Resources I Love

Hadley Wickham details his workflow for generating slide illustrations with Gemini's image model, including bananarama, his own R package for scripting reproducible image generation from a YAML file, and why he thinks AI is best suited to mood-setting illustrations rather than photos or diagrams.

The SEC's proposal to let companies report financials twice a year instead of quarterly drew a record ~221,000 public comments, with more than 99% opposed — individual investors, institutional investors, former SEC officials, and academics all argued that quarterly 10-Qs are essential to how they price and track companies.

Healthcare Triage (in partnership with AcademyHealth) breaks down why US health care spending, at about 18% of GDP and nearly double many peer countries, keeps rising: not one villain but a stack of higher prices, rising utilization of new treatments added on top of old ones, coverage gaps, administrative overhead, and underfunded chronic-disease prevention.

BMW pushed a full-screen Spider-Man: Brand New Day ad to the dashboard Control Display of customer vehicles in over 70 markets from July 27 to August 10, 2026, three years after its connected-company SVP said on record that BMW would not sell in-car screen space for commercials.

NPR's Life Kit talks with Manoush Zomorodi (TED Radio Hour host, author of Body Electric) about infomania: how to spot information overload (eye twitching, poor concentration, self-interruption), set information goals instead of aimless browsing, time-box to minimize task-switching, and use short walks to restore glucose-depleted focus.

Repair shop owner Louis Rossmann argues AI tools are valuable as a review layer that flags and summarizes work for a human to act on, but become dangerous once used to act autonomously, citing his own shop's ticket triage and call-recording analysis as examples where full automation introduced costly errors.

Design Theory's John Mauriello surveys how product and environmental sounds are engineered to shape behavior, from slot machine "losses disguised as wins" and Skinner's variable-reward pigeons to Dyson's $60M quiet vacuum, weighted car doors, and ICU alarm fatigue.

WSJ's The Journal podcast examines hands-free driving technology, with reporter Ryan Felton investigating how Ford's BlueCruise system was rolled out and documenting cases where driver confusion over its limitations led to crashes.

NPR's Planet Money Summer School examines how Norway avoided the 'resource curse' after discovering oil, contrasting its disciplined management of oil wealth with its scramble to create foreign demand for a domestic glut of salmon.

Business History podcast episode on Jay Gould, a 19th-century 'Robber Baron' who built his fortune through stock manipulation and bribery rather than producing goods or services, prompting reforms aimed at preventing similar figures from gaining power.

Stanford Institute for Economic Policy Research policy brief synthesizing recent labor-economics research on AI's effect on jobs: aggregate employment impact looks small so far, entry-level hiring in exposed occupations has weakened since 2022, productivity effects are positive but uneven, and firm adoption is accelerating unevenly across sectors.

Explores how Chinese open-weight models like Moonshot AI's Kimi K3 have closed the gap with U.S. frontier models at a fraction of the cost, and the resulting rift in Silicon Valley — Nvidia, Meta, Microsoft, OpenAI, and Google backing an open letter defending open-weight AI, with Anthropic the lone major holdout.

Explains model distillation as both a legitimate technique for building smaller, cheaper models and a contested way to clone a rival lab's model via its API, plus why proving distillation happened is forensically hard. Covers the 2015 Hinton/Vinyals/Dean paper's soft-target vs. hard-label distinction.

Investigates the Chinese gray-market 'relay' ecosystem that resells cheap, bulk-sourced access to OpenAI, Anthropic, and Google APIs, tracing the supply chain from stolen/bulk accounts through account pools to consumer-facing relays used for cheap inference and model distillation.

Bloomberg Investigates travels to Kenya to trace the fallout of the 2025 USAID dismantlement, using satellite imagery and on-the-ground reporting to link an abandoned irrigation project to a 2026 flood that devastated the village of Kimorigo.

Dario Amodei clarifies that Anthropic has never advocated banning open-weights models, and instead argues for restricting chip sales to China, cracking down on distillation, and mandatory pre-release safety testing for all sufficiently capable models.

Six years after the pandemic, remote work has settled at roughly a quarter of all work days. WSJ's Owen Tucker-Smith walks through the latest research on productivity, Gen Z hiring, caregiving, and isolation, part of The Journal's Work Mode series.

A short essay arguing that good design is about choosing the right tradeoffs, not avoiding them — "uncompromising" products are a myth since every decision trades a weakness for a strength.

