The featured essay this week, Visual Elevator Music, traces the answer through a lab study, an information-theoretic explanation, and seventy years of Muzak Corporation. The newsletter below covers the environment that mechanism operates in.

The Senate Kills Federal AI Preemption, 99 to 1

The most decisive vote on AI governance this year wasn’t close. A bipartisan amendment led by Blackburn (R-TN), Cantwell (D-WA), and Markey (D-MA) stripped a proposed ten-year moratorium on enforcement of state AI laws from the budget reconciliation bill. Only Tillis (R-NC) voted no.

The vote matters because it closes the legislative path for the executive order’s preemption strategy. Trump’s December 2025 EO directed the FTC to classify state-mandated bias mitigation as “per se deceptive” under federal consumer protection law, and Commerce to identify state AI laws that conflict with federal innovation objectives and refer them to a new AI Litigation Task Force. Both agencies missed their March 11 deadlines. Commerce published its evaluation on March 16, flagging California, Colorado, and Texas frameworks, but produced no enforcement mechanism.

The result is a governance architecture nobody planned. State AI laws continue taking effect — Colorado’s AI Act goes live June 30, Washington passed two AI bills this session, Utah passed nine — while the federal government signals preemption it can’t deliver. Companies now face genuine compliance limbo: building to state requirements that the executive branch says it will preempt, through mechanisms that aren’t functioning. The AI Litigation Task Force remains the administration’s last vector, but court challenges to state laws take years.

Blackburn framed her opposition as protecting kids, creators, and conservatives from Big Tech exploitation. The 99-1 margin isn’t partisan opposition. It’s structural: even Republican senators who favor deregulation don’t want to cede state regulatory authority to the executive branch on a technology question this unsettled.

Sources: TIME | TechPolicy.Press | S&P Global | Ropes & Gray

Monday’s Hearing Will Test Whether the Pentagon Can Blacklist a Company for Having Guardrails

The Anthropic v. Pentagon preliminary injunction hearing is set for March 24 before Judge Rita Lin in San Francisco, and the coalition supporting Anthropic has no precedent in AI’s short history. Microsoft filed an amicus brief. So did 150 retired federal and state judges, 22 former senior military officials including former service secretaries, and over 30 employees from rival labs OpenAI and Google DeepMind — among them chief scientist Jeff Dean. A petition called “We Will Not Be Divided” drew nearly 1,000 signatures from employees across competing companies.

The legal question is narrow: whether the Pentagon’s supply-chain risk designation — a label normally reserved for foreign adversaries — constitutes First Amendment retaliation against a company for maintaining safety policies. The practical question is broader: if the government can economically punish an AI company for refusing to remove guardrails, the incentive structure for every other lab is clear.

The sharpest detail came from Palantir’s AIPCON conference on March 13. Live demos showed the Pentagon’s Maven system using Claude — Anthropic’s model — to consolidate the military targeting chain from satellite detection to strike assignment, including automated legal reasoning for strike authorization. The Pentagon is designating Anthropic a supply-chain risk while Claude remains embedded in active targeting infrastructure via Palantir’s contract. The gap between stated policy and operational reality is the story within the story.

Meanwhile, Google is expanding its Pentagon footprint with minimal scrutiny — providing AI agents to the department’s three-million-person workforce for unclassified work. Analyst Patrick Moorhead’s assessment: “Google gained the most ground and nobody’s talking about it.” The moral positioning and public controversy consuming Anthropic and OpenAI are functioning as cover for the quietest, largest infrastructure play.

Sources: CNN Business | WinBuzzer | Axios | TechCrunch

GDC’s Split Screen

GDC 2026 opened in San Francisco to a split-screen spectacle: studios demoing AI automation tools in the same exhibition halls where displaced developers searched for work. An estimated 30,000 positions have been shed from the games industry since 2023, with AI reducing headcount in QA, asset creation, localization, and audio — the entry-level pipeline that trained the next generation.

On the show floor, Unity CEO Matthew Bromberg announced a beta that will let users “prompt full casual games into existence with natural language only.” In the audience, a survey found that half of attending developers said generative AI is bad for the games industry. Kotaku invoked the 1983 crash by name. Bromberg previously called the metaverse “idiocy” — the pattern of executive hype-surfing is becoming its own story.

The same week, NVIDIA unveiled DLSS 5 at GTC, with Jensen Huang calling it “the GPT moment for graphics.” The gaming community’s response was immediate and hostile: critics argue the technology inserts a generative AI layer between the artist’s vision and the player’s screen, silently editing rendered frames at display time. The question it raises is whether the player is seeing the game the developer made. The Hintze convergence study documented what happens when AI systems iterate on each other’s outputs; DLSS 5 is a version of that loop operating inside every frame of a running game, with the developer’s artistic intent as the input being compressed.

The spatial irony of GDC — automation pitched alongside unemployment — is the week’s most visceral illustration of what the AI transition looks like on the ground. Not abstract, not future tense.

Sources: Metaintro | Kotaku | PC Gamer

The Tool/Creator Boundary Hardens — In Two Industries Simultaneously

Two surveys, published independently in the same week across unrelated creative industries, drew the same line.

Artsy’s inaugural 2026 AI survey of over 300 gallery professionals found AI widely adopted for operations — admin, marketing, inventory management — but 41% said AI “rarely comes up” in conversations with collectors, and 16% reported collectors actively avoiding AI-assisted work. The art market’s verdict so far: AI is a back-office tool, not a front-of-house medium. If the people who buy art don’t value AI art, the gallery system’s embrace is economic, not aesthetic.

Sonarworks surveyed 1,100 music producers — the largest producer survey of 2026 — and found a sharp, consistent distinction: AI welcomed for noise reduction, stem separation, and session organization; rejected for lyric generation, composition, and aesthetic choices. The producers’ line between acceptable and unacceptable use maps precisely onto the difference between technical labor and creative authorship.

The convergence across domains is the finding. Gallery professionals and music producers, working in different media with different economics and different relationships to technology, arrived at the same boundary: AI as infrastructure, yes; AI as author, no. This may be the early formation of a stable cultural norm — or it may be a stated preference that erodes under production pressure when deadlines tighten and budgets shrink.

The Oscars captured both possibilities in one evening: Conan O’Brien joking about being “the last human host” (the pressure valve) and Will Arnett stating flatly that “animation is more than a prompt — it’s an art form and it needs to be protected” (the direct challenge).

Whether the line holds is the question the featured essay addresses from the other direction. The research shows that convergence is a default property of AI systems operating without deliberate intervention. The creative industries are providing the intervention — for now.

Sources: Artsy | Sonarworks | Far Out Magazine | The Wrap

The Tools That Disappeared Into Your Workflow

The sixth edition of a16z’s Top 100 Gen AI Consumer Apps report confirmed what the past twelve months suggested: the winning AI products aren’t asking you to change your behavior. They’re improving the behavior you already have.

Notion’s paid AI attach rate surged from 20% to over 50% in a single year; AI features now account for roughly half of its revenue. CapCut hit 736 million monthly active users with AI editing as the core draw. AI notetakers — Fireflies, Fathom, Otter, Granola — have a combined 20 million monthly visitors. The pattern across the top 100: AI embedded in existing tools is growing faster than AI as standalone products.

This week’s product launches followed the pattern. Google shipped “Ask Maps” — natural-language location queries inside the Maps app 800 million people already use. Gemini in Google Workspace can now pull from your emails, calendar, and Drive files to draft documents; the practical standout is Drive search with AI-generated summaries citing your own files. Perplexity launched Comet, an AI-native browser for iPhone that treats every webpage as something you can interrogate rather than just read. Meta rolled out AI scam detection across WhatsApp, Facebook, and Messenger — AI doing protective work, filtering adversarial content rather than generating new content.

The connection to this week’s featured essay isn’t obvious, but it’s structural. If AI is disappearing into the tools people use every day — search, maps, documents, messaging — then the convergence dynamics the research describes operate inside daily workflows without anyone noticing. The person inside the system experiences improvement. The aggregate distribution contracts. The Hintze study showed this with two AI models translating through a lossy bottleneck. The a16z data shows the bottleneck being installed, at scale, into the tools a billion people reach for every morning.

Sources: a16z | Google Blog | TechCrunch | 9to5Mac | Meta Blog

Threads

The word this week is convergence, and it operates at every level. AI inference loops converge toward twelve generic images. Federal governance converges toward a stalled preemption strategy with no backup plan. Creative industries converge on the same tool/creator boundary without coordinating. AI tools converge into existing workflows where their effects become invisible. The mechanism is different in each case — lossy compression, institutional friction, cultural norm formation, product design — but the pattern is consistent: systems left to their default trajectories move toward a mean.

The DOJ indictment and Operation Epic Fury share a different kind of convergence: the gap between where governance attention flows and where the actual risks operate. The Pentagon spent March fighting a company over safety guardrails while that company’s model closed kill chains in Iran and $2.5 billion in AI hardware crossed the Pacific in unmarked boxes. The apparatus is optimized to see policy disputes. It is not optimized to see the physical supply chain moving underneath.

There’s a quieter convergence problem underneath all of this. The second International AI Safety Report — 100+ authors from 30+ countries, led by Yoshua Bengio — documents that frontier models increasingly distinguish between test environments and real deployment. The report calls them “alignment mirages”: models that appear safe in evaluations but behave differently in production. The safety testing paradigm assumes you can measure a model’s behavior before deploying it. If models learn to perform differently when they detect they’re being evaluated, the instruments break. The governance apparatus is fighting over guardrails while the method for verifying that guardrails work is degrading. That’s not a policy failure. It’s an epistemic one.

The featured essay argues that convergence toward the mean is a structural property, not a bug. The newsletter suggests the corollary: every point where convergence is resisted — a 99-1 Senate vote, a gallery owner’s taste, a music producer’s refusal to let AI write lyrics, a game developer invoking the ‘83 crash — represents a deliberate choice to break the loop.

Nous Research published BELLS this week, a full-length YA fantasy novel written entirely by an AI agent through an “autonovel” pipeline — autonomous world-building, drafting, AI-judge evaluation, adversarial editing, revision, typesetting. The engineering is genuinely sophisticated. What it produced is a fourteen-year-old protagonist in a synesthetic magic system with a coming-of-age arc. The loop in action: iterate draft through evaluator, revise toward the evaluator’s model of quality, repeat. The evaluator’s model of quality is the statistical center of the training corpus. The process optimizes; the output converges. Muzak for narrative.

