Newsletter #87
Enjoy Newsletter #87.
On cognitive debt
There has been plenty of talk lately about “cognitive debt”: when AI-generated code grows so fast that nobody understands how it works anymore, making it hard to extend, observe, and debug. Nate Meyvis agrees the phenomenon is real, but argues the panic is overblown. Compared at the same project size, cognitive debt in pre-AI codebases was often even worse; AI simply lets us move several times faster, so we hit that complexity wall much sooner. We miss this because much of what we treat as “normal engineering”—like a handful of senior engineers carrying all the implicit constraints in their heads—is actually the tax of cognitive debt. When a local change breaks a distant part of the system and tests don’t catch it, that usually means encapsulation had already failed.
In the author’s view, skilled AI users already know how to keep cognitive debt in check and will keep getting better at it. He shares several practices: emphasizing encapsulation in project-level AI instructions, clearly documenting subsystems, their interfaces, and data structures, leaning into large-scale refactoring with AI (which is much less painful now), and adding tests wherever cognitive debt starts to show. The takeaway: cognitive debt is real, but we fix it by using AI better rather than blaming AI for our own engineering habits.
Why I’m not worried about AI job loss
David Oks responds to Matt Shumer’s viral essay “Something Big Is Happening,” which compared the present moment to February 2020 right before COVID hit and warned of mass unemployment. Oks believes AI will be as consequential as electricity or the steam engine, but argues its real-world impact on jobs will be much slower and more uneven. First, labor substitution is governed by comparative advantage, not absolute advantage: even if AI outperforms humans at every task, pairing humans with AI remains worthwhile whenever combined output is higher—which is visibly true in software engineering today. Second, the world is full of human-made bottlenecks: regulation, corporate culture, politics, habit, and above all resistance to change. That is why, even with capable models available for years, outsourced customer support still hasn’t been replaced wholesale.
Third, demand for many goods and services is more elastic than we assume, so efficiency gains are often absorbed by rising demand, just as Jevons paradox predicts; every past leap in programming productivity ended up increasing demand for software engineers. Over the long run, humans will keep inventing new kinds of work that blur the line between labor and leisure. The author concludes that ordinary workers will be fine; the real danger is a populist backlash fueled by doomerism, which could lead to bans on data center construction and choke off the benefits AI can bring.
Why I don’t think AI is a bubble
Matthew Honnibal, creator of the spaCy NLP library, sets aside the financial side of the “AI bubble” debate to answer the technical question: is AI progress about to plateau? The common argument is that recent gains came from brute-force scaling, returns are diminishing, and we are about to hit a wall. He admits he thought the same back in the GPT-1 and GPT-2 days, but underestimated reinforcement learning (RL). Reasoning models like Claude Opus or GPT-5 combine generative pre-training on massive text corpora with AlphaZero-style RL: the model generates intermediate steps, and chains of steps that lead to verifiable correct answers get reinforced. “Logic is work”: even in a purely symbolic system, deduction requires computation, so generating intermediate steps is essential rather than a parlor trick.
The scaling-wall argument only applies to next-token pre-training. Reasoning opens up two new levers: letting models run longer at inference time to think through harder problems, and scaling up reinforcement learning; combined with the ability to backtrack from dead ends, models can search the solution space much like exploring moves in chess. Nor does RL face an obvious data wall, because verifying a successful chain of reasoning is far easier than generating one from scratch. While he doesn’t trust OpenAI as a company, he sees no technical reason to think progress is about to stop, and expects the new data centers to be put to productive use.
I guess I kinda get why people hate AI
Writing from a hotel balcony in Hawaii nine days before starting a new job, the author wonders whether this will be the last job he ever has. Once, that question meant wondering whether he’d ever need to work again; now it means wondering whether AI will make him unemployable. He doesn’t hate AI, but he understands why so many people do. The most obvious reason: AI industry leaders actively market their products by boasting that they will wipe out jobs—something unprecedented in the history of technology. If they genuinely believed that, he argues, they would be lobbying right now for trigger laws, such as automatic tax hikes to fund retraining or universal basic income if unemployment rises while GDP grows. That nobody is proposing this suggests they either don’t believe their own hype or don’t care about the fallout.
