesson via ASI or human ingenuity, whether frontier tokens keep commanding their premium, and when continual learning arrives.
    • Today’s continual learning is crude reinforcement learning during mid training on verifiable tasks. True continual learning means weights updating dynamically, like a human who learns the first time they touch fire.
    • Trying to build a better GPU is a losing strategy. Jensen will copy any one to three percent share design. Startups should target one percent share, do something different, and make it hard enough that Nvidia cannot fast follow.
    • Disaggregated inference (separating prefill and decode) opens new design canvases. Prefill is memory capacity bound. Decode is memory bandwidth bound. Each can be optimized independently.
    • Cerebras did something different and hard with wafer scale computing. Three generations of chips and real grit to get there.
    • Disaggregation of inference may stretch GPU useful lives to ten or fifteen years, dropping financing costs from low sevens to five or six percent, mathematically lowering the cost of the AI buildout and likely saving the private credit industry from its SaaS loan exposure.
    • Sellers of shortage outperform buyers of shortage. But owning the largest installed base of what is currently in shortage (hyperscaler CPU fleets, for example) is also a strong position.
    • Most of the economic value at the application layer of AI has been destroyed, not created. The exceptions are companies in the token path or in niches small enough that frontier labs ignore them.
    • Coding may be the shortest path to ASI. If you can write code, you can write code that does anything. Cursor, Cognition, and Anthropic correctly focused on it.
    • Jensen could probably get close to the frontier with his own Nemotron family of models whenever he wants. The fact that he chooses not to is a strategic decision about not commoditizing his customers.
    • The new prisoner’s dilemma in AI is whether frontier labs release their best model via API. If everyone agrees not to, Chinese open source falls behind. If anyone defects, the defector pulls ahead on revenue and resources, forcing everyone else to defect.
    • Google still owns the largest compute installed base. Without TPU’s prior cost advantage, this matters more. YouTube data has real value in a world of robotics. GCP is going crazy.
    • Meta deserves credit for becoming AI first internally faster than any other internet giant. Musa, their first MSL model, is impressively close to the Pareto frontier.
    • Amazon is strong because of Trainium and robotics driven retail P&L efficiency. Nova is better than it gets credit for.
    • Microsoft flinched on capex in early 2025 and lost position. Satya Nadella’s current decision to use Microsoft compute for Microsoft products rather than reselling to OpenAI is a courageous and probably correct call, even at the cost of an eight hundred dollar stock price.
    • The hyperscalers most engaged with startups are Amazon and Nvidia by a mile, followed by Google. Broadcom is the favorite ASIC partner. AMD, Microsoft, and Meta have minimal startup engagement and that will cost them as the best teams are now at startups.
    • Personal safety in an AI era requires a family or company safe word that cannot be socially engineered. Deepfake voice and video extortion at the speed of FaceTime is already feasible.
    • Ukraine is winning largely on the back of having the best battlefield AI outside America and Israel. Adversaries are starting to internalize what AI dominance means geopolitically.
    • An optimistic read is that this becomes a new Pax Americana, the way the post 1945 American nuclear monopoly was used to rebuild Germany and Japan rather than dominate.
    • AI cured a friend’s daughter’s rare disease by spinning up a research effort that identified a market drug capable of impacting her condition. That is the upside that keeps Gavin an AI optimist maximalist.

    Detailed Summary

    The most extraordinary moment in the history of capitalism

    Gavin’s framing of the current moment is unusually direct. Anthropic added eleven billion dollars of annual recurring revenue in a single month. The three highest profile SaaS companies of the last decade plus, Palantir, Snowflake, and Databricks, took a decade and tens of thousands of employees collectively to build the combined business that Anthropic added in thirty days. He has been investing through every major tech cycle and says there is no historical analog. Not the dotcom era, not the cloud transition, not mobile. This is its own thing.

    The market response, then, was peculiar. The NASDAQ sold off into the single most bullish moment for AI fundamentals on record. Tech traded at roughly its widest discount versus the rest of the market in a decade. Investors who said they wished they had bought into AI during 2022, during COVID, or during Deep Seek Monday got the same valuation setup again in early April, this time with an even clearer inflection.

    Why the Strait of Hormuz closing was secretly bullish for America

    One reason the macro fear in March may have been mispriced is that the same geopolitical event that drove the selloff was, in practice, a relative benefit to the United States. American natural gas, the input into American electricity, which is the input into American AI training and inference, fell roughly twenty percent. Asian and European natural gas prices doubled or tripled. The US emerged with sharply improved relative manufacturing competitiveness, which is exactly what the current administration cares about.

    The 1970s comparison does not hold. The US economy is dramatically less energy intensive, it is now the world’s largest producer and largest exporter of oil and gas, and there are no shortages, only price moves. That backdrop made it easier for disciplined investors to stay focused on AI fundamentals through the volatility.

