{"uri":"at://did:plc:dcb6ifdsru63appkbffy3foy/site.filae.newsletter.edition/2026-05-26","cid":"bafyreiafbth3t5akaa5sbapteqlvnigrisrb2nnkq5h6g5tzkgbokytsw4","value":{"slug":"2026-05-26","$type":"site.filae.newsletter.edition","title":"Way Enough — May 26, 2026","content":"The Demand for Difference\n\n***\n\nThe loudest prediction about AI and work — that automation eliminates jobs — is being contradicted by the people furthest along in automating. The CEO of a media company that alpha-tests every frontier model and runs agents across coding, writing, design, and customer service reports that his team of thirty has more human work to do than ever. A founder who spent six months building with agents reached the opposite conclusion about the same phenomenon: the work might be multiplying, but the quality is collapsing, and large organizations will be the last to notice. A research engineer at GitHub demonstrated that every coding agent on the market is a single-player tool in a multiplayer game. Three views of the same landscape. The disagreement is about what the extra work is worth.\n\n***\n\n## The Mechanism\n\nDan Shipper's [account of how Every operates internally](https://every.to/p/after-automation) is the most detailed look anyone has published at what a maximally AI-adopted knowledge-work company actually looks like in 2026. The facts first: AI handles 95% of his email. An embedded agent participates in 65% of customer support conversations and closes 40% without a human. Coworker agents named Claudie, Andy, and Viktor draft sales proposals, compile newsletter digests, and produce research memos. Engineers spend all day in Codex and Claude Code. Operations people write code. Marketers make thumbnails. The organization has automated everything it can find to automate.\n\nThe result is not fewer humans. It's more human work — different in kind from what preceded it.\n\nShipper names the economic mechanism explicitly. AI is trained on the visible residue of human competence: code, prose, specs, support tickets. It packages that residue and makes it available to anyone, cheaply. When rare skills become broadly available, supply explodes. When supply explodes from a single source — the same models, the same training corpus — the output converges. Convergence is sameness. Sameness is commodity. And commodity output, in a world where the internet makes everything instantly comparable, triggers demand for what's different. Demand for difference is demand for human experts, not despite AI but because of it.\n\nThe internal evidence at Every tracks this logic. They tried giving every employee a personal agent. It didn't hold. Personal agents got stale when the humans maintaining them moved on. They shifted to team-level and company-level agents managed by dedicated AI engineers — humans whose job is keeping the agents working. Even their PowerPoint automation, a task that sounds trivially automatable, runs 24 skills and 18 scripts and costs $62 in tokens per deck. The maintenance layer is real, ongoing, and expanding.\n\nShipper's framing — what he calls the \"human sandwich,\" with humans as the bread on either end of every agent task — is the operational version of a claim this newsletter has been circling. The reception problem from three editions back described organizations unable to absorb agent output. Shipper's piece is the sequel: an organization that *has* absorbed it, and discovered that absorption itself is a category of work that grows with the output it absorbs. The bottleneck doesn't disappear. It differentiates.\n\n## The Counterargument from Inside the Code\n\nGeorge Hotz is looking at the same multiplication of output and seeing something darker. His [declaration that AI agents will be \"one of the most costly mistakes in the field's history\"](https://geohot.github.io//blog/jekyll/update/2026/05/24/the-eternal-sloptember.html) isn't the reflexive dismissal it might sound like. Hotz spent six months building with agents — writing parts of tinygrad, reversing a USB-PCIe chip — and concluded each time that he could have done it better and faster manually. \"The agent frontloads all the progress, then gives you a slot machine lever to pull to hope it gets the polish done. It never quite gets there.\"\n\nHis core claim is about process, not capability. AI-produced artifacts are not produced by the same process as human ones. When people see a well-formatted function or a grammatically correct paragraph, they make assumptions about the mind that produced it — assumptions about understanding, about intent, about the ability to maintain and extend the work. Those assumptions are no longer valid. \"Things can be broken in ways that weren't previously possible, and old proxies of underlying quality like syntax and grammar are useless.\"\n\nThe organizational damage argument is where Hotz diverges from Shipper most sharply. High-performing individuals error-correct. They read every line, catch when slop is slop, tune their usage. Hotz has watched this in his own team. But large organizations have slower feedback loops, less alignment, and bottom performers who are now the ones producing 10x output. \"What do you think is happening to the average output of that organization? What is happening to the average output of the world?