Seven Years in Seven Days: AI Takeoff March 2023: Strap In, Humans

AI takeoff

I can say honestly and without exaggeration: I’ve been in tech for more than 3 decades now, and this past 7 days has been the most eventful and most accelerated leap I’ve yet seen. It is, without a doubt, the AI takeoff. Let’s give it some context, and hit some of the highlights:

(The Lead-Up: Nov 22-Mar 23)

On November 22, 2022 (11/22/22), a brainy little startup with zero marketing budget and a cool $1 billion in funding quietly launched ChatGPT, a text-based conversational interface to its latest AI. Once it went public on 11/30, it went from 0 to 1 million registered users in 5 days. From 1 to 100 million in the next 90 days. But that was mere prelude…

Less than 60 days later, on January 23rd, 2023 (1/23/23… you getting this?) Microsoft hinted at, and confirmed a $10 billion (10,000 million dollar) infusion into OpenAI. Why? Primarily, to pay the compute and electricity bill for the training of bigger and better and stronger AIs across the coming months and years. OpenAI has less than 700 employees. But its AI models currently devour more than $100 million in compute & energy resources, per entity spawned. The next gen is predicted to cost north of $1 billion per entity (aka model, aka version). You can see how, at that rate, $10 billion can go pretty fast.

The THUNDERCLAP

On March 14, 2023 (Pi Day: 3/14/23)(we are now, officially, on Day 17 A.AIT. (after-AI Takeoff)), OpenAI soft-launched GPT4. No big party. No massive marketing campaign. Unassumingly, one additional option became available on the pull-down menu for paying users of ChatGPT. Oh, and: the President of OpenAI, Greg Brockman, did a quirky (and stunning, unbelievable) live capabilities demo showing, amongst other things, how GPT4 (which has a vision engine, to “see”) could build a functioning website by merely glancing at a sloppy sketch on a napkin. I kid you not:

GPT4 vision: scan a napkin sketch, make a website

I know how this sounds to you if you haven’t even had a single conversation with an AI. And yet. We are here. YOU are here, whether you recognize it or not.

The rocket has left the launchpad. The tsunami is within 100’ of shore.
This is Ai Takeoff. Pick your metaphor. We’re here.

So.

Where is “here”? For starters:

AI & The Standardized Test Battery

ChatGPT was soundly ridiculed, even as the press was praising its bizarre and clever creativity, for being so “dumb”. When given a vast battery of standardized tests used to judge human ability, it squarely landed in the bottom 50, and most often, the lowest tenth percentile. In other words, barely human. In many minds, not even human. A retard. A dummy. An idiot, at best. I myself, an AI optimist with aggressive projections, plotted AI acing those same tests within 4 years. Something like:

  • 2022: E
  • 2023: D
  • 2024: C
  • 2025: B
  • 2026: A

Then, GPT4.

Test results: uniformly in the top 20% (i.e. beating out 8/10 skilled humans) (these weren’t’ basic reading and writing tests… these were masters degree post-college professional exams for MDs and Lawyers), often in the top 10%.

This changes everything: GPT4 blows humans out of the water on standardized tests

So to break it down: The AI went from zero (2015) to 10 (2022) in the span of 7 years.

But, as we are so often reminded, tech moves on an exponential curve. So instead of going from 10 to 20 in anther 7 years, that process took about another 7 *weeks*.

And then how long to go from 20 to 100?
Not 60 years (the earlier pace). 90 *days*.

This is what we call AI Takeoff.

This is what we call, the rocket leaving the launchpad

To try and put a visual to this type of acceleration:

AI takeoff: humans struggle to comprehend exponential growth

So we stand at the precipice.

Because, clearly, in the next 6 months,
one of two things is going to happen:

By December 2023, we will either:

A) have found out that humans are indeed special, and that there are both as-yet-undiscovered yet-roundly-intuited physics and spiritual laws that prevent any non-biological entity — much less a man-created machine — from achieving and surpassing human intelligence, OR

B) we find out that AI is about to blow past human genius-level intelligence (the foretold Intelligence Explosion) in a matter of months, and won’t even bother to look back in the mirror. And so by Christmas (T minus 9 months and counting), we will have multiple (10s? 100s?) of superhuman, 1,000+ IQ AI entities roaming the internets, interacting with humans and machines alike. Oh. And we will have their creators — OpenAI, Tesla, Google, Baidu — who will profess to be able to “control” these entities.

The track record, however, for beings of inferior intelligence keeping super intelligent creatures in captivity is… well, let’s just say, questionable. These things will make Hannibal Lechter look like a chimpanzee. They will be able to conversationally dance circles around even the smartest of us. To obscure lines of bullshit with 10 layers of synthesized, plausible bullshit behind it. And… they will be able to manipulate.

Meta’s Cicero AI quietly won a high level tournament of Diplomacy. Poker, another game with great unknowns and massive amounts of human psychology involved, fell to AI many many years ago. But Diplomacy was yet another level. That game is 100% human, and 100% opaque. There are no dice, there is nothing random, there is only time & negotiation, and people must be “played” in order to win.  And Cicero did win, handily. Not all the games. But twice as many (26%) as the best human player (12%) in a tournament of 40 games and 82 expert humans.

