Before an AI writes an answer, it helps to first agree on the list of questions a good answer needs to cover. If the AI then writes one short, sourced sentence for each question on that list, you get an answer that covers what matters and lets you trace every claim back to where it came from. This paper builds an AI system that works exactly that way and it beats the usual approaches.

Who is this for

Anyone who relies on AI to pull together facts from a trusted pile of documents (a work knowledge base, a set of reports, a research library) and needs to actually trust the result. If it matters for you to be able to point to the source behind each sentence, and not just hope that the confident-sounding paragraph is right, then this is for you.

Background

“Incorporating Q&A Nuggets into Retrieval-Augmented Generation” is a 2026 research paper by Laura Dietz (University of New Hampshire) and colleagues at the University of Pennsylvania, Yale, the University of Amsterdam, and Johns Hopkins. It introduces an AI system called CRUCIBLE.

To follow along, you only need one idea. When an AI answers a question using a set of documents, researcher

s like to check its work by writing down the handful of small facts a good answer ought to contain — each one as a tiny question-and-answer pair, like “Who makes Roundup weedkiller? — Bayer.” These little fact-checks are called nuggets (small, self-contained facts). Normally nuggets are used afterward, as a grading checklist: count how many facts AI’s answer got right. The clever twist in this paper is to use that checklist at the start, as the plan for writing the answer — not just the test at the end.

If you saw our earlier post on GINGER, this is the next step in that same story.

The issue

The standard way AI answers a fact question is to first go fetch relevant documents, then write based on them. In such approaches, academia has identified many potential issues. This paper focuses on two:

First, repetition. The same fact often appears in many documents, so the AI sees the same point again and again and misses other points worth focusing on.

Second, and more important, you lose the paper trail. To write a smooth summary, most systems blend all the source material together into one mush before writing. The result reads nicely, but you can no longer tell which document backs any particular sentence. For anything that matters, a claim you can’t trace is a claim you can’t really trust.

Read more

I spend most of my time here telling you to be suspicious of AI. Slow down. Check the confident answer. Don’t hand over your thinking to a machine that’s wrong with a straight face. So consider this a plot twist: today I want to stick up for it.

Not because I’ve gone soft; because the pendulum has swung so hard the other way that the criticism has turned into exactly what it claims to hate, a polished, confident story that nobody bothered to verify.

Here’s the cycle we just lived through. First came the salesmen. Every founder with a model and a pitch deck promised you the end of work, the cure for everything, a robot god by Tuesday. And look — I get it. When you’ve raised at a valuation that only makes sense if you reinvent the global economy, you kind of have to talk like you’re reinventing the global economy. OpenAI’s valuation has been speculated in the hundreds of billions of dollars (Reuters). You don’t justify a number like that with “it’s a pretty handy writing tool.”

So the dreams got sold. And predictably, the dreams curdled. Now we’re in the backlash, and the backlash has its own greatest hits, which you’ve heard at every dinner table and in every group chat: It doesn’t actually work. It’s too expensive. The companies are hemorrhaging money, so they’re never going to be real businesses, and nobody’s even using the stuff at work anyway.

Some of that is fair, and some of it is the same lazy thinking as the hype, just wearing a skeptic’s jacket. Let me take them one at a time, because the details matter, and the details are where both the cheerleaders and the doomers get caught oversimplifying.

“It just doesn’t work.”

This is the one I have the least patience for, because it’s usually said by someone who typed one vague prompt into a free chatbot in 2024, got a wrong answer, and closed the tab forever. Meanwhile, ChatGPT’s weekly active users surpassed 400 million in early 2025 (Reuters), and by May 2026 it was pulling a record 626.9 million monthly active users (Similarweb on X). Claude, Gemini, Grok, Perplexity, and Microsoft Copilot have all surged in recent months too (AI Search Traffic Report 2026). You can quibble over whether any single figure is a little off, but the scale is too big to wave away.

We also know it’s landing across the board within companies. Deloitte found 25% of organizations already have at least 40% of their AI experiments in production, with 54% expecting to reach that level within six months (Deloitte: State of AI in the Enterprise 2026). Controlled studies keep finding real gains on specific tasks: a large NBER study found generative AI raised customer-support productivity by about 14% on average, and up to 34% for newer and lower-skilled workers (NBER: Generative AI at Work). Anthropic’s own analysis of 100,000 Claude conversations estimated the studied tasks were completed about 80% faster with AI assistance (Anthropic: Estimating productivity gains). That is not a toy that “doesn’t work.”

