Insights from the Perfect Question at the Perfect Time
It's common for me to get to a milestone and then ask people to take a look at what I've done. Understandable that the common reaction is "oh... neat," and a polite smile. I have taken to calling this my "train set" for lack of a better term because some of the other questions are "What are you going to do with it?" or "How are you going to sell this?" which I often answer as "Learning... and stuff" with the latter often reminding me to make sure my train set has an actual purpose. Those really are great questions and understandable given the context of how AI is being seen right now.
This time, in talking with my very intelligent (and beautiful. Yes, she reads this blog, why do you ask) wife, I got asked a different question:
"If you had to build this in 1995, how many people would it take?"
That one threw me and I didn't want to take the lazy way out and answer of "You couldn't have built this in 1995."
Teasing Out the Underlying Meaning
How do you answer a question that has such a satisfying lazy answer? You buck up and figure out what the question is actually asking.
Reframed: "What would it take to design a system of equal capacity and then run it if this were 1995?" So the answer would have to presuppose that both the experts exist and the will is present to emulate what I'm building with people and technology available at a time when Windows 95 and the Pentium were cutting edge.
I can work with that.
WAGging the Dog
In true XKCD "What if?" fashion, I tried to break things down in a logical manner while keeping to the spirit of the question. This meant breaking things into two parts:
The Design. My estimates* assumed that the world contained 10 of the right people that could design the end-result that currently lives on a scratch pad in my computer/head. Like Apple starting in the garage, you'd need a governance architect, knowledge/library scientist, process/ops designer, domain generalist (or small team), and everyone would have to have the "red-team" mindset. Let cook for about 5 years and you'd have the processes and people in place to really start cooking.
The Execution. Continuing along a line of best-guess assumptions, the actual execution would take about 60-80 frontier-level experts working shifts 24/7. One orchestration cell, a pool of scribes writing every turn and writing very very fast to produce structured notes, a comprehensive library that contained a team of experts to keep the system "memory" up to date, audit clerks, routing office (I'd choose ye-olde pneumatic tube system for fun and a satisfying "thwoompf" as every message is dispatched/received. Yes I know email existed. No I won't change my mind. "thwoompf">"You've Got Mail!"), and security not only looking at the physical but the information contained within.
This would run about as fast as the DMV in 1995 on a Monday morning. Things would get done but you'd have to wait.
For a quick comparison of 1 and 2 and a bit of a disclaimer...
This is not a humble brag, but a state of the current world situation. I'm one guy running several AI's that have written:
- A system core with 9,300 lines of Python code (41 modules and 24 test files)
- One notebook that helps keep me (the human. remember that guy?) organized and on track which is about 8K lines of JavaScript code
- Retro User Interface eXperience (RUIX) that is just over 3,500 lines of code
- 124 design documents (and counting) that are approaching 30K lines (most of it moderately relevant)
- 284 git commits
And I pretty much started on this in earnest a month(ish) ago.
*read "mildly educated wild-a$$ guesses"
But There's a Twist!
Here's the kicker: MOST of the minimum 60 people are only there to keep the other 59 synchronized. Using Brooks' Law of n(n-1)/2 in calculating potential communication channels, we have a whopping 1770 and that's not counting any possible external connection.
Before you think "well, yeah. computers are efficient at doing that" so "why u no use them," remember this is 1995. It was a bridge between the way things were done and the way things are done. Oddly enough, a similar question was probably being asked back in 1983 with people's Commodore 64. Look where that got us.
So I would say that maybe "Productivity" is the wrong word here.
Vertigo in the Face of the Right Word
The AI ends up doing three jobs at once and I think it's often undersold. The tool itself multiplies build speed, supplies a reasoning core (this really didn't exist before), and acts as a bench of experts. Combined.
Productivity really only answers the first part.
This is where the weight hit in only the way vertigo can. This may not be something like Apollo's scale of almost 400 thousand people and enough money to also need a team of accountants to keep track of the accountants.
It is something that is the same shape, though, being repeated by one person (many times over, you should see some of the stuff on huggingface.co). What once took an institution has now been replaced with a single person (albeit one standing on the shoulders of several institutions).
How did I even come to these conclusions?
With AI.
Yep. I'm sure you saw that one coming.
Putting It All Together
I actually consulted two. The first was to my local AI that was inside the system being built. Eggy took at look at the plan, estimated from what daily labor was accomplished, and then arrived at an estimation on what it would take to actually be an AI with that capacity.
Of course, Eggy missed the literal question of "Build" and assumed "Operate," but this helped surround the question better than I ever could have. Any intelligence can only answer from experience or learned understanding which made this question so much weightier in my mind. In this case, the play was the thing.
My other question was posed to what I have taken to call "AdvAIsor" (I hear your groans... they sustain me). This is a frontier model that caught the missed question of "build vs be" and gave me that piece of the puzzle.
The richest part was that this was three perspectives arriving at a richer answer than any one could have achieved.
ParlAIment
Yep, misspelled.
On purpose.
This is a small instantiation of the governed structure I'm building. Local and scoped intelligences reason from within. AdvAIsor reasons from without. The human takes responsibility for any external action. A ContrAIrian tells them how full of it they are and to think harder on point A subsection 3 (I'll talk about that more later). The fulcrum is that they are all checking each other's work.
I hope this answered her question.
One person that can equate to an institution's level of capability must also bring into being guardrails that protect the analysis: The governance, shared vetoes, audits that keep recording even when models go dark (looking at you Fable). Being cautious like this stems from the same place of how astonishing it is: the lever is real and with that fulcrum, one person can now truly move the world.