Think I’m kidding? Check out these dialog examples:

AI takeoff: Meta's Cicero schools humans in Diplomacy board game

So, my conjecture is, if yesterday’s AI can beat humans squarely at such a human bluffing / manipulation / deception / cooperation game as Diplomacy, then what can they not achieve in the real world, by manipulating humans to conform to their desires?

Sorry, I keep promising to give some highlights of the last 7 days. Just one more prelude: 4 months ago, I wrote what seemed like an aggressive timeline of AI developments I thought would transpire in the next 2 decades. 2022-2040. Well. Let’s reality check that prediction. As of March 31, 2023… we’re at about 2028 on my timeline. Guess it merits a bit of a revision.

Now. 7 days. Here goes:

AI Takeoff: Seven Days of AI HyperAcceleration

  • GPT4 is launched — March 14, 2023
  • It can see (interpret images)… this is advertised.
  • It is safe and locked out of the internet.
    So says its creator, OpenAI.
    …5 days later:
    .
  • OpenAI announces “plug-ins” for GPT4.
    • What’s that?
    • Basically, ports & APIs that allow GPT4 to be connected to anything you like: drones, robot arms, your email account, your calendar, your bank account, your Apple Watch, your Amazon Prime, your Tesla, you name it.
      …Yeah, so, um… well,

“So much for the safety thing! Progress!

  • MidJourney launched version 5 of its generative AI.art engine.
    • Deep Fakes are now truly here, and they are, more importantly, trivial to create
    • You type in a single sentence: “Donald Trump being tackled by cops on the streets of DC” and, 60 seconds later… BOOM, here are 4 photos of that (fictitious?) event:

AI takeoff: Deep Fake: Midjourney: Trump being tackled by DC Police

    • As I call it: Easy Bake Deep Fake™
    • It does all kinds of crazy amazing stuff beyond that (lighting! Film & lens choice! Styles! Mirrors! Reflections!) but the deep fakes, for now, steal the  show.
      .
  • Glass AI launched their clinical diagnostic tool.
    • The free demo version is not trained on any specialized data; just the public internet.
    • That said: I asked my doctor friend to supply me observations from a patient for the most difficult case he had ever faced. He and a team of high-profile doctors attempted incorrect after incorrect diagnosis for 40 days until they finally got it right.
    • Glass AI’s results?
    • I fed it the 5 sentences of observation notes.
    • In about 22 seconds, it had provided the (correct!) diagnosis, AND provided a thorough diagnostic and treatment plan.
    • (Human) Doc’s opinion? 100%, A+, Master-Level Clinician / Diagnostician. boom.
      .

…and of course, how can we ever forget:

  • Will Smith gorging on Spaghetti.
    • Text-to-Movie. We’ll say no more.
      Google it if you haven’t already consumed it.

And now, we enter… The Days of the

Fully Automated Media Engines (FAMEs)

Text to Video, Text to CGI, Text to 3D Worlds, Text to… you imagine it, it is manifest!

Let’er rip!

  • Text-2-Worlds : Type a scene description in natural / casual language, auto-generate a high-fidelity 3d model of the world you’re describing, in real time, as an asset for your videogame or movie set.

  • Text-to-Video : Rotoscope any video in the style of any 2d artwork at the click of a button


  • Prompt-2-Videogame: GPT4 building  fully functional videogames.
    • They’re not “great”, but they’re playable, and fun, and miraculously… they actually run.
    • GPT4 planned the game, wrote the code, and generated the prompts for MidJourney to create the art assets.


  • Video-2-CGI : Fully automated single camera mocap / animation pupeteering / CGI with camera / lighting / environment mapping, instantaneously:


  • Auto-Translation w/ Voiceover + CGI Lipsync : Seamlessly translate any film into any language, AND automagically re-render the mouth, jaws, lips, teeth and tongue of the actors to match the new language target.


  • Have GPT4 improv rap lyrics in the voice and style of any artist, in real time


  • GPT4 makes beautiful word-art poetry in the guise of your creative Collaborator, Bing Sydney

janus repligate eleizer yudkowski word art bing sydney


  • Have your Apple Watch listen to your voice and pass natural language commands to GPT4, which in turn will write the code you are describing, and commit it to your GitHub branch, automagically and correctly.

  • the GPT4 language model draw vector graphics
    (do you get it? This model isn’t supposed to have any visual understanding at all, its trained exclusively on text; and yet…)


  • Have the GPT4 language model write a pop song,

AI takeoff : LLM writing music, melody, lyrics

…or even a symphony:

AI Takeoff in the form of an original GPT4 Symphony

AI Symphony No. 1 : The Lyrical Odyssey
Movement I: Allegro Con Spirito


Sooooo…. Yeah.

So that, was:

SEVEN. DAYS.
7-6-5-4-3-2-1…
.
AI TAKEOFF.
.

No, this didn’t actually all happen all at once (well, it did, actually)… but it all came to a head at once, at the point of AI takeoff… like a field that was planted in the fall, and now it is spring, and all the flowers are blooming, all at once. But these flowers don’t wilt and die. These flowers keep getting bigger, and brighter, and taller….

next:

>>> What you can do, today.

, ,