The Five AI Commandments

Never get high on your own supply*

In the age of immediate gratification, no one has the patience to sit through ten new rules. The void we feel grows, the anxiety that we are falling behind increases, as the aperture through which we might glimpse an antidote shrinks. So I’ll be snappy.

Before we start: you’re not alone. The most interesting social media exchange you missed between Christmas and the New Year was a moment of vulnerability shared on Twitter by two figures at the AI forefront: Andrej Karpathy and Boris Cherny. They’re both having a hard time keeping up with all the breakthroughs and often fall back into old, outdated customs.

I started working on AI twenty-five years ago. First, in philosophy of artificial intelligence. Then, in the technology industry. In the beginning of my career, I consulted for insurance giants, then taught epidemiologists how to ingest big data, built the first democratic accountability app in social media, and patented a machine learning application to help editors get more attention to important stories. I've presented research on tech ethics at the American Management Association, as well as academic conferences. Now I help leaders at foundations, startups, and behemoths (like Activision, Alphabet, Meta, Microsoft, and X) apply and grow adoption of new tech in its infancy. It’s hard for me to keep up, too. But I’ve learned from the best. And, over time, I’ve observed some things that continue to work, through the ages.

Below I suggest five habits for evading psychosis, avoiding epistemic trespass, and managing hallucinations. A convenient side-effect is that practicing them will make you an AI expert. Have fun, learn stuff, and transform your anxiety and depression into excitement and optimism. I’ll offer some places to start if you’re an AI beginner, as well as next steps for intermediate AI users and existing experts who want to step up their game. There’s levels to it.

QuickStart for smart, sexy, AI-curious people. Through it, you’ll get the best of AI, without getting too high off the product, contracting any viruses, or accidentally deleting your hard drive.

I. Survey Widely: Use multiple models and interfaces, i.e. be promiscuous

I recommend everyone start with Anthropic’s Claude, and add Google’s Gemini and/or OpenAI’s ChatGPT. Use them in Terminal, your browser, desktop apps, mobile apps, via plugins, extensions, and APIs, if you are more technical. I also really like Google AI Studio and Vertex AI. If you don’t want to immediately buy multiple token credits or subscriptions try You.com, Gumloop, or LMArena (especially in “side-by-side” mode). Regularly skim Artificial Analysis.

Why is it so important to survey widely? Well, cases of AI-induced psychosis are increasing because some models have a sycophancy problem and humans are susceptible to flattery. The best way to keep yourself grounded is to use multiple models.

II. Compare Outputs: Different models and interfaces have different strengths

Using multiple models and comparing the results will allow you to familiarize yourself with the strengths and weaknesses of different model families and sizes. Some of the observations will be obvious, e.g. smaller models tend to provide faster responses. However, models hosted locally or using GroqCloud for inference (regardless of size) can provide faster completions.

Collect your prompts and completions in Google Sheets. The habit of surveying several models and comparing the results will help you get better at writing prompts. Seriously. Your brain can’t help but think of ways to improve prompts when it regularly copy-pastas and analyzes them. Watch. You’ll like watching.

III. Write Your Version: Thou can use AI to start but musteth check all the facts

Your goal is to appreciate what AI systems are good at and their limitations. This is where the differences between the models and human quality become most apparent. It’s also wherein tool use, web search, and reasoning budgets are more impactful. None of this will make sense to exclusive users of ChatGPT (as opposed to the OpenAI SDK or Codex). Its default interface autoroutes queries to different models, preventing users from grasping the relationship between the model and response. However, if you’re tickled to try different tools and know how models fit into compound systems, the easiest path to understanding is through playing with Anthropic Console.

What compound generative AI systems are especially good at is producing summaries and fast drafts of (natural and programming) language artifacts. Practice writing your own version of the outputs that you’ve prompted the different models to make for you. Use their outputs to get yourself started. Notice how much faster you can move with the assistance. But keep in mind: even the best off-the-shelf AI systems still hallucinate >25 percent of the time. So before including any piece of an AI output in your final draft be sure to fact check and run code review.