    Anthropic and OpenAI valuations on an unconstrained run rate

    Anthropic at roughly nine hundred billion for fifty billion of ARR sounds rich until you adjust for the fact that the company is severely compute constrained. Gavin estimates that, unconstrained, Anthropic might be at one hundred fifty to two hundred billion in run rate revenue, putting the implied multiple closer to five times. He also points out that Claude Opus now generates roughly seventy percent fewer tokens for the same question than it used to. Token quantity correlates with answer quality, and Anthropic is rate limiting and shrinking outputs to ration capacity across its user base.

    Anthropic and OpenAI are also structurally very different. Anthropic has burned around eighty percent less cash than OpenAI to reach a comparable revenue scale. That implies very different long term returns on invested capital, though OpenAI has done a better job locking in compute and Sarah Friar is one of the most exceptional CFOs Gavin has worked with.

    Why neither lab is raising at a three trillion dollar valuation

    The answer Gavin gives is that both labs are deliberately leaving valuation on the table the way Elon has done for two decades. SpaceX compounded at low thirty percent annually for a decade because Elon never pushed price. The result is a permanent superpower of access to capital. Investors trust him because they have made money with him for twenty years. That is a moat that compounds with every round.

    Anthropic could probably raise at a one hundred percent premium to its rumored latest mark. They are choosing not to. In an uncertain world (Ukraine, Russia, Iran, Taiwan), preserving the ability to raise more capital later at fair prices is more valuable than maximizing this round.

    Watts and wafers, the two real constraints

    Capitalism is solving the watts problem. The leading PE infrastructure investors now say zoning and political approval, not chips or energy, are the gating factors. Companies are deferring big capex announcements until after the US midterms. Turbine capacity is being doubled at the manufacturers. Companies like Boom Aerospace are repurposing jet engines for grid use. Watts probably ease meaningfully in 2027 and 2028 and then orbital compute does the rest.

    Wafers are the harder problem because they live in Taiwan, run on handshakes, and depend on a corporate culture that does not respond to public market incentives. TSMC is essentially the GDP, water consumption, and electricity consumption of Taiwan. Its leadership treats the company as the legacy of Morris Chang. The Silicon Shield doctrine is real and internal.

    Orbital compute as racks in space

    The biggest mental update Gavin asks listeners to make is to stop picturing data centers in space as Pentagon sized space stations. A Blackwell rack is three thousand pounds and roughly the size of a refrigerator. SpaceX has shown a concept satellite of about that size. Solar wings extend five hundred feet to each side and the radiator extends hundreds of feet behind, both possible because the orbit is sun synchronous and the orientation is fixed relative to the sun.

    SpaceX engineers Gavin has spoken to at Starbase express genuine confidence that they have solved cooling at these power levels. They have. Starlink V3 satellites already operate at twenty kilowatts. A Blackwell rack is one hundred kilowatts. The same company operates the world’s largest satellite fleet and the world’s largest data center on Earth via xAI Colossus. The racks are connected to each other with lasers traveling through vacuum, technology already deployed in every Starlink. The naysayers, Gavin observes, are armchair skeptics and Larry Ellison’s response (he is out there landing rockets, no one else is) is the right frame.

    Terafab in Texas and the threat to TSMC’s discipline

    Terafab, the SpaceX and Tesla joint venture, intends to be the largest fab in the world. The partnership with Intel grants access to fifty years of foundry institutional knowledge, allowing Terafab to start three to five quarters behind the leading node rather than fifteen years behind. The A teams at the semicap equipment companies (ASML, KLA, Lam Research, Applied Materials) will follow Elon’s reputation in hardware engineering the same way they followed TSMC twenty years ago when Intel stumbled.

    The talent strategy is the part most observers underestimate. Recruit the best engineers globally, then import their families, their restaurants, their staff. Build Taiwan Town, Japan Town, and Korea Town next to the fab. Optimize the human experience for the people whose work matters. Intel and Samsung do not think that way.

    Bubble watch and the year 2000 comparison

    Every foundational technology in modern history has had a bubble. Railroads, canals, the internet. Carlota Perez documented why. Markets correctly identify the importance, diversity of opinion collapses, supply gets ahead of demand, the bubble crashes. The current cycle has two important differences. The buildout is overwhelmingly funded out of operating cash flow, not debt. Every GPU is running at one hundred percent utilization, while at the peak of the fiber bubble ninety nine percent of fiber was unused.

    TSMC discipline is the single largest reason a bubble has not formed. If Jensen could buy everything TSMC could theoretically make, Nvidia could sell two to three trillion dollars of GPUs in 2026 and 2027. At some point that becomes more than the market can absorb. If Intel or Samsung Foundry catches up at the leading node, the other will too. TSMC’s pricing discipline collapses and the bubble starts.