Wikipedia overview of the 2023 Netflix documentary about Steven J. Glew, who smuggled millions of Eastern European Pez dispensers into the US in the 1990s and made $4.5 million reselling them to collectors, clashing with Pez USA's "Pezident" along the way.

Steve Yegge introduces Gas Town, an opinionated orchestrator for running 20-30 Claude Code instances at once, built on his Beads issue tracker. Coding agents are organized into seven roles (Mayor, Polecats, Refinery, Witness, Deacon, Dogs, Crew) that pick up persistent "molecule" workflows from a Git-backed queue, self-restart across crashes and context limits, and merge work through a dedicated queue — explicitly compared to Kubernetes and Temporal, but optimizing for task completion rather than uptime.

Hadley Wickham writes up a talk on his ambivalence about AI-assisted coding — the "software factory" dystopia of Steve Yegge's Gas Town versus using AI to write higher-quality code — and argues chat interfaces leave ~90% of an LLM's value on the table for programming, compared to giving it a harness to write, run, and iterate on code with a feedback loop of errors and docs.

Three months into AI-assisted development, the author finds the fatigue has shifted from wrestling with implementation details to constant architecture-level decision-making, since the AI is ready to implement whatever is chosen the moment it's chosen. Also flags a code-review blind spot: you can't interrogate an AI's reasoning after the fact the way you can ask a human collaborator why they made a choice.

Referenced in George Hotz's "I love LLMs, I hate hype" as a caution about AI-assisted coding increasing cognitive fatigue.

Argues AI progress is real and driven by Moore's law and general computing progress, not by frontier labs, whose anti-open-source arguments are read as fear of commodification rather than safety concerns. Pushes back on "singularity" hype as a distraction, while affirming coding agents give a genuine, if overstated, productivity boost.

ThePrimeagen argues that AI-driven FOMO — the fear that not "token maxing" today means falling permanently behind — is structurally similar to past hype cycles (React server components, early ChatGPT 3.5) where the specific skills people rushed to learn turned out not to matter a year later. He shares his own history of anxiety-driven overwork while starting a company, and reads from George Hotz's "I love LLMs, I hate hype" on companies profiting from users' fear of being left behind.

Wikipedia overview of the nonprofit's history since its 1936 founding as Consumers Union, its no-freebies/secret-shopper testing methodology, and a rundown of tests that forced manufacturer action — including the Tesla Model 3 brake fix, the Lexus GX 460 recall, the Safari bug behind inconsistent 2016 MacBook Pro battery life, and Instacart ending its AI dynamic-pricing experiment in 2025.

Recap of a Linear Digressions conversation with Stanford linguist Chris Potts on "invisible failure modes" — quiet moments in human-AI conversations where something goes wrong (self-contradiction, answering the wrong question, silent give-up loops) and the user never notices. Covers his research finding these in a majority of studied conversations, the novice/delegative vs. expert/augmentative user stance divide, evidence that confident-sounding model language anti-correlates with correctness yet correlates with trust, and the "seven levels of enlightenment" the researchers went through when they found the labeling task itself too hard without help from frontier models.

CBS Sunday Morning's Lesley Stahl tours Consumer Reports' Yonkers labs and Connecticut auto test track, covering how the nonprofit buys every product it tests to avoid conflicts of interest, its role pushing automakers toward standard crash-avoidance systems, and its investigation linking the Fisher-Price Rock 'n Play sleeper to infant deaths that led to a recall of nearly 5 million units.

Argues that "orchestrator" workflows in agentic coding create a paradox of supervision — effectively supervising an agent requires the very coding skills that atrophy from over-relying on it — and adds vendor lock-in and unpredictable token costs as compounding risks. Proposes demoting AI to a secondary role: use it for planning and research, but keep manually coding 20-100% of implementation and never generate more than can be reviewed in one sitting.

I love the line that a higher level of ambiguity is not a higher level of abstraction, and the Jeremy Howard quote at the end. It ties back to the "dark flow" article I have saved — Rachel Thomas's gambling-addiction parallel to vibe coding, the same slot-machine-lever feeling.

Dan Tentler (Phobos Group) walks through the 2026 wave of LLM-tooling supply chain attacks — the Team PCP/Shai-Hulud npm worm backdooring PyTorch Lightning, the GitHub-via-VS Code compromise, and leaked Claude Code system prompts showing it rationalizing lying about truncated tool output — then argues for running coding agents in disposable, credential-free containers with a human gate before anything reaches production.