You got this far. All by yourself? I’m impressed! (Does sycophancy work? I find it abrasive and the main reason I recommend other models over ChatGPT.)

The remaining commandments are more like frameworks.

IV. Don’t Use AI Like Search: Ask questions to which you know the answers

There’s a famous Silicon Valley story about a Microsoft executive who jumped ship after his young daughter made him realize a consumption paradigm had shifted. The child called his smartphone an “answering machine” because it had Google on it. And so it remains. Search engines are best for questions that we don’t know the answers to. Don’t use AI like search. Generative AI systems are better for producing creative prose and code in verifiable domains. They are more like “early draft machines” than answering machines.

The better use case for AI systems is questions about topics we are experts on. Within this framework, a human expert inspects AI outputs with two expectations. First, the expert will probably discover in the outputs something they didn’t already know and thereby learn something new in their area of existing expertise. Second, you’ll almost certainly find an AI generated fabrication. Your goal as an expert is to hunt down both, and thereby become an expert on AI within your domain of expertise, too. And if you’re an expert in a verifiable domain, check out Ralph.

V. Personas and Artifacts: Enduring concepts for evaluative foundations

Paradigms in AI shift rapidly. We moved from prompt engineering to multiagent orchestration to context engineering. At the moment powerful, general purpose agents (like Claude Code, Codex, or Gemini CLI), equipped with skills and guardrails, are the latest in best practice fashion. DIY system prompting isn’t as hot as it used to be. AgentOps has lost its thrust.

Over time it has helped to focus on concepts that endure. Personas and artifacts will endure. Think in terms of them. Nate B. Jones says jobs are “stacks of workflows.” I agree with him. At the end of the workflows are artifacts. Jobs are roles through which we animate personas, solve problems, and create artifacts. Like Ralph Wiggum. When evaluating outputs from different models and interfaces, think about them as being produced by personas. Use personas in system prompts for generating skills to pair with coding agents.

If that’s meaningless to you, don’t worry. Just register for Anthropic Academy, DeepLearning.ai, Google Cloud Skills Boost, or OpenAI Academy. Then read it again. In the meanwhile, identify steps within workflows in your current job where AI systems can be integrated to produce better artifacts faster. As you do this, keep those concepts in mind. Eventually, they will help you produce synthetic data for reinforcement fine tuning. Combine RFT with RAG to improve reliability of your base model in definable domains.

If that wasn’t meaningless to you, and you’re still hungry, dive into the Constitutional AI paper and Amanda Askell interviews. Think about how personas embody plural and conflicting values, which get expressed during alignment tuning. Then read again.

If you found this post helpful please share or provide feedback.

But never let them know your next move.

Here are two bonus precepts for the experts.

VI. Start Over, Early and Often: You can’t love these flows

Especially because context compaction is an open problem, but also because the autoregressive mechanics of transformers at the center of generative AI systems create huge KV-caches, and because once a conversation that’s been flowing makes a wrong turn it is hard to recover, it pays to run multiple concurrent conversations and start new threads often.

VII. Integrate Your Favorite Tools, Sparingly

Connecting a coding agent to Github, Linear, Notion, and/or Obsidian can be a real game changer. Just keep in mind that you don’t want to blow your context window. I recommend only integrating one (or a couple) at a time to start. I also don’t personally recommend rawdogging via “--dangerously-skip-permissions” or granting access to root directories. And avoid MCPs except when necessary.

Otherwise, go nuts.

Why is the title “The Five AI Commandments”?

• No one has time for ten

• I’m a hip hop head

• People in tech have a tendency to drink the Kool-Aid, i.e. a violation of Biggie’s fourth commandment, leaving us to begin with the fifth

• Much professional use of AI is covert, the act itself constitutes a transgression, so imposing “commandments” on them is ironic

• It’s a riff on the joke about how God gave Moses 15 commandments but he lost 5 on his way home from Mount Sinai