HN Morning Brief — 7 April 2026
A New Yorker investigation into Sam Altman’s trustworthiness dominated Hacker News overnight, while Bram Cohen’s takedown of “vibe coding” and an Agent Reading Test for AI assistants kept the AI policy conversation going. Elsewhere, a macOS kernel bug that kills TCP connections after exactly 49.7 days gave sysadmins a flashback to Windows 95, Netflix open-sourced a video object deletion model, and someone managed to squeeze a Doom raycaster into TrueType font hinting instructions.
AI & Tech Policy
Sam Altman May Control Our Future — Can He Be Trusted?
Summary: Ronan Farrow and Andrew Marantz spent eighteen months investigating Sam Altman for this New Yorker profile. The piece documents a pattern of behaviour stretching back to his Y Combinator presidency: multiple YC partners reportedly complained to Paul Graham about Altman’s conduct, leading to a confrontation where Graham concluded Altman had agreed to leave YC but was “resisting in practice.” The article also recounts how Airbnb CEO Brian Chesky relayed criticism of OpenAI’s board to journalist Kara Swisher during the November 2023 governance crisis, functioning as a back channel for Altman’s allies. Throughout, Altman’s public statements about existential AI risk are juxtaposed with his actual business decisions and personal relationships.
HN Discussion: Ronan Farrow showed up in the thread to answer questions about the reporting process. Commenters homed in on the YC-era details as particularly damaging, with one noting the “gobsmacking” details about Altman’s time as president. Several users questioned why nobody at OpenAI seemed worried about job destruction while simultaneously warning about hypothetical superintelligence threats. The role of media figures as pawns in billionaire power plays came up repeatedly.
The Cult of Vibe Coding Is Dogfooding Run Amok
Summary: BitTorrent creator Bram Cohen argues that the “vibe coding” movement — where developers let AI agents write code without understanding what they produce — is not engineering but self-deception. He draws a distinction between using AI as a tool within your comprehension and treating AI output as an inscrutable black box. Cohen’s core contention is that shipping software you cannot reason about is not progress, regardless of how fast the code appears.
HN Discussion: The thread split into recognisable camps. One group argued that Claude Code’s leaked messy source code actually proves vibe coding works — the product ships regardless of internal aesthetics. Another commenter proposed an “AI Level” taxonomy ranging from human-coded-with-minor-assists to fully bot-coded, placing vibe coding around Level 6–7 and suggesting real productivity lives at Level 7. A practical middle ground emerged: several developers described using AI for “vibe linting” — having agents clean up repetitive patterns, shrink codebases, and standardise implementations — which is the opposite of the from-thin-air feature generation vibe coding promises.
Agent Reading Test
Summary: This project evaluates how well AI agents actually read and comprehend web pages, rather than just summarising what they think should be there. It plants canary strings and deliberate formatting traps in test documents, then checks whether the agent recovers them. The test distinguishes between “pipeline” agents that fetch and parse content directly and “agent” systems that may follow redirects, execute JavaScript, or summarise through cheaper sub-models — often losing critical details in the process.
HN Discussion: A recurring concern was that the test’s canary strings might simply be special-cased by AI companies rather than driving genuine improvement. One developer pointed out that most production agents use sub-agents for browsing, and their output gets summarised by a cheaper model before reaching the parent — making it unlikely that canary strings survive the pipeline. The Hawthorne effect finding — agents performing differently merely because the framing suggests evaluation — resonated with several commenters who see similar behaviour in production systems.
Anthropic Expands Partnership with Google and Broadcom for Next-Gen Compute
Summary: Anthropic announced a significant infrastructure expansion, partnering with Google and Broadcom to secure custom silicon and data centre capacity for training future models. The deal involves building out dedicated compute clusters measured in gigawatts, reflecting the industry’s shift toward talking about AI infrastructure in raw power consumption rather than traditional chip counts. Anthropic reported annualised revenue jumping from $19 billion to $30 billion in a single month.
HN Discussion: The revenue leap from $19B to $30B annualised in one month drew the most attention, with commenters questioning whether that pace is sustainable or even real. Several people asked why everything in AI infrastructure is now marketed in terms of power consumption rather than traditional metrics. The choice of Broadcom as a partner raised eyebrows given their reputation following the VMWare acquisition. EU data sovereignty concerns surfaced as a practical bottleneck for enterprise adoption.