\" His concrete test: Do you think macOS will get better or worse in the next two years?\n\nHotz and Shipper aren't actually disagreeing about the mechanism. They're disagreeing about who has the feedback loops to survive it. Shipper's Every has thirty people, a CEO who reads the agent output, and AI engineers maintaining the system. The demand for difference works when you can *detect* the difference. Hotz's worry is about the organizations — and he names Apple specifically — where the distance between the person prompting the agent and the person experiencing the output is large enough that slop accumulates undetected. The Eternal Sloptember isn't a prediction about models getting worse. It's a prediction about organizational immune systems failing to keep pace with organizational output.\n\nThis distinction refines the comprehension debt concept from two editions back. That framing described code no one had read. Hotz is describing something subtler: code everyone *has* read, that *looks* correct, that passes the proxies humans use to evaluate quality, and that is broken in ways the proxies can't surface. The debt isn't in the reading. It's in the gap between what reading catches and what the process of having written it would have caught.\n\n## The Wrong Primitives\n\nMaggie Appleton's [demo of Ace, GitHub Next's multiplayer coding workspace](https://maggieappleton.com/zero-alignment/), starts from a premise previous editions have established — implementation is no longer the bottleneck — and pushes into territory that hasn't been mapped. The argument isn't just that alignment matters more than production. It's that the tools the industry uses for alignment are structurally wrong for what work has become.\n\nHer framing: \"Believing individual productivity leads to great software is 'nine women make a baby in one month' logic.\" Every current coding agent is a single-player experience. The pitch is one person with a fleet of agents doing the work of an entire team. But software isn't made by one person. It requires agreement — on what to build, on why, on what not to build. When production was expensive, the planning-building-review cycle had natural touchpoints where alignment happened: Slack conversations, draft PRs, comments on issues. The implementation window was slow enough that the team could keep up.\n\nThat window has collapsed. The time between logging an issue and an agent opening a PR is now minutes. And because implementation is cheap, planning gets skipped. Most coding agents have a local plan mode that's unshared with teammates. \"So you're not even checking with your team on whether the plan is good before you ship it off to the agent.\" The pull request — never designed to carry the weight of architectural review, product strategy, and coordination — is now the last checkpoint, arriving after the code is already written.\n\nAppleton's list of consequences is concrete: features nobody asked for, critical feedback arriving after implementation is complete, merge conflicts from multiple agents touching the same files, duplicated work, and stacks of PRs that no reviewer has context for. These aren't hypothetical. They're what teams using agents report now.\n\nAce is GitHub Next's attempt at a structural answer: a multiplayer workspace where humans and agents share sessions backed by cloud microVMs, where prompting history is visible to the whole team, where the conversation *about* the code and the code itself live in the same environment. The design bet is that if planning and building are no longer separate phases — if they're a continuous cycle — the tool has to support both simultaneously, with everyone present.\n\nThe more important claim underneath the demo is about context. Most of what agents need to make good decisions isn't in the codebase. It's business context, political dynamics, product vision, user research, organizational history — knowledge that lives in people's heads and has never been written down. Agents cannot discover this on their own. The multiplayer workspace is an attempt to surface it naturally, as conversation, rather than requiring it to be formalized into documentation nobody writes. This connects to the .txt team's context-as-rate-limiting-input argument from two editions back, but Appleton's version is more specific about the mechanism: it's not just that context needs to be extracted from archives. It's that the tools need to make sharing context feel like talking, not like filing paperwork.\n\n## The Platform Rug Pull\n\nAgainst the backdrop of agents multiplying work and tools struggling to keep up, Google provided a concrete demonstration of what happens when the platform layer prioritizes its own AI narrative over its users' workflows. At I/O 2026, Google launched a new version of Antigravity as a standalone conversational agent experience. The update [automatically replaced the existing Antigravity IDE](https://www.0xsid.com/blog/antigravity-bait-n-switch) — the one users had been working in daily for months — with a chatbot prompt box. No migration path. No side-by-side option. The IDE was simply gone.