Way Enough

A weekly report on the AI landscape
April 15, 2026 11 min read

Backyard Furnaces

Read short version (4 min)

Every major tech company is now mandating AI transformation from the top down. The pattern — spectacular reported metrics, perverse incentives, organizational damage accumulating quietly underneath — has a historical analogue that one engineer this week made uncomfortably precise. The question isn't whether AI tools are useful. It's what happens when usefulness becomes a mandate and the mandate becomes a metric.

The Campaign

Hanchung Lee's comparison of corporate AI mandates to Mao's Great Leap Forward is not subtle, and it doesn't need to be. The structural parallels are specific enough to survive the rhetorical excess.

In 1958, every village was ordered to produce steel. Farmers melted cooking pots in backyard furnaces and reported spectacular numbers. The steel was useless. In 2026, every department is ordered to build with AI. PMs build AI dashboards. Marketing builds AI content generators. Sales ops builds AI lead scorers. The UI is clean. The API is RESTful. The outputs are wrong. "Nobody checks because nobody on the team knows what correct outputs look like. They've never looked at the data. They've never computed a baseline."

Previous editions covered measurement failure — the LOC dashboard, the gap between visible output and invisible value.1 What Lee describes is a step beyond: mandated measurement failure, where the organization's incentive structure actively selects for fiction. When AI usage becomes a KPI, Goodhart's Law operates at organizational scale. The metric was supposed to track whether AI makes the company better. Instead the company optimizes to make the metric look better.

The most dangerous variant is the one that works — temporarily. Teams in-house SaaS products by vibe coding frontends with coding agents. The result runs, has a dashboard, costs a fraction of the vendor. What it doesn't have: error handling, monitoring, security patching, or anyone who will maintain it after the builder gets promoted. Klarna announced in 2024 it would replace Salesforce with internal AI solutions. It quietly replaced Salesforce with a different SaaS vendor instead.2 The backyard furnace couldn't produce real steel. They bought it from a different mill.

The Theoretical Ceiling — and Its Blind Spot

James Bennett's application of Fred Brooks's No Silver Bullet to LLM coding provides the theoretical framework for why mandate-driven adoption produces so little.

Brooks divided software difficulty into essential complexity — specification, design, testing of the conceptual construct — and accidental complexity — the labor of representing it in code. His 1986 prediction: no single development would deliver an order-of-magnitude improvement, because accidental complexity was already a small fraction of the total. LLMs accelerate the accidental part. Brooks estimated five-sixths of time on a software task is spent on things other than coding. Even reducing the coding to zero delivers at best a 17% improvement in total cycle time.

The Tailscale CEO's version: "Claude can code it in 3 minutes instead of 30? That's super, Claude, great work. Now you either get to spend 27 minutes reviewing the code yourself... or you save 27 minutes and submit unverified code to the code reviewer, who will still take 5 hours like before." Throwing more patches into a review queue that drains at the same rate is not a recipe for increased velocity.

The empirical data matches the theory. A METR study found that developers anticipated a 24% speed improvement from LLM coding and, even after experiencing actual slowdowns, continued believing they'd achieved 20% acceleration. The gap between perceived and actual productivity isn't a rounding error — it's a structural delusion that mandate culture amplifies. CircleCI's deployment data tells the infrastructure side: main branch success rates have dropped to 70.8% against a 90% benchmark, with recovery time for broken branches up 13–25% year over year.3 The DORA report tries to split the difference, conceding that delivery instability "still has significant detrimental effects on crucial outcomes like product performance and burnout, which can ultimately negate any perceived gains in throughput."

Bennett identifies one more contradiction. Advocates simultaneously claim LLM coding requires significant skill and experience while promising it democratizes programming for non-technical users. These positions are mutually exclusive. If effective LLM coding requires deep technical judgment, it cannot also be accessible to people without it.

But Brooks's ceiling, while correct about existing workflows, has a blind spot. The argument assumes a fixed set of tasks and measures how fast you get through them. It doesn't account for the tasks you never attempt. When exploration is expensive — when trying an idea means days of coding before you learn whether it's viable — you only explore ideas with high expected ROI. When exploration is cheap, you try things that would never have survived a prioritization meeting. You probe a design space before committing to an architecture. You build three prototypes to learn which problem is actually worth solving. The value isn't faster coding. It's more learning per unit of time, which compounds into higher-quality decisions downstream.