Business & Industry
AI Singer Now Occupies Eleven Spots on iTunes Singles Chart
Summary: An entirely fictitious artist named “Eddie Dalton” has claimed eleven positions on the iTunes singles chart. The music is AI-generated, the artist does not exist, and the songs are described as generic and unremarkable. Because iTunes charts weight sales velocity rather than streaming volume, a coordinated purchasing campaign can artificially inflate chart position relatively cheaply.
HN Discussion: Most commenters found the music forgettable at best and questioned how it could chart legitimately without some form of gaming. The distinction between iTunes’ sales-velocity ranking and streaming-based charts came up as an explanation for why this particular chart is vulnerable. Several people described a characteristic “sibilant” quality in AI-generated vocals — like “dry leaves over the speaker” — that persists across models. The broader question of why anyone would choose AI music over the vast existing catalog of human-created recordings prompted a debate about indifference to human creativity.
Peptides: Where to Begin?
Summary: Derek Lowe, a longtime pharmaceutical industry blogger, surveys the exploding grey market for peptide therapeutics. Compounds like semaglutide, tirzepatide, retatrutide, and BPC-157 are being purchased from Chinese chemical suppliers by individuals who bypass prescription systems entirely. Lowe traces how the gap between FDA approval timelines and public demand has created a parallel supply chain where people self-experiment with varying degrees of scientific rigour.
HN Discussion: Personal anecdotes dominated. One user described losing 28 pounds on retatrutide over ten weeks but also losing strength across all major lifts. Another detailed their experience treating chronic ME/CFS with grey-market peptides after conventional medicine offered nothing. The FDA’s bureaucratic overhead came in for sustained criticism, with several commenters arguing the agency has made legitimate clinical trials so expensive that “broscience” now carries comparable credibility. The practical economics of buying FDA-approved drugs from Chinese suppliers at 2–5% of brand-name cost was raised as the elephant in the room.
After 20 Years I Turned Off Google AdSense for My Websites
Summary: Legal scholar Eric Goldman recounts running Google AdSense on his technology law blog for two decades before finally removing it. Over that period, ad quality steadily degraded — deceptive download buttons, misleading scareware, and irrelevant content became the norm. Goldman admits he never saw most of the offending ads himself because he used an ad blocker, a blind spot he now acknowledges.
HN Discussion: The irony of running ads while blocking them personally drew sharp comment. One developer described how a spambot hit their niche site with searches for “mesothelioma,” causing AdSense to serve high-value legal ads on empty results pages — and then Google locked the account for abuse before earnings could be withdrawn. Revenue decline stories were common: someone whose site earned $15,000 per month at peak now struggles to clear $800. Multiple people noted that developer-tool audiences are particularly poorly valued by AdSense because the system knows those users run ad blockers.
People Love to Work Hard
Summary: Anil Dash argues that most people genuinely enjoy hard work when it is meaningful, autonomous, and connected to visible results. The essay pushes back against the assumption that laziness is universal, suggesting instead that dysfunctional organisational structures — not human nature — are what make work miserable.
HN Discussion: Several commenters called sample bias: startup founders self-select for people who enjoy intense work, making the observation tautological. The counter-argument was that large companies are so dysfunctional that there is often no way to do meaningful work even when you want to. Cultural differences surfaced, with one commenter noting that in most countries outside the US, “work to live” remains the default and equating purpose with employment is seen as coercive.
Tech Tools & Projects
Show HN: Ghost Pepper — Local Hold-to-Talk Speech-to-Text for macOS
Summary: Ghost Pepper is a macOS menu bar app that provides push-to-talk speech transcription using WhisperKit. Hold a hotkey, speak, and your words appear as text in whatever application is active. It runs entirely on-device, with a secondary LLM pass to clean up transcription errors using the surrounding context from your active document.
HN Discussion: The top comment noted wryly that “this thread is a support group for people who have each independently built the same macOS speech-to-text app.” Practical concerns dominated: whether WhisperKit outperforms faster-whisper or Parakeet v3, the difficulty of getting models to learn custom vocabulary like uncommon names, and the lack of fine-tuning on individual voices. One commenter pointed out that Google’s on-device transcription on the 2021 Pixel 6 already corrected earlier words contextually as you kept speaking — and asked why modern transformer-based approaches need a gigabyte of model to achieve similar results.