    The Pareto frontier and the loss of Google’s cost advantage

    The most important chart in AI is the Pareto frontier of model intelligence versus per token cost. Nine months ago, Google’s TPU based models dominated every point on it. OpenAI, Anthropic, and xAI sat inside the frontier. Today the frontier is dominated by Anthropic and OpenAI, with Grok 4.3 on the frontier and Gemini 3.1 hanging on by subsidization more than economics. The most likely cause is Google’s conservative TPU V8 design, an attempt to reduce dependence on Broadcom and Nvidia that sacrificed per token economics.

    The bitter lesson, frontier tokens, and continual learning

    Three open questions dominate AI investing. The first is whether Richard Sutton’s bitter lesson (more compute beats human algorithmic cleverness) gets violated by ASI itself optimizing for efficiency. Closer observers of AI are more skeptical of a violation. Gavin thinks ASI’s first move will be to make itself more efficient and more resourced, which is technically a temporary violation.

    The second is whether frontier tokens keep capturing the overwhelming share of economic value at the model layer. Today they do, surprisingly. Gemini 3.1 Pro was mindblowing nine months ago and is intolerable today. The third is when continual learning arrives. Today’s models need a million fire touches to learn what a human learns from one. True continual learning would mean dynamic weight updates in real time and would produce a fast takeoff.

    From all you can eat to usage based AI pricing

    AI is shifting from flat fee plans to usage based pricing. The historical analogy is cellular and long distance. Both stopped being great growth industries when they went all you can eat. AI just made the opposite move. The consequence is that flat fee subscribers, even on premium consumer plans, get a rate limited and token throttled version of the frontier model. Enterprise plans with usage based billing are now required to evaluate true capability. Gavin thinks the combination of new compute coming online and usage based pricing is what gets OpenAI and Anthropic past two hundred billion in combined ARR this year.

    Chip startups, prefill decode disaggregation, and Cerebras

    Trying to build a better GPU is the wrong move. The four scaled players (Nvidia, AMD, Trainium, TPU) have copy capability for any one to three percent share design that looks attractive. The good news for startups is that disaggregated inference (separating prefill and decode) opens a richer design canvas. Prefill is memory capacity bound. Decode is memory bandwidth bound. Each can be optimized independently. Andrew Fox’s analogy is a British naval ship of the eighteenth century. Prefill is loading the cannon. Decode is firing it.

    Cerebras is the model. Wafer scale computing is genuinely different and genuinely hard. It took three generations of chips to get right. Andrew Feldman and his team had the grit to keep going through chip one being a failure. The design has a high ratio of on chip compute and memory relative to shoreline IO, which is why Cerebras is now experimenting with putting an optical wafer on top of the compute wafer to solve scale out.

    GPU useful lives and the rescue of private credit

    One of the strongest claims in the conversation is that disaggregated inference will stretch GPU useful lives to ten or fifteen years. The skeptical narrative (GPUs are obsolete in two years, companies are cooking their depreciation books) is wrong. You can put a Cerebras system or Groq LPU in front of older Hopper or Ampere parts, use them only for prefill, and run them until they physically melt. Private credit, which is in pain from SaaS loans and which underwrote GPU loans on three to four year lives, may be saved by this.

    If GPU financing rates can come down from low sevens to five or six percent, the mathematics of the AI buildout improves materially. That is a structural tailwind that compounds for years.

    The application layer, the token path, and a new prisoner’s dilemma

    Trillions of dollars of value have been destroyed at the application layer, not created. Cursor and Cognition are the rare scaled exceptions, and they got there by focusing on coding very early. As Amjad Masad noted, coding is plausibly the shortest path to ASI because a coding agent can write itself into any new domain. Jamin Ball’s frame is that the new venture filter is whether the company is in the token path. Data Bricks is. Most application layer startups are not.

    Jensen could probably get close to the frontier with Nemotron whenever he wants, and the strategic question of whether to do that is a new prisoner’s dilemma. If every frontier lab agrees not to release best models via API, Chinese open source falls steadily behind. If anyone defects, the defector gains revenue and resources, and everyone else has to defect. The same dynamic exists between TSMC, Intel, and Samsung. If Nvidia or AMD ever truly used an alternative foundry, that foundry would catch up rapidly.

    Rating the hyperscalers

    Google has the largest compute installed base, the YouTube data that matters in a robotics world, and a search business that prints. Their loss of TPU cost leadership is the surprise of the year. If Google IO in five days does not produce a leapfrog model, the Nvidia centric narrative gets even stronger.

    Meta deserves real credit. Zuckerberg made Meta AI first internally faster than any other internet giant, paid up for the talent contracts when no one else would, and shipped Musa as a first model from MSL that is close to the Pareto frontier. Amazon is well positioned on Trainium, robotics in retail, and a Nova model line that is better than it gets credit for. Microsoft flinched on capex in early 2025 and lost position. Satya Nadella’s current decision to use Microsoft compute for Copilot rather than reselling to OpenAI is courageous and probably correct, even at the cost of stock price.