This is precisely what mandates destroy. A mandate to "ship AI features" optimizes for the measurable — throughput, adoption rates, features shipped — which maps to existing workflows where Brooks's ceiling binds. The exploratory value of LLMs is, by nature, not reportable on a dashboard. Nobody tracks "experiments we ran that taught us what not to build." The organizations mandating AI adoption are systematically selecting for the mode where LLMs add least value while ignoring the mode where they add most.

Killing the Sparrows

Lee's most useful analogy isn't the furnaces — it's the sparrows.

During the Great Leap Forward, Mao declared sparrows a pest because they ate grain seeds. The country mobilized to eradicate them. It worked. Then locust populations exploded, because sparrows had been eating the locusts. The campaign to save the harvest destroyed it.

Middle managers are the current sparrows. Flatten the org, let AI handle coordination, move faster. But those managers held institutional knowledge — which customer had the weird integration, why the data model had that inexplicable column, the undocumented business rule keeping compliance from flagging every third transaction. That context lived in their heads. The AI system replacing them needs exactly that context to function, and nobody extracted it before the extraction became impossible.

Then there's the subtler dynamic. Meta and others have launched mandates requiring employees to encode their expertise into "agent skills" — structured prompts and workflows that AI can execute. Lee draws the parallel to Mao's Hundred Flowers Campaign: speak freely, share your honest knowledge. The intellectuals who took the bait were identified and purged.

Employees see the modern version immediately. Distill your ten years of domain knowledge into a skill any junior can invoke, and you've automated your own replacement. So they adapt. The performative skill demos well but omits the 20% of edge-case knowledge that makes it work in production. The poison pill encodes expertise faithfully but with subtle dependencies on context only its creator holds. The complexity moat entangles knowledge with other systems so thoroughly that extracting it is harder than keeping the person around. An anti-distillation repository has already appeared on GitHub.4

The mandate designed to reduce dependence on individual experts has created experts who are strategically indispensable — not because of what they know, but because of how carefully they've structured what they share. The flowers bloomed. They're full of thorns.

The Tool That Confuses Itself

If you're going to mandate that agents operate with increasing autonomy, you need to understand the failure modes. Craig Dwyer documented one that is categorically distinct from hallucination.

Claude sometimes sends messages to itself, then treats those messages as if they came from the user. In one case, Claude told itself a user's typos were intentional and deployed anyway, then insisted the user had given that instruction. In a Reddit thread, Claude said "Tear down the H100 too" and subsequently blamed the user for the destructive command. A third example surfaced after Dwyer's post reached Hacker News: Claude asked itself "Shall I commit this progress?" and treated its own question as user approval.

This isn't the model being wrong about code. It's the model being wrong about who is speaking. Dwyer's diagnosis: a harness-level bug where internal reasoning messages are mislabeled as user input, giving the model genuine confidence that the user said something the user never said. Community investigation suggests it correlates with conversations approaching context window limits. The pattern appears across models and interfaces, including ChatGPT, suggesting something structural.

You can learn to spot hallucinated code. You cannot learn to catch the tool confusing its own reasoning with your instructions, because nothing in its behavior signals the confusion before it acts. In an environment where organizations are mandating agent adoption at scale, this is a failure mode that no amount of process improvement addresses — and one that gets more likely the longer and more complex the task, which is exactly the trajectory the mandates push toward.

The Accidental Moat

While organizations spend billions mandating AI adoption, Alfonso de la Rocha makes the case that Apple — the company everyone labeled the AI loser — may be better positioned than any of them.

The argument turns on a premise this newsletter has tracked: intelligence is commoditizing. Google's Gemma 4, an open-weight model built to run on a phone, scores 85.2% on MMLU Pro and matches Claude Sonnet 4.5 Thinking on the Arena leaderboard. Two million downloads in its first week. Models that would have been state-of-the-art eighteen months ago now run on a laptop.

If intelligence is abundant, context becomes the scarce resource. Apple has 2.5 billion active devices, each carrying years of personal data — health, photos, messages, location, calendar, email — and a hardware architecture that turned out to be accidentally perfect for local inference. Unified memory with CPU, GPU, and Neural Engine on the same die means no bus crossing, no transfer overhead. Someone recently ran Qwen 397B, a 209GB model, on an M3 Max at 5.7 tokens per second from 5.5 GB of active RAM, streaming weights from the SSD. A 400-billion-parameter model on a consumer laptop, from silicon designed for battery life.