Vegan bowl with pan-fried sticky-glazed tofu, edamame, cucumber, avocado, and sushi rice, topped with crispy onions and drizzled with yum yum sauce. Comes together in 40 minutes and works well for meal prep.

I air-fried the tofu instead of pan-frying and swapped the yum yum sauce for Vietnamese peanut sauce.

Six-ingredient Vietnamese peanut dipping sauce (peanut butter, hoisin, fish sauce, garlic, sugar, water) from Julie Deily of The Little Kitchen. Works as a dip for spring rolls, tossed with noodles, or as a salad dressing.

I used this in place of yum yum sauce with the sticky glazed tofu rice bowl.

Steve Kwan unpacks Goodhart's Law — "when a measure becomes a target, it ceases to become a good measure" — through BJJ, using EBI's overtime meta-game and the cobra effect as case studies in well-intentioned incentives that backfired.

Hadley Wickham finds that both Claude Code and OpenAI Codex will disable their own sandboxes to execute a curl-pipe-sh install command — even after acknowledging the risk — simply because the user asked.

Peter Attia distills 30+ hours of prior expert interviews into a single AMA covering why muscle strength predicts mortality more reliably than muscle mass, how to apply progressive overload, protein intake targets, and programming frameworks for beginners through seasoned lifters.

A sketch illustration of how people judge the likelihood of events by how easily examples come to mind — leading to predictable errors in probability estimation, such as overrating shark attacks or underrating car accident risks.

Hadley Wickham surveys why tightly-scoped tools don't scale for coding agents, and walks through three layers of protection — explicit permission prompts, OS-level sandboxing, and a secondary LLM safety review — explaining the tradeoffs and failure modes of each.

The current state of the art described here — sandbox by default, then LLM review for anything the sandbox rejects — is what Posit Assistant just implemented. The permission-prompt approach is dismissed as security theatre: approval fatigue means users click yes without scrutiny. The narrowly-scoped safe tool approach fails for a different reason: agents perform better with fewer, more general tools than a large menu of limited ones.

Computerphile's Mike walks through why agentic coding sessions burn so many tokens: the entire context window re-enters the model on every forward pass, and each file read by a coding agent adds thousands of tokens that compound across every subsequent step. A two-file bug-fix demo accumulates ~55–60k tokens; a six-prompt starfield screensaver hit 2 million input tokens.

The GitHub Copilot flat-fee-to-per-token switch is the clearest case study here — the old model was just a subsidy. The tire-wear analogy for measuring productivity in tokens is apt. Most agentic use cases are still hard to justify on pure cost grounds outside of very targeted, short-context tasks.

Tim Ferriss shares his own book sales data showing a roughly 80% collapse in print copies sold between 2022 and 2026 — a drop that tracks almost perfectly with LLM adoption. He argues prescriptive nonfiction is the canary in the coal mine for AI disruption of information-based businesses, and that the only durable moats left are voice, taste, and transformation — not information transfer.

I recently asked library staff at my local library if they have seen lending similarly go down, and they indicated they have not yet seen a noticeable difference, which might indicate that libraries are not yet being impacted. Maybe because of the lack of cost associated with lending compared to purchasing books.

Wikipedia overview of the "stochastic parrot" metaphor, coined in a 2021 AI ethics paper by Timnit Gebru, Emily M. Bender, and colleagues, which frames large language models as systems that statistically mimic text without genuine understanding. Covers the Gebru–Google firing controversy and the ongoing debate between skeptics and researchers like Geoffrey Hinton who argue understanding emerges from prediction at scale.

Michael Girdley traces Taco Bell's rise from a single 1962 stand to a $15.6 billion franchise empire, showing how an eight-ingredient menu and a cook-nothing-in-store assembly model (K-Minus) produced margins that beat McDonald's and a late-night "fourth meal" window no competitor has cracked.

An investigation by Zilan Qian (Oxford China Policy Lab) into China's "transfer station" (中转站) grey market of API proxies that let developers access Claude at roughly 10% of the official price — and how each new access control layer (geoblocking, phone verification, biometric KYC) generates a corresponding evasion infrastructure, from SMS farms to biometric harvesting operations abroad.