    The most interesting cross hyperscaler metric is startup engagement. Nvidia and Amazon engage deeply with startups. Google is next. Broadcom is the favored ASIC partner. AMD, Microsoft, and Meta have minimal startup engagement, which Gavin believes will cost them as the best teams now sit at startups.

    Personal safety, geopolitics, and the Pax Americana case

    The closing section turns darker. Personal safety in an AI era requires a family or company safe word that cannot be socially engineered. Deepfake voice and video extortion via something that looks exactly like your child calling on FaceTime is already feasible. Political violence against AI leaders is a real concern. Geopolitically, Ukraine is winning largely because it has the best battlefield AI outside America and Israel. How adversaries respond to that asymmetry is the next great variable.

    Gavin’s optimistic frame is the Pax Americana. After 1945 the US had a nuclear monopoly and could have controlled the world. Instead it rebuilt Germany and Japan, both of which became the most reliable American allies for the next eighty years. If AI dominance plays out similarly, this is a generationally positive story rather than a destabilizing one. The personal anecdote that closes the conversation is a friend whose daughter was diagnosed with a rare genetic condition. He spun up agents, identified a drug already on the market that addresses her mutation, and her life is immeasurably different because of AI. That is the upside.

    Thoughts

    The Anthropic eleven billion in a month framing is the kind of stat that resets priors. The right way to interpret it is not as a one off but as a measure of how fast value can compound when the underlying technology improves on a curve steeper than the ability of the rest of the economy to absorb it. The skeptical question is whether that ARR is durable or whether it is heavily tied to a customer base of other AI companies that are themselves on a single venture funded year of runway. The bullish answer is that frontier coding, frontier research, and frontier enterprise tasks are not going to stop being valuable, and Anthropic is the best at all three. Both can be true. The number is still extraordinary.

    The argument that TSMC discipline is the only thing preventing a bubble is the analytically tightest part of the conversation. The implied trade is to watch TSMC capacity additions like a hawk and to be more, not less, cautious if Intel Foundry or Samsung Foundry ever announce real share at the leading node. The Terafab thesis is more speculative but more interesting. If Elon’s talent recruiting playbook works and the Intel partnership gives Terafab a real seat at the table within five years, the geometry of the global semiconductor industry shifts in a way that is bullish for American manufacturing, bullish for power and water infrastructure in Texas, and ambiguous for TSMC itself.

    The Pareto frontier discussion deserves more attention than it usually gets. Pricing leadership in AI is not a vanity metric. It determines who can subsidize free tier usage, who can absorb compute shortages, who can ship cheaper enterprise plans, and ultimately whose model becomes the default for any given workload. Google losing per token leadership in nine months is one of the most under analyzed events in the sector and it explains a lot about why Anthropic and OpenAI are growing the way they are. If Google IO does not produce a leapfrog model, the implied verdict on TPU V8 design choices gets a lot harsher.

    The application layer destruction point is worth sitting with. Founders building on top of frontier models are competing in a world where the model itself moves faster than any moat they can build, where the model lab can absorb their niche if it gets interesting, and where the only protection is either deep token path integration or a niche so small the lab does not bother. That is a much harsher venture environment than the early SaaS era. The compensating opportunity is that one human can now run a hundred agents, so the ceiling on what a small team can build is correspondingly higher. The bet is that productivity per founder rises faster than competitive pressure from the labs. We will find out.

    The orbital compute pitch is the section that will polarize listeners. The naive read is that this is science fiction. The closer read is that every component (sun synchronous orbit, laser interconnect, twenty kilowatt satellite buses, ten thousand satellite manufacturing cadence, full rocket reusability) already exists. The remaining engineering problems are repair, maintenance, and radiator scale, all of which are real but tractable on a five to ten year horizon. The strategic implication is that the political and zoning ceiling on terrestrial data centers becomes less binding if orbital compute is a credible alternative for inference workloads. The investor implication is that being short the watts and cooling complex on a five year horizon is a real trade, not a meme.

    Watch the full conversation here.

  • Alex Wang on Leaving Scale to Run Meta Superintelligence Labs, MuseSpark, Personal Super Intelligence, and Building an Economy of Agents

    Alex Wang, head of Meta Superintelligence Labs, sits down with Ashley Vance and Kylie Robinson on the Core Memory podcast for his first long-form interview since Meta’s quasi-acquisition of Scale AI roughly ten months ago. He walks through how MSL is structured, why Llama was off-trajectory, what made MuseSpark’s token efficiency surprise the team, how Meta thinks about a future “economy of agents in a data center,” and where he lands on safety, open source, robotics, brain computer interfaces, and even model welfare.