But here’s where I won’t lie to you, because lying to you is the whole thing Toottee exists to not do: the skeptics aren’t hallucinating either. That widely cited MIT-based report suggested that most corporate generative-AI pilots (~95%) had yet to show measurable bottom-line impact (Fortune/MIT). My favorite gut-check of the whole debate comes from METR, who took experienced open-source developers, gave them AI tools, and measured them on big, messy, mature codebases. The devs felt about 20% faster. They were actually 19% slower (METR study).

Sit with that gap for a second, because it’s the most Toottee fact in this entire piece. The tool works. Your sense of how well it’s working is unreliable. “Does AI work?” is the wrong question. The right one is: for which task, for which person, and checked how? It crushes greenfield projects, drafting, and helping a junior climb a learning curve. It quietly drags when an expert already holds the whole codebase in their head. “It doesn’t work” isn’t a finding. It’s a refusal to ask the second question.

“It’s too expensive.”

This is the easiest one to wave away, and the numbers aren’t close. The cost of intelligence is in free-fall. The price to reach a given level of quality has been dropping by a median of about 50× a year, and as much as 900× a year for some capabilities (Epoch AI). It’s falling on an unusually steep cost curve. Whatever you priced AI at the last time you checked, it’s wrong now, and it’s wrong in your favor.

The honest caveat, because there’s always one, is the token-cost illusion (Artefact). Cheaper per token does not mean a smaller bill. Usage explodes, reasoning models chew through tokens like crazy, and the frontier labs are still spending fortunes on training and data centers (Sequoia: AI’s $600B question). So: “the unit price is collapsing” is true. “AI is cheap” is not. Both things exist at once.

“They’re losing money, so they’ll never be real businesses.”

Now we’re at the big one, and this is where the backlash does its sloppiest thinking. The premise is true. By analyst estimates, OpenAI is on track to lose something in the tens of billions in 2026 and doesn’t expect to turn a profit until the end of the decade (Reuters). That’s a serious estimate, and I’m not going to pretend it away.

But “losing money while scaling” is not the same sentence as “will never be a business.” Amazon lost money for the better part of a decade on purpose. The tell is that the doomers lump every AI company into one undifferentiated money-pit, and the moment you stop doing that, the story falls apart. Anthropic, by contrast, appears to have ramped up revenue quickly into the multi-billion-dollar range and has been reported to be targeting positive cash flow by 2028 (Reuters). One company burning cash on a moonshot timeline and another sprinting toward breakeven are not the same data point, and treating them as one is exactly the kind of confident oversimplification I’d want you to catch in anyone else.

Here’s the part where I switch sides, though, because the strongest version of the bubble argument deserves its due, and it’s a lot scarier than “they lose money.” The industry has lined up hundreds of billions of dollars in projected infrastructure spending against a far smaller base of current revenue, and critics argue the ecosystem is knotted up in interdependent deals where the same dollars get passed between chipmaker, cloud, and lab in a way that can make demand look bigger than it is (NPR). Even Daron Acemoglu — a Nobel economist, not a hater — says plainly that “much of what we hear from the industry now is exaggeration” (NPR). There may well be a financial reckoning. Some of these valuations are going to eat dirt.

Notice the move I just made, because it’s the whole point of this newsletter: the technology being real and the financing being a bubble are completely separate claims. The fiber-optic cable laid during the dot-com bubble didn’t stop being useful when the stocks crashed. A market can be overpriced and the underlying thing can still be one of the most important tools you’ll ever learn. Both. At. Once.

“Nobody’s actually using it at work.”

Half right, and the half that’s wrong is instructive. By any historical standard, adoption isn’t slow — it’s among the fastest any technology has ever spread. Enterprise generative-AI use roughly doubled, from about a third of companies to nearly two-thirds in two years (McKinsey). What’s actually lagging is value realization: many organizations report real friction getting it to pay off, and the most careful macro estimate we have projects the whole-economy productivity boost will be modest — on the order of half a percent over the next decade (MIT: The Simple Macroeconomics of AI). That’s not “nobody’s using it.” That’s “everybody grabbed the tool and most of them haven’t learned to use it well yet.”

Which, if you’ve been reading Toottee for more than a week, you’ll recognize as the entire reason this thing exists.

So where does that leave us?