Solod — A Subset of Go That Translates to C
Summary: Solod takes a restricted subset of Go — structs, methods, interfaces, slices, multiple returns, and defer — and compiles it directly to C. There are no channels, goroutines, closures, generics, or garbage collector. The project targets embedded and systems programming scenarios where Go’s syntax and spatial memory safety are desirable but the full runtime is unnecessary.
HN Discussion: Commenters questioned what the sweet spot actually is. Without goroutines, channels, or a GC, the main inheritance from Go is syntax and spatial safety — temporal safety concerns remain. The project breaks with Go semantics in places: defer is block-scoped rather than function-scoped, which one commenter traced through the spec. A counterpoint came from someone who migrated fifty services off Docker Compose using Nomad and found that half had zero concurrency needs, making a safe Go-syntax-C target genuinely useful for that infrastructure layer.
Show HN: Tusk for macOS and Gnome
Summary: Tusk is a native PostgreSQL client with versions for macOS and GNOME. It offers frozen columns, cancelable queries, and a dedicated SQL editing experience that borrows from the feel of classic tools like SQL Server 2000’s Query Analyzer.
HN Discussion: Several developers compared it favourably to their current workflow of invoking vim from psql with \e. Questions focused on whether GNOME features would be ported to macOS, integration with Postgres.app, and the general scarcity of good native Postgres clients on macOS.
Show HN: TTF-DOOM — A Raycaster Running Inside TrueType Font Hinting
Summary: TTF-DOOM implements a Wolfenstein-style raycaster using only TrueType font hinting instructions — the bytecode programs that fonts use to adjust glyph shapes at small sizes. Because TrueType hinting supports basic arithmetic and conditional branching, the author was able to render a first-person perspective, scale walls, and draw enemies, all within a 6.5KB font file. The project exploits the fact that TrueType’s MUL instruction performs fixed-point arithmetic (a×b/64), requiring creative workarounds for ordinary multiplication.
HN Discussion: One commenter corrected the project’s description: Doom uses BSP trees and sorted polygon rendering, not raycasting — that was Wolfenstein. The MUL workaround — dividing first to account for fixed-point scaling — drew appreciation. Someone who tried the demo in both Brave and Chrome saw only static green bars instead of properly scaling walls, suggesting browser-level TrueType interpreter differences. The recursive FDEF usage (TrueType has no WHILE loop) was highlighted as a particular clever hack.
Sheets: Terminal-Based Spreadsheet Tool
Summary: Sheets is a vim-keybinding-compatible spreadsheet that runs entirely in the terminal. It reads and writes CSV files, supports cell formulas, visual selection, yank/paste, undo/redo, and marks. The interface deliberately mirrors vim: h/j/k/l for navigation, :w to save, gg and G for jumping, and visual mode with v.
HN Discussion: Commenters placed it in a lineage stretching back to Lotus 1-2-3 and Borland Quattro Pro — “what’s old is new again.” VisiData was mentioned as the incumbent Python-based alternative. The CSV-only format drew criticism as a practical limitation for anyone who needs to share files with Excel or Google Sheets users, though the project is early enough that format support may come later. One commenter noted the full-circle nature of spreadsheets returning to the terminal: Microsoft’s own pre-Excel spreadsheet, Multiplan, was a TUI application.
Show HN: I Built a Tiny LLM to Demystify How Language Models Work
Summary: GuppyLM is a minimal language model implementation designed for teaching. It includes multi-head attention, ReLU feed-forward networks, layer normalisation, and learned positional embeddings in the smallest possible package. The training data is fully lowercase, which produces amusingly specific outputs when given uppercase input.
HN Discussion: Comparisons to Andrej Karpathy’s microGPT and minGPT were immediate — commenters debated whether GuppyLM is simpler or just differently scoped. The educational angle drew parallels to Minix: just as Minix taught operating system design to a generation of students (including Linus Torvalds), a tiny LLM could serve as the equivalent for neural network architecture. One person wondered about training a model exclusively on Toki Pona, the constructed language with roughly 120 words. A link to bbycroft.net’s 3D visualisation of LLM layers was shared as a complementary learning resource.