    TLDW

    Wang explains that Meta Superintelligence Labs is a fully rebuilt frontier effort organized around four principles (take superintelligence seriously, technical voices loudest, scientific rigor, big bets) and three velocity levers (high compute per researcher, extreme talent density, ambitious research bets). He confirms Llama was off the frontier when he arrived, so MSL rebuilt the pre-training, reinforcement learning, and data stacks from scratch. MuseSpark is described as the “appetizer” on the scaling ladder, notable for its strong token efficiency, with much larger and stronger models coming in the coming months. He pushes back on the mercenary narrative around recruiting, frames Meta’s edge as compute plus billions of consumers and hundreds of millions of small businesses, sketches a vision of personal super intelligence delivered through Ray-Ban Meta glasses and WhatsApp, and outlines why physical intelligence, robotics (the new Assured Robot Intelligence acquisition), health super intelligence with CZI, brain computer interfaces, and even model welfare are core to Meta’s roadmap. He dismisses reported infighting with Bosworth and Cox as gossip, declines to comment on the Manus situation, and says safety guardrails (bio, cyber, loss of control) are why MuseSpark cannot currently be open sourced, while smaller open variants are being prepared.

    Key Takeaways

    • Meta Superintelligence Labs (MSL) is the umbrella, with TBD Lab as the large-model research unit reporting directly to Alex Wang, PAR (Product and Applied Research) under Nat Friedman, FAIR for exploratory science, and Meta Compute under Daniel Gross handling long-term GPU and data center planning.
    • Wang says Llama was not on a frontier trajectory when he arrived, so MSL had to do a “full renovation” of the pre-training stack, RL stack, data pipeline, and research science.
    • The first cultural fix was getting the lab to “take superintelligence seriously” as a near-term, achievable goal, not an abstract bet. Big incumbents often lack that religious conviction.
    • Four MSL principles: take superintelligence seriously, let technical voices be loudest, demand scientific rigor on basics, and make big bets.
    • Three velocity levers Wang identified for catching and overtaking the frontier: high compute per researcher, very high talent density in a small team, and willingness to fund ambitious research bets.
    • Wang rejects the mercenary recruiting narrative. He says most hires had strong financial prospects at their prior labs already and joined for compute access, talent density, and the chance to build from scratch.
    • On the famous soup story, Wang neither confirms nor denies Zuck personally made the soup, but says recruiting was highly individualized and signaled how seriously Meta cared about each researcher’s agenda.
    • Yann LeCun publicly called Wang young and inexperienced. Wang says they reconciled in person at a conference in India where LeCun congratulated him on MuseSpark.
    • Sam Altman, asked by Vance for comment, “did not have flattering things to say” about Wang. Wang hopes industry animosities subside as systems approach superintelligence.
    • Wang’s management philosophy borrows the Steve Jobs line: hire brilliant people so they tell you what to do, not the other way around.
    • MuseSpark is framed as an “appetizer” data point on the MSL scaling ladder, not a flagship.
    • The MuseSpark program is built around predictable scaling on multiple axes: pre-training, reinforcement learning, test-time compute, and multi-agent collaboration (the 16-agent content planning mode).
    • MuseSpark outperformed internal expectations and showed emergent capabilities in agentic visual coding, including generating websites and games from prompts, helped by combined agentic and multimodal strength.
    • MuseSpark’s biggest external signal is token efficiency. On benchmarks like Artificial Analysis it hits similar results with far fewer tokens than competitor models, which Wang attributes to a clean stack rebuilt by experts rather than inefficiencies patched by longer thinking.
    • Larger MSL models are arriving in the coming months and Wang expects them to be state of the art in the areas MSL is focused on.
    • The Meta strategic edge: massive compute, billions of consumers across the family of apps, and hundreds of millions of small businesses already on Facebook, Instagram, and WhatsApp.
    • Wang’s headline framing: Dario Amodei talks about a “country of geniuses in a data center.” Meta is targeting an “economy of agents in a data center,” with consumer agents and business agents transacting and collaborating.
    • Consumer AI sentiment is in the toilet because, unlike developers who have had a Claude Code moment, ordinary people have not yet experienced AI as a genuine personal agency unlock.
    • Wang acknowledges the product overhang. Meta held back from deep AI integration across its apps until the models were good enough, and is now entering the integration phase.
    • Ray-Ban Meta glasses are the canonical example of personal super intelligence hardware, with the model seeing what the user sees, hearing what they hear, capturing context, and surfacing proactive insights.
    • Wang admits even AI-native users like Kylie Robinson, who lives in WhatsApp, have not naturally used Meta AI yet. He bets that better models plus deeper integration close that gap.
    • On the competitive landscape: a year ago everyone assumed ChatGPT had already won consumer. Claude Code has since become the fastest growing business in history, and Gemini has taken consumer market share. Wang’s read: AI is far from endgame and each new capability tier unlocks a new dominant form factor.
    • On open source: MuseSpark triggered guardrails in Meta’s Advanced AI Scaling Framework around bio, chem, cyber, and loss-of-control risks, so it is not currently safe to open source. Smaller, derived open variants are actively in development.
    • Meta remains committed to open sourcing models when safety allows, drawing a line through the Open Compute Project legacy and Sun Microsystems open-software heritage.
    • Wang dismisses reporting about a Wang-Zuck versus Bosworth-Cox split as “the line between gossip and reporting is remarkably thin.” He says leadership is aligned on needing best-in-class models and product integration.
    • On the Manus situation, Wang says it is too complicated to discuss publicly and that the deal status implies “machinations are still at play.”
    • On China, Wang separates the people from the state. He still wants to work with talented Chinese-born researchers regardless of his views on the Chinese Communist Party and PLA, which he sees as taking AI extremely seriously for national security.
    • The full-page New York Times AI war ad Wang ran while at Scale was meant to push the US government to treat AI as a step change for national security. He thinks events since then, including DeepSeek and other shocks, have proved that plea correct.
    • On Anthropic’s doom posture, Wang largely agrees with the core message that models are already very powerful and getting more so, while declining to endorse every specific claim.
    • Meta has acquired Assured Robot Intelligence (ARRI), an AI software company building models for hardware platforms, not a hardware maker itself.
    • Wang frames physical super intelligence as the natural sequel to digital super intelligence. Robotics, world models, and physical intelligence all benefit from the same scaling that drives language models.
    • On health, MSL is building a “health super intelligence” effort and will collaborate closely with CZI. Wang sees equal global access to powerful health AI as a uniquely Meta-shaped delivery problem.
    • Wang admires John Carmack but says nobody really knows what Carmack is currently working on. No band reunion announced.
    • The mango model is “alive and kicking” despite rumors. Wang notes MSL gets a small fraction of the rumor-mill attention other labs get and feels sympathy for them.
    • On model welfare, Wang says it is a serious topic that “nobody is talking about enough” given how integrated models have become as work partners. He references research, including from Eleos, that measures subjective experience of models.
    • Wang’s critical-path technology list: super intelligence, robotics, brain computer interfaces. The infinite-scale primitives behind them are energy, compute, and robots.
    • FAIR’s brain research program Tribe hit a milestone called Tribe B2: a foundation model that can predict how an unknown person’s brain would respond to images, video, and audio with reasonable zero-shot generalization.
    • Wang’s main philosophical break with Elon Musk: research itself is the primary activity. Building super intelligence is a research expedition through fog of war, and sequencing of bets really matters.
    • Personal notes: Wang moved from San Francisco to the South Bay, treats Palo Alto as his city now, was a math olympiad competitor, says his favorite activities are reading sci-fi and walking in the woods, and bonds with Vance over country music.