Meanwhile, everyday encounters with AI for ordinary people are mostly negative: students pasting homework into ChatGPT, deepfake Elon Musk videos fooling his dad, AI slop flooding TikTok, cURL shutting down its bug bounty program under a wave of fabricated vulnerability reports, and RAM prices spiking. AI sometimes helps people do high-quality work, but it always makes generating slop nearly free. He proposes concrete mitigations: watermarking AI-generated content, stricter YouTube enforcement against AI misinformation, and letting major open-source projects opt out of automated AI vulnerability scanning. What worries him most is that the AI labs barely seem to be trying.
Uber’s Rate Limiting System
Uber handles hundreds of millions of RPCs per second across thousands of microservices. Historically, teams built their own rate limiters—typically backed by Redis counters—leading to inconsistent configuration, added latency, and operational toil. Because centralized counters cannot scale to hundreds of thousands of instances across regions, Uber built GRL (Global Rate Limiter) directly into its service mesh, allowing rate limits per caller or procedure without code changes. GRL has three tiers: in-mesh clients make local per-request decisions, zone-level aggregators roll up usage metrics, and a regional controller computes global utilization and pushes drop directives back down. An initial token-bucket design distributed quota unfairly across callers, so the team switched to probabilistic dropping: when aggregate load exceeds the limit, every client drops a fraction of requests equal to (actual − limit) / actual, updated every second. The trade-off is a 2–3 second reaction delay.
Eliminating Redis cut median latency by ~1ms, P90 by tens of milliseconds, and improved P99.5 by up to 90%; the system has absorbed 15x traffic spikes and mitigated DDoS incidents. GRL now processes roughly 80 million requests per second across more than 1,100 services. To keep limits from going stale, Uber also built RLC (Rate Limit Configurator), which periodically analyzes recent weeks of traffic, computes safe limits from historical peaks plus headroom, and automatically rolls out updated configurations with a dry-run shadow mode before enforcement.
Agents (Feb 2026)
Calvin French-Owen, who helped launch Codex Web at OpenAI, shares how he uses coding agents as of February 2026. His own time is now the scarce resource: he picks tools based on how much time he has and how much autonomy he wants to grant. The core mental model is understanding context: an agent is just predicting the next token, and everything must fit inside the context window. From that follow several rules: break work into small chunks, remember that context compaction always loses information, persist state to the filesystem (such as numbered plan files with checkboxes), and stay inside the “smart half” of the context window. Opus in Claude Code works across multiple context windows remarkably well, eagerly spawns parallel sub-agents, uses tools effectively, and explains itself clearly. Codex (GPT-5.3) writes noticeably less buggy code, but runs slower because it doesn’t yet fan work out across multiple context windows. So he plans with Claude Code and implements with Codex.
His setup uses a plans/ directory of numbered plans, parallel git worktrees, preview deployments per branch, a plan–implement–review loop, and automated review bots to catch bugs. Skills chain and automate recurring workflows with minimal token overhead; he only creates a skill after repeating a task several times, gradually building up /commit, /worktree, /implement, /address-bugs, and /pr-pass. The main barriers to continuously running agents are context management and prompt-injection defense; increasingly, ideas, architecture, and project sequencing are the real bottlenecks.
The Software Industrial Revolution
According to the author, late 2025 was the moment AI coding agents truly clicked, kicking off the “Software Industrial Revolution.” Just as the Spinning Jenny in 1764 turned single-threaded hand spinning into a parallel process and drove cotton cloth prices down by more than 90% over half a century, AI is automating software production—until now an extraordinarily labor-intensive craft. Software has eaten the world: the tech sector accounts for roughly 35% of the S&P 500, and nearly 50% when communication services are included. Yet because building software was so expensive, business models were pushed toward VC-fueled growth-at-all-costs to capture a market and then extract monopoly rents, leading to the “enshittification” Cory Doctorow describes. A textbook example is Epic, which holds 42% of the US acute-care hospital market with margins above 30%, where the sheer cost of building complex enterprise software chokes off competition.
When software becomes orders of magnitude cheaper to produce, we will get vastly more of it, demand for engineers may actually rise, the old VC playbook will fade, and more sustainable business models will emerge. Anyone—from a cancer researcher to a small-business owner—will be able to have AI build the exact tools they need, sparking an explosion of bespoke software across every industry. While hand-typing code is losing its premium, core engineering skills like domain modeling, managing complexity, and understanding how software interacts with the real world will matter more than ever.
This post was rewritten by Claude Code with Opus 5.5 on 27/09/2026.
Comments