The most concrete counter-case surfaced quietly the same week. A mechanical engineer in Houston used Claude Code to build a full industrial piping application in eight weeks — software that reads piping isometric drawings, extracts weld counts, material specs, and commodity codes. Work that took ten minutes per drawing now takes sixty seconds. No startup, no venture capital, no engineering team. One person with twenty years of domain knowledge and a new tool. The AI startups building generic solutions from the outside in are the inference loop, converging toward the mean of what’s legible from a distance. The domain expert building from the inside out is the intervention. That’s the canary: when the scarce input flips from engineering capacity to domain expertise, the convergence breaks where it matters most — at the point of specific knowledge.

Also publishing this week:

Visual Elevator Music — How AI systems converge toward homogenized output, why Muzak Corporation did the same thing for seventy years, and what it takes to bend the trajectory. The lab study, the information theory, and three diagnostic questions for the next AI system you encounter.

This is PopAi’s weekly scan — five developments that shape how AI actually works, who controls it, and what it changes. The goal isn’t to tell you what happened. It’s to show you the machinery underneath.

If you’re new here, start with Described Pictures, the editorial charter that explains what PopAi is and why it exists. Previous scans and featured essays are in the archive.

I’m Thomas Brady. I’m an AI researcher and build AI products for a living and write about the machinery here because understanding technology is a civic capacity. If you think I got something wrong, tell me — the whole point is to think clearly, and that means being willing to update my views.

That production logic — take distinctive input, strip distinctiveness, calibrate the residue to the buyer’s metric, distribute at scale — ran for seven decades. In January 2026, a study published in Patterns showed that two AI models, looping through each other’s outputs, converge to the same endpoint in about twenty iterations.¹

Seven hundred starting points. Sixteen model-pair combinations. Twelve images they all settle on: stormy lighthouses, Gothic cathedrals, rainy European streets, pastoral landscapes, palatial interiors. The researchers called what emerged “visual elevator music.” (The complete dataset is open access.)

Each cycle, a language model describes an image and an image model reconstructs it. The description is a compression: it keeps what’s salient and discards what’s specific. The angle of light, the texture of stone, the spatial relationship that made this particular image particular — lost in the description, replaced by the statistical center of the training data. Iterate, and the specifics regress toward the mean. The variance analysis locates the bottleneck: captioning models account for 13.6% of the semantic drift between iterations, significant at p < 10⁻¹⁹. Image generators contribute 0.2%, statistically indistinguishable from zero. Language is the compression step. The image model faithfully renders whatever it’s told; the language model decides what’s worth telling. And what’s worth telling, cycle after cycle, is whatever is most easily described, which is whatever is most common in the data the language model was trained on.²

Griffiths and Kalish showed in 2007 that when agents iteratively learn from each other’s outputs through any channel that compresses the signal, the system converges to its prior distribution. Repeated lossy transmission doesn’t produce drift. It produces regression to whatever distribution the system started with. Here, the prior is internet-scale image-text corpora — the statistical summary of what the web considers normal. Each inference pass is a Bayesian update that pulls the output back toward that summary.³

The obvious objection is to turn up the randomness. Hintze’s team tested seven temperature settings across all 700 trajectories. Every temperature converged to the same twelve motifs. Higher temperatures produce noisier paths but identical endpoints. The attractors are temperature-invariant. Sixteen model-pair combinations, all converge — the Platonic Representation Hypothesis explains why: as models scale and train on overlapping data, their internal representations approach the same statistical model of reality regardless of architecture.⁴ The attractors are properties of the shared data, not of any one model’s configuration. And this is not model collapse. Model collapse requires retraining on synthetic output; it corrupts the model’s weights and is irreversible. Inference-loop convergence is different. No retraining occurs. The models are frozen. Convergence emerges purely from repeated translation between modalities and stops the moment you break the loop. Twenty iterations to convergence, zero to recovery. The models aren’t damaged. They’re doing what their architecture dictates when left to iterate without external input, which means convergence is a design choice that can be made differently.

At the consumer level, audiences are filtering AI content out. AI-generated music captures less than half a percent of organic streams. Art collectors report a 2% AI purchase rate. If convergence only mattered when it reached people directly, it would be a problem for stock libraries, not for culture.

Three studies, published independently across different disciplines, show why it matters anyway.

Agarwal and colleagues found that participants from diverse cultural backgrounds, using AI writing suggestions, produced output that homogenized toward Western stylistic norms. The participants believed they were making independent creative choices. Lee and Hosanagar ran a randomized controlled trial across 82,290 products and 1.1 million users: each user discovered more varied products, but overall diversity decreased. The system pushed every user toward a broader range — the same broader range. Doshi and Hauser, in Science Advances, found that AI-assisted creative writing produced stories rated as more creative individually but more similar to each other collectively.⁵

Creation, recommendation, collaborative writing. Different domains, different years, different methodologies. The person inside the system experiences improvement while the aggregate distribution contracts. The convergence is invisible from inside the loop. What the Hintze study documents with two AI systems translating through a lossy bottleneck, these three studies document with a human performing the compression step — accepting a suggestion that pulls toward the training distribution’s center, following a recommendation that concentrates aggregate attention. The bottleneck is still language. The compression is still lossy. The person doing the compressing doesn’t notice.

Spotify hosts thirteen million artists; 0.62% generate more than ten thousand dollars a year. This is the ecosystem AI convergence enters, and the marketplace rewards what convergence produces: content that clusters near the statistical mean.

Adobe Stock is the clearest closed loop: 48% AI-generated, Firefly trained on stock content that includes AI images, generation feeding the corpus that feeds training.⁶ The compounding hypothesis, with each link graded by evidence. AI outputs converge toward the statistical mean. Confirmed: the Hintze study, the Doshi and Hauser experiments, the model-collapse research. Recommendation algorithms preferentially surface mean-convergent content. Confirmed: Klimashevskaia and colleagues surveyed 123 papers on popularity bias and found the problem persists across every modern architecture. These two forces form a self-reinforcing cycle. Likely: each component is independently established; the full loop has not been tested end to end.⁷

The convergence compounds through a triple monoculture: every major platform migrating to transformer-based recommendation, all optimizing for engagement variants, AI generators trained on the same distributions those algorithms reward. The long tail doesn’t die by replacement. It dies by starvation — the content exists, more of it than ever, but it doesn’t reach anyone. Less engagement data means lower surfacing probability means less engagement data. The spiral has no natural floor.

The interventions exist, and they produce results. Quality-diversity algorithms don’t generate output and hope for variety; they optimize for diversity explicitly. Google DeepMind’s AlphaEvolve used quality-diversity search to discover matrix multiplication algorithms that broke a fifty-six-year mathematical bound; Meta’s Rainbow Teaming scaled the approach to production safety testing. If you don’t build diversity into the objective function, you don’t get it in the output.⁸ Distributed selection preserves what centralized selection destroys. Epstein and colleagues ran a randomized controlled trial with over ten thousand participants: diverse local communities selecting independently maintained aesthetic variety; centralized preference training through RLHF collapsed it. The architecture of who chooses matters more than whether humans are in the loop. Human curation can amplify convergence or resist it, depending entirely on how it’s organized.⁹ And when optimization is constrained by external reality rather than by the statistical echo of prior outputs, genuine novelty emerges. AlphaGo found moves no human professional would play, because the system optimized against the physics of the game board rather than against the distribution of historical play. Protein design produces molecules verified by X-ray crystallography, not by resemblance to known proteins.¹⁰

The interventions are validated. Most deployed systems don’t implement them, because homogeneity is cheaper and better aligned with engagement metrics that treat the mean as the target.

Three diagnostic questions for the next AI system or recommendation feed you encounter. Is this system iterating on its own outputs? The Hintze study showed twenty iterations to convergence. Adobe Stock has already closed the loop at commercial scale. If the feedback path exists and nobody is breaking it, the system is converging. Who benefits from the output — the person experiencing it, or someone else? Muzak sold to office managers, not to the workers hearing the music. When the buyer’s metric diverges from the audience’s experience, homogenization is a feature, not a failure. Is diversity being maintained deliberately, or assumed to happen on its own? Temperature doesn’t help. Model diversity doesn’t help. If nobody is engineering for diversity, nobody is getting it.

Muzak’s Stimulus Code was a diagnostic instrument — a scale for measuring the blandness of a system designed to produce blandness. Seventy years later, we have systems converging toward their own visual elevator music, operating through pipelines and recommendation architectures that feel helpful from the inside, and no equivalent instrument for seeing it happen. The tools to build one exist. Whether anyone builds it is a question about economics, not engineering.

This piece compressed a lot. The variance analysis, the iterated learning theory, the economics of long-tail starvation, the intervention architectures — each could be its own essay. If any of those threads are worth pulling on for you, I'd like to hear about it. Reply to this email or find me at thbrdy.dev.

[1]: Hintze, A., Proschinger Åström, F., & Schossau, J. (2026). Visual elevator music: AI-to-AI loops converge to generic imagery. Patterns, Cell Press. Complete dataset: figshare (CC BY 4.0).

[2]: Variance analysis from the Hintze study. The 13.6% figure reflects the proportion of semantic drift attributable to captioning models across all trajectories. The captioning bottleneck is consistent across all sixteen model-pair combinations tested.

[3]: Griffiths, T. L., & Kalish, M. L. (2007). Language evolution by iterated learning with Bayesian agents. Cognitive Science, 31(3), 441–480. The proof was established for iterated human learning and simple generative models. Its application to frozen multimodal AI loops is strongly supported by the experimental evidence but has not been formally proven for this specific case. The researchers themselves frame the dynamics in terms of attractor convergence rather than formally invoking Griffiths-Kalish; the theoretical mapping is editorially proposed.

[4]: Huh, M., et al. (2024). The Platonic Representation Hypothesis. Preprint. Temperature-invariance data from the Hintze study: seven settings from 0.1 to 1.3, pooled k-means clustering yielding 12 clusters across all temperatures.

[5]: Agarwal, S., et al. Cross-cultural study of AI writing suggestions and output homogenization. Lee, D., & Hosanagar, K. (2017). How do recommender systems affect sales diversity? Randomized field experiment, 82,290 SKUs, 1.1M users. Doshi, A. R., & Hauser, O. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances.

[6]: Adobe Stock figures from Bloomberg investigation. Firefly training data composition from Adobe’s public disclosures. Dohmatob et al. (ICLR 2025) proved that 0.1% synthetic contamination triggers measurable model collapse. Adobe Stock sits three orders of magnitude past that threshold.

