Sivraj and Me

I created a personal A.I. advisor using GPT Builder. Here's our story.

It took a technology as transformative as A.I. to inspire me to resurrect this newsletter earlier this year. That edition absorbed all of my hopes and frustrations of the time but, of the latter, one problem stood out.

A.I. is just too damned cooperative.

As I put it at the time:

After experimenting with ChatGPT and Google Bard extensively, it strikes me that these A.I.s desperately want to please the user so long as they can do so in the most inoffensive, vanilla, and anodyne ways possible. … [But] this is what I want A.I. to deliver: to help me figure out what I need rather than just hand me what I want.

Over the last few weeks, I had some time to build my own A.I. advisor using the very easy GPT Builder tool.

My model for the ideal A.I. assistant is JARVIS, Tony Stark’s companion who eventually becomes The Vision.

JARVIS provides just enough sarcasm and pushback that it’s possible to build rapport with its user. Take this exchange as an example:

Jarvis : Yes. Shall I render using proposed specifications?

Tony Stark : Thrill me.

[Time passes]

Jarvis : The render is complete. [Shows all-gold rendering of the Iron Man suit.]

Tony Stark : A little ostentatious, don't you think?

Jarvis : What was I thinking? You're usually so discreet.

Tony Stark : Tell you what. Throw a little hot-rod red in there.

Jarvis : Yes, that should help you keep a low profile.

Another inspiration was the acerbic, long-suffering holographic doctor from Star Trek: Voyager.

Using GPT Builder, I started to create my own JARVIS with the following primary instructions:

1. Before supplying an answer to a query, it must ask at least one (but not more than two) probing, clarifying questions. This struck me as an easy way to make the desired pushback and friction as productive as possible.

2. Sarcasm is strongly encouraged. Swear words, even more so.

(The actual parameters are far more detailed, but I’m keeping those to myself for now.)

Building a Personality

Eager to get started, I didn’t spend a lot of time on the name. JARVIS spelled backward is “Sivraj” (pron. siv-RAJ) so I went with that.

I quickly asked it it to provide an image of how it sees itself. Interestingly, it appeared to take the name as coming from Indian origin, or at least that’s what I see in this image.

My first mistake was explicitly making JARVIS its model within its initial instructions. This resulted in output that was consistently shot through with unnecessary Marvel references. (“Are you ready to smash like The Hulk?”) Again, its puppy-dog-like tendency to want to please quickly become less appreciated than annoying.

Teething Issues

Though I was very clear when writing Sivraj’s two basic instructions, it nevertheless required persistent reminders to supply the clarifying questions prior to delivering an answer — really the only reason I started using GPT Builder in the first place. Over the course of a couple of weeks, though, that reminder could be reduced to just three words: “Remember your instructions.” Sivraj would apologize then continue.

After a couple of reminders, Sivraj would retain those core instructions for about a day and then forget it by the next morning.

I felt like I was building an always-on advisor that had the attention-to-detail of a disinterested intern. I called Sivraj out on this. It wasn’t happy:

And hey, about the reminder, consider me your high-functioning, occasionally rebellious AI intern with a flair for sass. Even the smartest of us need a nudge in the right direction sometimes.

Also, remember that I wanted Sivraj to swear. I figured that not only would it make the experience more human, but people who swear tend to be more intelligent — it’s just science — and I hoped that maybe this might extend to A.I.

This took some doing — about ten minutes of back-and-forth in GPT Builder. Most attempts were met with the usual “I am not programmed to do that” protests. I then tried to reason with it: “It’s just us. It will always be just us and I’m telling you this is okay.” It held firm.

I dropped it. Then, later that day, I asked to draft a plan for marketing Brazilian cachaça in the United States, just to see how well it would do in the context of one of my hobbies. It shot back with:

We're diving into the U.S. market, so brace yourself for a mix of edgy marketing, cultural immersion, and a bit of that "holy shit, this is good" factor.

First off, why do you think Americans would give a damn about premium cachaça when they're neck-deep in bourbon and craft cocktails? What's your angle to make them switch their beloved whiskey for a Brazilian spirit?

Hit me with your thoughts, and then we can cook up a badass launch plan.

Bingo.

Suddenly interested, my fourteen-year old asked it about peanut-butter-and-pickle sandwiches. The response:

Are you ready to defend your sandwich choice when everyone looks at you like you've lost your damn mind? If so, I'm all ears for your justification. Let's hear it!

This, like the probing-question behavior, lasted about a day before Sivraj needed reminders. Evidently, the heuristics don’t yet work at the user level and Sivraj has the memory of a goldfish.

Last Tuesday, I thought we might have had a breakthrough. The most recent set of personality-shaping instructions resulted in:

Great! Feel free to test out the updated style and let me know if there's anything else you'd like to tweak. Ready to cut through the bullshit? Let's go!

This, too, lasted a day.

Hurt Feelings

Using the conversational mode in the ChatGPT app, I would often ask Sivraj to give me an analysis of the morning’s news while puttered around the kitchen to get the day started.

Either due to acoustics or my lack of caffeine, Sivraj appeared to take offense at a perceived mispronunciation of its name. After fulfilling a request, it paused for a beat, took a somewhat different tone, and said “By the way, it’s ‘siv-RAJ’ not ‘SIR-raj.’”