Right back at the founding idea. The first wave of people went all-in on AI without understanding it. The second wave is writing it off without understanding it. Same mistake, opposite jersey. The hype was a story sold to you with a straight face, and you learned — correctly — to doubt it. The backlash is also a story being sold to you with a straight face. It just flatters you more, because doubt feels smarter than enthusiasm. It usually isn’t. It’s just the other flavor of not checking.

Give the technology its due. It works, unevenly. It’s getting cheaper, fast. The businesses are a genuine mess, and some of them are a bubble, and none of that tells you whether the tool in front of you is worth learning. That’s still your job to figure out, task by task, check by check.

Don’t go all in. Don’t write it off. Do the harder, less tweetable thing: actually understand the damn thing, and decide for yourself.

That’s the whole game. Always has been.

An AI answer is only as good as the information it has access to. Give it too little and it guesses; give it too much and the one fact you needed gets buried in the middle, where models reliably lose track of that fact.

Who is this for

Anyone using AI to answer questions based on a bundle of trusted materials. Think company documentation, legal, medical, or regulated industries with strict rules about what counts as true.

Background

GINGER (Grounded Information Nugget-Based Generation of Responses) is a 2025 research paper by Weronika Łajewska and Krisztian Balog at the University of Stavanger. It tackles Retrieval-Augmented Generation (RAG), the now-standard setup where an AI first retrieves documents and then provides an answer based on them. The system was evaluated on TREC RAG’24, a standardized test from the long-running Text REtrieval Conference, where it reached the best results anyone had achieved at the time.

Read more

Don’t judge an AI’s answer as one block, and don’t let the polish fool you. Check it one claim at a time.

Who is this for

This is for anyone who leans on AI for factual writing and wants a sharper way to talk about accuracy than “it seemed right.” That could be a student, a professional, or frankly anyone who uses AI and cares whether its answers are accurate.

Background

FActScore, short for Fine-grained Atomic Evaluation of Factual Precision, was introduced in a 2023 paper published at EMNLP, one of the top research conferences in natural language processing (NLP). The fast pace of computer science research already makes it old, yet its core ideas are still in use today.

The issue

When an AI tool writes a paragraph, it rarely lands on “all true” or “all false.” A single answer can combine facts that check out with facts that are either irrelevant or incorrect. Evaluating the whole thing with a thumbs up or thumbs down misses the point. The question is not whether it’s accurate, but how much of it is accurate.

A common way to approach this is to break the generated text (the AI’s response) into atomic facts, the smallest standalone claims. Take this generated text as an example:

“He was born in 1961 into a household of 5 living in Kansas City. His family later moved to South Dakota because of his father’s job in the mining industry.”

This paragraph can be split into 5 atomic facts.

1. He was born in 1961.

2. He grew up in a household of 5.

3. He lived in Kansas City.

4. His family later moved to South Dakota.

5. His father worked in the mining industry.

The solution

FActScore builds on this idea of atomic facts. First, it uses a language model to break an LLM response into atomic facts. Each claim is then checked against a trusted knowledge source (Wikipedia in this case) to see whether the source supports it. The FActScore is simply the percentage of facts that hold up.

They initially used humans to evaluate the claims, but that is slow and expensive. So their main contribution is an automated version that retrieves passages from the source, and has a model judge each claim, mirroring human judgment with under 2% error. That made it cheap enough to score 6,500 generations from 13 models, work that would have cost about $26,000 by hand.
The findings are pointed. A few clear patterns emerged in how the models failed.

1. Even ChatGPT, a leading model at the time, reached only 58% factual precision, so a large share of its claims were not supported.

2. The more obscure the subject, the more errors crept in.

3. Claims buried later in a response were more likely to be wrong than the ones near the start.

4. Models that rarely declined to answer tended to score worse, since always guessing produces more unsupported claims.

Interestingly, even tools that search the web and cite their sources (like Perplexity) were less reliable than their reputation suggests.

Towards Practical Everyday Solutions

1. FActScore gives you a vocabulary and a high-level framework for evaluating an LLM response (breaking answers into atomic facts, evaluating the factual precision of each claim, and detecting which claims are supported, not-supported, or irrelevant).

2. The results show that you should be skeptical of claims about niche topics, and of the details that show up deep in a long answer.

3. As a mental model, if you are using the text produced by these models, make sure to independently check each claim and verify them if you are unsure.

The habit underneath all of this is simple. Don’t trust the polish, check the pieces.

