Invaluable Inference
The Cost of Compute
AI is now doing radically advanced research that’s fast, cheap and groundbreaking.
Why is this good news for most people’s job prospects, at least in the short term?
Because we’re almost in the era when AI becomes too valuable to do your job, even if it can.
They simply asked Sol to make itself more efficient, and it did.
OpenAI has now just used their latest model, Astra, to solve 10 major “unsolved” problems in mathematics, quantum complexity and theoretical computer science.
But one of the most-interesting things about it - aside from the fact that we now have AI doing major scientific and technological research - is that for these 10 problems, OpenAI used no more than $200 in computational costs for each one.
In other words, once the model existed, they spent $2,000 solving 10 major questions.
They have also admitted there are plenty of major outstanding questions they could not solve by investing no more than $200 in compute into the effort.
They didn’t mention if those same questions were already answered with $2,000 or $20,000 in investment, but let’s maintain our perspective here.
We’ve just solved critical questions which have defied top researchers for many years, simply by feeding them to the latest AI with a minimal budget of computation to work them.
And while LLMs are showing the most success in math and programming - both areas in which they struggled only a few years ago - we’ve also seen major progress in biotech and chip design, and can safely assume they will be contributing to other branches of scientific, technological, engineering and medical (STEM) research in the near future.
What does this mean?
If a huge array of breakthroughs already exist in math, computer science and a few other fields for just $200 each, then we’ll see revolutions in those fields.
From the $200 discoveries, the $2,000 epiphanies, the $20,000 magnum opuses and so on.
Remember, we’ve had brilliant people spend whole careers trying to crack some of these questions.
So everything important we can solve at $200 or $200,000 we will.
Here’s the irony. Making revolutionary changes behind the scenes that feed into all of our other scientific and technological progress is much more dramatic in the long term.
But it doesn’t have the public impact as the guy at the counter in your local McDonalds or KFC disappearing and being replaced by a robot in a paper hat.
But while someone may decide to burn money on such an initiative, the vast majority of compute is going to be siphoned into the world’s real priorities.
You have the AI race itself and everything the top AI labs are spending computationally on keeping all their business and consumer accounts operating.
Then everything the second and third tier AI companies are spending trying to catch up or build their own niches, while keeping their accounts serviced.
Then every government building their own internal “sovereign AI,” if only so they can analyze sensitive intelligence and proprietary data internally instead of sending it to a private US company or the US government to handle. Or worse yet, China, with its reputation for data and IP theft.
Then every company and academic researcher sprinting along, trying to exploit these models to the maximum during this window of opportunity.
That last one might sound like a footnote, but depending on how quickly companies and countries catch on to what’s happening, it’s actually an existential race for everyone who realizes they’re in it.
Why?
I’ve mentioned options a few organizations have to try and create their own frontier AIs, but the truth is that building an AI on a par with, surpassing or even close to those of OpenAI, Anthropic or Google DeepMind is an option for very few players.
And a narrow window which is closing very fast.
But that makes this last race all the more important.
If you can’t create the AIs themselves, but you can leverage existing AIs in your own research to create multiple patents giving you a lock on critical products or your own industry niche…
Then you have a chance to protect and even build your business while the giants around you are still clashing over ultimate control over AI and ultimate power.
That may not sound like much, but when the stakes are - at least in business terms - existential, you do what you must. Or look up at the last minute, like the dinosaurs, just before the meteor hits.
So yes, we’re about to see unheard-of demand for inference compute - the computation that powers AI research.
Assuming no one vaporizes the Internet and the global economy by flooding us with trillions of scarcely-controlled agents with superhuman hacking skills, of course.
Or anything similarly insane.
Unfortunately, the idea that reducing labor to zero somehow sends profits to infinity has apparently crept its way into the pitch decks presented to too many venture capitalists.
Let me blunt, if the AI industry were spending trillions of dollars now in the hopes of spending trillions in the future in order to put people working in sweatshops for a few dollars a day out of work, we’ll all go bankrupt.
The same can be said of menial tasks costing several dollars an hour in labor as well.
That doesn’t mean these jobs don’t get disrupted.
Absolutely the opposite. But more will be affected by increased overall efficiencies or an industry itself going away.
As a company, you’ll spend to keep your assembly line running, or to make your process more efficient, or to expand your market share. And so on.
But we’re still at a point where massive capital investment specifically to eliminate a trivial expense is not in the cards for anyone who wants to stay in business.
Improving those efficiencies? Certainly.
Making sure you’re in the next emerging field, or the most-profitable version of the one you’re in now? Without a doubt.
Handing a humanoid robot a hammer and telling it to pretend to be blacksmith so you can threaten skilled artisans with your own “handmade crafts?”
At a ludicrous cost and with no viable market? Probably not.
Hence, even if we spectacularly increase the efficiency of our AIs, their very power insures every strategist will be pouring that compute into every challenge that matters.
Not into every paper hat.
