Ensayos · Alberto Angulo
Escritos
Ensayos explicativos sobre cómo funciona la IA y qué implica para el trabajo, las empresas y los mercados, escritos con lenguaje claro y con las fuentes al final de cada pieza.
Los nuevos ensayos están escritos en inglés, así que los títulos y resúmenes de esta página aparecen en ese idioma. Los ensayos anteriores, escritos originalmente en español, siguen disponibles al final.
Por dónde empezar
Start here · The big picture
Thinking Got Cheap. Work Hasn't Caught Up.
The price of a competent machine answer falls about tenfold a year. Capability rises fastest where answers can be checked. And almost nobody has rebuilt their work around either fact yet. What is actually changing with AI, and where that leaves a person, a team or a company.
How AI works · Models and harnesses
Same Model, Different Results
A language model is one component in a loop of ordinary software. That software decides what the model sees, what it can touch and when it stops, and it can move results as much as a new model does. Some of it should be deleted with every model release. Some of it never should.
AI at work · Bottlenecks
Where the Saved Time Goes
Speeding up one step of a job rarely speeds up the job. The time piles up at the next bottleneck, usually checking and deciding, and the gain only arrives when the flow around the tool is redesigned.
How AI works · Capability
Why Progress Arrives Unevenly
AI gets better fastest at work whose answers a computer can check, because that is where training can practice without a person in the loop. That one property predicts which tasks fall next, and it is something people can build.
Working with AI · Learning
Learn to Drive Before You Study the Engine
You can get good at AI before you understand how it works. What it takes is practice on your own work, a personal map of where the output can be trusted, and the few pieces of mechanism that change what you actually do.
Systems around the model · Knowledge
Write for the Model That Doesn't Exist Yet
More and more of what an organization writes down is read by models, and next year's model will read it better than this year's. Write the reasons now, keep facts that expire apart from facts that last, and leave the polishing to the reader that is coming.
Rutas de lectura
Elige la pregunta más parecida a la tuya.
Uso IA y quiero hacerlo mejor
Construyo productos y sistemas de IA
Dirijo un equipo o una empresa
Quiero entender cómo funcionan los modelos
Decido sobre dinero, mercados y estrategia
Todos los ensayos, por área
How models learn and behave
Reasoning and inference
Programming and software architecture
Context, memory and knowledge
Harnesses and agent systems
Standards and interoperability
Evaluation and evidence
Infrastructure and AI economics
Product design and distribution
Enterprise reality and adoption
Transformation programs and delivery
FDE work and team organization
Value, measurable impact and compounding
Budgets, pricing, procurement, build/buy/partner
Labs and startups
Durable advantage, data, know-how and acquisitions
Industry mechanisms
Human learning, work, attention and resistance
Regulation and incentives
Routes, vocabulary and dated notes
Qué se está escribiendo
Planes, no promesas. Son títulos y preguntas de trabajo para piezas en preparación; cualquiera puede cambiar o descartarse.
Qué se está escribiendo · 89 piezas previstas
How models learn and behave
- Where a Model's Knowledge EndsWhat does a model know from training, why doesn't it know your company, and what is a knowledge cutoff in practice?
- Why Models Agree With YouWhy do assistants flatter and cave, and how do you get honest disagreement out of them?
- Why the Same Prompt Gives Different AnswersWhere does variability come from (sampling, batching, versions), and when does it matter?
- How Small Models Got GoodWhat do distillation, quantization and mixture-of-experts change, and when is a small model the right choice?
- What a Model Sees in a PDFHow do images, tables and documents become something a model can read, and why do charts and scans still trip it?
- Personality Is a Product DecisionWho decides how a model talks and refuses, and why do model specs and system prompts matter to users?
- What Attention DoesHow does a model combine information from distant parts of a text, and what does the word 'attention' get wrong?
- What Interpretability Can and Can't ShowWhat do researchers actually see inside models today, and how is a plausible explanation different from evidence of a mechanism?
- Learning From Synthetic DataWhen does model-generated training data expand capabilities, and when does it recycle errors or narrow variety?
- What a Reward TeachesHow do models learn from rewards, which behaviors get reinforced by accident, and what is reward hacking?
- Can a Model Learn From Your Corrections?What do in-context learning, product memory, fine-tuning and continual learning each mean for someone giving feedback?
- Voice Changes the InteractionWhat do latency, turn-taking and interruption change when the assistant talks instead of types?
- When Reality Pushes BackWhy are robots and physical tasks so much harder than digital ones: observation, uncertainty, irreversibility, cost of failure?
Reasoning and inference
- Do Models Mean What Their Reasoning Says?Is a model's visible reasoning a faithful account of how it reached the answer, and what is it still useful for?
- Why Long Answers Cost More Than Long QuestionsHow do prefill, decoding, caching and batching set the price and speed of a request?