🔗 Sources

• https://aclanthology.org/2023.emnlp-main.741/

• FActScore GitHub repository

This is the final part of our series on navigating AI. In Part 1, we looked at how AI’s polished tone tricks us into trusting it with trivia. In Part 2, we looked under the hood at why AI hallucinates when the stakes are higher. Now, let’s talk about what to do about it.

(A quick note before we dive in: you will notice a lot of anthropomorphism in this piece. While AI obviously doesn’t have human feelings or intentions, giving it human traits is simply the most effective mental model for learning how it acts and works).

The golden rule for dealing with AI is this: Treat it like a smart, hardworking, but overconfident intern. It is incredibly fast, remembers (almost) everything it was taught, and is great at following clear instructions. The caveat is that it desperately wants to please you, which means it would rather guess the answer than admit it doesn’t know.

To manage this intern, we need to understand a few quirks about how it works:

• It is overly verbose: It loves to write paragraphs upon paragraphs, sometimes talking in circles, to answer a simple question.

• It needs bite-sized instructions: If you feed it a paragraph with four different commands, it will get lost. If you give it one instruction at a time, it performs brilliantly.

• It needs to “think”: Asking an AI to work “step-by-step” or to reflect on its own answer actually improves its logic and reduces mistakes.

• It has biased memories: While it acts as massive stores of information, it easily remembers concepts that were heavily prevalent on the internet, but struggles with niche or less-documented topics.

Given these traits, how do we actually verify what the AI tells us? Let’s look at three practical scenarios.

Scenario 1: The Closed Sandbox (Using AI for First Drafts)

The safest way to use AI is to give it the answers upfront. Large language models are incredible at synthesizing information if they have the right context.

Give the model your own Excel sheet, your class PowerPoints, or your Tableau outputs, and ask it to write a summary or pull insights, similar to asking an intern to draft a report solely using a task-specific stack of files.

The wording might not be perfect, and it might be a bit exaggerated or repetitive, but the model gives you a solid first draft in seconds.

The Verification strategy: Because you provided the source material, verifying is easy. You already know the data. You just skim the draft to ensure it didn’t misrepresent your own files. Polish the wording and you’re done.

Scenario 2: The Librarian (Using AI for Database Retrieval)

Often, we can’t feed the model everything we want because of “context limits” as it can’t read the entire Library of Congress in one prompt.

To solve this, developers use a variety of tricks like Retrieval-Augmented Generation (RAG). Think of how Perplexity works: you ask a question, the system Googles the keywords, finds 10 relevant web pages, feeds only those specific pages to the AI, and asks it to summarize the answer. Enterprise tools like Microsoft Copilot (for internal files), AlphaSense (for finance), and Harvey (for law) do the exact same thing with corporate documents.

Understanding this process is the key to verifying it. If the retrieval system fails to find the right document, the AI will likely hallucinate an answer to fill the void.

The Verification Strategy: Check the specific citations. For example, imagine a corporate finance director asks their internal AI: “What drove the Q3 revenue shortfall in the European division?” If the Q3 data hasn’t been uploaded to the database yet, the AI might pull a report from Q2 and confidently claim the shortfall was due to “supply chain delays in Germany.” The text sounds perfect. But if you click the footnote and see it links to a Q2 document, you instantly know it’s a hallucination. Always check the receipt.

Scenario 3: Researching the Unknown (The Cross-Examination)

When you don’t have your own files or a corporate database to rely on, you are relying entirely on the AI’s internal memory. This is where you need to know its blind spots and how to interrogate it.

Know Where AI Struggles

AI is brilliant at explaining well-documented concepts, summarizing common knowledge, breaking down complex topics, and generating ideas. However, it has massive blind spots. It is fundamentally weak when dealing with recent information, niche or obscure facts, precise statistics, lesser-known people, and local or regional specifics. If your question falls into these categories, assume the AI is guessing until proven otherwise.

Red Flags to Watch For

Even when our “overconfident intern” is hallucinating, it sounds incredibly convincing. Keep an eye out for these warning signs:

• Too-convenient details: Did it give you an exact number, a perfect quote, or a hyper-specific data point without a clear link? Be suspicious.

• No hedging: Real experts use phrases like “it depends,” “there is debate on this,” or “traditionally.” AI models are programmed to sound authoritative and rarely hedge unless explicitly prompted.

