85 Things I've Learned About AI
In 1,000 days of ChatGPT
Experts are all figuring out this technology as they go, so their claims are more biased by preference, ideology, or group membership than by truth.
AI is “spectacular when it works but regularly fails,” as independent researcher Colin Fraser says; it should be “boring and reliably works every time” instead.
The price for AI products won’t come down forever, so we’re living through a temporary golden window of affordability.
Not even the best researchers know their AI models well enough to make claims about whether they can do X or can’t do Y with 100% certainty.
Your relationship with AI will change more than you think, even if you actively resist it; the customs and social currents shift beneath our feet.
Taste, imagination, and agency > talent, intelligence, and ability.
It is the chronically curious and the neurodivergent who find AI’s deepest secrets.
Interpretability research has the highest interestingness-profitability ratio; it is fascinating, but it doesn’t pay off financially.
Those with too little at stake deny AI’s utility.
Those with too much at stake deny AI’s flaws.
AI writing is everywhere; you only notice the noticeable bits—you can only find human writing in the places you already know you can find human writing.
The new status symbol is AI-free sites and AI-free content (although it’s even cooler to be AI-free and not have to say it because no one doubts it).
Context—human or historical—is the last bulwark against AI; its outputs are “all necklace but no neck,” as psychologist Adam Mastroianni puts it.
Some write for the AIs, others write against them; I have no idea what’s better, but I’d say principled indifference wins both.
AI’s frontier of ability is jagged in an unfamiliar way; what it cannot do is just as fascinating as what it can.
Emotional education will always be more valuable than AI education.
There’s no edge in having used AI since day one, except the habit: start now.
ChatGPT can be your best friend for a day, not for life.
Scott Alexander’s maxim in the AI age: “Beware the man of one [AI] study.”
Google DeepMind keeps winning in the technoscientific domain, but Sam Altman keeps winning the popularity contests.
AI companies provide the best use practices; never trust an AI influencer selling you a course or “the best 1000 prompts.”
China has caught up in both innovation and adoption (more so if the US pushes out Chinese elite AI researchers, as Secretary of State Marco Rubio wants).
Em dashes are generally accepted to be the most suspicious sign of AI, but if you pattern-match any em dash to AI writing, you’re dumber than ChatGPT.
Knowing a lot of AI ≠ being rich overnight.
Whatever belief you have about AI, there’s plenty of confirmation for it; be more wary of claims that match what you think than otherwise.
Every writer whose process I’m familiar with uses AI to some degree (e.g., Grammarly); where do you draw the line of what’s acceptable?
Custom instructions in ChatGPT are great if you don’t forget about them; otherwise, the default model is ok.
As Princeton professor Arvind Narayanan argues, you’re better off treating AI as a “normal” technology rather than as a herald of either apocalypse or utopia.
You’re never the target of hype; shareholders are.
I suspect a superior intelligence wouldn’t chase relentless progress but inner peace; we’re just not smart enough to fix the source of our restlessness.
Serious testing of AI tools has a negligible cost, yet huge potential upside; you should familiarize yourself with the technology, even if only to criticize it.
AI’s flaws are inscrutable and becoming more, not less so.
When the bottleneck to useful superintelligence stops being intelligence, it will move elsewhere (i.e., the natural frictions of society and bureaucracy).
In times of turmoil and controversy, listen to the quiet ones if you want the truth, not those trying to grab your attention actively.
You don’t need to have an opinion on AI.
In AI, everything (even the name) is and isn’t marketing at the same time.
AI requires regulation, but not at the expense of innovation; AI requires innovation, but not at the expense of human well-being.
Extreme over-hypers and anti-hypers are often cut from the same cloth.
Automation can’t lift the human toll of working terrible jobs fast enough, yet it rushes to snatch the good ones.
Skeptics are as crucial for progress as they are misunderstood, for they exert a balancing force over the often reckless optimists.
Prompt engineering will probably never die completely, insofar as the UI of AI relies on interacting with another entity.
Some people predicted the present 20 years ago; you should follow them.
Find the right mix of AI sources for what you care about and dismiss the rest; treat the news as a river, not a bucket.
AI keeps conquering territory, but we’ll always have the kisses.
To know if an AI tool works, don’t read news headlines; try it yourself.
No one knows how AI works for the same reasons no one knows how the human brain works: their complexity surpasses even the brightest of minds.
It’s never too late to learn AI; when I started 10 years ago, I thought I was late.
There are no AI masterpieces because no one has dared go far enough in exploring the possibilities of this technology.
AI skills are the new “soft skills”; the old soft skills are a must.
Only those who own the relationship with their clients will thrive in an era where anyone can go to ChatGPT and try to compete with you.
People want to hate slop, but we absolutely love it.
Society should trust its students, and they should be responsible; if it doesn’t work, the problem is of incentives, not technology.
AI apps should have a higher minimum age; AI apps that can flirt or engage in erotic roleplaying should have an even higher minimum age.
Blogger Gwern Branwen: “Sampling can prove the presence of knowledge but not its absence.”
ChatGPT is the largest social experiment in human history, and it’s revealing the uglier sides of both technology and humanity.
Keep reading with a 7-day free trial
Subscribe to The Algorithmic Bridge to keep reading this post and get 7 days of free access to the full post archives.