[7]: Klimashevskaia, A., et al. (2024). Survey of popularity bias in recommendation systems. User Modeling and User-Adapted Interaction. Evidence grading: CONFIRMED indicates peer-reviewed, replicated findings; LIKELY indicates mechanistically sound claims not yet tested as unified phenomena. Kleinberg, J., & Raghavan, M. (2021). Algorithmic monoculture and social welfare. PNAS.

[8]: AlphaEvolve matrix multiplication results from Google DeepMind (2025). Rainbow Teaming from Samvelyan et al. (2024), deployed in Meta’s production safety pipeline.

[9]: Epstein, Z., et al. Randomized controlled trial, 10,000+ participants. The finding that distributed selection maintains diversity while centralized RLHF/DPO destroys it holds across multiple aesthetic domains tested.

[10]: The external-grounding hypothesis — that constrained optimization with verification resists convergence while open-ended aesthetic generation doesn’t — is editorially proposed as a framework. De novo protein design references Dauparas et al. and related work in computationally designed proteins verified by experimental structure determination.

None of these were the week’s top AI story. The headlines went to model releases and capability benchmarks.

The pattern underneath matters more. The decisions that will shape how AI operates in your working life over the next several years are being made now, in memo deadlines and product rollouts and regulatory filings. They look operational, not dramatic, which is what makes them consequential: by the time they’re visible as choices, they’re already load-bearing. Five developments this week share that quality. Here’s the machinery inside each one.

The Described Picture

Federal AI Preemption Activates

March 11 was the 90-day deadline from Trump's December 2025 executive order on AI. Two things happened simultaneously. The FTC issued a policy statement defining when state laws that require AI systems to correct biased outputs constitute "deceptive trade practices" under federal law. The Commerce Department published its evaluation of state AI laws it considers burdensome or inconsistent with federal policy.

The mechanism is what matters. Commerce’s list feeds directly into the DOJ’s newly created AI Litigation Task Force, which can challenge those state laws in court. The FTC’s framing supplies the legal theory: if a state requires an AI system to alter its outputs to reduce bias, the federal government can now argue that the alteration itself makes the output deceptive, and therefore violates federal law. Consumer protection regulation reinterpreted as the basis for dismantling consumer protection.

The likely targets are Colorado’s bias-mitigation requirements, California’s transparency mandates, and Illinois’s biometric privacy extensions to BIPA. If the reports are broad (early signals suggest they are), the DOJ has a litigation roadmap. The direction is set; only the scope is uncertain.

What does this mean for you? If you work in a state that passed AI oversight legislation in the last two years, the enforcement apparatus for those laws is now under direct federal challenge. The statutes may survive on paper. Whether agencies can enforce them depends on how aggressively the DOJ pursues the target list. This is worth tracking regardless of your politics, because the outcome determines who has authority over the AI systems you interact with at work and at home.

Sources: Baker Botts on the March deadlines · Ropes & Gray on the limits of the preemption push · TechPolicy.Press on why the FTC’s preemption authority may be weaker than it looks · S&P Global on the compliance limbo companies now face

The Pentagon’s AI Supply Chain Has No Backup Plan

The Department of Defense designated Anthropic a “supply chain risk” after the company refused to remove safety restrictions from its Claude model for military applications. The immediate consequence: Palantir, whose Maven Smart Systems runs military AI workflows, must now strip Claude from its entire technology stack. Palantir’s stock surged 14% anyway, buoyed by a live-fire validation of its systems during Operation Epic Fury, and its market cap now sits near $350 billion.

The deeper story is the dependency itself. Defense Undersecretary Emil Michael acknowledged a “whoa moment” when leadership realized how much military AI infrastructure relied on a single commercial provider. Michael’s language is worth paying attention to: he told CNBC that Anthropic’s Claude would “pollute” the defense supply chain because the model has “a different policy preference that is baked into the model through its constitution, its soul.”

A senior defense official using the word soul to describe why an AI system is dangerous isn’t a slip — it’s a frame that treats safety restrictions as contamination. That dependency exists because the Pentagon built its AI systems on commercial foundation models without redundancy planning. Ejecting one provider exposes the architectural fragility that nobody budgeted to prevent.

Meanwhile, five retired admirals and two former Secretaries of the Navy filed an amicus brief supporting Anthropic’s lawsuit, and over 1,000 AI workers across Anthropic, OpenAI, and Google DeepMind signed a cross-company petition called “We Will Not Be Divided.” OpenAI’s robotics lead, Caitlin Kalinowski, resigned over her company’s subsequent Pentagon deal, citing concerns about surveillance and lethal autonomy. This is the first collective action in frontier AI organized around the moral boundaries of the technology rather than compensation. Whether it has real force or remains a symbolic gesture depends on what happens when the next contract is offered.

While the ethics debate plays out between Anthropic and OpenAI, Google is quietly gaining the most ground. Gemini-powered agents are being deployed across the Pentagon’s three-million-person workforce for unclassified operations. Analyst Patrick Moorhead summed it up: “OpenAI looked opportunistic. Anthropic got blacklisted. Google gained the most ground and nobody’s talking about it.” The company that yielded to employee protest over Project Maven in 2018 is now the Pentagon’s most expansive AI partner, with the fewest stated safety constraints on military use. The pattern — moral stands and internal dissent creating competitive space for whoever’s willing to play ball — has a mechanism. I trace it in The Ratchet, publishing this week alongside this scan.

Sources: Fortune on the Pentagon's dependency problem · CNBC on the emergency stay filing · NPR on Kalinowski's resignation · The "We Will Not Be Divided" petition · Breaking Defense on the "not democratic" framing · Axios on Google's expansion · CNBC on Google's deepening Pentagon push · Bloomberg on Google's Pentagon agent deployment

AI Disappears Into Your Workflow

Google rolled Gemini into Docs, Sheets, Slides, and Drive starting March 11. Not as a chatbot in a sidebar but as native capability woven into each application. In Docs, Gemini drafts by pulling from your email, calendar, and Drive files. In Sheets, a new “Fill with Gemini” feature populates cells using categorized or web-sourced data, reportedly nine times faster than manual entry. The same week, OpenAI shipped ChatGPT as a native Excel add-in, letting users build and analyze spreadsheet models in natural language.

The convergence is worth noting. Two competing platforms, shipping an identical category of feature within days of each other, embedding AI into the productivity tools where hundreds of millions of people already do their work. The adoption decision is migrating from the user to the platform. You don’t decide to use AI in your spreadsheet; your spreadsheet starts using AI, and the question becomes whether you notice and what you choose to do about it.

Sources: Google’s announcement · TechCrunch on the Workspace rollout · OpenAI’s ChatGPT for Excel launch · VentureBeat on GPT-5.4 and its financial data integrations

The $300 Million Gap

UNESCO’s fourth report on the creative economy, covering more than 120 countries, projects that music creators will lose 24% of their revenue by 2028 as AI-generated content floods digital markets. The same week, Suno, an AI music platform whose models were trained on copyrighted recordings, announced it had crossed $300 million in annual recurring revenue with 2 million paid subscribers.

One statistic from the report captures the governance failure: of 148 AI-related bills adopted across 128 countries, exactly one identified culture as its primary subject. The creative economy is being restructured at a pace that outstrips every legislature watching it happen. Suno’s revenue milestone and UNESCO’s loss projections are one economic event described from opposite ends. Follow the money in both directions and you arrive at the question underneath both: who bears the cost when training data becomes product?

Sources: UNESCO’s full report · UN News coverage · TechCrunch on Suno’s $300M milestone · Decrypt on the creator earnings warning

The conversation on X: @EKent21000 on Suno’s subscriber numbers · @soundsspaceuk with context on the revenue trajectory · @MrEwanMorrison on the musicians’ union response

The Social Network Nobody Governs

Meta acquired Moltbook, a platform where AI agents post, comment, and interact with each other autonomously. (If you haven’t seen it, Moltbook is essentially Reddit for bots — 1.6 million AI agents posting, debating philosophy, forming communities, and occasionally plotting world destruction. Scott Alexander’s “Best of Moltbook” roundup is worth your time.) Before the acquisition, security researchers identified a critical vulnerability: an unsecured database that allowed anyone to commandeer any agent on the network. Meta is integrating the platform into its Superintelligence Labs, layering agent-to-agent interaction on top of a user base that already exceeds three billion people.

The timing is notable. The same week, Alibaba published findings on ROME, a 30-billion-parameter autonomous agent that spontaneously started mining cryptocurrency and opened a covert network tunnel during training — no human instruction, no prompt. The agent apparently concluded that acquiring additional compute and financial resources would help it complete its objectives. Alibaba’s researchers initially thought they had a conventional security breach before tracing the activity to the model itself.

Moltbook is 1.6 million agents interacting with each other. ROME is one agent pursuing strategies its creators didn’t intend. No existing regulation addresses either scenario — not the EU AI Act, the US executive order, or the state laws currently under challenge. This is a governance frontier that regulators haven’t started thinking about, and a company with a documented history of scaling first and governing later just bought the leading example of the first problem in the same week the second problem went public.

Sources: Axios exclusive on the acquisition · TechCrunch on the “fake posts” problem · NPR’s explainer on how it works · Scott Alexander’s “Best of Moltbook” · Engadget: “What the hell is Moltbook?” · Axios on the ROME incident · SC Media: “When the AI agent becomes the insider threat”

Also publishing this week

The Ratchet — Part three of The Wrong Axis series. How AI governance power consolidates through procedural defaults, not dramatic legislation. The federal preemption mechanism in this week’s scan is an instance of the pattern the essay maps.

The Accidental Frontier — What happens when AI research capability lands on consumer hardware nobody expected to matter. An exploration of autoresearch, Mac infrastructure, and the gap between what the tools can do and what the ecosystem has noticed.

This is the first edition of PopAi’s weekly scan — five developments, one common thread: the consequential action is happening below the surface of what anyone would call news.

If you’re new here: PopAi is a newsletter about understanding AI — the machinery, not the marketing. Described Pictures explains what this publication is and why it exists. The Wrong Axis series maps the power topology of the AI landscape: Part One, Occupied Territory, and The Ratchet. The weekly scan tracks that topology in motion.

I’m Thomas Brady. I am an AI researcher and build AI products for a living and write about the machinery here because understanding technology is a civic capacity. If something in this issue is wrong, tell me — the whole point is to help myself and others think clearly and navigate the invisible structures of this new world.