• Fabricated sources: AI will happily invent books, articles, or studies complete with real authors and plausible publication years. Always Google the title or click the link to ensure the citation actually exists before you trust it.

Creative Verification Strategies

Here is where it gets interesting. You can actually use AI’s own technological tendencies to test its answers.

• Play devil’s advocate: Challenge the response. Ask: “Is there any evidence against this?” or “What are the alternative explanations?” You can even feed it a false premise. For example, ask about the gold standard, and then follow up with, “Wait, didn’t the US actually stay on the gold standard until the 1980s?” If it immediately apologizes and agrees with your false statement, you’ve exposed a hallucination. It is just trying to please you.

• The consistency test: AI developers frequently use this trick to catch errors. Ask the model the exact same question in three slightly different ways. If the core facts such as the dates, names, or percentages, change between the responses, the model is guessing.

• Check consensus across models: Think of different AI chatbots like different people at a dinner party. Ask the exact same question to ChatGPT, Claude, and Gemini. If they all give roughly the same answer, you are probably on solid ground. If they diverge significantly, you’ve found a knowledge gap and need to do traditional research.

• Force a “fact-check list”: Prompt engineering best practices recommend asking the AI to verify its own work. Add this instruction to the end of your prompts: “List the three most critical factual claims in your response, evaluate your confidence in each, and provide the sources.” This forces the model to start evaluating its own logic.

The Bottom Line

AI is a powerful tool, not a trusted authority. The more confident it sounds, the more critical it is to verify. We are replacing the old habit of “search and investigate” with “ask and verify.” To survive this shift, build the habit of the cross-examination now, before the convenience becomes complacency.

By now, you understand the “trust trap.” To understand why AI lies to us (and how to catch it), we need to look under the hood. We’ll start with the What, move to the How, and end with the Why.

The Mechanism of the Lie: Hallucinations

What they are

In technical terms, the AI chatbots we use today are empowered by “deep learning autoregressive models.” That’s a mouthful, but it just means they are designed to look at a sequence of text and predict the next best text to follow.

How they work

These models are built on massive architectures trained to guess the next “token” (roughly, a word or part of a word).

Take the sentence: “The CEO of Apple is…”

The model looks at its training data, billions of articles, blogs and forums, and calculates that “Tim” is the most likely next word. Then it looks at “The CEO of Apple is Tim” and predicts “Cook.”

It didn’t “know” Tim Cook is the CEO in the way you or I do. It just knew that in the history of the internet, those words appear in that order statistically often.

Why this leads to hallucinations

After doing the math, the AI doesn’t just have one answer; looking at the sequence of words “The CEO of Apple is”, it has a ranked list of probable next words.

• Option A: “Tim” (90% probable)

• Option B: “Steve” (5% probable)

• Option C: “Sundar” (1% probable)

To make the AI sound natural and creative, it isn’t programmed to always pick the #1 answer. It has a “temperature” setting that introduces randomness. Occasionally, it grabs Option B or C just to vary the sentence structure. This is why you can ask ChatGPT the same question twice and get two different answers.

With facts, at least the mistake is findable. Harder to catch is when the model doesn’t just swap a name; rather, it swaps a consequential word, shifts a tone, or changes your underlying strategy without instruction.

Take a look at the following two responses. The prompt is identical, but the output is different. No hallucination. No wrong answer. Just two responses that handle a sensitive customer interaction in entirely different ways. One invites further debate; the other firmly closes the door.

The Danger

Here is why this matters: The model is prioritizing probability, not truth. It is trying to complete a pattern, not answer a question. Imagine you are a legislative aide using AI to draft a policy memo. In legal writing, the difference between “the agency shall enforce the rule” and “the agency may enforce the rule” is the difference between a federal mandate and a mere suggestion.

To a human, those are opposing concepts. To an AI, “shall” and “may” are just two statistically similar words that both appear frequently in legal texts. If the model’s “randomness” coin flip lands on the wrong one, you have a sentence that reads perfectly but creates a legal disaster.

What this looks like in practice

When patterns override facts, you get:

• Fake Citations: AI inventing academic papers complete with plausible titles, real authors, and fake publication years because that structure “looks” like a citation.

• Zombie Facts: Stating outdated information as current truth (referencing a CEO who stepped down or a restaurant that closed years ago).

• Phantom Statistics: Generating precise-sounding numbers (e.g., “73% of users...”) that have no source but fit the sentence flow.