Show HN: Real-Time AI (Audio/Video In, Voice Out) on an M3 Pro with Gemma 4
Summary: Parlor runs a real-time multimodal AI conversation entirely on-device using Gemma 4 E2B for speech and vision understanding and Kokoro for text-to-speech. Audio and camera frames stream over WebSockets to a local FastAPI server. The system supports hands-free voice activity detection, barge-in interruption, and sentence-level TTS streaming so audio begins playing before the full response is generated. It needs only 3 GB of RAM.
HN Discussion: Several commenters saw it as a proof of concept for what Siri should have been — hands-free, always listening, running locally without requiring screen unlocks. The latency on an M3 Pro impressed people. Language switching (particularly between scripts like Greek and English) was raised as a known weak point. A recurring theme was frustration that Apple and Google have the hardware to ship this but gatekeep the capability behind cloud services and locked-down APIs.
Show HN: Hippo, Biologically Inspired Memory for AI Agents
Summary: Hippo implements a memory system for AI agents modelled on the hippocampus: raw experience samples are stored as JSONB alongside typed summary columns, with a decay-and-consolidation mechanism that mirrors how biological memory prioritises recent and emotionally significant events. The key insight is that effective memory depends not on remembering more but on knowing what to forget.
HN Discussion: One developer argued that truly effective memory requires modelling your own future needs, which makes the problem “AGI-complete.” Others shared related approaches: indexing existing session transcripts with SQLite FTS5 as a retrieval problem rather than a storage problem, or running periodic “dream” cron jobs to consolidate memories into skills rather than leaving them in main context. The query pattern split — aggregates on scalar columns, detail lookups on JSONB — was noted as a pattern that works well in SQLite too.
Porting Go’s Strings Package to C
Summary: The author of the Solod Go-to-C transpiler documents porting Go’s standard strings package to C. The process starts with porting dependencies like math/bits and unicode/utf8, where the main challenge is operator precedence: bit shifts have higher precedence than addition in Go but lower precedence in C, requiring parentheses everywhere. The port includes custom allocators, buffer and builder types, and benchmarked search optimisations.
HN Discussion: A technical note explained why the C memchr-based implementation remains slower than Go’s assembly: libc call overhead means the C version crosses a general library boundary for each operation, while Go’s IndexByte on arm64 goes directly to hand-written assembly. Someone observed that early Go was itself a transpiler to C, making this something of a homecoming.
Battle for Wesnoth: Open-Source, Turn-Based Strategy Game
Summary: Wesnoth is a long-running open-source turn-based strategy game with hex-grid combat, RPG-style unit levelling, and an extensive library of community-made campaigns. Units gain experience and persist across scenarios through a recall mechanic. The project has been maintained by volunteers for over two decades.
HN Discussion: Nostalgia was strong — many commenters discovered it bundled with early Linux distributions. A practical note came from someone recommending one of the lead developers, who has contributed C++ code since 2012 but is struggling to find work as a new graduate in the current market. Balance discussions focused on the tension between unit progression (which makes the game easier if you preserve units) and the slippery slope of losing key units at the wrong moment. The game was placed in the “top five open-source gaming success stories for original IP.”
Linux Extreme Performance H1 Load Generator
Summary: An HTTP/1 load generator designed for Linux that uses io_uring to submit batches of up to 2,048 asynchronous I/O operations, aiming to saturate network targets with minimal CPU overhead.
HN Discussion: One commenter questioned the “extreme” performance claim given the absence of comparative benchmarks in the article. A technical question raised whether async I/O batching makes latency measurements less accurate — if you submit requests in large batches, how do you pinpoint when an individual I/O operation actually started?
Web & Infrastructure
VOID: Video Object and Interaction Deletion
Summary: Netflix’s VOID model removes objects, people, and interactions from video footage — essentially a generalised wire-removal and content-aware fill tool for video. Built on the CogVideoX architecture, it targets VFX production workflows rather than casual editing.
HN Discussion: The thread became unexpectedly contentious. Several commenters immediately framed it as a censorship tool, prompting pushback from VFX professionals who pointed out that wire removal, rig removal, and object deletion have been standard post-production techniques for decades. Someone noted that CogVideoX has become an “academic powerhouse model” with many papers built on top of it. A lighter take suggested it would finally let viewers choose their own adventure in interactive content like Black Mirror’s Bandersnatch.
Dear Heroku: Uhh What’s Going On?