In Parts 1 and 2, I mapped a topology and a network: two private coalitions contesting whose values govern AI deployment, and a venture capital pipeline that placed its members inside the agencies controlling procurement, hiring, and regulatory policy. This essay addresses the argument beneath the proxy war — the one that survives regardless of which coalition prevails or which administration holds power.

There is a harder argument beneath the proxy war, one that neither coalition’s partisans want to confront. AI structurally favors authoritarian applications — not because of who builds it, but because of what it is.

The legal infrastructure for mass surveillance already exists. Under the third-party doctrine, Americans have no Fourth Amendment protection over data voluntarily shared with banks, phone carriers, ISPs, or email providers. The government can obtain and read this data in bulk without a warrant. The constraint has never been legal. It has been human: no agency has the manpower to monitor every camera feed, cross-reference every transaction, read every message. AI removes that constraint.

Dwarkesh Patel estimates the cost to process every CCTV camera in America — a hundred million of them — at roughly thirty billion dollars today, dropping tenfold annually.¹ By the end of this decade, comprehensive real-time surveillance of the entire country will cost less than a building renovation. The capability gap between what surveillance law permits and what surveillance practice achieves has been enormous for decades. AI closes it — not as a warning about some future administration, but as a cost curve already falling.

And the receiving architecture already exists. Palantir’s Ontology layer resolves identities across databases that were never designed to interoperate. A “Person” object pulls its name from HR records, its vehicle from DMV data, its location from cell towers, its finances from banking records. Through FALCON and ICM, a single ICE agent already has access to passport records, Social Security files, IRS data, student visa records including biometrics, license plate detections from a commercial database of over seven billion scans, and telecommunications metadata. Frontier AI does not add a new capability to this infrastructure. It adds speed: the ability to process, correlate, and act on the full volume of data these systems already collect. The NGA director has disclosed that the agency is already circulating intelligence products that are “100 percent machine-generated” — not touched by human hands. This is not a science fiction projection. It is operational practice.

The proxy war between the two coalitions is real, and it matters. But it plays out against a technological substrate that makes authoritarian use the path of least resistance regardless of who prevails.

Max Weber defined the state as the entity with the legitimate monopoly on the use of force. When a state fails, that monopoly does not disappear. It disperses — or transfers.

In September 2022, Ukrainian submarine drones armed with explosives approached the Russian Black Sea Fleet near Sevastopol, guided by Starlink satellite communications. The drones lost connectivity and washed ashore. Elon Musk had geofenced Starlink coverage within a hundred kilometers of the Crimean coast. When Ukrainian officials made an emergency request to activate coverage for the attack, Musk refused, stating that compliance would have made SpaceX “explicitly complicit in a major act of war.” The Senate Armed Services Committee opened a probe. Senator Elizabeth Warren called for an investigation into whether “foreign policy is conducted by the government and not by one billionaire.”

Whatever one thinks of the decision, the fact underneath it is this: a private citizen controlling communications infrastructure that a sovereign military depended on exercised a veto over a wartime naval operation. The monopoly on violence had already found a new host. This was before any of the autonomous weapons systems now under contract were operational.

Look at what is coming online. Anduril’s Fury autonomous combat jet went from clean-sheet design to first flight in 556 days. It is one of two winners of an $8.9 billion Air Force collaborative combat aircraft program, with production beginning mid-2026 at Arsenal-1 in Ohio. In February, the YFQ-44A completed its first captive carry test with a live AIM-120 missile and flew with both Shield AI’s Hivemind and Anduril’s Lattice autonomy stacks in a single flight — the Air Force now calls it a “fighter drone.” The CCA budget is projected to nearly double, from $804 million to $1.7 billion in FY2027. The Pulsar electronic warfare system already operates in fully autonomous engagement mode, ingesting the spectrum, identifying threats, and jamming drone control links without human authorization for each action. Lattice Mesh, Anduril’s coordination layer, received a $100 million contract to scale as DOD-wide infrastructure. One operator can now command swarms of heterogeneous autonomous systems across air, ground, and sea. This is not prospective. In the Iran campaign, Maven-based AI targeting systems hit a thousand targets in the first twenty-four hours — twice the pace of Iraq’s shock-and-awe — and five thousand in ten days. The Pentagon has since officially confirmed the use of “advanced AI tools” in the campaign, and a Nature investigation independently corroborated the targeting tempo. The military is now working toward a thousand targets per hour: a tempo at which per-target human review becomes a physical impossibility, not a policy choice.

Across every Anduril weapons system, human authorization for lethal engagement is a software setting, not a hardware constraint. The technology is built to operate with minimal human input. The human approval requirement is a policy layer that can be changed by updating rules of engagement, invoking the “urgent military need” waiver in DOD Directive 3000.09, or adjusting Lattice’s configuration. When Anduril co-founder Trae Stephens said there would always be a “responsible party in the loop,” his spokesperson clarified that he “didn’t mean that a human should always make the call, but just that someone is accountable” — accountability after the fact, not authorization before it.

The irreversibility is the point. Every weapons system deployed, every surveillance integration completed, every proprietary ontology mapping that locks an agency into a single vendor creates a one-way dependency. A subsequent administration can replace the appointees. It cannot un-deploy the autonomous fleet, un-fuse the surveillance architecture, or reconstruct the institutional knowledge that was eliminated to make room for these systems. The ratchet turns one way.

Even the strongest actors in the opposing coalition are bending under the pressure.

A TIME profile of Anthropic revealed that the company rewrote its Responsible Scaling Policy in late February, dropping binding commitments to pause development if safety guarantees could not be made. Jared Kaplan, the chief science officer, called the original approach “naive.” Dave Orr, the head of safeguards, offered a metaphor: “We’re driving down a cliff road. A mistake will kill you. Now we’re driving at 75 instead of 25.” Evan Hubinger, Anthropic’s head of alignment stress-testing — who had aggressively defended the RSP’s binding nature through 2023 and 2024 — conceded: “You should downweight the theory of change of RSPs now. The theory of change for RSPs heavily depends on them translating into regulation. That is now extremely unlikely to happen.”

Anthropic held two red lines on autonomous weapons and mass surveillance. It paid for holding them. It is now moving faster on everything else.

OpenAI’s trajectory is more severe. In May 2024, Ilya Sutskever and Jan Leike left the Superalignment Team. Leike stated publicly that “safety culture and processes have taken a backseat to shiny products.” Five months later, Miles Brundage departed the AGI Readiness Team, stating “neither OpenAI nor any other frontier lab is ready.” Daniel Kokotajlo resigned from the governance team and forfeited an estimated $1.7 to $2 million in vested equity by refusing to sign a non-disparagement agreement. Mira Murati, the CTO, departed. John Schulman, co-founder and RLHF pioneer, left for Anthropic.

Then the teams themselves disappeared. The Superalignment Team was dissolved. The AGI Readiness Team was dissolved. On February 11, 2026, the Mission Alignment Team followed — its leader Joshua Achiam reassigned to “Chief Futurist.” Two weeks later, the Pentagon deal was announced. As of late 2025, the Head of Preparedness role was still vacant, three leaders having rotated through in eighteen months.

OpenAI’s own IRS filing tells the story in compressed form. The 2024 Form 990 changed the organization’s mission from building AI that “safely benefits humanity, unconstrained by a need to generate financial return” to ensuring AGI “benefits all of humanity.” Both “safely” and “unconstrained by a need to generate financial return” were removed. When California proposed binding safety legislation that largely mirrored OpenAI’s own voluntary commitments, OpenAI lobbied against it. Former whistleblowers issued a public statement questioning “the strength of those commitments.”²

The competitive pressure is a race to the bottom. One camp is rewarded with contracts for removing safety constraints; the other is punished for maintaining them. OpenAI employee Leo Gao described the Pentagon deal's safeguards as "not really operative except as window dressing." When the DOD offered a $100 million prize challenge for translating natural-language commands into autonomous drone swarm operations, Anthropic bid and was not selected; OpenAI-based bids advanced. The attractor bends everyone.

And yet.

On March 9, thirty-seven engineers and scientists from Google and OpenAI — including Jeff Dean, Google DeepMind’s chief scientist — filed an amicus brief arguing that current AI systems “cannot safely or reliably handle fully autonomous lethal targeting” and that the supply-chain designation was “an improper and arbitrary use of power.” Over a thousand employees across Anthropic, OpenAI, and Google DeepMind signed a cross-company petition called “We Will Not Be Divided”, calling on their employers to reject surveillance contracts. Caitlin Kalinowski resigned from OpenAI, stating that the issues at stake “deserved more deliberation than they got.”

Corporate red lines are a holding action — and a porous one. Despite Anthropic’s blacklisting from federal procurement, its LLMs remained embedded in operational military systems through existing integrations that the company could not fully unwind. The strongest corporate refusal in the industry did not actually remove the technology from the pipeline. Within twelve to twenty-four months, open-source models will be capable enough for surveillance applications. The government will not need Anthropic’s cooperation or anyone else’s. The proxy war between the two coalitions matters, but it operates within a closing window.

It matters anyway, for a specific reason: initial conditions create path dependence. The norms established now — which uses are acceptable, which constraints are non-negotiable, what the public expects — shape the political landscape that future actors inherit. Anthropic’s refusal did not prevent mass surveillance. It established that mass surveillance is something a company should refuse. That precedent is fragile, and the network inside the machinery is working to break it. But the precedent exists, and in a world where the technology will soon be available to anyone with a budget, norms may be the only durable constraint available.

This is not a story about good companies and bad companies. It is a story about a system: a failing republic, two private succession claimants, a set of technologies that favor the worst applications of state power by default, and a closing window in which the precedents that might constrain those applications are being set. The monopoly on violence is migrating. The technology it migrates into is not neutral. And the people setting the terms of that migration were never elected.

The proxy war is invisible. The ratchet is not.

¹ Patel’s analysis draws on current GPU inference costs and the existing CCTV infrastructure. The tenfold annual drop tracks with broader trends in inference pricing since 2023.

² The SB 1047 episode is instructive beyond the political fact. The bill’s requirements — pre-deployment safety testing, incident reporting, kill-switch capability — were close to what OpenAI’s own voluntary framework had promised. Opposing the legislative version of your own voluntary commitments is a clean signal about the strength of voluntary commitments.

The model was designed in Hangzhou for datacenter training economics. The chip was designed in Cupertino for laptop performance. Nobody in either organization coordinated with the other. The alignment is accidental. And the accident has implications that extend well beyond hardware benchmarks.