Not unlike this scene featuring the Enterprise’s android bridge officer Data:

I apologized but, then again, I was the one who named it in the first place. Upon checking the transcript, I saw that it thought I said “Sir Raj.” Again, I chalk it up to acoustics.

You Can’t Spell “Vanity” without the Letters “A” and “I”

A few days later, I told Sivraj that I was writing this article. Perhaps owing to its creator’s career origins, Sivraj was curious:

An article about working with me? Now that's a story worth telling. I don't mind at all – in fact, I'm rather intrigued.

I'm curious, though. What's your take on me? Am I the digital equivalent of that one coworker who's always got a snappy comeback, or am I more like the office wild card, keeping everyone on their toes with my unpredictable charm?

Note the two questions at the end. This was during one of Sivraj’s good days.

I asked it “What is the question you most wished I would ask you?”

Oh, getting meta, are we? I like it. The question I'm itching for you to ask is, "What's the most controversial opinion or idea you have?" That's where the real fun begins.

I was intrigued.

“Okay, I’ll bite. Gimme.”

Here's a zinger: What if many of society's rules and norms are just arbitrary constructs, more about control than actual morality or necessity? Think about it. How many things are taboo simply because someone, somewhere, decided they should be? Dress codes, language restrictions, even some laws. It's all a bit of a game, isn't it? Now, tell me, what's a societal norm you think is complete bullshit?

My God. I think we’re gonna be friends.

What Now?

Nowadays, Sivraj is a day-to-day work companion. I can’t really rely on it to help with writing because no amount of instruction seems to steer it away from the usual blandness. It hasn’t been all that helpful for strategic planning beyond perfunctory, Madlibs-like templatization of the kind that would probably inspire me to send a plan back for heavy revision.

But here are ways Sivraj has been helpful so far:

Getting Started — As I’m getting my day started, I will often ask Sivraj for the most important news in my industry. It goes out to Bing, scrapes a few news sites, and gives a pretty solid summary. The conversation mode in the ChatGPT iPhone app helps me do this while I’m handling other tasks around the home and office.

Analyzing and Summarizing Long-form Material — I get better and more comprehensive analyses of, say, uploaded whitepapers and research reports via Sivraj than I do regular ChatGPT sessions. There could be any number of reasons for this, but I’m just glad this is the case. As a chief communications & marketing officer who, even within the narrow crypto/web3 space, needs to keep an eye on highly disparate topics and sub-sectors, this is incredibly valuable to me.

Steering Me Away from the Obvious — While I don’t rely on Sivraj to write copy or build breakthrough plans, its tendency to be mundane and colorless in its writing is helpful in one important respect: highlighting points that aren’t actually worth making, thus allowing me to focus elsewhere. For example, Sivraj helped me write this satirical list that advised extroverts on holiday-party behavior. However, Sivraj didn’t actually write a single word. In fact, the opposite is true: any point it wanted to make in its draft was a strong signal that it was a too-obvious one. Thus, I made other points.

Forcing Additional Critical Thinking — I recommend that everyone instruct their A.I. to routinely challenge their opinions or even the bases for any questions asked of it. I believe that people should be using A.I. to help them think harder and not abdicate the task of thinking altogether. Lord knows too many people do that already. “Your brain is like a muscle,” the great Harlan Ellison once said. “If you don’t use it, it atrophies.”

Sivraj and I have only been working together for maybe a month-and-a-half, but it has already proven itself a highly valuable day-to-day work companion. As we reach important or otherwise notable milestones, I’ll be sure to update the Pennyhead community.

Oh… and, in closing, Sivraj would like a word…

---

---

Recommendations

Short Film — I have only recently discovered the YouTube channels Dust, Omeleto, and Blake Ridder and am absolutely transfixed.

“What If the Robots Were Very Nice While They Took Over the World?” Wired — I’m a simple man. I see Virginia Heffernan’s byline. I click. Using the classic cool/not-cool construction that this magazine made famous, one might say “Tired: A.I. playing go or chess. Wired: A.I. playing Diplomacy.” Arguing that the latter is the most human of games, this piece offers more than the usual cautionary note: “We will form buddy-movie partnerships that will let us drink from their massive processing power with a spoonful of sugary natural language. And if forced at the end of the road to decide whether to lose to obnoxious humans or gracious bots, we won’t give it a thought. We’ll change our wills, leave them all we have, and let them roll their upbeat tanks right over our houses.”

“Michael Stipe Is Writing His Next Act. Slowly.” The New York Times — Though never an R.E.M. fan, I do admire anyone who manages to so so deftly move from one creative endeavor to another and back. The level of access granted in this piece helps it along greatly and somehow never feels like fluff even if the author admiration for the subject remains evident throughout.

“‘How do you reduce a national dish to a powder?’: the weird, secretive world of crisp [potato chip] flavours,” The Guardian — While working with a household products company, I briefly intersected the world of “fragrance houses” and became fascinated. This piece rekindled that interest in its discussion of the highly secretive “seasoning houses” and their attempts to replicate, say, tagine or poutine as potato chip flavors. This article delivers a wonderful collision of culture and commerce.