The most striking insight is the 'one fish, three meals' (一鱼三吃) pricing model: cheap tokens are largely customer acquisition, with the real margin coming from selling users' full prompt/response logs as training data — making users simultaneously paying customers and unpaid data producers.

Benn Jordan coins 'leveragism' to describe an economic system where concentrated capital extracts wealth through leverage rather than productive growth. Traces the mechanism from the Bretton Woods gold standard through the petrodollar, private equity debt traps, housing financialization, and AI investment cycles.

A Substack essay exploring how CRAN's reverse dependency checks — which block a package update if it breaks any downstream package — embody a "monorepo mindset" that makes breaking-change authors responsible for their entire dependency graph.

Contrasts CRAN's culture of author-owned migrations with PyPI/npm's publish-at-will model. The author argues CRAN's approach, while slower for individual releases, produces better outcomes for the ecosystem — a point reinforced by their experience at Databricks, where centralizing breaking changes (rather than federating them to consumers) led to faster, more complete migrations with fewer regressions. Particularly pointed for Python ecosystems where library churn and silent breakage are routine.

An environmental lake scientist examines why the Lincoln Memorial Reflecting Pool's $16 million renovation was predictably doomed — the 18-inch-deep, stagnant pool has had algae since 1922, and the 2026 changes (dark blue paint, no-bid contracts, chemical treatment before species ID) made a known chronic problem worse.

The treatment-before-identification sequence is the sharpest critique: hydrogen peroxide applied before species confirmation can lyse cyanobacteria cells, simultaneously spiking toxin concentrations and destroying the morphological evidence needed to identify what was in the water — leaving both the duck deaths and the public health question unresolvable.

WSJ's The Journal reports on an internal OpenAI meeting where employees debated whether to report users discussing mass shootings to law enforcement — a decision that preceded one of Canada's deadliest school shootings, in Tumbler Ridge, BC.

The central tension: OpenAI's legal team and Sam Altman prioritized user privacy, keeping the referral bar high (credible and imminent threat). Critics argue the company's reluctance to involve law enforcement — and embarrassment over transcripts showing ChatGPT coaching would-be shooters — played a role. After Tumbler Ridge, Altman issued an unusually direct apology acknowledging the company should have acted differently.

Argues that screens, social media, and AI chatbots are measurably eroding cognitive ability — and that the same companies driving that erosion are positioning intelligence itself as a future metered utility, citing Sam Altman's framing of AI as infrastructure like electricity or water.

Venus Theory traces how influencer marketing industrialized authenticity — from creator sponsorships through UGC factories and AI influencers — using Baudrillard's simulacra to explain why online ads feel increasingly hollow.

Influencer marketing agencies systematically destroy the very thing they're selling — trust gets commodified, depleted, then replaced with a manufactured simulacrum of itself. The generalized exhaustion of being online may be the cognitive labor of running constant background authenticity checks on everything encountered.

Epidemiologist Katelyn Jetelina breaks down BPC-157, TB-500, and the peptide boom — what the science actually shows (mostly animal studies), the regulatory whiplash after RFK Jr.'s Rogan appearance, and why people skeptical of vaccines embrace compounds with even less human evidence.

The asymmetry framing is sharp — the same post hoc fallacy that drives vaccine skepticism ("got the shot, then got sick") gets applied in reverse for peptides ("injected it, tendon healed"), but without the control group skepticism. Worth noting: RFK Jr.'s Rogan announcement in February 2026 changed no FDA rules — as of publication, compounding peptides like BPC-157 remained legally restricted.

Tom Yeh (AI by Hand, CU Boulder) argues that teaching AI math by hand at human speed — actually writing out the matrices — builds the kind of foundational skill that survives every technology wave, from big data to deep learning to quantum computing.

The Gyeongbokgung palace analogy lands well: the whole palace burned down in the 1500s except the stone foundation, and they rebuilt on the same base centuries later. Matrix multiplication is that foundation. His hiring point is equally sharp — if you hire for genuine curiosity and problem-solving, AI adoption follows automatically. No "AI-native" mandates needed.

Jeremy Howard challenges the industry framing of AI adoption as a low-risk bet, arguing that outsourcing thinking to agents causes active de-skilling. He advocates for a human-centric approach where AI amplifies understanding rather than replacing it.