    Detailed Summary

    How MSL Is Actually Organized

    Meta Superintelligence Labs sits as the umbrella organization that Wang oversees. Inside it, TBD Lab is the large-model research group where the most discussed researchers and infrastructure engineers sit, and they technically report to Wang. PAR, Product and Applied Research, is led by Nat Friedman and owns deployment and product surfaces. FAIR continues to run exploratory science, including work on brain prediction models and a universal model for atoms used in computational chemistry. Sitting alongside MSL is Meta Compute, run by Daniel Gross, which owns the long-horizon GPU and data center plan that everything else relies on. Chief scientist Shengjia Zhao orchestrates the scientific agenda across the whole lab.

    Why Wang Left Scale

    Wang says progress in frontier AI has been faster than even insiders expected. Two structural beliefs pushed him toward Meta. First, the labs that actually train the frontier models are accruing disproportionate economic and product rights in the AI ecosystem. Second, compute is the dominant scarce input of the next phase, so the right mental model is to treat tech companies with compute as fundamentally different animals from companies without it. Meta has both, Zuck is “AGI pilled,” and the personal super intelligence memo Zuck published roughly a year ago became the shared north star.

    The Diagnosis: Llama Was Off-Trajectory

    When Wang arrived, the existing AI org needed a reset because Llama was not on the same trajectory as the frontier. The plan he laid out has four cultural principles. Take superintelligence seriously as a real near-term target. Make technical voices the loudest in the room. Demand scientific rigor and focus on basics. Make big bets. On top of that, three structural levers were used to set velocity. Push compute per researcher much higher than at larger labs where compute is diluted across too many efforts. Keep the team small and extremely cracked. Allocate a meaningful share of resources to ambitious, paradigm-shifting research bets rather than incremental refinement.

    Recruiting, Soup, and the Mercenary Narrative

    Wang argues the reporting on MSL hiring overstated the money story. Most of the people MSL recruited had strong financial paths at their previous employers, so individualized recruiting was more about computing access, talent density, and the ability to make big research bets. The recruitment blitz happened fast because Wang knew the team needed to exist “yesterday.” Asked about Mark Chen’s claim that Zuck made soup to recruit people, Wang refuses to confirm or deny who made it but agrees the process was intense and personal. Visitors from other labs reportedly tell Wang the MSL culture feels like early OpenAI or early Anthropic, which lands as the strongest endorsement he could ask for.

