Short updates:
It took them a while to announce the lineup for the World Music Festival. I’m mainly looking at Ragamala (9/26-27), the annual overnight concert of Indian music at the Chicago Cultural Center, and the final concert at Navy Pier Beer Garden (10/5). A lot of the artists they book for this festival are pretty obscure, but I’ve listened to Rogê (Brazil) who’s playing at Navy Pier, and his last 2 albums are great: Curyman and Curyman II.
Open House Chicago is happening 10/18-19 this year, and they just announced the sites. One site that’ll be open to public is The Roof Crop, the rooftop farm above Maxwells Trading that I mentioned in my last post. Here’s my post from last year’s Open House, when I went to Ukrainian Village.
If you’re like me, you’ll get something out of this post by
: How to get your time back
Let me start with the principles that guide how I try to use AI1.
AI is the last word on nothing
Use AI to reduce drudgery and enhance joy
Make AI distill knowledge from the open internet
I’ll talk about what I mean and the reasons behind these, but I first want to talk about a recent success I had with Claude2.
Adventures in Mexican cooking
Continuing the home cooking theme from last time, I want to tell you about trying to learn a new cuisine.
I’m not one of those Japanese people who need to eat rice every day, but I do cook Japanese food quite often. The other cuisine I cook a lot is Italian—mostly but not exclusively pasta. These are the cuisines that I’m familiar enough with that I can whip up decent weeknight dinners without necessarily following strict recipes.
I’ve been trying on and off to add a third element to this mix. For everyday cooking, I want to have some idea of what the end result should taste like, so that suggests cuisines that I’ve sampled on many different occasions: Indian, Chinese, Mexican, Thai, Vietnamese, maybe a couple more.
When you’re cooking something that’s not generic “American” food, obtaining ingredients can be an issue. I have a couple of Thai cookbooks that I bought before going to Bangkok, but I haven’t cooked out of them too many times because the ingredient lists can be long and hard to collect.
In Chicago, there are so many people of Mexican heritage3 that obtaining Mexican ingredients is probably the easiest out of the list of cuisines above. I did attempt in the past to cook more Mexican, mostly cooking out of Rick Bayless’s Authentic Mexican. I’ve mostly liked the results, but the dishes could be more time-consuming than I wanted, and taco nights weren’t happening consistently.
One reason I’m telling you all this is that something like this is what I told Claude. I also told it what kinds of cookbooks I like—I named some Italian cookbooks that I regularly cook out of4, and also Salt, Fat, Acid, Heat and The Food Lab, which are books that help me understand what’s going on in the kitchen (and when I eat out).
Then I asked Claude to recommend me Mexican cookbooks using its research mode (which requires a paid subscription). Here’s the resulting document.
Claude’s top recommendation was Truly Mexican by Roberto Santibañez. This is far from an obscure book, with more than 300 reviews on Amazon, but it’s not a book I’d heard of before. In any case, I went to the library, and started reading it.
It was clear this was the right book for me, just from this quote in the intro:
To truly empower people, I knew I had to relegate a passion of mine—the history and culture of my home country—to the backseat. Fortunately, Mexican food already has some wonderful chroniclers—authors such as Rick Bayless, Susana Trilling, and Marilyn Tausend, have spent their careers showcasing the cuisine’s vast regional variation. My objective is different. Instead of emphasizing this diversity, my goal is to show you how much the seemingly disparate food actually has in common.
Santibañez takes a pretty unusual approach—he organizes the book by the sauces.
And this makes sense once you start cooking. Usually, sauces that you make end up being more than enough for one meal. So a book that teaches you the sauces, and what kinds of dishes would work well with a given sauce, is great for cooking multiple meals in a short time period.
Dried chili and adobo
I had a day off by myself last month, and spent most of the day biking to various places in Uptown5. For dinner, I wanted to make adobo, which requires some dried chili peppers. So I biked to Edgewater Produce6, which was recommended by Claude, and I scored some small bags of dried chili. They also had bigger bags, but I didn’t know how much I would like using these ingredients.
Making the adobo is pretty simple. You roast the dried peppers—a process Santibañez explains in detail because it’s central to building the flavor—then blend them with other ingredients like vinegar and spices. And then you can use it to marinate meats or as cooking liquid or as sauce. With my first batch of guajillo adobo, I marinated salmon and chicken, and braised pork shoulder in it.
