AI as a Learning Machine: Simon Eskildsen’s Practical Operating System

The most useful AI workflows do not replace learning. They make curiosity cheaper, association faster, and expertise easier to borrow long enough to act.

Source: Every / AI & I, Dan Shipper with Simon Eskildsen, video ID D_4AYLr_hFI

Who Is Simon Eskildsen?

Simon Eskildsen is the co-founder of turbopuffer, a startup building a search engine that starts with vector search and tries to make it cheaper and easier to run at scale. Before turbopuffer, he became a cult figure among productivity nerds through Every’s “Superorganizers” series, where his 2020 piece on making yourself into a learning machine went viral.

His credibility comes from an unusual combination: startup intensity, deep technical taste, and a decade-plus of disciplined personal knowledge management. He is not theorizing about AI from the sidelines. He is trying to keep a company, a family, a body, a home, and a brain operating under pressure.

The New Learning Machine Is a Loop, Not a Library

Eskildsen’s core shift is subtle but important. The old learning-machine model was mostly about capture: read books, take notes, process highlights, create flashcards, and review them until knowledge became retrievable. That system still matters. He still reviews between 50 and 200 Anki cards a day, jokes about throwing a party when he hits 10,000 cards, and stores everything from PostgreSQL JSON-versus-JSONB distinctions to colleagues’ children’s names.

But AI changes the shape of the loop. Google was useful when you already knew roughly what you were looking for. People were useful when you needed associations: “I’m using this data structure; what else might I do?” Language models now occupy some of that conversational middle ground. They are not Simon’s oracle for multi-step reasoning. He describes them more practically as “an average of the internet” that can be steered toward a region of latent space and asked to riff.

“When you have a rough idea of the island that you want to land on, it can paint the picture for you really well.”

That framing explains nearly every tool choice. AI is not a replacement for memory, taste, or expertise. It is a low-friction association engine that makes more domains temporarily accessible.

Flashcards Still Matter Because They Create Real-Time Bandwidth

The most counterintuitive lesson is that AI has not made Eskildsen’s flashcards obsolete. If anything, it clarifies what they were for. Flashcards are not a ceremonial archive. They are a sink for knowledge he wants to resurface. When a fact or concept deserves to be available without lookup friction, it goes into Anki.

His current card format is aggressively simple: front, back, optional reverse, extra, and source. He used to maintain many card types; now he uses one. The reverse field matters because knowledge often needs to be retrieved from both directions. Dan Shipper connects this to debates about whether language models understand logical entailment: if a model learns “Dan thinks there are six glasses in a bottle,” will it know whose answer was “six glasses”? Simon’s practice offers a human answer: people also train reversibility.

The source field carries more than citation value. It adds memory, nostalgia, and emotional texture. A card might remind him of “Naj at Carbon in 2014” or a waiter with a deep radio voice from a restaurant that no longer exists. Over twelve years, the flashcard deck becomes partly a semantic memory system and partly a personal time capsule.

This is the point founders often miss about knowledge systems. The goal is not to store everything. The goal is to increase the bandwidth of association at the moment of work. The best collaborator is still a high-bandwidth human like his co-founder sitting nearby. A language model expands the range of domains he can converse with. Flashcards keep a personal substrate of facts and concepts warm enough to connect.

Raycast Makes AI Useful Because It Removes the Ceremony

About 80 to 90 percent of Eskildsen’s LLM use runs through Raycast, which he describes as Spotlight on steroids. The important feature is not model novelty. It is proximity. Command-space, type a question, press tab, get an answer. When his Blue Yeti microphone sounded bad before the conversation, he asked whether to speak into the side with the logo. That kind of question is too small for a formal research session and too specific for old search habits. Raycast makes it worth asking.

He also uses Raycast AI commands as reusable prompts. His recipe command encodes dietary restrictions, preferred formatting, optional substitutions, approximate calories, and his desire for a compact ingredient-first answer. He does not want a food blogger’s ancestral clay-tablet story; he wants a usable cooking scaffold. The model supplies the “Flavor Bible” effect: what goes with butternut squash, what combinations are plausible, what optional ingredient might make dinner more interesting.

His “Define” command is even more revealing. When he encounters a word, place, person, or concept, the command returns a definition, six educational example sentences, related words, and often an image prompt. “Affable” becomes Gandhi, Benjamin Franklin, and approachable user interfaces. “Lambent” becomes soft flickering light, auroras, and Newton’s candle. “Eigengrau” becomes brain-gray, visual perception, and the mind’s role in constructing sight.

The prompt is not merely defining terms. It is deliberately generating hooks. It turns a word into a small cluster of memorable associations, making the eventual flashcard easier to encode.

AI Is a Universal Translator Between Professional Dialects

One of the strongest business use cases is translation between dialects of English that professionals rarely acknowledge as dialects. Tech founder to lawyer. Operator to accountant. Homeowner to contractor. Product person to compliance reviewer. These are not just vocabulary differences; they are different expectations about precision, risk, framing, and what counts as a useful draft.

Eskildsen uses AI constantly to create first drafts for lawyers and vendors because professionals often respond better to something concrete. If turbopuffer needs to describe the exact algorithm for measuring uptime in a customer contract, it is easier to send a plausible paragraph and let counsel edit than to ask counsel to infer the technical system from scratch. The same applies to accounting terms, customer redlines, and crisp copy revisions.

“I have no idea how to access the legal latent space unless someone just puts me in it.”

This is one of the cleanest practical definitions of useful AI: a tool that drops you close enough to the right dialect that your own judgment can take over. The output is rarely the final answer. It is a location in language-space from which editing becomes possible.

