I’ve seen a lot of GPTs that do interesting things with image recognition and DALL•E, like analyzing an image to understand the scene and then recreating it from scratch as a painting.
But I wanted to do something different, like, “what if we could guess the stories behind the pictures?”. Obviously it’s for laughs, so this GPT is called “How We Got Here” and it theorizes a possible chain of events that took place before the picture to result in what you see. It works for any given scene or an object, but because of OpenAI’s restrictions it will refuse to comment on images of famous people.
Here’s an example:
A parliament session started with the introduction of a controversial bill: The Mandatory Karaoke Performance Act.
The opposition leader, known for his baritone renditions of “My Way”, vehemently opposes the bill, fearing unfair advantages for tenors.
During the debate, an amendment is proposed: All speeches must be sung to the tune of popular folk songs.
The finance minister, a secret shower-singer, is thrilled and immediately begins to belt out budget allocations in A-minor.
The amendment passes by one vote, thanks to a swing voter who misunderstood the motion, thinking it was about free concert tickets.
The session escalates as a counter-amendment is suggested: All sung speeches must be auto-tuned for fairness.
Chaos ensues when the auto-tune machine malfunctions, making everyone sound like rogue robots.
In a bid to shut down the malfunctioning device, a scramble erupts, with one member accidentally broadcasting his vocal warm-up scales live on national television.
The image captures the exact moment when the tech-savvy intern is summoned to fix the auto-tuner, while the rest try to cover up the incident by looking deeply concerned about fiscal policies.
After making the sci-fi adventure Chrono Quest, I thought my next GPT game should be all the way in the other direction, so Cruising For Love is a bit of a rom-com dating sim set on a cruise.
You are on a five-day cruise to try and find romance. You should have a new experience each time you play: new destinations, new activities, new people to meet, and hopefully new breakfast items at the buffet restaurant. You don’t have to tell the game your gender or what kind of person you’re into, but it doesn’t hurt.
You can simply play it like a choose-your-own-adventure game and pick from the multiple choices given to you at the end of each turn, or take the keyboard and start providing more detail about where you’d like to take things. You can double-time/triple-time, play hard to get, take someone shopping for diamonds, reveal your secret magic skills, or try to seduce the captain. Nearly anything you can dream of, as long as it’s related to finding love.
You may or may not encounter some surprises along the way, making your successful pursuit of a love interest not exactly a given. So turn on the charm, put your leisure suit on, and start cruising for love!
Christmas is creeping closer, but the Goodreads Challenge angel won’t be darkening my doorstep as I’ve redeemed myself with two weeks to go! James Hogan’s Thrice Upon A Time was the twelfth book of my year, and definitely one of the better ones. It’s a 1980s time travel story where no time travel takes place, but it grapples with ideas about how timelines are rewritten, plus some other global topics that seem quite prescient when read today. Stylistically, it’s aged, but in that classic sci-fi way I love, which takes me back to reading books in the library after school. I think those hours, that precious access back then to a ton of books I couldn’t wait to read, were the part of going to school I looked forward to most. Anyway I’ve started a dumb new book that I should be finishing this year for bonus credit: The Paris Apartment by Lucy Foley.
If you’re looking for reading material, it may interest you to hear that I somehow managed to finish B’Fast, the AI-generated breakfast zine project I mentioned last week. The InstaZine GPT I made to create the content is also available through the same link (I updated the page with some additional usage tips today). Now that it’s done, I’m planning to make a companion breakfast-themed zine called “B’Fast (Brandon’s Version)” which will be made entirely by me, in a way that an AI presumably would not. But probably not straight away.
===
Earlier this year, Hipstamatic redesigned and relaunched their Hipstamatic X app. The “X” was dropped, and they added a new social feed. It was the official replacement for their original app which became Hipstamatic Classic. Where the original was funded by in-app purchases for new filters (at a pace of roughly one new 99c release each month), the new Hipstamatic charges a $30/year subscription, doubling their income from faithful fans.