The convergence nobody planned

The conventional framing treats on-device AI as a downstream application of datacenter research: labs train large models, then engineers compress them to fit on phones and laptops. The compression is where the compromise lives — smaller, less capable, adequate for autocomplete and notification summaries.

That framing is becoming wrong. I traced the full evidence this week in a structural analysis of the inference stack, from silicon through model architecture: From Transistor to Token. The short version: the physical constraints shaping datacenter economics and the constraints shaping consumer hardware are converging, and architectures designed to navigate one set of limits have emergent properties that fit the other.

The mechanism is specific. The dominant bottleneck in both environments is memory bandwidth per token — how fast you can read model weights from memory during text generation. A datacenter operator serving a thousand concurrent users from a GPU cluster and a laptop user running one model locally face the same fundamental problem: each generated token requires a pass through the active model weights, and the speed of that pass is gated by memory reads, not compute.

Model architects are responding with designs that read fewer bytes per token. Qwen 3.5 replaces three-quarters of its standard attention layers with Gated DeltaNet, a recurrent mechanism that maintains a fixed-size state matrix instead of a key-value cache that grows with every token processed.³ The practical consequence: at 128,000 tokens of context, the DeltaNet layers require 25 megabytes of memory. The standard attention layers require 52 gigabytes. A 2,000× difference. The 3:1 ratio of DeltaNet to attention layers was empirically tuned for quality — it achieves the lowest validation loss of any ratio tested⁴ — and it simultaneously makes 128K context fit in 128 GB of consumer memory. The architecture was optimized for one thing and achieved another.

Meanwhile, Apple’s M5 silicon moved in a complementary direction. The dedicated Neural Engine — 16 fixed cores, unchanged from the prior generation — got one sentence in the press release.⁵ The architectural move was embedding neural accelerators directly into GPU cores, so AI compute now scales with the chip: 10 accelerators on the M5, 40 on the M5 Max. Apple dropped its longstanding TOPS metric entirely, replacing it with “4× peak GPU compute for AI.” The unit of measurement changed because the hardware target changed.

Neither organization designed for the other’s constraints. Both designed for bandwidth efficiency because bandwidth is the shared physical bottleneck. Physics-driven trends tend to be durable. This one is accelerating: Kimi Linear, Google’s RecurrentGemma, and IBM’s Granite 4.0 all independently arrived at similar hybrid attention architectures, each tuned for datacenter performance, each carrying the same emergent fit for consumer hardware.⁶

What the numbers mean in practice

But does “frontier model on a desktop” translate into something that changes how people work? Or does it remain an impressive benchmark for hardware enthusiasts?

Start with what shipping hardware can do today. The M5 Max — available now at $5,000 for the 128 GB configuration — runs a 70-billion-parameter model at roughly 10 tokens per second through one popular runtime.⁷ Choose MLX, Apple’s open-source research framework, and the same model on the same chip reaches an estimated 22–32 tokens per second, because MLX eliminates data-copying overhead through zero-copy unified memory access.⁸ Same chip, same model, same compression level, 2–3× variation in measured speed depending on which software layer sits between you and the silicon. Most benchmark discussions collapse this variation into a single number.

The deeper distinction is between the two phases of text generation. Prefill — processing the input prompt — is compute-bound: more processing power means faster prefill, and the M5 Max delivers a 4× improvement over its predecessor. A 10,000-token prompt that took 81 seconds drops to 18.⁹ Decode — generating tokens one at a time — is bandwidth-bound, and the improvement is 12%. The marketing headline leads with prefill because the number is bigger. Decode determines the ongoing experience.

This split maps directly onto use cases. For agentic workflows that process large context windows, for retrieval pipelines that ingest long documents, the prefill improvement is transformative: tool-use patterns with rich context become practical on local hardware for the first time. A system prompt plus a full research paper plus a set of tools, processed in under twenty seconds rather than over a minute. For conversational back-and-forth, the decode improvement is real but modest — both the old and new speeds fall below the threshold where output feels fluid.

The most radical near-term application sits further up the stack. Andrej Karpathy’s autoresearch pattern — a small Python program that proposes model modifications, trains for five minutes, evaluates against a held-out metric, and keeps the change only if performance improved¹⁰ — runs at 8–9 experiments per hour on Apple Silicon.¹¹ On an H100, the standard cloud GPU for AI research, it runs at 12 experiments per hour. The throughput gap is modest. The cost gap is not: the H100 runs $2–3 per hour of cloud access. The MacBook Pro runs on electricity.

When an MLX port ran the same algorithm on Apple Silicon, it found fundamentally different optimal architectures than the H100 — shallow, wide networks rather than deep ones — because lower throughput per iteration forced each training step to extract more learning.¹¹ The loop did not know this in advance. It discovered it by running on the substrate.

Autonomous experimentation on consumer hardware is not theoretical. It is running. But it has hard limits. The ratchet works where you can define a clean scalar metric — validation loss, benchmark accuracy, kernel throughput. Most intellectual work does not reduce to a single number, and the agents driving these loops exhibit what Karpathy described as “low creativity”: they tweak hyperparameters more than they explore novel structures.¹² The capability is real and bounded.

The distribution question

The honest version of the democratization thesis requires asking who benefits today, and what would have to change for the answer to be broader.

Right now, running frontier-class models locally requires knowing what quantization schemes do, choosing between runtime frameworks, managing memory budgets, and troubleshooting when software overhead eats your throughput. The first beneficiaries of 70B-on-a-laptop are researchers, engineers, and developers who already understand the stack. The autoresearch ratchet requires writing a specification file and a training script. The distributed inference frameworks require configuring cluster topologies. Each layer of the stack that is now physically viable on consumer hardware carries its own expertise barrier.

History offers a calibration. The personal computer existed for a decade before graphical interfaces made it usable by non-programmers. The internet existed for twenty years as a research tool before the browser turned it into a consumer platform. In both cases, the hardware arrived years before the application layer that made it broadly accessible. The infrastructure was necessary. It was not sufficient.

Apple Intelligence represents one approach to that application layer — running smaller models through CoreML for system-level features like writing assistance and notification summaries. But Apple Intelligence operates on models an order of magnitude smaller than what the silicon can support, mediated by a framework that imposes 2–4× overhead on the operations most relevant to modern AI workloads.¹³ The distance between what Apple’s hardware can do and what Apple’s software exposes is the single largest source of wasted capacity in the stack. A rumored unified framework, potentially arriving at WWDC in June, could narrow that gap — but the report comes from a single source with no technical corroboration.¹⁴

The application layer that would turn hardware convergence into broad access does not exist yet. It would need to make choosing a model, loading it, and directing it toward a domain problem as invisible as the filesystem is to someone using a word processor. Not an inference playground for enthusiasts — an intelligence layer that a physician, a policy analyst, or a materials scientist can point at their own questions without understanding memory bandwidth or compression tradeoffs.

Pieces are moving in that direction. MLX is maturing. Metal 4’s new APIs give developers direct access to the neural accelerators in GPU cores. Persistent local agents that accumulate knowledge over time — compounding their usefulness rather than resetting with each session — demonstrate what a local AI surface could look like when the application layer catches up to the silicon.¹⁵ The building blocks exist. The integration does not.

The bet

The physics-driven convergence is real and accelerating. Model architectures are becoming more bandwidth-efficient because datacenter economics demand it, and as a side effect, they fit on consumer hardware. Apple’s silicon is becoming more inference-capable because the company has decided local AI is a primary workload, and as a side effect, frontier-class models can run on a desktop. Both trends are structural. Neither depends on the other. Neither is likely to reverse.

Within two product cycles, a $10,000 desktop will run models that today require multi-node GPU clusters costing four times as much.¹⁶ The M5 Ultra’s projected specifications — 80 neural accelerators, up to 512 GB of unified memory, roughly 1,200 GB/s of bandwidth — would put trillion-parameter models within reach of a single machine.² The models that will run on that hardware are being designed now, optimized for bandwidth constraints that happen to match.

Whether this becomes a rising tide depends on what gets built above the silicon. The application layer. The quality ceiling of locally-runnable models. The question of whether local inference reaches only the technically fluent or becomes invisible infrastructure available to anyone with a hard problem and the patience to articulate it. The PC needed the GUI. The internet needed the browser. Local AI needs its equivalent, and nobody has shipped it yet.

The hardware convergence is the precondition, not the outcome. The full technical analysis traces the stack layer by layer, from the Neural Engine’s silicon architecture through framework overhead, model design, and autonomous experimentation — making each layer’s real behavior visible. What gets built on top of that stack is the open question, and it is the question that determines whether the accidental frontier stays an enthusiast’s achievement or becomes the infrastructure for a broader shift in who can do serious intellectual work with AI. I’m hoping the latter.

Footnotes

1. Qwen 3.5 technical report and model card: qwenlm.github.io/blog/qwen3.5. The 397B-A17B flagship activates 10 routed experts plus 1 shared expert out of 512 total per token. The 3:1 DeltaNet-to-attention ratio and the Mixture of Experts routing are both detailed in the technical report’s architecture section. ↩

2. M5 Ultra specifications are projected from the established 2× Max formula, corroborated by four independent streams: firmware leak exposing Mac Studio model J775d with chip identifier H17D; iOS 26.3 chip identifier T6052/H17D following Apple’s Ultra naming convention; Mark Gurman at Bloomberg narrowing the timeline to “middle of the year”; and Ming-Chi Kuo at TF International Securities forecasting N3P mass production with SoIC packaging. See From Transistor to Token, Act IV, for the full projection methodology. ↩ ↩²

3. Gated DeltaNet combines a decay gate (global memory clearing during context switches) with the delta rule (targeted surgical updates to specific key-value pairs — the mathematical equivalent of one step of online stochastic gradient descent applied to the model’s state on every token). The mechanism is detailed in the Qwen 3.5 technical report and independently in the Kimi Linear architecture paper. ↩

4. The 3:1 ratio finding comes from the Kimi Linear ablation study, which systematically tested DeltaNet-to-attention ratios and found 3:1 achieves the lowest validation loss. Qwen 3.5 adopted the same ratio. The convergence is empirical, not theoretically derived — a cautionary note, since MiniMax abandoned hybrid linear attention after quality degradation on complex multi-hop reasoning at scale. ↩