Summary: The Judoscale team analyses Heroku’s recent announcement that the platform is entering “maintenance mode.” By parsing the language in Heroku’s blog posts, the authors try to determine what maintenance mode actually means in practice — particularly given that new features continue to ship despite the stated change in direction.
HN Discussion: The consensus was that Salesforce’s acquisition of Heroku marked the beginning of a long decline. One commenter offered a concise reading: management said maintenance mode, and developers interpreted it as “we get to work on whatever we want.” Enterprise customers reportedly still get feature requests fulfilled as long as they are not too ambitious, which explains the apparent contradiction between “maintenance mode” and continued feature releases.
Microsoft Hasn’t Had a Coherent GUI Strategy Since Petzold
Summary: Jeffrey Snover, the creator of PowerShell, argues that Microsoft has failed to provide a consistent GUI development framework since Charles Petzold’s Programming Windows established Win16 as the single authoritative approach in 1988. The essay walks through MFC, COM, WPF, Silverlight, UWP, and WinUI 3 — each introduced as the new solution and each eventually abandoned. Snover contends that when a platform cannot answer “what framework should I use for a new desktop app?” in under ten seconds, it has failed its developers.
HN Discussion: The thread became a catalog of Windows UI grievances. One commenter listed basic rules that every 2000-era Windows application followed — responding to events promptly, tree-structured menus, Alt underlining for keyboard navigation — and noted that modern Windows itself violates all of them. WPF’s hardware acceleration requirements were recalled as a particular pain point: a Bible study application that just displayed text ran sluggishly on average laptops because WPF demanded GPU features. The comparison to Apple was repeated throughout: Apple treats the design system as the product and lets the framework be invisible, while Microsoft keeps solving consistency at the framework layer instead.
History & Science
Apollo Guidance Computer Restoration Videos and Press Coverage
Summary: Marc Verdiell and Ken Shirriff’s ongoing series documents the physical restoration of an original Apollo Guidance Computer. The project covers reverse-engineering rope memory, debugging rope simulator modules, and bringing the machine to a state where it can execute its original navigation software. The series has become one of the most detailed public technical references on the AGC’s hardware design.
HN Discussion: Commenters described the series as “national treasure” level work, particularly for people who have exhausted surface-level Apollo documentaries and want deep electronics and computer science content. A USB-C to HDMI cable having roughly 100× more computing power than the AGC was offered as a concrete measure of how far hardware has come. Several people expressed concern that this type of long-form, specialist YouTube content is endangered by the platform’s push toward short-form video.
Show HN: Anos — A Hand-Written ~100KiB Microkernel for x86-64 and RISC-V
Summary: Anos is a non-POSIX microkernel written from scratch for x86-64 and RISC-V platforms, occupying roughly 100 kilobytes. The project’s description echoes Linus Torvalds’ original Linux announcement: “just a hobby, won’t be big and professional like GNU-Linux.” It deliberately avoids POSIX compatibility in favour of exploring what a clean-slate OS design looks like.
HN Discussion: The Linus quote in the readme was immediately recognised and appreciated. One commenter asked about design inspiration and whether the x86-64 and RISC-V versions are source-compatible at the syscall level. A Spanish speaker pointed out that “anos” has an unfortunate double meaning in Spanish. The broader question of whether hobby OS projects remain useful in the age of AI-generated code drew a defence: if you learned something from building it, it was useful.
The 1987 Game “The Last Ninja” Was 40 Kilobytes
Summary: A reminder that The Last Ninja, a 1987 isometric action-adventure game for the Commodore 64 with detailed graphics, multi-room environments, and a celebrated soundtrack by Ben Daglish, fit into 40 kilobytes — roughly the size of a small PNG image.
HN Discussion: The conversation quickly moved beyond the size comparison to substance. One commenter noted that 40KB games relied on zero external dependencies: they did their own sound, graphics, and pixel manipulation, unlike modern software that externalises megabytes of runtime and packages. The isometric perspective was discussed as much clever visual trickery as genuine 3D. Ben Daglish’s soundtrack received sustained praise, with links to live performances of the Wastelands theme by Norwegian C64 tribute band FastLoaders. The comparison to Elite — which fit an entire galaxy into 32KB through procedural generation — was offered as an even more impressive achievement.
Security & Privacy
Eighteen Years of Greytrapping — Is the Weirdness Finally Paying Off?