I used the new app for some photos during my trip to Japan and mostly enjoyed the experience, but it was too buggy and the UI was still too cluttered and confusing (a longstanding problem with the original Hipstamatic app as well) for me to consider continuing with a paid subscription.
Their main problem is that there are now over 300 filters in the forms of “films” and “lenses” and “flashes” that you can combine to make infinite looks, and no good way to make the attractive human-curated combinations they recommend accessible and discoverable. In the last version, they tried to give a few of these combinations the tangible metaphor of being unique “cameras”, each one with a different skeuomorphic body you picked off a shelf, but essentially they were presets you could call up. But you can keep, what, ten of these aside in a little drawer before you couldn’t tell them apart? And so many of the other combinations were left out of sight and out of mind.
Now, after a week of teasing social media posts, wherein a “physical camera” was shown in videos — quite obviously a 3D model rendered in AR, but some people believed they were going to release a hardware product anyway — they’ve released a major update (v10) that tries to untangle the Gordian knot of their UX issues.
In this new version, they’ve tried to marry what worked in the original app with a new info architecture and set of metaphors to manage the library of looks they’ve accumulated over the last 14 years. You get just ONE skeuomorphic camera to call your own and customize the look of, and this camera is capable of loading up many presets. You can either let the camera detect the scene and choose a suitable preset for it (Auto mode), or specify the preset yourself (Manual mode). There are 9 possible scenes, such as Travel, People, and Still Life, but in a puzzling and unfortunate move, when you start using the app, each of these scenes has just one or two associated presets. That means you’re going to see the same looks over and over, when there are over a hundred more hidden away in a long list. This was presumably done to allow you, the user, to customize your experience and assign your favorite presets to the scenes.
There are two major problems at this point. One: leaving it up to the user to gain their own understanding of all the pre-existing “good combos” and assign them to 9 scene categories is insane. It’s a lot of work to hand off to a customer you hope will pay you money. The team should be doing the work of tagging each preset combo with a recommended use case, AND making it easy to assign them. It’s not currently easy. I had to move back and forth between two sections of the app looking at presets and memorizing their names to go assign to a scene, because these things aren’t placed together. Off the top of my head, it just needs an in-line list of suggested presets (from the aforementioned tagging exercise) on the same screen where you customize a scene’s presets. Perhaps this is coming. I’d argue it should have been in the MVP release of such a big redesign.
Two: as I mentioned, there are infinite possible presets given the number of ingredients they’ve accumulated. You can make your own combos, but there’s no great way to experiment and do this — there should be a sandbox where you can explore each lens/film/flash’s characteristics and try them out in real time to find a good combo. There used to be a section of the app called the “Hipstamatic Supply Catalog” where you could browse all these effects (it was only like a static magazine, but they could have made it interactive), and this now seems to be gone or I can’t find it anymore in the maze of menus and buttons. Perhaps they’re okay with most users just using the curated “good presets” and never making their own, but it seems like a missed opportunity.
I was feeling a mix of optimistic and bored, so I paid for an annual subscription anyway and will be trying to take lots of everyday silly snaps with this, and maybe even use it on my upcoming trip to Thailand. But if you know someone who works at Hipstamatic, please talk to them about taking on some external advice.
===
I finished watching Pluto on Netflix. It’s still a strong recommendation for me; a modern anime made with classic sensibilities and a story that really keeps you guessing. It’s also a very different Astro Boy story, suitable for people who hear “Astro Boy” and think it’s stuff for kids.
We started watching A Murder at the End of the World and I’m really liking it so far. Especially its star, Emma Corrin, who I’ve never seen in anything else before. They’ve got the most strikingly similar face to Jodie Foster, I was sure they were related.
New playlist! BLixTape #3 is done, made up of mostly new songs that I’ve been listening to since mid-October. Add it here on Apple Music.
I mentioned last week that I’d started making a zine about breakfast using AI tools to write all the text and make the pictures, leaving the final job of laying it out to my own (inexperienced) human hands.