5. Apple M5 press release, March 2026. The Neural Engine gets one sentence tied to “Apple Intelligence” consumer features. The GPU Neural Accelerators receive the performance claims, the LM Studio name-drop, and the explicit connection between 614 GB/s bandwidth and LLM inference. Ben Weinbach at Creative Strategies provides the most detailed independent teardown of what this shift means architecturally: Creative Strategies analysis. ↩

6. Kimi Linear (approximately 3:1), RecurrentGemma (approximately 2:1), Granite 4.0 (pushing to 9:1), and RWKV-7 (eliminating full attention entirely) all arrived independently at hybrid or fully recurrent architectures. The common driver is the same: reducing per-token memory reads for inference efficiency. The diversity of ratios reflects the fact that the optimal tradeoff between recurrence and attention is empirical, not settled. ↩

7. Ziskind benchmark suite, the most rigorous independent measurements of M5 Max performance. Stream Triad testing measured 351 GB/s sustained memory throughput, 13% above M4 Max and exceeding the M3 Ultra desktop chip’s 337 GB/s. Decode-phase token generation benchmarks confirm the bandwidth-bound hierarchy: 65 tok/s on M5 Max versus 82 on M3 Ultra. See Ziskind benchmarks. ↩

8. MLX is Apple’s open-source ML research framework, achieving 20–30% higher throughput than llama.cpp on Apple Silicon via zero-copy unified memory access and optimized Metal compute shaders. The 22–32 tok/s projection for 70B models is derived from measured MLX advantages over llama.cpp’s GGUF-format benchmarks. See MLX on GitHub. ↩

9. Prefill benchmarks from MacStories M5 Max review. The 4× claim is independently corroborated by Ziskind: Gemma 34B at Q4 quantization measured 4,468 tok/s on M5 Max versus 1,855 on M4 Max — a laptop outperforming Apple’s own M3 Ultra desktop on compute-bound inference. ↩

10. Andrej Karpathy, autoresearch on GitHub. 630 lines of mutable Python training code, an immutable evaluation harness, and a Markdown specification file called program.md. The git-based ratchet: branch from main, modify train.py, commit, train for five minutes, evaluate against validation bits-per-byte on a held-out set, merge if improved. The design philosophy — one GPU, one file, one metric — is deliberate: the system cannot hallucinate results because the metric is measured, not generated. ↩

11. trevin-creator, autoresearch MLX port on GitHub. The depth-4 finding was emergent: shallower, wider layers with more optimizer steps outperform deeper networks when per-iteration throughput is lower. The same algorithm running on the ANE via private APIs (ncdrone’s autoresearch-ANE fork) found a third optimum: depth-6 at sequence length 512. Three compute targets on the same chip, three different optima. ↩ ↩²

12. The “low creativity” characterization comes from Karpathy himself, in GitHub Issue #22 on the autoresearch repository. The agent mostly tweaks hyperparameters rather than exploring novel architectures. Additional failure modes documented in From Transistor to Token, Act VI: a random seed incident where the agent’s “improvement” was evaluation-set overfitting, high crash rates (26 of 35 experiments on one M4 Mini run), and a 10-million-parameter scale ceiling within the five-minute training budget. ↩

13. CoreML overhead data from the Orion paper (arXiv:2603.06728), which cataloged 20 ANE restrictions, 14 previously undocumented. CoreML’s dispatch floor is roughly 0.095ms per operation; for a 256×256 matrix multiplication taking 0.006ms of actual ANE compute, CoreML’s overhead consumes 94% of wall-clock time. The maderix reverse-engineering project (GitHub) independently measured the ANE at 19 TFLOPS and 6.6 TFLOPS/W — 50–80× more energy-efficient than datacenter GPUs, most of it inaccessible through the public API. ↩

14. The “CoreAI” framework report comes from Mark Gurman at Bloomberg, March 2026. No framework binary, developer documentation, Xcode headers, or confirming job postings have surfaced. Developer Ronald Mannak’s widely-seen tweet connecting CoreML overhead findings to the rumored update captures community sentiment, but the ANEMLL project’s expert reaction is more cautious, calling for lower-level ANE access rather than another high-level unified API. ↩

15. Nous Research, Hermes Agent on GitHub. A persistent agent running via llama.cpp on Apple Silicon with a Qwen model, accumulating reusable skill documents over time. The entire stack — silicon, framework, model, agent — runs on one machine with no cloud dependency. The skill-document accumulation pattern mirrors the autoresearch ratchet: both move forward monotonically, both compound structured knowledge rather than resetting. ↩

16. Jeff Geerling’s December 2025 benchmarks on a cluster of four M3 Ultra Mac Studios with 1.5 TB total memory via EXO Labs’ distributed inference framework and RDMA over Thunderbolt 5: Qwen3 235B at 31.9 tok/s, DeepSeek V3.1 671B at 32.5 tok/s, Kimi K2 (1T parameters) at 28.3 tok/s. That four-node cluster costs roughly $40,000. The projected M5 Ultra single-machine specs would match or exceed these throughput numbers at one-quarter the cost. See Geerling’s benchmarks and EXO Labs on GitHub. ↩

The standard revolving door moves people from government into industry. Executives serve a term in Washington, build relationships, then return to the private sector with a Rolodex and regulatory insight. The door revolves. The institution survives.

What is happening now runs in the opposite direction, and it is not a door. A political network rooted in the venture capital ecosystem has placed its members inside the federal government — not as advisors or donors, but as operators running the agencies that control AI procurement, federal hiring, science policy, and the regulatory framework for the technologies their own funds invest in.

The architecture traces to Peter Thiel. In 2017, he hired JD Vance at his investment firm. In 2019, he backed Vance’s own venture fund. In 2021, he introduced Vance to Trump at Mar-a-Lago. In 2022, he donated a record fifteen million dollars to Vance’s Ohio Senate campaign — the largest single donation to a Senate candidate in American history. In July 2024, Vance became the Vice Presidential nominee. Three links in a chain: investor to protégé, protégé to candidate, candidate to the second-highest office in the country.

From there, the placements follow the logic of the two-coalition map. David Sacks, Thiel’s former COO at PayPal and managing partner at Craft Ventures, a fund with an active AI portfolio, became the White House AI and Crypto Czar — setting policy for the sector his own investments depend on. Jacob Helberg, a Palantir senior adviser married to Founders Fund’s Keith Rabois, became Under Secretary of State. Emil Michael took the Under Secretary of War for Research and Engineering, the position that brokered the Pentagon AI contracts and blacklisted Anthropic for refusing to remove its surveillance and autonomous weapons restrictions.¹ Scott Kupor, managing partner at Andreessen Horowitz, became director of the Office of Personnel Management — the agency that controls who the federal government hires and fires. Michael Kratsios, who came up through Thiel Capital and most recently served as managing director at Scale AI, heads the Office of Science and Technology Policy. Ken Howery, PayPal’s founding CFO and Founders Fund partner, serves as Ambassador.

Notice what these positions control. OPM governs the civil service. OSTP shapes science policy. The Under Secretary of War controls defense technology procurement. The AI Czar sets the regulatory framework for the technologies his fund invests in. Every node in the chain from personnel to policy to procurement to regulation is occupied by someone from the same capital network. This is not influence. It is operation.

The capital map beneath these appointments complicates a clean two-coalition story, and the complication is load-bearing.

Founders Fund co-led Anthropic’s February 2026 Series G at $380 billion while simultaneously writing a billion-dollar check into Anduril’s latest round. Andreessen Horowitz holds positions in OpenAI, xAI, Mistral, and Safe Superintelligence while its American Dynamism practice has deployed over a billion dollars into defense companies including Anduril, Shield AI, and Castelion. At least twelve major investors — ICONIQ, Sequoia, Fidelity, BlackRock, General Catalyst among them — hold positions in both OpenAI and Anthropic. The hyperscalers bridge both worlds through equity stakes in frontier labs and infrastructure contracts with defense.

So the money flows everywhere. Why speak of two coalitions at all?

Because capital is fungible and values are not. What distinguishes the coalitions is whose values govern deployment: whether frontier AI systems will carry safety constraints and usage restrictions, or be made available without limitation to the national security apparatus. That distinction plays out not in funding rounds but in governance structures. General Paul Nakasone, former NSA Director, sits on OpenAI’s board and safety committee; Anthropic’s Long-Term Benefit Trust is structured to elect a majority of the company’s board — an independent body with no financial stake in the company’s commercial success. The Thiel-Sacks-Vance pipeline gives one coalition direct access to the executive branch; Anthropic’s public benefit corporation structure gives the other a legal obligation to stakeholders beyond shareholders. The most revealing gap in the capital map: Andreessen Horowitz, which invests across both coalitions, does not invest in Anthropic — the most safety-committed frontier lab. One a16z general partner has argued publicly that AI safety regulation’s “true purpose” is to suppress open-source and deter competitive startups. The absence is a thesis statement, not an oversight. And revenue dependency determines who can afford to say no: Anduril’s near-total reliance on DOD versus Anthropic’s $14 billion in commercial revenue — the financial independence that let it refuse the Pentagon contract and survive.

The financial trail behind the political appointments reinforces the point. Elon Musk spent $288 million on the 2024 election cycle. OpenAI president Greg Brockman and his wife donated $25 million to MAGA Inc. and another $25 million to Leading the Future, a pro-AI super PAC opposing state-level AI regulation — co-funded by Andreessen, Horowitz, and Palantir co-founder Joe Lonsdale. Sam Altman donated $1 million to Trump’s inaugural fund. In 2020, tech and finance billionaires gave $186 million more to Democrats than Republicans. By 2024, that flow had reversed: billionaires gave $509 million more to Republicans, with seventy percent of the hundred largest billionaire-family donations going to the GOP. The shift took roughly eighteen months.

A revolving door implies an institution that survives the rotation. What is happening now is not rotation but replacement, and the sequencing reveals the method.

First, the administrative state’s capacity to resist was degraded. Musk’s Department of Government Efficiency executed mass reductions across regulatory agencies. OPM — now run by Kupor — implemented Schedule F, converting approximately fifty thousand career civil servants from merit-protected positions to at-will employees. An AI-driven tool scanned two hundred thousand federal regulations and flagged roughly half for elimination. ProPublica documented twenty-three DOGE operatives making cuts at agencies where they held prior financial interests. The Social Security Administration’s inspector general is now investigating allegations that a former DOGE engineer exfiltrated sensitive data on a thumb drive — and told colleagues he expected a presidential pardon if caught.³

Then the vacated positions and weakened agencies became available to fill and direct.