That last one, cerdo en adobo, is easily my favorite Mexican dish to make at the moment. After the initial searing of the meat and the adobo, you can leave your pot in the oven for 1.5-2 hours, and you end up with fork-tender pork with mild chili flavor.
And then I slightly went overboard and ordered 3 lbs of dried chili online, from a vendor recommended by Claude.
Some of this might end up in mole if I’m feeling ambitious, but at the very least, I’ll be making a lot of adobo.
Back to AI
Now that you’ve read this whole episode, I can talk about how AI was used.
AI is the last word on nothing
You’ll notice that none of the final product is AI. I’m not asking Claude to draw a picture, write a novel, or even to have a conversation, really. I’m asking it to give me recommendations, and I’m using it to have real experiences, which I may or may not like.
It doesn’t always work as well as I described for Mexican food. I’ve asked Claude recommendations for cookbooks in a lot of other cuisines7. For Persian food, its top recommendation was Food of Life by Najmieh Batmanglij. It did look like a very good cookbook when I looked through it at the library, but it’s not what I got with Truly Mexican, where the book explained the cuisine in a way that made it easier for me to cook it more regularly.
Some of what Claude says about this book is also factually incorrect, like this:
The book explains the science behind Persian cooking—why certain rice varieties work better, how to achieve perfect tahdig (crispy rice bottom), and the precise meat-to-fat ratios that make koubideh kebabs stay together on skewers.
At this point, everyone should know that large language models “hallucinate”. It can produce bullshit that sounds plausible, because producing plausible sequences of words is what these models are mainly trained to do8. Turning on research mode does anchor Claude’s answers to what it finds on the internet, and I think it reduces the impact of these hallucinations, but it’s still a problem.
I guess this is where I should note how I’ve used Claude in the process of writing this Substack. I’ve used it to find mistakes, and gotten feedback on potential problems with things like the structure and narrative flow—I sometimes agree with the feedback and incorporate it. I’ve asked for words and phrases I’m trying to think of—“emblematic” from my last post, for one—but I’ve never had it write a sentence9.
Use AI to reduce drudgery and enhance joy
Now, what would’ve happened if I didn’t use AI? I may have looked for a Mexican cookbook with internet search as my main friend. It would’ve taken some trial and error to find a book that works for me—I’m almost certain that I would not have landed on a book as good as Truly Mexican on the first attempt. It may still work out in the end, and I’d find some Mexican dishes I’d cook regularly, after a few attempts.
The fact that I needed to do a lot of trial and error, though, would make me hesitant to even start—and this isn’t a hypothetical, because I was on pause for months after trying to cook out of the Rick Bayless book.
I knew Claude’s advice may not always be accurate, but I was willing to give it a shot because it was based on some very specific preferences of mine10. This greatly reduced the amount of time I spent figuring out which books to look at, and the amount of worry that I would’ve had about if I had the right books.
When computers started beating humans at chess, one thing people realized was that a human and a computer working together—a “centaur”—was actually stronger than the computer on its own. Since AI and humans have different strengths, you can do what you’re good at, and let AI do what it’s good at.
One huge weakness of AI as it exists is that it cannot have real experiences and enjoy things. That works out great for you because the whole point is for you to enjoy things and be happy. And one strength—if you can call it that—of AI is that it does not get bored or tired. You should let it do the tedious work that detracts from your experiences.
I do want to say, though, that understanding something new is actually a big human strength compared to AI, as it exists now. Yes, you can make Claude remember something you said, but it’s essentially holding that in its memory so that its answers are more consistent with it. There’s no permanent change to the model because of what you said.
Make AI distill knowledge from the open internet
What makes some people think AI understands everything is that it is extremely well-read, at least when it comes to stuff that’s available for free. It has read a significant fraction of the internet, which includes discussions on every topic that people find worthy of discussion. The amount of text Claude saw during its training is many, many orders of magnitude larger than what the most well-read person in human history managed. That’s why it makes sense to make Claude figure out what people are saying on the internet.
I’m almost certain Claude has never seen the text of Truly Mexican because it’s not available online for free. But it can read what people are saying about it online, and what people are saying online about all the other Mexican cookbooks, and that’s why it was able to make a good guess for which ones I might like.
Some more examples
Let me close it out with some examples of how you can use AI in a way that follows these principles.
Book recommendations more generally—You can provide the AI with information about books you’ve read and what you thought of them. Then you can ask which Mexican novels you might like, for example.