• The “Expert” Lie: Confidently explaining a complex concept incorrectly in a way that is impossible to catch unless you are already an expert in that specific niche.

The Productivity Paradox

Now, all of this might make you want to swear off AI entirely. If it lies, hallucinates, and makes up laws, why bother?

Here is the twist: You absolutely should use these tools.

The goal of this series isn’t to scare you into digital celibacy; it’s to help you move from being a passive consumer to an active pilot. When used correctly, AI is still the most powerful productivity multiplier we’ve seen in decades. It can summarize in seconds what takes you hours, draft emails that clear your inbox, and brainstorm ideas when you are stuck.

The danger lies only in the blind handoff.

Productivity isn’t about speed; it’s about effective output. If you save 30 minutes writing a memo but spend three days cleaning up a PR mess because the stats were fake, you haven’t been productive, you’ve been reckless. The sweet spot is a “trust, but verify” workflow where you use AI to do the heavy lifting, but retain the role of the editor-in-chief.

We need to build a new set of digital reflexes. Just as you check your blind spot before changing lanes, you need quick, low-friction habits to check AI before you hit send.

In Part 3, we’ll cover practical strategies for catching these mistakes before they cost you.

Yes, the capabilities vary, and plenty of people still aren’t using these tools, but the shift is underway; we’re replacing google searches with artificial conversations, and given how fast these services are weaving into daily life, we need to confront an important question: how do we verify what they are telling us?

I’ve taken a bit of a different approach to this post, turning it into a multi-part series tackling that problem:

1. Part 1: The Easy Wins (and the Trust Trap) — Why AI’s ability to nail trivia is the reason it’s so dangerous.

2. Part 2: When It Actually Matters — Work decisions, coursework, life advice—and what happens when AI confidently makes things up.

3. Part 3: A Verification Toolkit — Practical strategies for fact-checking AI, from playing devil’s advocate to using multiple models against each other.

Part 1: The Easy Wins (and the Trust Trap)

Let’s start with trivia, from questions like “when did the US stop using the gold standard?” to an obscure fact only your medieval history professor would know off the top of their head—AI handles these topics brilliantly.

This isn’t new territory. Google could answer these too, however, the experience was different: you’d get a list of links, click through Wikipedia, maybe skim an economics article, even read a post or two on Reddit, and piece together the answer yourself. You were doing the synthesis.

Now you skip all of that instantaneously. You ask, you receive. The AI summarizes everything and responds like a knowledgeable friend texting you back. No clicking, no scanning, no cross-referencing. Just the answer, formatted however you want it.

Take the gold standard question mentioned previously. When I asked Grok and googled the question, these are the responses I received:

The response is so polished, so authoritative, that questioning it feels almost unnecessary. When something gives you a well-formatted, knowledgeable-sounding answer, why would you doubt it?

Here’s where our instincts fail us. We learned internet skepticism the hard way: check who published this, compare sources, be wary of anonymous claims. But AI responses don’t trigger those alarms. They don’t look like a sketchy forum post or a clickbait headline. They look professional. Almost academic.

Our old skepticism was built for a different threat. We know to question a random website. We don’t instinctively question something that sounds like a well-read colleague who’s done the research for us.

The shift from “search and verify” to “ask and trust” is outpacing our instincts. We need new habits for this new tool.

For trivia, there’s a simple fix: check the sources. Some services make this easy (Perplexity, for example, shows the links it pulled its response from), and you can click through just like you would with Google results. For low-stakes questions, this works fine. You’re basically falling back on your old googling instincts: skim the sources, confirm the answer, move on with your day.

But what happens when the question isn’t trivia? When you’re not asking about the gold standard, but about something that affects your job, your grades, or your decisions? That’s when source-checking gets harder, and when the stakes of blind trust get real.

One of the biggest routines is the 8-hour workday. Think about it: you spend roughly 8 hours in school as a kid, then graduate into 40 years of 8-hour workdays. That’s one-third of every day. Do the math — just the working hours alone add up to 14 years of your adult life1, or about 121 days every single year2. That’s a lot of time to spend doing something that might not fulfill you, or doing it inefficiently when you could be doing something else.

This isn’t about replacing your thinking or turning into some AI-dependent robot. It’s about using the tools we have now, AI and LLMs, to make these routines work better for us. Let’s talk about how.