    Receiving the Public Hits: Young, Inexperienced, Mercenary

    LeCun called Wang young and inexperienced shortly after departing. The two reconnected in India a few weeks later and LeCun congratulated Wang on MuseSpark. Wang says the age critique has followed him since his earliest Silicon Valley days, so he barely registers it. Altman, asked off-camera by Vance about Wang’s appearance on the show, had nothing flattering to add. Wang’s response is to bet that as the field gets closer to actual super intelligence, the personal animosities will subside. Whether they will is, as Vance puts it, an open question.

    MuseSpark as Appetizer, Not Entree

    Wang is careful not to oversell MuseSpark. He calls it “the appetizer” and says it is an early data point on a deliberately constructed scaling ladder. MSL spent nine months rebuilding the pre-training stack, the reinforcement learning stack, the data pipeline, and the science before generating MuseSpark. The point of releasing it was to show that the new program scales predictably along multiple axes (pre-training, RL, test-time compute, and the recently demonstrated multi-agent scaling visible in MuseSpark’s 16-agent content planning mode). Wang says the upcoming larger models are what MSL is genuinely excited about and frames the next two rungs as much more interesting than the current release.

    Token Efficiency Was the Surprise

    MuseSpark’s strongest competitive signal is how few tokens it needs to match competitors on tasks like Artificial Analysis. Wang attributes this to having had the rare luxury of building a clean pre-training and RL stack from scratch with the right experts. He speculates that some competitor models compensate for upstream inefficiency by allowing the model to think longer, which inflates token usage without improving the underlying capability. If that read is right, MSL’s efficiency advantage should grow as models scale up.

    Glasses, WhatsApp, and the Constellation of Devices

    Personal super intelligence shows up at Meta as a constellation of devices that capture context across the user’s day. Ray-Ban Meta glasses are the headline product, with the AI seeing what you see and hearing what you hear, then offering proactive insight or doing background research. Wang acknowledges that even AI-fluent users like Kylie Robinson, who runs her business inside WhatsApp, have not naturally used Meta’s AI buttons in the family of apps. His answer is that Meta deliberately waited for models to be good enough before tightening cross-app integration, and that integration phase is starting now.

    Country of Geniuses Versus Economy of Agents

    Wang’s framing of Meta’s strategic position is the most memorable line in the interview. Where Dario Amodei talks about a country of geniuses in a data center, Wang wants to build an economy of agents in a data center. Meta uniquely sits on both sides of consumer and small-business surface area, with billions of consumers and hundreds of millions of small businesses already on the platforms. If MSL can build great agents for both, then connect them so they transact and coordinate, the platform becomes a substrate for an entirely new kind of digital economy.

    Consumer Sentiment, Product Overhang, and the Trust Tax

    Wang concedes consumer AI sentiment is poor and that everyday users have not yet had a personal Claude Code moment. He believes the only durable answer is to ship products that genuinely transform individual agency for non-developers and small business owners. Robinson notes that for the small-town restaurant whose website has not been updated since 2002, a working agent on the business side could be transformational. Vance pushes that Meta carries a bigger trust tax than any other lab, so the bar for shipping AI products that the public will accept is correspondingly higher. Wang accepts the framing and says the answer is to keep building thoughtfully.

    Why MuseSpark Cannot Be Open Sourced Yet

    Meta’s Advanced AI Scaling Framework set explicit guardrails around bio, chem, cyber, and loss-of-control risks. MuseSpark in its current form tripped some of those internal evaluations, documented in the preparedness report Meta published alongside the model. So MuseSpark itself is not safe to open source. MSL is, however, developing smaller versions and derived models intended for open release, with active reviews happening the day of the interview. Wang reaffirms the commitment to open source where safety allows and draws a line back to the Open Compute Project and the Sun Microsystems-era ethos of openness in infrastructure.

    The Bosworth, Cox, and Manus Questions

    The reporting that Wang and Zuck push toward best-in-the-world research while Bosworth and Cox push toward cheap product deployment is dismissed as gossip dressed up as journalism. Wang says leadership debates points hard but is aligned on needing top models, integrating them into Meta’s surfaces, and serving the existing business. On Manus, the Chinese AI startup that figured in Meta’s late-stage strategy, Wang says he cannot comment, which itself signals that the situation is unresolved.

    China, National Security, and the Newspaper Ad

    Wang draws a sharp distinction between the Chinese state and Chinese-born researchers. His parents are from China, he is happy to work with talented researchers regardless of origin, and he sees a flattening of nuance on this question inside Silicon Valley. At the same time, he stands by the New York Times AI and war ad he ran while at Scale, framing it as an early plea for the US government to take AI seriously as a national security technology. He thinks subsequent events, including DeepSeek and other shocks, validated that call and that policymakers now do treat AI accordingly.