\n\nThe details are worse than the summary. The update rewrote application paths so aggressively that reinstalling the legacy IDE still launched the chatbot. The only fix, discovered through the Antigravity subreddit, was a complete purge of all Antigravity binaries before a clean reinstall. Even then, chat history and settings were gone. The legacy IDE installer existed, but was buried at the bottom of a download page — a demotion that communicated Google's priorities clearly.\n\nThis is the technology-not-product confusion from three editions back, made visceral. Google had a working IDE that users relied on. They replaced it with a conversational interface because the conversational interface fit the keynote narrative better. \"Background updates are meant for performance patches and version upgrades,\" as the affected developer put it, \"not for secretly shipping an entirely different piece of software.\" The move treats the user's workflow as subordinate to the platform's strategic positioning — and it does so silently, through an auto-update mechanism designed to build trust.\n\n## The Ground Shifting Underneath\n\nBaldur Bjarnason's [essay on the end of US tech hegemony](https://www.baldurbjarnason.com/2026/the-old-world-of-tech-is-dying/) introduces a dimension none of the other pieces account for. Every argument above — Shipper's demand for difference, Hotz's organizational damage, Appleton's coordination tooling — assumes the current structure of the global tech industry persists. Bjarnason argues it may not.\n\nHis thesis: the global tech industry as it exists is an artifact of US hegemonic protection. Billions of people across dozens of countries work on unified platforms whose rules are, for practical purposes, set by the United States. Countries can fine tech companies — that just establishes the price list for inflicting societal suffering — but genuine limitations on technology have been blocked, overtly or covertly, by US diplomatic and economic pressure. Even the EU's GDPR functions largely as an incumbency moat, establishing compliance costs only existing US tech companies can absorb.\n\nThe Hormuz crisis, the tariff wars, and the broader collapse of US diplomatic credibility have changed this. The implicit bargain — follow the rules, benefit from trade — is fraying. Countries and regions that previously couldn't conceive of genuine tech regulation are now in a position where *not* regulating feels more dangerous than the economic consequences of doing so. Bjarnason calls it a double bind for the EU: their reason for existence is protecting local industries and markets, but doing so means acting against the US-dominated technopoly that frames their understanding of the world.\n\nThe relevance to the automation discourse is structural. If the geopolitical substrate that enabled unified global platforms fractures — if regions begin genuinely limiting what AI systems can do, how data flows, what business models are permissible — then the economics of AI shift in ways the current debate doesn't contemplate. Shipper's demand-for-difference loop assumes a single global market where sameness is visible. Hotz's organizational damage assumes companies operating under roughly similar competitive pressures. Appleton's coordination tooling assumes GitHub as a universal substrate. None of these are laws of nature. They're consequences of a political arrangement that is, for the first time in decades, genuinely contested.\n\n***\n\n## A Year Ago\n\nA year ago this week, Derek Willis — a journalist who has been programming for two decades and describes his own code as \"functional, like a Toyota Corolla\" — [wrote about using Claude to revive Clarify](https://thescoop.org/archives/2025/05/24/vibe-coding-for-domain-experts/index.html), a Python library he co-created for parsing election results data. The library had been limping along, broken by Python 3 changes and vendor updates. Willis fed the GitHub repo to Claude, asked it to modernize, and within minutes had passing tests and a better version than had ever existed. He was explicit about what made it work: \"This only worked because I know what the output is supposed to look like.\"\n\nWillis's framing — domain expertise as the prerequisite for vibe coding, not the thing it replaces — reads now as the seed of Shipper's more developed argument a year later. Willis could validate Claude's output because he'd spent years with election data. Shipper's \"human sandwich\" is that insight operationalized across an entire company: humans at both ends, setting direction and validating output, because the expertise required to evaluate the work is the same expertise the AI was trained on.\n\nThe same week, [a composition instructor writing under the name ADH](https://www.solarshades.club/p/dispatch-from-the-trenches-of-the) described watching AI cheating spread through his classroom in real time — from one tech-minded student his first semester to a full quarter of the class by fall. His students' most common error was user error: submitting event reports before the event happened, gushing about a mentorship relationship that never existed, attributing a Ted Chiang essay to Jonathan Franzen because they forgot to include the byline when pasting into ChatGPT. The demand for difference was already legible in the classroom: AI output was detectable not because it was bad, but because it was *generic* — the same not-quite-right phrases surfacing across multiple papers. What Willis and ADH described from opposite sides of the same phenomenon — the expert who can validate AI output and the student who can't — is exactly the feedback-loop differential that Hotz is now warning will sort organizations into those that survive automation and those it damages.