AI is already beyond the reach of the regulators, and that’s good news for the rest of us
Something important happened this week and it was nothing to do with squabbles between factions of Russian thugs, or even Hunter Biden.
Databricks, a privately held software company, with circa $1 billion annual revenues, announced it would pay $1.3 billion to acquire an utterly obscure artificial intelligence (AI) start-up called Mosaic ML.
This matters because it is one more sign that while pundits still debate government control of AI, it will soon be beyond the reach of regulators from any but the most repressive regimes, the sort that can shoot you just for owning a computer.
That’s because AI, which yesterday was a science project, tomorrow will be a technology.
How can you tell the difference? The same way you judge a wine—by the price.
People who do science projects don’t much count the cost. They just want to show it can be done.
The science project becomes a technology only when it meets the market and wins.
NASA’s Mercury, Gemini, and Apollo were science projects.
SpaceX is a technology.
When Jack Kilby made the first integrated circuit (IC) in his Texas Instruments (NASDAQ: TXN) lab, it was an important science project. He proved that every component of an IC could be made in silicon.
Kilby’s IC wasn’t a technology because the final wiring had to be soldered in place by hand, making mass manufacture implausible. The IC became a technology only when Fairchild Semiconductor’s Robert Noyce used Jean Hoerni’s “planar process” to eliminate those dangling wires so the device could be manufactured by the millions, cheaply.
Alpha GO, the Google (NASDAQ: GOOGL) AI that beat human world champion Lee Sedol in 2016 was a science project.
To date, so is ChatGPT.
Sure, you can pay Open AI $20 per month and run your own little science project on Microsoft’s (NASDAQ: MSFT) cloud. But currently it’s still a stunt, or at best a super-expensive service available only through giant cloud services such as Microsoft’s, or Amazon Web Services‘s or Google Cloud, all of which have big government bullseyes on their backs.
Have $390 million handy?
Most AI super-computers run on GPUs (graphical processing units) provided by Nvidia (NASDAQ: NVDA) or Advanced Micro Devices (NASDAQ: AMD). These chips typically cost $10,000-15,000 a piece. And a handful won’t do: Google’s new A3 GPU supercomputers for Google Cloud calls for 26,000 of the things. That would come to as much $390 million, not counting other components. Plans are being made for a supercomputer that will use 60,000 GPUs.
For decades, the computers kept getting bigger as the AI models, in their science project phase, kept getting more ambitious. Today’s AI models are measured in “petaflop days” a petaflop being a million, billion (1015) floating point operations per second. For the last few years, model size has been doubling every few months.
In 2018, GPT-1 employed 100 million “parameters” (variables) to reach its conclusions. GPT-3 used as many as 175 billion. GPT-4, just released, may be using 1 trillion.
MSFT: “Please don’t buy ChatGPT, we can’t afford to sell it.”
OpenAI reportedly plans to charge at least 15 times as much for the new version as its predecessor. According to industry analysts at Trading Places Research, this may be in part to discourage people from using it, since Microsoft is struggling to put together the hardware to do the job.
Worse yet, big GPU arrays suck so much power that the industry is already seeing system failures from excessive heat. These costs did not matter for the science project phase. The only thing Google cared about when it built Alpha GO was to prove it could be done.
For AI to become a technology, costs must reverse direction and come down by orders of magnitude. It’s happening.
Slimming down the super models
The first path to lower cost is economizing on model size, which is what Databrick’s acquisition of Mosaic ML is about. Mosaic claims it can build models of the same utility as those currently costing tens of millions for tens of thousands. It is far from alone in that struggle. Even big-spending Google has announced a next-gen, large language model which is both one-third smaller and better than its predecessor.
The Open-Source movement is rife with model shrinkers, thanks in part to firms such as Meta which have released their large language model source code to the community.
Meanwhile, on the hardware side, alternatives to those $15,000 GPUs are coming fast. Of particular interest are so called FPGAs or field programmable gate arrays. FPGAs have the unique property that their hardware itself can be reconfigured remotely, even after being deployed.
Though GPU-driven devices excel at training the models, such reconfigurable FPGAs, when applying the models to specific tasks, consume orders of magnitude less capital and energy, and can be embedded in devices smaller than your smart phone.
FPGA-based hardware will accelerate the movement of AI capability out of the huge data centers to the “Edge” of the network, where the end-users live. (That would be you.)
AI: Power to the people
Once distributed to the Edge via user-friendly apps, AI becomes impossible to regulate in a democratic society. It becomes almost invisible, fading into the landscape. Every app will incorporate AI, which is the best camouflage
Currently, the biggest FPGA makers are AMD (via its recent acquisition of Xilinx) and Intel (NASDAQ: INTC) via its acquisition if Altera in 2015. Both will deploy their FPGA skills to compete with Nvidia on the edge.
But even these firms may not be the ultimate winners. Newbie FPGA firms are coming fast. Some, like QuickLogic (NASDAQ: QUIK), are exploring radically new, more versatile, and less expensive FPGA technologies that have even a bigger edge at the Edge.
Forget the regulators. Just go figure out how cheap, universal AI can be your servant rather than your replacement.
P.S. Join me at FreedomFest, “the world’s largest gathering of free minds,” just a month away! I, along with my fellow Eagle Financial gurus, Mark Skousen, Bryan Perry and Jim Woods, will be speaking. The full agenda — speakers, panels, debates and breakout sessions — is now posted online. Go to www.freedomfest.com/agenda to check it out. You will be amazed! You can also click on the name of each presenter to see when and on what topic they will be speaking. Click here to find out more. When registering, use code EAGLE50 to receive a discount. I hope to see you at FreedomFest in “Music City,” Memphis, Tennessee, July 12-15.
P.P.S. Join my Eagle colleagues and me on an incredible cruise! We set sail on Dec. 4 for 16 days, enjoying a memorable journey that combines fascinating history, vibrant culture and picturesque scenery. Hear seminars on the days we are cruising from one destination to another, as well as dine with members of the Eagle team. Places we’ll visit include Mexico, Belize, Panama, Ecuador and more! Click here now for the details.
P.P.P.S. Hey, we are not the only Eagle Financial wise guys with insights into AI investing. For an insightful summary of what some of our colleagues are thinking, see our editor Paul Dykewicz’s weekly columns on stock and income investing.