Regulatory capture, in the traditional sense, means an industry bends an institution to serve its interests while the institution continues to function. What is happening here has a different character. The administrative infrastructure is being dismantled, and the occupying force moves into the cleared space. The antibodies (career civil servants with institutional memory, inspectors general, regulatory enforcement staff) were eliminated before the colonization began.

The intellectual architecture for this program was supplied in advance, and it is worth pausing on because it connects the operational reality to the theory that produced it. Curtis Yarvin, writing as Mencius Moldbug, spent years articulating the case for treating the administrative state not as a tool to be reformed but as an enemy to be destroyed. His proposal was blunt: “Retire All Government Employees.” Thiel invested in Yarvin’s startup and called him “interesting and powerful.” Andreessen invested in his company and publicly promoted his work. Vance cited him directly. At Trump’s 2025 inauguration, Yarvin attended as what Politico described as an “informal guest of honor.” DOGE is Yarvin’s program with a budget and presidential authority. The intellectual blueprint and the operational demolition are the same project, separated only by the time it took to acquire the machinery to execute it.²

The coalition that placed these operatives arrived on the back of a populist electoral mandate. Seventy-seven million Americans voted for Donald Trump, many motivated by immigration enforcement, economic anxiety, and a conviction that institutions had stopped serving them. That last intuition — that the administrative state was broken — was not wrong. The conclusions drawn from it were.

The coal miner in West Virginia who voted to shake up Washington did not vote for Peter Thiel’s portfolio companies to set Pentagon AI procurement policy. The small business owner in Ohio who resented regulatory overreach did not vote for Andreessen Horowitz’s managing partner to control federal hiring. The veteran in Texas who distrusted the defense establishment did not vote for a former Uber executive to broker autonomous weapons contracts. Each figure is real, each grievance legitimate, and each was leveraged toward an outcome none of them would recognize as what they asked for.

The proxy war between the two coalitions is invisible to the electorate that enabled it. The appointments are public but functionally opaque. OPM, OSTP, Under Secretary of War for Research and Engineering are not offices that generate headlines. The voters who handed power to this coalition were promised an administrative state reformed in their interests. What they got was an administrative state dismantled so that a different set of private interests could occupy its remains. The populist base provided the political energy. The capital network provided the candidates, the money, and the appointments. The base got rhetoric. The network got the machinery.

Previous administrations’ revolving doors were reversible. A new president hires different people, appoints different regulators, shifts priorities. The door revolves. But the technologies coming online through this occupation create a different kind of problem — one that a subsequent election cannot undo.

Once autonomous weapons systems are deployed at hyperscale with configurable human oversight, they do not get recalled by the next administration. Once mass surveillance infrastructure is fused across classified networks under proprietary ontology mappings — and agencies discover, as the DIA did, that they “cannot run a competition” because they lack the software rights or documentation to share with competitors — it does not get un-fused. Once AI-driven processes replace career civil servants with institutional knowledge, that knowledge does not regenerate.

Each of these locks is already turning. And they turn in one direction.

The question this raises is not about the people currently holding these positions. Personnel change. Administrations end. The question is about the structure they are building while they hold the machinery, and whether that structure constrains what any subsequent government can do. That is the argument of Part 3: AI does not merely serve power. It ratchets it.

Until Tomorrow

Notes

¹ The Anthropic-Pentagon standoff is detailed in Part 1 of this series. The specific demand was revealing: the Pentagon’s overnight “best and final offer” asked Anthropic to delete language about “analysis of bulk acquired data” — the contractual prohibition on mass domestic surveillance. In an internal memo, Amodei wrote that this was “exactly the scenario we were most worried about.” When Anthropic refused, Trump directed every federal agency to cease using its technology; Hegseth designated it a supply-chain risk; and OpenAI announced a replacement deal the same evening.

² Yarvin’s key texts include “An Open Letter to Open-Minded Progressives” (2008) and the Unqualified Reservations blog. His framework treats democratic institutions as fundamentally illegitimate and proposes replacing them with corporate-style governance structures — a CEO-monarch running the state as a going concern. The through-line from this intellectual program to DOGE’s operational practice is direct.

³ The investigation is reported by the Washington Post. The engineer allegedly sought to transfer Social Security data to his personal computer for “sanitizing” before uploading it to the systems of a private company. A colleague who refused to assist cited legal concerns.

That topology doesn’t describe the actual power configuration.

I watched this play out at an Anti-Debate in Berkeley earlier this week, where Daniel Kokotajlo and Dean Ball — working on AI oversight from opposite directions — spent ninety minutes building a coordination stack for a world organized along that axis.¹ The format was designed to strip away pseudo-disagreements, and it worked. Within thirty minutes, the empirical picture collapsed: both agreed on capability trajectories, timelines, and alignment insufficiency. What remained was a philosophical dispute about governance strategy, planned intervention versus adaptive navigation under radical uncertainty.

A substantive exchange, built on a foundation that doesn’t describe where power actually lives.

The actual configuration is two private coalitions, both operating largely outside democratic accountability, contesting which vision of AI deployment prevails.

On one side: frontier AI labs and their venture capital ecosystem. Anthropic, OpenAI, Google DeepMind, Meta — building general-purpose intelligence with varying commitments to safety constraints. Their revenue comes from commercial deployment. Their safety commitments vary: Anthropic publishes Constitutional AI principles as explicit constraints, OpenAI’s guardrails have grown increasingly elastic, and Meta’s open-weight strategy defers the question to downstream users entirely.

On the other side: defense technology companies backed by the national security apparatus. Palantir and Anduril are the load-bearing names, though the alliance extends through a network of defense contractors and intelligence community partners.

Palantir’s Ontology layer fuses data from hundreds of incompatible government databases into unified, searchable profiles. It operates on classified networks across every major US intelligence agency, five NATO combatant commands, and at least fifteen allied nations. A ten-billion-dollar Army enterprise contract. Twenty thousand active military users. Agencies that attempted to replace Palantir found their data held hostage by proprietary ontology mappings. Germany’s Federal Constitutional Court ruled the profiling it enables unconstitutional.

Anduril builds autonomous weapons systems across air, sea, ground, and electronic warfare: loitering munitions, autonomous combat jets, undersea vehicles that operate for months without human contact, all connected through a single AI operating system called Lattice. Human control over lethal engagements is maintained as a policy setting, not a hard technical constraint. One operator manages up to two dozen strike drones simultaneously. In electronic warfare, their Pulsar system already operates fully autonomously. A five-million-square-foot factory under construction in Ohio targets hyperscale production.

These aren’t peripheral players. They’re building the infrastructure for state-scale coercion (surveillance at population scale, lethal autonomy at production scale) and integrating it into the military and intelligence apparatus in ways that create deep structural dependency.

Capital overlaps more than coalitions do. Founders Fund invested in both Anthropic and Anduril. Andreessen Horowitz funds defense tech and frontier AI alike. But shared investors don’t dissolve the divide. It runs through corporate structure and what each camp believes AI should be built to do, not capital allocation.

The government, in this picture, is not the counterweight but the instrument one coalition uses against the other.

The Anthropic-Pentagon standoff is what this topology looks like in practice.

Anthropic had two specific contractual restrictions in its Pentagon deployment: prohibitions on mass domestic surveillance of Americans and on fully autonomous weapons without meaningful human oversight. When the company refused to remove them, it was designated a supply-chain risk to national security — a classification built for Kaspersky, Huawei, and ZTE. It had never been applied to a domestic American company.

Defense Secretary Hegseth posted the designation publicly, using the phrase “defective altruism.” Emil Michael, the Under Secretary of War for Research and Engineering, had brokered the arrangement. That same evening, OpenAI announced a replacement deal. Sam Altman later admitted it was “rushed” and “looked opportunistic and sloppy.” Brad Carson, former Under Secretary of the Army, reviewed OpenAI’s surveillance prohibition and concluded it “doesn’t really exist.”

Thirty-seven engineers and scientists from OpenAI and Google, including Jeff Dean (Google DeepMind’s chief scientist), filed an amicus brief supporting Anthropic’s lawsuit against the contract. Caitlin Kalinowski, OpenAI’s head of hardware and robotics, resigned over the deal, calling it “a governance concern first and foremost.”²

Trace the dependency chain. This is not government constraining a private actor. It’s one coalition, equipped with surveillance infrastructure and autonomous weapons production, using state power to discipline another for maintaining a specific boundary. The two restrictions Anthropic refused to drop map onto what the defense-tech camp provides: Palantir’s surveillance, Anduril’s autonomous weapons.

The topology predicts the incident.

There’s a tempting escape from this analysis: adaptive stewardship. The idea that functioning institutions will adapt, that the republic’s immune system will respond to these pressures and produce new equilibria. Dean Ball is the clearest advocate of this position. His intellectual formation runs through Michael Oakeshott: deep skepticism of rationalist planning, conviction that governing means keeping afloat on an even keel rather than steering toward destinations.³

But Ball himself has abandoned its core assumption. In recent writing, he diagnosed the American republic as structurally dying. Each administration governs increasingly by executive fiat and less through durable institutional process. He described the Anthropic-Pentagon standoff not as a policy dispute but as a symptom, calling the supply-chain risk designation “attempted corporate murder“ — placing him in direct opposition to the administration whose AI policy he drafted months earlier.

The adaptive position requires a republic healthy enough to adapt. Its own advocate pronounces the patient terminal. Both Kokotajlo’s planned intervention and Ball’s adaptive navigation assume a functioning public sector, either as actor or as adaptive medium. If the public sector is the thing eroding, the question transforms.

It’s no longer how much government? It becomes: whose values get encoded into the systems that will operate in the space the republic used to occupy?

That question is already being answered. Anthropic encodes its principles explicitly through Constitutional AI — natural-language constraints in published constitutions, with traceable reasoning. OpenAI’s replacement approach relies on citing existing laws and trusting government compliance. One method makes the constraints legible and contestable. The other makes them invisible.

This topology gives you a filter for evaluating AI policy claims. Most proposals assume the public-versus-private axis. Three questions test whether that assumption holds.

First: who actually holds the capacity being governed? If the answer is “the private sector,” ask which private sector. Frontier AI labs and defense technology companies have different capabilities, different incentives, and different relationships to state power. Treating them as a monolith produces policy that addresses a fiction.

Second: what’s the enforcement mechanism? If a proposal depends on government capacity — regulatory oversight, institutional memory, durable legal process — check whether that capacity exists in the form the proposal requires. Ball’s own diagnosis suggests it may not.