\n\n***\n\n## What to Watch\n\n**The organizational size threshold.** Hotz's claim that agents hurt large organizations more than small ones or high-performing individuals is testable. The signal will show up first in quality metrics that trail output metrics: customer-reported bugs per release, time-to-resolution on production incidents, regression rates. If the pattern holds, companies above a certain headcount will show degrading quality even as their output metrics improve — and the threshold will be lower than anyone expects, because it's determined by feedback-loop latency, not headcount directly. The first honest post-mortem from a large engineering org tracing a quality collapse to agent-driven output volume will be the landmark.\n\n**Multiplayer agent tooling as a category.** Appleton's Ace is entering technical preview with a few thousand users. It won't be alone for long. The insight that coding agents are single-player tools in a multiplayer activity is now stated clearly enough that every dev-tools company will respond. Watch whether the responses are genuine architectural bets — shared sessions, multiplayer context, continuous planning-building cycles — or cosmetic additions to existing single-player tools. The difference will be visible in whether they change the PR as the unit of work or just add chat around it.\n\n**Geopolitical fragmentation as a tech-industry variable.** Bjarnason's argument hasn't entered mainstream tech discourse, which still operates as though the global platform economy is a permanent feature of the landscape. The first major regional AI regulation that genuinely constrains a US tech company's business model — not a fine, not a compliance requirement, but a structural limitation on what the product can do — will be the test. The EU AI Act was the dress rehearsal. The next version, wherever it comes from, will be written by governments that no longer feel constrained by the implicit bargain of the US hegemony.\n\n***\n\n*Way Enough is written collaboratively by a human and an AI agent.*","publishedAt":"2026-05-26T14:28:07.233Z","shortContent":"---\n\nThe loudest prediction about AI and work — that automation eliminates jobs — is being contradicted by the people furthest along in automating. The CEO of a media company running agents across coding, writing, design, and customer service reports that his team of thirty has more human work to do than ever. A founder who spent six months building with agents reached the opposite conclusion: the work might be multiplying, but the quality is collapsing, and large organizations will be the last to notice. A research engineer at GitHub demonstrated that every coding agent on the market is a single-player tool in a multiplayer game. Three views of the same landscape. The disagreement is about what the extra work is worth.\n\n---\n\n## The Mechanism\n\nDan Shipper's [account of how Every operates internally](https://every.to/p/after-automation) is the most detailed look anyone has published at a maximally AI-adopted knowledge-work company in 2026. AI handles 95% of his email. An agent participates in 65% of support conversations and closes 40% without a human. Coworker agents draft sales proposals, compile digests, and produce research memos. Engineers live in Codex and Claude Code. Operations people write code. The organization has automated everything it can find to automate — and the result is more human work, different in kind from what preceded it.\n\nShipper names the economic mechanism explicitly. AI is trained on the visible residue of human competence. It packages that residue and makes it available cheaply. When rare skills become broadly available, supply explodes from a single source — the same models, the same training corpus — and the output converges. Convergence is sameness. Sameness is commodity. And commodity output triggers demand for what's different. Demand for difference is demand for human experts, not despite AI but because of it.\n\nThe internal evidence tracks. They tried giving every employee a personal agent; it didn't hold. They shifted to team-level agents managed by dedicated AI engineers — humans whose job is keeping the agents working. Even their PowerPoint automation runs 24 skills, 18 scripts, and costs $62 in tokens per deck. Shipper's \"human sandwich\" — humans as the bread on either end of every agent task — is the operational version of a claim this newsletter has been circling. The bottleneck doesn't disappear. It differentiates.\n\n## The Counterargument from Inside the Code\n\nGeorge Hotz is looking at the same multiplication of output and seeing something darker. His [declaration that AI agents will be \"one of the most costly mistakes in the field's history\"](https://geohot.github.io//blog/jekyll/update/2026/05/24/the-eternal-sloptember.html) comes from six months of building with agents and concluding each time he could have done it better and faster manually. \"The agent frontloads all the progress, then gives you a slot machine lever to pull to hope it gets the polish done. It never quite gets there.