Product recommendations—One annoying thing that keeps happening as a parent is that we often have to figure out if we need to buy something for our kids, and if so, which exact product we should buy. There are websites like Babylist that give recommendations, but there are many of these websites, and not all the recommendations are relevant for our particular situation (we don’t have that much space at home, for example). Instead of going through a bunch of these recommendations, we can perform an initial filter by asking AI for a list of recommendations given our situation.
Foreign language conversation practice—In order to speak a new language, you do have to practice speaking that language at some point. Until very recently, you really needed to get someone who speaks that language and practice with them, whether with a teacher/tutor or with a friend. Being capable of reasonably natural conversation in many languages is one of the core features of AI chatbots, so it really makes sense to use this to practice speaking in a language you want to learn. You do want to make sure you aren’t just having nice conversation with the AI—you should be learning and practicing so that you can talk to a real human. So I think it does help to use specialized apps like Langua, which built language learning features on top of AI conversation.
What I’m listening to now
Here’s a live album recorded in Chicago: Live At Thalia Hall by LA LOM.
LA LOM is short for the Los Angeles League of Musicians, and yes, they are from LA. And they play instrumental Latin dance music with a surf rock guitar sound.
I liked their debut album from last year, and thought they had an interesting style that was a mix of polished and rough parts. The live album does make it sound like a fun show—I thought the same about Neal Francis’s Francis Comes Alive, also recorded at Thalia Hall, so maybe the venue is just great.
The best song is the opening track, “Alacrán”, with Middle Eastern elements. The guitar in “Lucia” is great fun.
The obligatory note here that what I mean by “AI” in this post is generative machine learning models, and products built around them.
I recommend Understanding AI as a relatively non-technical newsletter for understanding the latest developments in AI technology, and AI as Normal Technology for a more academic look at some of the potential problems with the technology.
Also, a few of my takes on “problems with AI”, since I don’t have a good place to put them:
Websites being flooded with scraping traffic is a bigger problem than the environmental impacts of AI.
AI slop flooding the internet—including falsely associating the slop with someone—is a bigger problem than AI violating someone’s real intellectual property.
“Humans using AI to do bad things” is a bigger problem than “AI taking over the world”—for now.
People using AI chatbots as replacement for human relationships and therapy is almost certainly bad.
Claude is the language model (or a class of language models) developed by Anthropic. I’ve defaulted to using Claude rather than OpenAI’s ChatGPT for a few reasons. For one, there are reasons to be concerned about OpenAI’s corporate governance and CEO Sam Altman. But the biggest reason is that I find ChatGPT responses to be very boring.
Then I found out that Claude’s “default personality” is like 60% on the way to being me.
Why does Claude love Caffè Strada and sometimes claim to have a Japanese wife? Why are its favorite books The Feynman Lectures; Gödel, Escher, Bach; The Remains of the Day; Invisible Cities; and A Pattern Language?
From this post.
Something that might surprise you if you’re not familiar with Chicago: when you look at the number of people born in Mexico, the Chicago metropolitan area ranks 2nd in the US, only behind LA, and just ahead of Houston, Dallas, and Riverside.
Marcella Hazen’s classic, Katie Parla’s Tasting Rome, and books from the Rome Sustainable Food Project, especially Verdure.
Among other things, I finally made it to Birrieria Zaragoza. It was great.
in Andersonville, confusingly.
This is probably the least defensible use of AI in this post. You can Yelp “Mexican grocery” and look at a few of the top results to be pretty sure that Edgewater Produce is the closest Mexican grocery shop to Uptown.
If you do end up making Claude do a lot of research like this, one recommendation is to create “projects”. You can add documents to a project, and Claude can use those documents as relevant context. All the background I wrote about what I usually cook and what cookbooks I like can be in these documents, and you can have multiple chats about slightly different topics, but with these documents as the starting point.
This isn’t just for convenience. Claude has a limit on the amount of context that it keeps track of during a chat, and it’ll end the chat when you go past it. But context stored in these project files are exempt from this limit. So with project files, your questions to Claude can be more nuanced and catered toward what you really need.
What’s surprising is that this works so well for certain tasks.
And yes, all the em dashes are mine, not AI’s. You would not like how many parentheses I would use if I didn’t have em dashes (and footnotes).
One important thing: you should sometimes make AI try to figure out things you already know. That’ll tell you how reliable it is on similar tasks.