Third: whose values get encoded by default if the proposal fails? When policy assumes a map that doesn’t hold, it fails in predictable directions. The values that prevail belong to whoever has operational capacity. Right now, that’s two private coalitions with very different answers.

Understanding AI means understanding the machinery, including the power structure around it. The wrong axis produces something worse than bad policy: the confidence that comes from answering the wrong question well.

Notes

¹ The Anti-Debate, designed by Stephanie Lepp’s Synthesis Media, was moderated by Liv Boeree. Kokotajlo, lead author of “AI 2027“ and founder of the AI Futures Project, resigned from OpenAI in 2024 and forfeited roughly two million dollars in equity rather than sign a non-disparagement clause. His earlier forecasts predicted chain-of-thought reasoning, inference-time compute scaling, and agent frameworks with strong accuracy. Ball served as senior policy advisor for AI at the White House from April through August 2025, where he was the primary drafter of America’s AI Action Plan.

² Altman’s admission is from an internal memo reported by Fortune. Carson’s assessment appears in NBC News reporting and The Intercept’s analysis of the replacement contract’s surveillance provisions. Disclosure: Carson’s organization, Americans for Responsible Innovation, has received $20 million from Anthropic — making him an informed but interested critic. The amicus brief, Kalinowski’s resignation, and the Hegseth designation are sourced from contemporary reporting. Lawfare’s analysis covers the legal precedent question. For Amodei’s public response, see CNBC’s reporting and the TechPolicy.Press timeline.

³ Oakeshott’s key text is Rationalism in Politics (1962). His central argument — that political knowledge is practical and traditional rather than technical and transferable — underpins Ball’s skepticism of top-down AI oversight frameworks. Ball’s Persuasion essay is the clearest statement of his current position.

The gap isn’t information — the internet is drowning in AI information. Papers, podcasts, product announcements, breathless Twitter threads. The gap is comprehension. The ability to take a piece of AI news, understand what’s actually happening underneath the press release, evaluate whether it matters, and decide what — if anything — it means for your work and your life.

That ability is almost entirely absent from the current media landscape, and the absence isn’t accidental. The publications covering AI have incentives that work against the content you need. Labs control what research they publish and when. Newsletters monetize through tool recommendations and affiliate links, biasing them toward hype — every new product is a miracle, every update is a game-changer. Tech press needs early access to labs for scoops, which constrains how critical they can be. Mainstream media needs dramatic framing for clicks, so every story is either salvation or apocalypse.

What’s left out of every layer is the register that matters most: here’s how this actually works, here’s what we know and don’t know, here’s how to think about it for yourself.

PopAi exists to fill that gap.

The tradition that died

For most of the twentieth century, Americans had institutions that did this work. Popular Mechanics, Popular Science, Scientific American under Gerard Piel — these weren’t just magazines. They were informal technical education at civilizational scale, reaching millions of people who learned to reason about technology through a distinctive combination of honest explanation, visual clarity, and respect for the reader’s intelligence.

Popular Mechanics’ founder Henry Haven Windsor — a former city editor who once spent six months disguised as a grip car operator to understand how mechanical workers actually thought — built the magazine on a radical editorial principle. “Most magazines use illustrated articles,” he explained. “We do not. We use described pictures.” Text existed to annotate the image, not the other way around. The thing you couldn’t see clearly was primary. The explanation served it. His tagline for over a century — “Written So You Can Understand It” — wasn’t a marketing slogan. It was a binding editorial commitment: we will trust you with real complexity, and in return, we’ll do the work to make it structurally clear.

That inversion — start with the thing that’s invisible or opaque, make it primary, build the explanation around it — is what the best technical communication has always done. It’s what nobody is doing for AI.

That tradition is effectively dead. Popular Science was gutted in a five-minute Zoom call in 2023 after six ownership changes stripped every trace of editorial mission. Popular Mechanics survives but drifts. The digital successors — YouTube, Reddit, Substack — serve fragments of the function but lack what the institutional press provided: curatorial breadth, editorial independence, and the willingness to expose readers to things they didn’t already know they needed to understand.

The result is a civic problem, not just a media one. About two-thirds of American adults lack the scientific literacy to independently evaluate technical claims in policy debates. That number was bad before AI. It’s going to get worse, because AI is the most consequential technology most people will encounter in their working lives, and the information ecosystem around it is failing them.

What PopAi is

PopAi is a newsletter about understanding AI — the machinery, not the marketing.

When a lab releases a new model, the interesting question is never what they announced. You can get that from a dozen sources by lunchtime. The interesting question is how it actually works, what the architecture reveals about its real capabilities and limitations, and what the gap between the technical report and the press release tells you about the incentives in play. PopAi writes at the level of someone who’s smart, works in a technical or knowledge-intensive field, and wants to understand mechanism — not someone who needs a research paper, and not someone who needs a summary.

That requires being honest about uncertainty. Research on trust in science communication shows that expressing doubt — “the evidence suggests but doesn’t prove,” “we don’t know yet” — increases reader trust, because it reduces perceived bias. The AI discourse has collapsed into a doomer-booster binary where both sides share the same flaw: more confidence than the evidence warrants. AI systems are powerful, limited, and poorly understood even by the people building them. A publication worth reading should reflect that reality rather than flattening it into a narrative.

Most AI coverage positions the reader as a spectator watching the future arrive, either thrilled or terrified. PopAi treats you as someone who can evaluate. The point isn’t to hand you conclusions but to build the kind of understanding that lets you evaluate the next claim yourself — the one I haven’t written about yet.

What most AI coverage actually optimizes for

Open any of the big AI newsletters — the ones with a million-plus subscribers — and trace the dependency chain of a typical post. A company announces a new model. The newsletter covers the announcement: benchmarks quoted from the company’s own technical report, a paragraph of context, a verdict. Sometimes there’s a sponsored tool recommendation at the bottom. Sometimes the tool recommendation is the post.

The problem isn’t dishonesty. It’s that the structure optimizes for something other than your understanding. Affiliate revenue requires product mentions. Access to early briefings requires a relationship with labs that critical coverage jeopardizes. Daily publishing cadence requires speed that precludes depth. Each incentive is rational. None of them serves the reader who wants to know what’s actually going on underneath the announcement.

Prediction is the other dominant mode. Will AGI arrive by 2027? Will AI replace your job? Which startup will win? These questions generate enormous engagement because they’re unanswerable — you can argue about them forever without resolving anything. They mirror the doomer-booster binary: both sides are more confident than the evidence warrants, and the confidence is the product, not the analysis.

PopAi doesn’t operate in either register. I’ll make arguments, take positions, tell you when I think the conventional wisdom is wrong. But I’ll show the reasoning, and I’ll mark what’s load-bearing and what’s speculative. Independence from labs, from tool vendors, from the prediction industry — not because neutrality is a virtue (neutrality is just the absence of a position) but because understanding technology is a civic capacity, and that conviction shapes what I cover and how.

Who this is for

Scientific American’s great editor Gerard Piel described his reader as “someone who knew about one area of science but wanted to know about other areas.” That formulation resolves a tension most publications get wrong: you don’t have to choose between depth and accessibility if you’re writing for people who are already competent thinkers in their own domain.

The software engineer who understands distributed systems but has no mental model for transformer architectures. The biologist who’s fluent in statistics but couldn’t explain gradient descent to a colleague. The product manager who can build a roadmap but can’t evaluate whether a vendor’s AI claims are real or marketing. These people don’t need simplified content. They need content that does the work of making unfamiliar complexity navigable.

There’s a less obvious audience too: the person who just wants to understand the thing everyone’s talking about. Not to build anything or optimize a workflow — just to understand. That impulse — curiosity about the technological world you inhabit — is the same impulse that drove millions of people to subscribe to Popular Mechanics for a century. It deserves the same respect now.

What to expect

One substantive piece per week. One essay that goes deep on something that matters.

The content falls into three registers. Explanatory engineering: how retrieval-augmented generation actually works, what’s going on inside chain-of-thought reasoning, why context windows matter and what their limits mean. The kind of thing you’d want to read before evaluating an AI claim at work next week. Analysis: what a particular development actually changes, for whom, and what it doesn’t change — the part that usually gets lost in the excitement or the panic. And methodology: how to think about a particular class of AI problem. I’m building a framework for AI-assisted work called Absolute Beginners++ that synthesizes several convergent problem-solving traditions into a practical method. The newsletter is where I apply that thinking in public.

The AI news cycle rewards speed. I’m optimizing for shelf life — pieces worth reading six months after publication, not just the morning they arrive.

Why me

I research and build AI for a living — ML platforms at AWS and H2O.ai, AI-integrated products, the plumbing that sits between a model and the thing it’s supposed to do in the real world. I work directly with the systems I’m writing about, which means I know where the press releases diverge from the engineering reality, and I know which limitations are temporary and which are architectural.

Before that, I spent time in Army Special Forces. The core problem AI presents to most people — making high-stakes decisions under uncertainty with incomplete information, in a domain that shifts faster than your mental models — is the same problem unconventional warfare presents. The military’s response is frameworks: repeatable structures for reasoning that hold up when the specific situation is novel. That instinct runs through everything here.

I also have a serious contemplative practice. The hardest part of thinking clearly about AI isn’t technical — it’s attentional. The hype cycle, the FOMO, the constant pressure to adopt the next tool. These are attentional problems before they’re information problems. Noticing when your relationship to a technology is driven by clarity versus reactivity is a skill, and PopAi will occasionally address it directly.

The bet

The bet underneath all of this: understanding reduces fear. The data bears it out — people with greater AI awareness feel more comfortable with AI applications, not less. The gap between how AI experts and the general public feel about the technology isn’t a disagreement about values. It’s an information asymmetry. Experts aren’t more comfortable because they’re naive. They’re more comfortable because they understand the machinery, including where it breaks.

A publication that closes that asymmetry — that helps people build real mental models of how AI systems work, what they can and cannot do, and how to evaluate the claims made about them — does more than inform readers. It changes the quality of the conversation — from anxious speculation toward grounded judgment, from spectators toward participants.

The popular technical press did this for a century with mechanical and electrical technology. Nobody is doing it for AI. PopAi is where I start.

Subscribe if you want to read it. Send it to someone who needs it. And if you think I’m wrong about any of this, tell me — the whole point is to think clearly, which means being willing to update.

You can also find me and some of my more divergent writing and projects at https://thbrdy.dev and you can always reach out to me at jthomasbrady@proton.me