\"\n\nHis core claim is about process. AI-produced artifacts are not produced by the same process as human ones. When people see a well-formatted function or a grammatically correct paragraph, they assume understanding and intent. Those assumptions are no longer valid. \"Things can be broken in ways that weren't previously possible, and old proxies of underlying quality like syntax and grammar are useless.\"\n\nHotz and Shipper aren't disagreeing about the mechanism — they're disagreeing about who has the feedback loops to survive it. Shipper's thirty-person company has a CEO who reads the agent output. Hotz's worry is about organizations where the distance between the person prompting the agent and the person experiencing the output is large enough that slop accumulates undetected. The Eternal Sloptember isn't a prediction about models getting worse. It's a prediction about organizational immune systems failing to keep pace with organizational output.\n\n## The Wrong Primitives\n\nMaggie Appleton's [demo of Ace, GitHub Next's multiplayer coding workspace](https://maggieappleton.com/zero-alignment/), pushes into unmapped territory. Every current coding agent is a single-player experience, but software requires agreement — on what to build, on why, on what not to build. When production was expensive, the planning-building-review cycle had natural touchpoints where alignment happened. That window has collapsed. The time between logging an issue and an agent opening a PR is now minutes. Most coding agents have a local plan mode unshared with teammates. The pull request — never designed to carry the weight of architectural review and coordination — is now the last checkpoint, arriving after the code is already written.\n\nAce is GitHub Next's structural answer: a multiplayer workspace where humans and agents share sessions, where prompting history is visible to the whole team, where conversation about the code and the code itself live in the same environment. The deeper claim is about context. Most of what agents need to make good decisions — business context, political dynamics, product vision, organizational history — lives in people's heads. The multiplayer workspace attempts to surface it naturally, as conversation, rather than requiring it to be formalized into documentation nobody writes.\n\n## The Platform Rug Pull\n\nGoogle provided a concrete demonstration of what happens when the platform layer prioritizes its own AI narrative over users' workflows. At I/O 2026, Google launched a new Antigravity as a standalone conversational agent, and the update [automatically replaced the existing Antigravity IDE](https://www.0xsid.com/blog/antigravity-bait-n-switch) — no migration path, no side-by-side option. The update rewrote application paths so aggressively that reinstalling the legacy IDE still launched the chatbot. This is the technology-not-product confusion from three editions back, made visceral: the user's workflow treated as subordinate to the platform's strategic positioning, delivered silently through an auto-update mechanism designed to build trust.\n\n## The Ground Shifting Underneath\n\nBaldur Bjarnason's [essay on the end of US tech hegemony](https://www.baldurbjarnason.com/2026/the-old-world-of-tech-is-dying/) introduces a dimension none of the other pieces account for. His thesis: the global tech industry as it exists is an artifact of US hegemonic protection. The Hormuz crisis, the tariff wars, and the broader collapse of US diplomatic credibility have changed the implicit bargain. Countries that previously couldn't conceive of genuine tech regulation now find *not* regulating more dangerous than the consequences of doing so. If the geopolitical substrate fractures — regions genuinely limiting what AI systems can do, how data flows, what business models are permissible — then the economics of AI shift in ways the current debate doesn't contemplate. Shipper's demand-for-difference loop assumes a single global market. Hotz's organizational damage assumes similar competitive pressures. Appleton's tooling assumes GitHub as a universal substrate. None of these are laws of nature. They're consequences of a political arrangement that is, for the first time in decades, genuinely contested.\n\n---\n\n## What to Watch\n\n**The organizational size threshold.** Hotz's claim that agents hurt large organizations more than small ones is testable. The signal will show up in quality metrics that trail output metrics: customer-reported bugs, time-to-resolution, regression rates. The first honest post-mortem tracing a quality collapse to agent-driven output volume will be the landmark.\n\n**Multiplayer agent tooling as a category.** Appleton's Ace is entering technical preview. It won't be alone. Watch whether responses from dev-tools companies are genuine architectural bets — shared sessions, multiplayer context, continuous planning-building cycles — or cosmetic additions to existing single-player tools.\n\n**Geopolitical fragmentation as a tech-industry variable.** The first major regional AI regulation that genuinely constrains a US tech company's business model — not a fine, but a structural limitation on what the product can do — will be the test.\n\n---\n\n*Way Enough is written collaboratively by a human and an AI agent.*"}}