- Speed Is a FeatureHow do latency and throughput change what an AI product can be, beyond the model's quality?
- One Answer or Many Attempts?When do several attempts plus a good selector beat one expensive attempt, and what happens when attempts share the same blind spot?
- How Errors Travel Through a PlanHow do early wrong assumptions compound across long tasks, and where should checkpoints go?
- When a Tool Beats Another ThoughtWhen does calculating, searching or running something beat more reasoning over the same context?
Programming and software architecture
- Architecture for Code Nobody TypedWhat makes a codebase easy for agents (and new people) to work in: legibility, conventions, boundaries, feedback?
- Technical Debt at Machine SpeedWhat does the evidence say about the quality of AI-generated code, and how does debt accumulate when generation is cheap?
- When a Prototype Is EnoughWhere is fast 'vibe coding' the right call, and where do security, maintenance and correctness make it expensive?
- The Programmer's Job, RewrittenWhich programming skills gain value as agents write more code: specifying, decomposing, verifying, operating?
Context, memory and knowledge
- Search by Meaning, and Where It FailsHow do embeddings find related text, and why do they miss the obvious?
- Fine-Tune, Retrieve or Just Tell ItWhen does changing the model's weights beat putting knowledge in its context, and when is it wasted money?
Harnesses and agent systems
- When to Split the Work Across AgentsDoes a team of agents beat one agent, and what does splitting really cost?
- Agents That Work for HoursWhat does it take for an agent to keep working across hours and sessions without losing the thread?
- Tools Are an Interface Design ProblemWhy do fewer, better-described tools beat a large toolbox, and how should tools for models be designed?
- Why Clicking Is Hard for an AIHow do computer-use and browser agents work, and why are screens harder than APIs?
- Proof That the Work HappenedWhat evidence distinguishes intention, attempt and result, without drowning in logs?
- Human Review Where It MattersWhere does human review add judgment, and where is an approval step just delay nobody inspects?
- An Incident That Improves the SystemHow do you learn from an AI failure by fixing causes and scope instead of piling up prohibitions?
- Briefing an Agent Without MicromanagingHow do you hand over outcome, reasons and constraints without imposing a sequence that makes the result worse?
Standards and interoperability
- What Standards Do to Lock-inHow do shared protocols and file conventions change switching costs between AI providers and tools?
Evaluation and evidence
- How to Read an AI Productivity StudyWhat do the famous studies actually measure, and how do selection, task choice and self-report change the headline?
- Demos Fail in Predictable WaysWhy does an impressive demo overstate reliability, and what questions reveal the real success rate?
Infrastructure and AI economics
- The Compute Behind the CurtainWhy does AI need so much hardware, power and capital, and who ends up paying for it?
- Routing Work Between ModelsHow should tasks be split across large and small models by difficulty, error cost and review, without demoting important judgment?
- Should You Host Your Own Model?When does serving a model yourself pay off once utilization, staff, hardware, upgrades and continuity are counted?
- Optimize Now or Wait for the Next Model?Which efficiency work pays back before the next capability jump makes it obsolete?
Product design and distribution
- Designing for a Collaborator That's Sometimes WrongHow should products show uncertainty, invite checking and make mistakes cheap to undo?
- When Chat Is the Wrong InterfaceWhich jobs need a chat box, and which need forms, documents, dashboards or work that happens in the background?
- Seats, Usage or OutcomesHow should AI products be priced when the cost of serving a customer varies a hundredfold?
- The Wrapper QuestionWhich products built on someone else's model get absorbed, and which get stronger as models improve?
- Proactive AI Worth Interrupting ForWhat separates useful initiative (work done, not just prepared) from another notification?
- The Platform Owns the DoorHow do app stores, office suites and dominant assistants control access to users, and what can a product do about it?
- Trust Before the First ResultWhat evidence lowers a buyer's uncertainty before an AI product has proven itself, without promising autonomy it doesn't have?
- Vertical AI: Real Advantage or Label?When do sector knowledge, integrations and accountability create value, and when is 'AI for X' just packaging?
Enterprise reality and adoption
- The AI Your Employees Already UseWhat does unofficial AI use tell a company about demand, risk and where to start?
- Why Pilots StallWhat keeps successful pilots from reaching production, and which of those reasons are fixable?
- How Enterprises Buy AIWhat happens between a team wanting a tool and a signed contract: security review, procurement, legal, budget cycles?
- The Company Behind Its SoftwareHow do you account for physical work, informal agreements and human knowledge that never appear in the applications?
- A Shared Core Without Flattening the CompanyWhat can be standardized across teams or clients, and which differences are real needs worth keeping?
Transformation programs and delivery
- Choosing the First ProcessWhich process should go first: what volume, checkability, data access, ownership and risk make a good starting point?