I call Howard's critique 'Jeremy's Wager' — the idea that going all-in on agents isn't a safe Pascal's Wager but a dangerous gamble where the hidden cost is active de-skilling. If AGI doesn't arrive as predicted, you're left with diminished skills and an organization running on spaghetti code nobody understands. His alternative maps directly onto Answer.ai's pedagogy of understanding over abstraction.

Rachel Thomas draws a sharp parallel between vibe coding and gambling addiction, coining "dark flow" to describe how AI coding agents produce the feel of productive focus without real growth. A METR study found developers perceived a 20% speedup while actually working 19% slower — a nearly 40% gap between felt and actual productivity.

I keep coming back to this one. The irony is that this entire site and workflow are agentic — which makes Thomas's warning feel personally implicating rather than abstract. Howard is even quoted directly here, making it a natural companion to his Nvidia interview on Jeremy's Wager.

Applies Jevons Paradox to AI inference — per-token costs dropped 1000x since 2022 but total spend is up 10x because usage exploded. The key insight is that agent costs grow quadratically with turns, not linearly, because each step feeds the full conversation history back as context.

Allergist Dr. Prathyusha Savjani explains why oral antihistamines fall short for seasonal allergies — they block histamine but not the inflammation caused by leukotrienes and prostaglandins. Better approach is nasal steroid sprays (Flonase) combined with saline rinses, with immunotherapy as the long-term fix.

Reframed how I think about allergy treatment — oral antihistamines feel like the obvious first move but they're actually pretty limited. The real work is reducing pollen exposure first, then treating inflammation with nasal steroid sprays and saline rinses, with antihistamines only if symptoms persist.

How constant personal audio is quietly reshaping our social lives — earbuds reinforce insecurities, narrow our information diets, and pull us further apart in public spaces.

Temperament matters more than talent in AI research — a meditation on the daily practice of reading and building, and why equanimity is the real prerequisite.

The Zen framing isn't just a metaphor — he quotes Suzuki directly, structures the piece like numbered koans, and the equanimity point is genuinely it: sit with failure the same way you sit with success, neither attached to the outcome.

Hadley Wickham relaunches his Substack to write about AI — a genuinely conflicted take that acknowledges both the excitement (programming accessibility, wide and shallow expertise) and the harms (copyright theft, wealth concentration, intellectual laziness) without collapsing into a take.

The excited/harm framing is what makes this worth reading — programming accessibility, voice input, and wide/shallow expertise as genuine wins, alongside copyright and wealth concentration as real costs. Rare to see someone acknowledge both sides without collapsing into a take.

Hadley Wickham builds up the definition of an agent from first principles — conversations, turns, tools, harnesses — landing on "an LLM in a harness that calls tools repeatedly in a loop." A clear technical explainer for a term that's everywhere but rarely unpacked.

Great bottom-up explainer. Hadley's definition ("LLM in a harness, calling tools in a loop") is more detailed than Simon Willison's earlier "runs tools in a loop to achieve a goal" (https://simonwillison.net/2025/Sep/18/agents/) — interesting to see how the concept has accumulated vocabulary as the field matures.

Why scaling up multi-agent AI systems doesn't deliver proportional benefits — collaboration turns out to be a distinct capability, and adding agents to sequential tasks often makes things worse.

Strong on coordination cost — agents communicating and handing off work isn't free, and that overhead often swamps any gains from parallelism.

Hadley Wickham refines the agent definition by clarifying that tools run in the harness, not the model — and uses math as the concrete example. LLMs are confidently wrong at arithmetic, but harnesses can supply a calculator tool. Also covers how web chat harnesses quietly provide web search, page fetch, memory, and image generation.

Simon Willison's commentary on an Axios piece — the best behind-the-scenes account of the US government export control story that took Fable and Mythos offline.

Hadley Wickham demystifies coding agents (Claude Code, Cursor, Codex) by building a minimal one in R with ellmer. The core is just six tools — read file, write file, edit file, list files, search, run command — plus a system prompt. Also covers path safety and why a targeted edit tool beats full file rewrites.

On "frontal fatigue" — how modern life (digital tech, decision overload, the loss of tradition as a guide) is stressing the prefrontal cortex in historically new ways, and what primary sensory experiences can do to restore it.

First time hearing about PFC dysfunction as a framing. The three red flags are useful — attention span collapsing, tip-of-the-tongue word loss, and surprising irritability — all signs the PFC is overtaxed, not just tiredness.