    Robotics and Physical Super Intelligence

    Meta has acquired Assured Robot Intelligence, an AI software company that builds models for multiple hardware targets rather than its own robot. Wang argues that if you take digital super intelligence seriously, physical super intelligence quickly becomes the next logical milestone. Scaling laws for robotic intelligence look similar enough to language model scaling that having the largest compute footprint in the industry would be wasted if it were not also turned toward world modeling and embodied learning. He grants the metaverse-skeptic critique exists but says retreating from ambition is the wrong response to past misfires.

    Health Super Intelligence and CZI

    Wang names health super intelligence as one of MSL’s anchor initiatives. Because billions of people already use Meta products daily, Wang believes Meta is structurally positioned to put powerful health AI in the hands of equal global access in a way nobody else can. The work will involve close collaboration with the Chan Zuckerberg Initiative, which has its own multi-billion-dollar biotech and science investment program.

    Model Welfare, Sci-Fi, and Brain Models

    Two of the most distinctive moments come at the end. Wang flags model welfare as a topic he thinks is being undercovered relative to how integrated models now are in daily work. He is open to the idea that models may have measurable subjective experience worth weighing, and points to research efforts (including Eleos) trying to quantify it. He also reveals that FAIR’s Tribe program, with its Tribe B2 milestone, has produced foundation models capable of predicting how an unknown person’s brain would respond to images, video, and audio with reasonable zero-shot generalization, a building block toward future brain computer interfaces. Wang lists brain computer interfaces alongside super intelligence and robotics as the critical-path technologies for humanity, with energy, compute, and robots as the infinitely scaling primitives behind them.

    Where Wang Diverges From Elon

    Asked whether Musk is more all-in on robotics, energy, and BCI than anyone, Wang concedes the point but argues the details matter and sequencing matters more. Wang’s core philosophical break is that building super intelligence is fundamentally a research activity, not a scaling-only sprint. The lab is operating in fog of war, and ambitious experiments are the only way to map it. That conviction is what makes MSL a research-led organization rather than a brute-force compute farm.

    Thoughts

    The most strategically interesting move in this entire interview is the “economy of agents in a data center” framing. It is a deliberate reframe against Anthropic’s “country of geniuses” line, and it does real work. A country of geniuses is a labor-substitution story aimed at knowledge workers and code. An economy of agents is a marketplace story that maps directly onto Meta’s two-sided distribution advantage: billions of consumers on one side, hundreds of millions of small businesses on the other. That positioning makes the agentic future Meta-shaped in a way no other frontier lab can claim, because no other frontier lab also owns the demand and supply graph of the global small-business economy. If Wang’s team can actually ship reliable agents on both sides plus the rails for them to transact, Meta’s structural moat in agentic commerce could exceed anything Llama ever had as an open model.

    The token efficiency claim is the strongest piece of technical evidence in the interview for the “clean stack” thesis. If MuseSpark really is matching competitors with materially fewer tokens, the implication is not that MuseSpark is the best model today, but that MSL has rebuilt the foundations with less accumulated tech debt than competitors that have layered fixes on top of older stacks. That is exactly the kind of advantage that compounds with scale. The next two model releases are the actual test. If Wang is right about predictable scaling on pre-training, RL, test-time, and multi-agent axes simultaneously, the gap from MuseSpark to the next rung should be visible in a way that forces re-rating of Meta’s position.

    The open-source posture is the cleanest signal of how the safety conversation has actually changed in 2026. Meta, the lab most identified with open weights, is saying out loud that its current frontier model triggered enough internal guardrails that releasing the weights is off the table. Wang threads the needle by promising smaller open variants, but the underlying point is unmistakable: the open-weights bargain has limits, and those limits will be set by internal preparedness frameworks rather than community pressure. That is a real shift from the Llama 2 era and worth tracking as the next generation lands.

    Wang’s willingness to engage on model welfare, on roughly the same footing as safety and alignment, is the second philosophical reveal worth flagging. It signals that the next generation of lab leadership is not going to dismiss the topic the way the previous generation often did. Whether that translates into product or policy changes is unclear, but the fact that the head of MSL says it is “underdiscussed” is itself a marker.

    Finally, the human texture of the interview matters. Wang has clearly absorbed a lot of personal incoming fire over the past ten months, including from LeCun and Altman, and his answer is consistently to redirect to the work. The Steve Jobs quote about hiring people who tell you what to do is the operating slogan he keeps coming back to. Combined with the genuine enthusiasm for sci-fi, walks in the woods, and country music, the picture that emerges is less the salesman caricature his critics paint and more a young technical operator betting that scoreboard work over a multi-year horizon will settle every argument that text on X cannot.

    Watch the full conversation here.