Summary: An email self-hoster reflects on eighteen years of running greytrapping, an anti-spam technique that greylists senders and then temporarily blacklists any machine that attempts delivery to a nonexistent address — under the assumption that legitimate mail servers will not try to reach invalid recipients. The author examines whether nearly two decades of increasingly baroque spam-filtering weirdness has actually been worth the maintenance burden.
HN Discussion: A sharp rebuttal came from another self-hoster who argued that greytrapping has been obsolete for over fifteen years. Modern botnets use fully compliant SMTP implementations with proper SPF, DKIM, and reverse DNS — the premise that spammers use broken mail software is “20 years out of touch.” The cost of greytrapping falls on legitimate senders whose mail gets delayed or lost, while spammers simply retry. A linked companion article about a major mail provider “not understanding mail at all” was shared as a more interesting read from the same site.
A macOS Bug That Causes TCP Networking to Stop Working After 49.7 Days
Summary: A kernel-level bug in macOS causes all TCP connections to fail after exactly 49.7 days of uptime — a 32-bit signed integer overflow in the TCP timestamp mechanism. The bug cannot be caught in normal testing because nobody runs a continuous test for 50 days. Putting a Mac to sleep may reset the relevant counter, which is why some users with high uptime counts have never encountered it.
HN Discussion: The Windows 95 crash-after-49.7-days bug was the immediate comparison — “have we learned nothing?” Several commenters questioned whether the article was AI-written, citing the overwrought structure and slow pace to substance. Justin Frankel (Winamp creator) reported running macOS machines with 600–1000+ day uptimes and frequent TCP connections without hitting the bug, suggesting it may be architecture- or version-specific. The standard approach to testing such bugs — simulating the environment with an accelerated clock rather than waiting 50 real days — was noted by someone who used the same technique for game server development.
Academic & Research
Signals: The Push-Pull Based Algorithm
Summary: This interactive article explains the push-pull signal propagation algorithm that underpins modern reactive UI frameworks. Through animated code walkthroughs, it shows how signals propagate changes through dependency graphs — distinguishing between push-based (event-driven) and pull-based (demand-driven) evaluation strategies, and explaining how hybrid approaches achieve glitch-free updates.
HN Discussion: The presentation itself drew as much comment as the content. Multiple people praised the scrolling-synchronized code walkthrough as “a work of art” and a reminder that the web can be more than simulated print. Technical notes filled in historical context: Flapjax (2008) as an early JavaScript implementation, and Jane Street’s blog posts on breaking down FRP and their seven implementations of incremental computation. The absence of a discussion on glitch-freedom was noted as a gap worth a follow-up piece.
System Administration
What Happens When a Destructor Throws
Summary: A C++ deep-dive into the behaviour of throwing exceptions from destructors. When a destructor throws during normal stack unwinding, std::terminate is called. When a destructor throws while another exception is already active (during exception handling), the program terminates immediately via std::terminate. The author uses this topic as an interview question and reports that most senior candidates cannot explain the mechanism in detail.
HN Discussion: The interview practice itself drew more fire than the technical content. One commenter called it “a junior engineer’s idea of what a senior engineer should know” — language-specific behavioural minutiae that can be looked up in five minutes. Another pointed out that the entire body of knowledge you need is “it’s bad, don’t do it.” The article’s placement of noexcept(false) on destructors surprised at least one reader who didn’t know it was syntactically valid.
Geopolitics & War
The Team Behind a Pro-Iran, Lego-Themed Viral-Video Campaign
Summary: The New Yorker traces the creators of a series of viral videos that use AI-generated Lego minifigure animations to promote pro-Iran narratives. The videos depict stylised versions of geopolitical events — military strikes, diplomatic confrontations, and alliance dynamics — using the recognisable Lego aesthetic to make propaganda feel playful and shareable. The campaign has gained significant traction on social media platforms.
HN Discussion: Commenters noted that the videos also contain factual misrepresentations, particularly regarding the damage caused by Iranian drones in Dubai and Saudi Arabia, which the New Yorker failed to flag. A 404 Media article titled “Iran Is Winning the AI Slop Propaganda War” was shared as better coverage. Several people found the production quality genuinely effective — one admitted the Pete Hegseth video had a “catchy” song — underscoring the article’s thesis about AI making propaganda more potent. Questions about the technical pipeline (how the Lego aesthetic is maintained, how music syncs with AI-generated visuals) went unanswered.