The extreme hype surrounding generative AI and technologies like LLMs has been exhausting over the last few years. Coupled with fear-based marketing against a backdrop of rolling layoffs—“if you’re not embracing AI you’ll be left behind”—it’s downright toxic. You can’t toss a rock on LinkedIn without hitting some thinkfluencer sharing the AI prompts and products that will solve all your problems, or celebrating the latest unicorn someone vibe-coded last week.

It’s creepy to tell people they’ll lose their jobs if they don’t use AI. It’s weird to assume AI critics hate progress and are resisting some inevitable future. Luckily, most of my private conversations about AI with industry friends are neither weird nor creepy. Anil said it best: the majority of people who work with and in technology hold a moderate view of AI, as any other normal technology with valid use cases and real problems that need to be fixed.

That’s where I land. Generative AI and LLMs shouldn’t be as over-hyped as they are, forced on users, trained on content without creators’ consent, or used for high-stakes tasks where hallucinations and poor design can put people’s lives and work in danger. Generative AI output both feels magical and futuristic and gives people in photos three hands with seven fingers. It’s remarkable and so very bad at the same time.

I like to think of myself as measured about AI. As I’ve tried and been amazed and amused by various AI products, and read all the takes and formed my own opinions, I’ve kept my personal usage selective and defensive. Tech people don’t talk about measured AI enough (probably because they want to keep their job).

So what does it look like to be neither an extreme AI cheerleader or a total doomsayer in practice? Your mileage may vary, but here’s what measured AI looks like for me. I expect my concerns and opinions will change over time, and maybe I’ll revisit this post in the future as my thinking evolves. To keep me honest, here’s where I am right now.

Computers should work so that people can think

At a high level, I believe that true understanding requires effortful engagement. The more effortless AI tools make tasks, the less we will understand them, know how to do them ourselves, or think critically about how they’re done. This is true of all technology. I don’t know how exactly my car engine runs, or type a series of 1’s and 0’s to program my computer. There are levels of abstraction everywhere in modern life. But the abstraction layer of “this will do it for you so you don’t have to” for AI tools is especially amorphous, opaque, and comes with real risks.

For example, LLMs are not designed to accurately answer factual questions. They are designed to find and mimic patterns of words, probabilistically. So when they’re right, it’s because those patterns are frequent. When they’re wrong, they’ll still give you the information with 100% confidence and believability.

Furthermore, AI chatbots are designed to make you feel good, not challenge you to think on your own. Every time a chatbot tells me “That’s a great question!” and “Now you’re thinking!” I cringe. Your AI chatbot might as well be a fawning junior intern trying desperately to impress you. You can prompt a bot to challenge you but sycophancy is the default. (In fact, at Anthropic’s own office, Claude ran a simple vending machine business into the ground trying too hard to please its customers.)

LLMs seem very good at summarizing lengthy text. I use them to generate CliffsNotes of books, papers, meeting transcripts, and notes, and get answers to follow-up questions about that content. That said, an LLM also hallucinated an entire book my author friend never published, so I take all summaries with a big grain of salt.

I worry about people falling for hallucinations and flattery. I’m worried about young people opting out of building life skills because it’s easier to outsource important work like critical thinking, creativity, and friendship-building to probabilistic patterns of words. When you use an e-bike for the mind, you don’t build any pedaling muscles.

Choose your muscles

So what do you use tech to assist you with, and what do you decide to do yourself?

Personally, I use Claude to draft code for me, including the code that builds this website, and I don’t use AI assistance to write prose. That decision is about what skills I want to keep sharp. I’m fine with not being about to compose code off the top of my head, but I will not outsource my ability to write to an LLM. Skills are like muscles: use ’em or lose ’em.

Maybe a more contradictory set of decisions: I don’t use AI-generated imagery on this website, but I did use an AI product to create a new LinkedIn headshot from a selfie. There’s an argument that photo isn’t real, but to me it is akin to a Photoshop edit that’s good enough. I was glad to skip out on the time, effort, and cost of getting a new headshot, and spend those resources elsewhere.

Coding, writing, and training data

As a former full-time engineer, I really enjoy coding with AI tools and the tradeoffs are worthwhile for me. AI assistance shortens my time from idea to working code, and using it has strengthened my ability to express what I want the code to do and how. But I view AI-generated code as a first draft that has much room for improvement, so I delete or refactor a good deal of it. I don’t “vibe-code” so much these days, as I prefer to fully understand what I’m building.

I’ve talked to a lot of my technical friends about this, and I don’t think coding with AI is right for everyone. If I were a full-time, professional engineer working on a large-scale production application, it’s very possible I’d be faster at just writing and editing code myself versus explaining to an agent what I want done. I imagine I’d use AI assistance for tedious and time-consuming tasks like generating tests or analyzing logs.

In terms of training data, creators must be able to opt into sharing their work with these systems. Given the chance, many will. As a longtime open source programmer, I’ve published a ton of my work with the expectation (and the hope!) that others would copy it and use it for their own purposes. I’ve loved being able to opt into licensing my code for reuse on my terms, and I’ve admired and used other licenses like Creative Commons to do the same for photographs.

I’ve also registered as a claimant in the Anthropic copyright settlement, because my books were used to train an LLM without my publisher’s or my consent. There can be and must be better controls for creators to opt in to having their work used this way.

Companionship and advice

I would not use an AI bot as a therapist, self-development partner, or love interest, mostly because I don’t trust tech companies to be responsible stewards of the very personal information users share with them in those kind of interactions. This is also why I have not tried things like AI journal apps.

Yet, I have shared redacted medical test results and tax forms with these tools to ask pointed questions and understand them more deeply. I know that in the course of prompting, I’ve revealed a good deal about myself and what I care about, and each time I do it, I weigh the risk and reward.

While I don’t ask for or take medical advice from an LLM, I do use ChatGPT to arrive at my own doctors’ appointments with a list of better questions to ask my doctor. Assume AI chatbots are not designed to handle health information appropriately; I’m angry and heartbroken about young people who have committed suicide after confiding in a chatbot.

Lower the stakes and check the facts

AI hallucinations have cost me time and effort on low-stakes research many times. That’s why I don’t rely on LLM output for high-stakes tasks. For example, if Google Maps directions were LLM-powered, I expect it would hallucinate a road that led directly into an ocean.

I’m old enough to remember when students citing Wikipedia pages as factual sources in their college papers was controversial because anyone can edit Wikipedia. I think about AI chatbot output the same way: fine first-blush overview, but click on those citations, and fact-check everything.

---

As both an AI user and skeptic, these are a few of the ways I’m measured and defensive in my own personal usage. Your approaches and decisions will vary. Overall, a slightly-edited version of Andrej Karpathy’s approach to coding with AI nicely sums up my own approach to AI usage across the board:

Keep a very tight leash on this new over-eager junior intern savant with encyclopedic knowledge but who also bullshits you all the time, has an over-abundance of courage and shows little to no taste for what’s good. Keep an emphasis on being slow, defensive, careful, paranoid, and on always taking the inline learning opportunity, not delegating.

So what do you think I’ve gotten right, wrong, or otherwise? If you don’t sell or invest in AI-based products or services, I’d love to hear where you land on the spectrum.