The Best Consumer Use Cases Are Tiny, Specific, and Embarrassingly Practical

Eskildsen’s examples are not grand AGI theater. They are the kinds of problems that used to leak time, money, and attention.

At a cabin in rural Quebec, he wanted to convert an old freezer into a drinks fridge. A model suggested the canonical home-brewing hack: buy a cheap temperature controller, plug the freezer into it, put the probe inside, set the target to 5°C, and let the controller switch the freezer on and off. A $20 Amazon device replaced a possible $1,000 fridge purchase.

For legal and government-adjacent life in Quebec, his wife uses AI to translate into Quebecois French. For physical therapy, he tried ChatGPT-guided wrist curls for tennis or golfer’s elbow and saw the issue disappear. For chronic neck and shoulder tightness, a model suggested strengthening muscles rather than endlessly softening tissue; two weeks of exercises helped more than years of vague ergonomic intuition. For recurring episodes of blurry vision, it helped him identify the possibility of ocular migraines and notice aspartame as a potential trigger.

None of this should be read as medical advice or expert replacement. The pattern is narrower and more useful: when the alternative is doing nothing, postponing, or paying for the wrong first conversation, AI can generate a plausible first map. It gives a motivated person enough structure to test, ask better questions, or approach the real expert with more context.

Voice and Context Will Lower the Cost of Curiosity

The most exciting frontier for Eskildsen and Shipper is not simply better text chat. It is AI that sits close enough to a human activity that asking becomes almost frictionless. Shipper describes using advanced voice mode while reading: when he encounters an unfamiliar word, historical figure, battle, or philosophical idea, he asks immediately. The surprising part is how many half-known things become askable once the cost drops.

Eskildsen connects that to children. His daughter is growing up in a world where she may ask “Wally the walrus” endless questions without exhausting a parent. She may also need help preserving Danish in Canada, where Danish-speaking exposure is scarce outside family calls. An AI toy or interface that speaks Danish, Mandarin, or Thai on different days changes the texture of early language exposure.

That thought leads to a bigger claim: tools reshape humans. Reading reorganizes the brain and changes analytical perception. Language models may do something similarly profound, not by being implanted into brains, but by changing the environment around curiosity. A three-year-old who can always ask why, get an answer, and step into a generated scene is developing under different conditions than previous generations.

Context Is the Product Surface

Notion AI interests Eskildsen because it sits where his notes already live and can pull context from the workspace. Shipper uses similar workflows to prepare for guests like Tyler Cowen: collect book summaries, connect them to intended points, and rapidly form a run-of-show with enough grounding to sustain a real conversation. Eskildsen sees this as part of turbopuffer’s world: semantic search pulling relevant context from years of documents, other people’s work, and document evolution.

The same idea shows up in journaling tools like Dot and social agents like Shapes, both mentioned as turbopuffer users. The future value is not merely a longer context window. It is the ability to retrieve the right pieces from billions of documents, personal history, and shared workspaces while respecting security and safety boundaries.

Dan’s proposed communication primitive captures the same direction: send a structured summary, but attach enough hidden conversational context that the recipient can ask follow-up questions. The message ships with a TL;DR and a small FAQ agent. Simon’s instinct is that this could work if it feels authentic and knows when to escalate rather than over-answer.

The Practical Framework: Build a Learning Loop Around Friction

Eskildsen’s operating system can be reduced to a few repeatable moves:

Key Lessons

The strongest lesson is that AI rewards people who already maintain loops. Eskildsen reads, reviews, asks, tests, stores, and reuses. AI amplifies that because it plugs into behavior that already exists. Without the loop, it is just another chat box.

The second lesson is that memory and AI are complements. Flashcards preserve what should be instantly retrievable. Language models supply broad associative reach. Search finds known targets. Humans provide high-bandwidth judgment. The mistake is expecting one layer to replace all the others.

The third lesson is that the best workflows are often unglamorous. The killer app is not a single cinematic demo. It is a hundred small reductions in friction that make you more likely to ask, understand, draft, decide, and remember.

Why This Matters for Diffie

For Anand and Diffie, the obvious analogy is not “add AI to everything.” It is put the agent exactly where the frontend engineer already feels friction. Eskildsen’s Raycast workflow wins because it is available at the moment of need. Diffie should aim for the same posture in browser testing: one action away from “what broke?”, “why did it break?”, “what changed?”, and “how do I reproduce this?”

Diffie’s ICP work should also borrow the dialect-translation insight. Frontend engineers, QA leads, product managers, and founders describe the same bug in different languages. A valuable AI testing product can translate among those dialects: DOM diff to human-readable failure, console error to likely user impact, reproduction trace to Linear/Jira-ready issue, flaky visual mismatch to probable CSS or data-loading cause. The product should not merely report failures; it should put each stakeholder into the right latent space to act.

The flashcard lesson maps to product memory. Diffie should remember durable patterns across a team’s app: recurring selectors, known flaky flows, common auth states, previous false positives, component-specific testing heuristics, and the engineer’s preferred debugging style. That memory should resurface at the moment of test generation or failure triage, not sit in a dashboard no one opens.

For GTM, the practical move is to show tiny, embarrassingly concrete wins. Do not sell an abstract “AI browser testing platform.” Show the $20 freezer-controller equivalent: “paste this user flow, get a reproducible bug report,” “turn this visual diff into a PR comment,” “convert this failing checkout run into a minimal reproduction,” “ask why this test is flaky and get the three likely causes.” The product narrative should make the buyer feel the same shift Simon describes: fewer unresolved questions, fewer round trips, and a faster path from curiosity to action.