Type Uno: Learning New Skills and Changing Careers

A lot of us reach a point where we begin to question our careers and even consider switching to another path. Yeah… yeah… I know we all know that colleague who always says that but keeps doing the same Excel or PowerPoint work year after year without change. However, many of us are genuinely unhappy with our current career status, and no time has been better to use our time wisely to switch paths and gain insight into a new career.

One of the best things you can do is treat these LLM chat interfaces as your personal tutor. It’s true that some LLMs hallucinate and offer made-up information, but for learning new topics and understanding them at a high level, the occasional hallucination doesn’t matter nearly as much as you’d think. You’re not building a rocket ship based on Claude’s advice—you’re trying to understand what product managers actually do, whether you’d enjoy data analysis, or how venture capital firms evaluate startups. For that, these tools are phenomenal.

Here’s how I use it: Let’s say I’m curious about transitioning into product management. I don’t need to spend $2,000 on a bootcamp or read five books to understand if I’m interested. Instead, I open Claude, ChatGPT, Gemini, or Grok and ask: “explain product management to me like I’m a software engineer who’s never done it before. What does a typical day look like? What frameworks do PMs use? What’s the difference between B2B and B2C product work?”

Within 10 minutes, I have a solid understanding. Then I can go deeper: “walk me through how you’d use the RICE scoring framework for feature prioritization” or “give me a mock scenario where I need to decide between two features and explain your reasoning.” You’re essentially getting a patient, infinitely available tutor who meets you at your level and adjusts explanations based on your questions.

Want to understand machine learning but don’t have a PhD in computer science or math? Ask Claude to explain gradient descent using only high school algebra. Curious about finance? Have a conversation about cap tables, dilution, and liquidation preferences. The point is that you can explore career paths in hours instead of months, and you can do it during your lunch break or on your commute home.

This doesn’t replace doing the actual work—if you want to become a data scientist, you’ll still need to write code and build projects, but it dramatically lowers the barrier to exploration. You can figure out what interests you before investing serious time and money.

Type Dos: Working Smarter to Reclaim Your Time

Now let’s talk about your current job — the one you’re already dedicating 40 hours a week to. How can you leverage AI to get your work done faster and better, freeing up time for the things that actually matter?

The honest truth is that most knowledge work involves repetitive tasks that don’t require genius-level thinking: writing emails, summarizing meeting notes, drafting reports, formatting presentations, analyzing data in predictable ways. These tasks eat up hours of your week, and they’re exactly the kind of work that LLMs excel at.

Here’s a practical example: Let’s say you need to write a project update email to your team. Normally, you’d spend 15 to 20 minutes crafting it, making sure the tone is right, the information is clear, and nothing important is missed. Instead, you can paste your rough notes into Claude and say: “Turn this into a professional project update email. Keep it concise, highlight the three main accomplishments, and flag the two blockers we’re facing.” Boom — done in less than 90 seconds. You review it, tweak a sentence or two, and send it.

Or let’s say you’re analyzing sales data and need to create a summary for your manager. Instead of spending an hour in Excel building charts and writing commentary, you can upload your data to Claude, explain what insights you’re looking for, and get a first draft of the analysis in minutes. You still need to verify the numbers, perhaps dig deeper yourself, and add your own interpretation, but you’ve cut the grunt work significantly.

The key insight here is that AI doesn’t replace your judgment — it accelerates the mechanical parts of your work so you can focus on what requires actual human thinking. You’re not outsourcing strategy; you’re outsourcing tedious execution.

And here’s the beautiful part: if you save 30 minutes a day through these optimizations, that’s 2.5 hours per week, which is over 10 hours per month3. That’s time you can use to learn that new skill from Type Uno, spend with your family, work on a side project, or just relax without guilt.

The tools are here. The question is whether you’re going to use them.

On a side note, there are still companies talking about privacy issues and how employees cannot use outside GenAI resources. By this point in 2026, most companies are working through their bureaucratic hurdles to bring these tools into the hands of their employees, and frankly they’re realizing the benefits. If your company isn’t? Perhaps that’s a signal it’s time to brush up that resume and LinkedIn page of yours and start applying elsewhere.

Type Tres: Small Changes That Add Up

Not everyone is looking to change careers or revolutionize their workflow. Maybe you’re generally content with where you are, but you still want to grow in small, manageable ways. This is where AI can be incredibly powerful for low-stakes exploration.