Well, here it is. It was a lot of fun to put together, and I got a crash course in print design while trying to evolve it from a word processor document into something a little more creative and fun. You can sort of see the gradual improvement from beginning to end, as things become a little neater and more coherent — much like the mind in the morning as you make your way through breakfast.
I’d say the final content is 98% AI generated; I made some cuts and changed just a handful of specific phrases: in one place to avoid possible offense, and most of the others were in a poem at the end, to make it suck less.
I did it with the help of my custom GPT: InstaZine, which I encourage you to check out if you want to do something like this. It will brainstorm a bunch of different article ideas for your approval, then write them in a range of different authorial voices, hopefully giving the final product more diversity and interest than if you just did it the normal way with ChatGPT.
You can read the zine below in the embedded Issuu viewer, or download the PDF. Let me know how you like it!
Some additional tips for using InstaZine: Start by telling it what the subject of the zine should be, and describe what it should be like if you know. Ask it to create a list of article ideas — it should suggest a bunch of titles with a short abstract of each one, usually with a fictional author’s name and some description of the style and approach it will take. If you’re happy with this list, you should copy it off to the side in your own notes and play these descriptions back to the GPT before asking it to write each one. The reason is… the context window still isn’t great with ChatGPT as of Dec 2023, and it won’t necessarily do a great job if you don’t remind it of the brief after 20 messages of writing other articles and generating the images for them. By default, it creates colorful minimalist illustrations suited for a certain kind of magazine I like, but you should override this with whatever style you want.
As usual, I find myself in disbelief that another year is nearing its end. My Goodreads Reading Challenge count stands at 11 out of 12, and I’m halfway through a book right now, so I guess I’ll just make it before New Year’s. Which, incidentally, I’ll now be spending overseas thanks to some last minute plans. I’ll say where and post some photos after I’m back.
On reflection, it’s a bit of a shame that almost all the books I’ve read this year were just 3-star affairs. It’s like I’ve held back from tackling the big names on my reading list, choosing lighter and more inconsequential fare. In some ways, this has been a calmly chaotic year, with instability in the wider world putting everyone on edge, and that may have influenced my need for soothing, low-stakes entertainment. I saw a mention somewhere that the self-care industry is “sedating women”, making them focus on trying to fix something in themselves instead of fixing the problems out there. I can relate.
The holiday overeating has begun (although I may have forgotten to stop after last year), which I think is linked to a feeling of letting go and treating yourself in the evenings as work slows down (or seems less important) at this time of the year. We ended up eating out a fair few times, and as I write this I’m looking forward to another trip to Maji Curry this evening.
It’s not just fat cushioning my bones — while at Tokyu Hands this week (now simply called Hands), I saw a $75 wavy seat cushion and decided I had to have it for all the sitting around I do when working from home. Does it do anything for me? I don’t really know! But I’m treating myself. And then on the weekend we wandered into some kind of fancy organic bedding store and walked out with a pair of new pillows. Kim unfortunately may have chosen the wrong height/density for her sleeping style, but after one night I can cautiously report that mine cradles my noggin just fine.
===
Where’s the usual AI garbage, Brandon? I can hear you thinking it! Well okay, so Peishan mentioned she’d made a new zine, which reminded me of a project idea I’d written down and filed away. It was to make a zine on the subject of “Breakfast”, but using only AI-generated words and images.
If you’re thinking that sounds like a pretty mediocre zine, then you understand the challenge here. We’re now at a point where generative AI’s infinite supply threatens to drive down the perceived value of all but the best; content vs. art. So I’d like to see if my human labor of directing an AI worker to deliver above-average quality and packaging its output as a coherent product, can create something worth looking at. The only way to find out is to make it! And now that we can do custom GPTs, I decided to start by making one that acts like a diverse team of writers and artists, with a range of different styles, which can then be applied to a zine on any topic you like.