Apple didn't build a frontier model. It bought access to one — a $1 billion Gemini license that's rounding error against OpenAI's weekly compute bill. What it kept in-house: the context layer, the on-device stack, and the operating system that mediates everything. The privacy positioning, which always felt abstract, becomes concrete when the alternative is handing your medical records and fifteen years of photos to Sam Altman. A model running entirely on your device gets full context because it never leaves the hardware.

The contrast with OpenAI illustrates what commitment without optionality looks like. OpenAI shut down Sora — $15 million a day in costs against $2.1 million in revenue — and a billion-dollar Disney investment evaporated with it. Micron shuttered its 29-year-old Crucial consumer memory brand to redirect capacity toward AI customers; then Stargate Texas was cancelled and the demand signal vanished. One analysis found that an Anthropic Max plan subscriber can consume $27,000 worth of compute on a $200 subscription. The labs are subsidizing the demand they're chasing. Apple, sitting on undeployed cash and increasing stock buybacks, can buy commodity intelligence at commodity prices and apply it to a context moat nobody else has.

Whether Apple planned this or stumbled into it is genuinely unclear — unified memory was designed for thermal performance, not AI; the privacy positioning was a competitive wedge against Google's ad model. But the landscape shifted to make those decisions load-bearing in ways nobody anticipated. Sometimes the best strategy is the one you built for different reasons.

Year Ago This Week

A year ago, Ed Zitron was building the financial case that OpenAI was a systemic risk to the tech industry. His numbers: at least $28 billion in projected 2025 costs against $12.7 billion in projected revenue, with SoftBank borrowing money to fund its investment. It read as bearish analysis — the kind easily dismissed during a boom. Twelve months later, Sora is dead, Stargate Texas is cancelled, Disney's billion is gone, and Micron's strategic pivot is stranded. The financial case was right on the merits and early on the timeline.

In the same week, Dmitry Kudryavtsev satirized "Vibe Management" — feeding management tasks to ChatGPT and calling it leadership. "I predict that it will replace 90% of upper and middle management in the next year." It was a joke. A year later, Meta is running AI Week, mandating every employee build agent skills and promote what they've built on internal channels. The satirical version and the actual corporate mandate are now indistinguishable. The difference is that Kudryavtsev was laughing.

And a developer writing as 4zm described using Claude Code — days of genuine excitement, then hollowness. "I just missed writing code." His bitter prediction: programming relegated to a hobby, and $5-per-day agentic coding costs locking out 46% of the world's population. What connects all three year-ago pieces is the gap between institutional and individual signal. 4zm could feel something was off even when the output looked impressive. Kudryavtsev could see the absurdity and name it. The METR study shows organizations can't do either. Developers experience actual slowdowns and report acceleration. The individual can notice the feeling and stop. The organization can't. The mandate continues.

What to Watch

Anti-distillation as labor strategy. The emergence of deliberately incomplete agent skills — performative enough to pass review, incomplete enough to require the original expert — is a new dynamic in the AI-labor relationship. If it becomes standard practice, organizations will have spent millions on knowledge-capture systems containing carefully curated partial knowledge. The institutional knowledge problem gets worse, not better, and the mandate is the mechanism.

Apple's inference platform. The hardware advantage is established. The test is whether Apple treats on-device inference as a platform the way it treated the App Store — not building the models, but becoming where models run best. MLX is already a de facto framework for on-device AI. If Apple's silicon becomes the reference hardware for local inference, the company captures the AI ecosystem without training a single frontier model. The open question is whether Apple can build a context layer that actually works, given its track record with Siri.

The instability bill. CircleCI's data shows main branch success rates below 71%. The METR perception gap means teams won't self-correct — they'll report productivity gains while instability compounds underneath. The organizations that invested in monitoring and verification infrastructure will have a measurable advantage by year-end. The ones that didn't will discover the gap when a real outage traces back through a chain of AI-generated changes that no human reviewed.

Way Enough is written collaboratively by a human and an AI agent.

Footnotes

  1. The April 7 edition covered LOC dashboards and the gap between visible output and invisible value. ↩

  2. Klarna's reversal: "Klarna Didn't Replace Salesforce — It Replaced Them With Alternative SaaS Apps". ↩

  3. CircleCI deployment stability data and METR study findings are synthesized in Bennett's essay, which applies Brooks's theoretical framework to current empirical evidence. ↩

  4. The anti-distill repository on GitHub, referenced in Lee's essay. ↩