Think about all those things you’ve been “meaning to try” but never quite get around to: learning Spanish, playing board games, understanding basic investing, reading more books, picking up photography. These aren’t career changes—they’re small enrichments to your life that you’d enjoy if you could just get started.

The problem is that starting is hard. You don’t know where to begin, you’re worried about looking stupid, and committing to a class or course feels like too much pressure or money. This is where treating AI as a casual learning companion changes everything.

Want to learn Spanish? Don’t sign up for Duolingo yet. Just open Claude and say: “Teach me 10 useful Spanish phrases for traveling. Then quiz me on them.” Five minutes later, you’ve learned something. Tomorrow, ask for 10 more. Next week, ask it to explain basic verb conjugation. There’s no pressure, no grades, no feeling like you’re failing if you skip a day.

Curious about games like backgammon or checkers? Ask Claude “I want to play backgammon with you” and watch it create an interactive game for you. You can actually play while asking questions and getting guidance. Claude explains why certain moves are good or bad. You’re learning at your own pace, on your own time, with a tutor who never gets impatient.

The same applies to puzzles, cooking techniques, personal finance, meditation practices—anything where the barrier to entry is simply not knowing how to start. AI lowers that barrier to almost zero.

And here’s the thing: most of these explorations won’t turn into anything serious. You’ll try Spanish for a week and realize you’re not that interested. Fine. You’ll play around with your backgammon games for a month and move on to something else, also fine. But one or two of these experiments will stick, and suddenly you’ve added something genuinely enriching to your life without the pressure of a major commitment.

The best part? If something does catch your interest, you can lean in harder. That casual Spanish practice turns into actual lessons. That backgammon curiosity becomes a daily habit. You’ve found a new direction not by making a big scary decision upfront, but by dipping your toes in the water until something felt right.

The Bigger Picture

Look, I’m not going to pretend that AI is going to solve all your problems or make you love your job if you genuinely hate it. But what I am saying is that we have tools now that can genuinely improve the day-to-day experience of being alive—if we use them intentionally.

Your routines are going to happen whether you optimize them or not. You’re going to spend those 121 days per year at work regardless. The question is: are you going to spend them feeling stuck, overwhelmed, and unfulfilled? Or are you going to use the tools at your disposal to learn faster, work smarter, and explore more?

The technology is here. It’s free or cheap. It’s accessible from your phone or laptop. The only thing missing is your decision to actually use it.

So what routine are you going to improve first?

Today, I want to explore how we as humans interact with the tools that supposedly make our lives better. From the Industrial Revolution to post-war marketing campaigns, we’ve been promised countless innovations that would enhance our existence. Whether these tools actually provide meaningful marginal utility remains debatable, but that’s not my focus here.

I want to write about AI and what it promises us. In many ways, this is just another tool with great potential and some peril—great peril? Honestly, I don’t know. Anyone claiming they know exactly where this is headed is probably just looking for another subscription.

In my day-to-day experience, even before AI, we became accustomed to searching for information—but searching never eliminated the need to think. We could find information more efficiently, but we still had to synthesize it, form our own opinions, draw our own conclusions. That ability to think for ourselves seems increasingly at risk with the advent of agentic systems and their crisp, polished outputs. That’s really what this is about: we’re lazy, and finally there’s a tool that lives up to its promises. Not AGI—I mean the practical ability to edit a document instantly, get summaries formatted seventeen different ways without lifting a finger. It’s seamless, it’s more than enough, and it’s beyond what we could have imagined three years ago.

What seemed impossible now seems inevitable. You can use AI to genuinely improve your life in ways that were once out of reach. But remember: it’s just another tool in your toolbox. It’s not your companion. It’s not your girlfriend, and it’s not you. It’s like your computer, your car, your mailbox key. The difference is its ability to mimic natural language and communicate whatever you want, however you want it. (I have to stop myself here—even I’m anthropomorphizing this tool.)

What I want to explore in my writing are the ways I use these tools—Claude’s chat interface, Cursor, Perplexity—and how they help me accomplish things I couldn’t before. They’re not replacing my thinking or how I live my life. They’re newer, shinier tools I leverage to expand my capabilities. Through this Substack, I’ll share ways you can learn more, build more, read more, write more. I want to show how not just companies, but individuals can improve their lives with these tools.

One last thing: using these tools effectively requires understanding how they function. Not the technical details of logits in a deep learning model, but rather the probabilistic behavior these models were designed to exhibit. Understanding that will make all the difference.

Happy reading!