I’m still testing it out, but so far I’ve gotten a handful of articles. And in doing so, I’ve realized that I know nearly nothing about print layouts or how to design an attractive zine. I’ve read my share of mags, of course, but without effectively taking in their details. I’m making it with Pages on my Mac, and using its “Free Layout Mode” has been the best approach I’ve found. It’s sort of like a digital version of making a physical zine: I’m moving chunks of text and cut-out imagery around on A4 canvases; almost like scrapbooking. I just need more fonts and more imagination and more time.
===
I listened to no new music this week.
I didn’t turn my Switch on once.
I haven’t seen any films.
We did start Season 2 of Bosch Legacy though, and that’s still as great as ever. Not just the modern noir vibes and great jazz soundtrack. It’s a show that respects its audience and their time, without overelaborating on plot points or explaining every term or acronym that comes up. We’re already on episode 7 of 10, and I’ll be sad when it’s over. Thankfully a third season has been confirmed!
(This week’s featured image was taken while walking around Tiong Bahru, edited with a Ricoh GR Positive Film effect simulation preset I made.)
Why do people never rinse their mouths out after brushing their teeth in the movies?
Isn’t it misleading and bad for oral hygiene education if directors leave it out for pacing reasons?
What’s the recipe for a Vesper martini?
How might a wealthy Indonesian put their billions of IDR to work beyond investing? (asking for a friend, I swear)
Aside: ‘Asking the internet’ used to be our go-to phrase, but in an era where AI might be the one answering, does the term need revising? We used to be able to say ‘asking the internet’ but what about when you’re really asking an AI? They live on the internet and were certainly trained on internet content, but the old definition meant looking up content and new replies created by people; what do we call it when the answers AI generated? Keeping in mind that these answers may well be wrong, and in ways different from how a human might be wrong, it doesn’t feel like we should use the same terms. Or maybe we’ll keep referring to any hive mind as the internet?
Can I go a week without talking about AI? I think those days may be behind us.
Even iA Writer, the Markdown text editor I use for these updates, released a new update with an AI-related feature. No, it’s not automatically finishing your sentences or summarizing your essays. They’re all about the writing experience and process, and so they’re embracing how people use ChatGPT as a writing assistant, but helping them to preserve their own authentic voices. Text pasted from ChatGPT can be visually differentiated from text you wrote yourself, so you can see the Frankensteinian stitches on your monster. It also saves this info in the metadata for provenance.
This suggests that many users send text back and forth to ChatGPT so often that they end up forgetting which bits they wrote themselves, which isn’t a problem I’ve had so far, but going forward, who knows? It’s a good idea and one I’m glad they’re testing, but iA Writer has always been a niche tool for a certain kind of user. I thnk word processors with integrated AI are going to be so widely used and loved by the end of 2024 that no one will care about who did what part.
===
I made and released a new GPT, a game called Chrono Quest where you go back in time to improve humanity’s chances of beating an alien invasion. You can read more about it in my post here, but there are many ways to succeed, limited only by your imagination and problem solving inclinations. As a kid playing text adventure games, I never thought I’d see the day they could write themselves as you played. It even creates illustrations along the way, although those aren’t strictly necessary.
I’ve got some other game ideas I’ll probably get on soon over the Christmas break. OpenAI announced yesterday that the GPT “App Store” meant to open in November was being delayed until early 2024. I guess that gives me more time to learn.
===
It’s not accurate to say I didn’t get anything during the Black Friday/Cyber Monday sales. Pixelmator Pro for Mac was 50% off, and I thought it was time I upgraded to it from regular ol’ Pixelmator, which I must have bought over ten years ago for personal use as a substitute for Photoshop. Those were the days of per-once, use-forever software. To the developers’ credit, Pixelmator Pro is still offered through that model, although their newest app, Photomator, prefers a subscription pricing plan. The latter just won Apple’s Mac App of the Year award, by the way.
Pixelmator Pro is more than just a Photoshop-type editor now, it’s also a video and vector image editor, and comes with lots of templates for creating posters, logos, and so on. And like Photomator, it has useful ML-based features for correcting color, removing noise, and increasingly resolution of photos. These are pretty old-school and conservative by the standards of generative AI — see the recent development of Magnific AI, a tool confusingly billed as “upscaling” when it’s really closer to hallucination. It can subjectively improve the quality of photos by generating plausible (but inaccurate) pixels.