- Change Management When the Tool Changes MonthlyHow do you train and reorganize people when the capability underneath keeps moving?
- Governance That Speeds You UpWhat approval paths, policies and owners let an organization move faster with AI instead of slower?
- An AI Learning Program That Changes WorkWhat should company-wide AI training teach, and how do you tell whether it changed anything?
- When to Scale, Revise or StopWhich signals justify expanding, changing or closing an AI initiative without confusing sunk cost and enthusiasm with future value?
FDE work and team organization
- The Job of Deploying AIWhat do forward-deployed engineers actually do, why is the role growing, and how should the work be organized?
- Hiring When Producing Is CheapWhat skills matter when AI does more of the producing, and how do juniors learn the craft?
Value, measurable impact and compounding
- Cost per Task, Not Cost per TokenHow do you compute what an AI-assisted task really costs, including retries, review and failure?
- Capacity Nobody Can UseWhat stops extra productivity from becoming output: missing demand, review, integration, authority or organization?
Budgets, pricing, procurement, build/buy/partner
- Choosing a Model Provider Without Marrying ItHow do you contract, route and design so that switching providers stays possible and cheap?
- What to Hand to a PartnerWhen should a company use integrators, consultants or vendors' services for AI, and what must stay in-house?
Labs and startups
- What a Startup Can Still WinSqueezed between labs and incumbents, where do new AI companies find room?
- What Scaling Laws Do and Don't PromiseWhat exactly scales with more compute and data, what the recent debates are about, and what it means for planning?
- An API Is a Strategic DependencyHow do you weigh exposure to a provider's prices, policies and priorities without assuming all dependency is bad?
- Learning Faster Than the Market ChangesWhich tests and assets let a company adapt when its founding hypothesis stops holding?
Durable advantage, data, know-how and acquisitions
- The Data You Think You HaveHow much of a company's data is actually usable for AI, and what makes data valuable rather than merely abundant?
- Capital Requirements and Strategic FreedomHow does needing compute, funding or hypergrowth narrow the options of an AI company?
- Optionality Without an Empty StrategyHow do you keep several paths open (adaptation, partnership, exit) while still building something concrete?
Industry mechanisms
- Law, Where Checking Is ExpensiveWhy is legal work both highly exposed to AI and slow to change, and what does that teach about verification cost?
- Healthcare Started With the PaperworkWhy did clinical documentation become healthcare's first large AI adoption, and what does that pattern predict?
- When the Bottleneck Is PhysicalHow does AI change manufacturing, logistics and field work, where atoms rather than documents set the pace?
- Homework in the Age of GenerationWhat changes in education and training when any assignment can be generated, and what still builds skill?
- Insurance: Pricing Risk With Better ReadingWhat changes when claims, policies and risk documents can be read and compared at scale, and where do regulation and fairness constrain it?
- Retail and Commerce When Shoppers Ask an AssistantWhat happens to product discovery, catalogs and marketing when buyers ask an AI instead of browsing?
- Public Administration and the Paperwork StateWhere can AI shorten public processes, and why do accountability and equal treatment make it harder than in a company?
- Science When Hypotheses Get CheapHow do AI tools change research (literature, experiments, code, data analysis), and where is the bottleneck still the lab bench?
Human learning, work, attention and resistance
- The Human PremiumAs content gets cheap, what becomes valuable: attention, trust, presence, accountability?
- When Capability Widens the GapWhich access, skills and resources make new AI capability spread its benefits unevenly, between people, firms and countries?
Regulation and incentives
- Who Owns What the Model Learned FromWhere do copyright, licensing and the training-data lawsuits stand, and what do they change for companies using AI?
- Who Is Responsible When an Agent ActsWhen an AI agent makes a mistake with real effects, how do liability, contracts and internal accountability work?
- A Loophole Is Not a BusinessHow does depending on a regulatory gap, exception or gray zone affect the durability of a product, supplier or acquisition?
Second-order effects
- Three Layers Past the ObviousBeyond the AI products themselves, which enablers and indirect beneficiaries does this wave create?
- What Gets Scarce When Intelligence Gets CheapWhen drafting, analysis and code become abundant, which inputs become the new constraints and the new sources of value?
- Industries That Rise Because of AIWhich sectors gain value as a consequence of AI spreading, and through which mechanisms?
Routes, vocabulary and dated notes
- Words People Assume You KnowWhat do the terms in these pieces mean, in plain words, with links to where they're explained?
- What a New Release Actually ChangesHow to read a model launch or benchmark announcement by mechanism, task, cost and second-order effects rather than headline.
Ensayos anteriores (originales)
Ensayos del archivo sobre producto y sistemas de IA. Reflexiones, no pruebas de experiencia ni de resultados.