Check out this “upscaling” of Tomb Raider 1.
⚡ OK, this is impressive 🤯
I also wanted to try upscaling Lara Croft from PS1 to super high resolution! Mini step by step tutorial 👇
Stumbling into the New York jazz scene by accident, I found two jazz artists I’d like to recommend: Brandee Younger and Samara Joy. Both already have a couple of albums out.
Younger is a harpist who blends genres and leans modern. You’ll hear some hip-hop production, and it’s really not what you think when you hear the word “harp”. Her new album is Brand New Life, and is apparently based on and inspired the work of legendary harpist Dorothy Ashby, who I was also ignorant of. This is a weird observation, but hear me out. The opening track, the previously unrecorded, Ashby-written piece You’re A Girl For One Man Only, has a haunting melodic fragment that I think I recognize from… the soundtrack of the Japanese game/anime Steins;Gate of all things?!
Samara Joy is a much more traditional vocalist, but what an incredible instrument her voice is. I’ll leave a video of her covering Lush Life below and you’ll see what I mean. I’m about to put her Christmas EP, A Joyful Holiday, on and get some lunch. See y’all later.
(This week’s featured image was taken at a new mall that’s sprung up in Holland Village this weekend. It’s disconcerting; the massive development has been hidden behind hoardings for the past few years, and now revealed, it’s an unexpected contrast to the other old buildings and shopfronts, like a bionic arm of mediocre high street brands slapped onto an aging body.)
A thrilling adventure through time to give humanity a fighting chance!
Chrono Quest is an interactive fiction game, like the Choose-Your-Own-Adventure stories of old. There are occasional graphics generated by DALL•E, just like how those books would have illustrations every now and then. Unlike those books, it’s a GPT that writes a different story each time, one that responds to your inputs and imagination. You can stick to the provided multiple choice options, or respond freely with your own ideas.
The set up: An alien invasion threatens earth, and recently discovered time travel technology is our only hope. You must go back in time to change history and prepare humanity to meet this challenge.
Things I’ve tried (mild spoilers): I’ve done some ridiculous things in this game. I’ve gone back to find Jesus and recorded a video message of him appealing to people of the future to work together against the invaders. I prevented Hitler from taking power by stealing his playbook and being an even bigger Hitler. I imbued primitive cavemen with modern human DNA, to accelerate our evolution (and found myself the dumbest and shortest man in the world upon my return to the future). I also stopped the Fall of Constantinople with Semtex plastic explosives, prevented the Library of Alexandria from burning by building a firebreak , and became Leonardo da Vinci’s best mate after stumbling into his workshop with a fake stab wound. My colleague Brian played it and beat the aliens by uniting the world under a single market economy and rewriting property laws.
===
This is the first game I’ve made since the launch of custom GPTs a few weeks ago. I played a bunch of text adventures as a kid, and just with books versus movies, they can be more immersive and fun than AAA games made for millions of dollars.
AI Dungeon blew my mind when it came out a few years ago, unstable and liable to forget or misunderstand context as it was, it fulfilled many childhood dreams by being a flexible “dungeon master” that could take a story almost anywhere you asked it to. Wanted to pull out a gun in a medieval story, or use diplomatic words to get out of a situation? It was up to you. That was three years ago and built on GPT-2. It’s now three years later, and anyone can make their own custom story/game powered by GPT-4 — with virtually every shortcoming of AI Dungeon solved. Stories are incredibly coherent, natural, and well written.
You can use normal ChatGPT to play interactive fiction games, and the story of Chrono Quest was one of the first I experimented with earlier this year. You can have a pretty good time even with GPT-3.5! But the advantage of custom GPTs is that you can craft a game world and share it with other people to play, specifying a certain style, and keeping an element of surprise when it comes to how the game plays out and interacts with the player.
I might make more, I don’t know! Your thoughts are welcome.
Oof, this was a tiring week. Navigating change, physically recovering from a vaccine’s side effects, and having conversations about Christmas gifts, that shit will wear a blogger out. Helping me navigate all this was my astrology GPT, Co—Sign, who would tell me how to deal with challenging situations and to trust in my own nature, even if some decisions felt like bad ideas by other people’s standards.
Of course, the week started off with more twists in the OpenAI management saga, which seems to have ended now with Sam and Greg back at the company, not at Microsoft, and the formation of a new board underway. This pleases me because OpenAI was a company/brand with significant momentum and destroying value of that sort for no good reason doesn’t sit well with me (this may or may not be a reference to something else happening in my life). I’m glad I won’t have to port my collection of GPTs to another product like Poe or whatever.
I used Gen AI to bone up on Silicon Valley mind viruses like EA (effective altruism) and e/acc (effective accelerationism), which I’d kinda grokked in passing but not spent any time specifically reading up on. I have to say I feel myself falling in the accelerationist camp — given current world events, we’re not exactly proving ourselves a worthy species, so we may as well hasten our demise or salvation in every aspect from the upheaval of labor to economic principles. Incidentally bitcoin hit a new high for the year this week, going over $38,000 for a brief period.
I’ve buried the lede, but AI voice recognition and synthesis technology has enabled my favorite gag of the year: this Chinese dubbing of Van Morrison performing Caravan with The Band. It’s sublime; a French chef’s French kiss. It caught me in a moment of weakness and I couldn’t watch the whole thing through because I was dying of laughter. Van shouting “turn your radio up” in Mandarin will live rent free in my head from this week forth. This video is so precious to me I’ve saved a copy on personal cloud storage just in case the tweet goes down.
===
I was excited to hear that the Muji store in Plaza Singapura had reopened after an extensive renovation, now twice its original size and the largest in South-East Asia. We went down on the weekend to take a look, and it had things normally seen only in Japan: plants, bicycles, an embroidery service, renovation services, a wider range of furnishings, frozen food, and regional specialty goods (including $350 silk scarves). Not your average Muji! In fact, it’s billed as a “Global Flagship Store”, and I hope their gamble pays off and Singaporeans vote for more of this with their wallets.
===
Up to episode 4 of Pluto on Netflix now and it’s really the best anime series of the year for me. Just on those late 90s cybernoir SF mystery vibes. This is what the Ghost in the Shell Standalone Complex reboots should have been like.
Also up to episode 4 of Blue Eye Samurai and I’ll admit it’s gotten better. More complexity, somewhat interesting stakes, but the CG anime look with fake low-fps stuttering is getting a little annoying. If I had a modern TV I might actually turn on motion smoothing just to make it a more authentic experience.
I tried real hard to avoid buying a 4K HDR TV during the Black Friday sales and succeeded. Gotta save up funds for the dark days ahead.
Speaking of premonitions, we saw episode 2 of Lessons in Chemistry on Apple TV+ (mild spoiler alert) and let me tell you, my cinematic Spidey sense tingled and I called the event that happens at the end of the episode minutes before it happened. I’m usually quite bad at anticipating TV twists, but something about the atmosphere and pacing and shots just told me what was coming.
We also saw the first two episodes of Monarch: Legacy of Monsters on Apple TV+ and were left impressed. It’s the new series involving Godzilla and other monsters that either belong to Toho or Warner Bros now? Really, the monster custody is unclear to me. But halfway through, as new characters were being introduced in flashbacks, Kim assumed an Asian woman was the mother of a younger Asian woman we’d already seen. And I said, “maybe it’s just an unrelated Asian woman”, to which she laughed, “there are never unrelated Asians”. Which I took to mean in American shows, not the real world, and sadly she’s right.
Many people have been raving about Andre 3000’s new instrumental album with flutes, New Blue Sun, and I’ve tried to get into it, I have. But it’s literally put me to sleep a couple of times, so I’m assuming it’s just not for me.
(This week’s Featured Image is a Dall•E representation of a Chinese Van Morrison impersonator performing in a Muji store.)
sangsara.net 