Artificial Intelligence: The Pattern Is Broken

Technology inflection points are led by the young. The micro-computers of yore famously went from basements and garages to Wall Street and Main Street thanks to a young cohort ready and willing to build, tinker, and program. Similarly, dial-up modems and bulletin board systems connected the young and enthusiastic, contributing to the growth of the internet.  The same story played out for personal digital assistants such as the Palm Pilot, the Casio BOSS, and the Apple Newton. First dismissed as toys by mature and serious adults who preferred leather-bound day planners, the handheld devices evolved into smartphones, largely due to a young and enthusiastic community.

A very different story is playing out for Artificial Intelligence. The young are not enthusiastic; they are, instead, booing their pro-AI commencement speakers.  Meanwhile, the mature and serious are embracing and investing in AI with considerable fervor and unfathomable amounts of money. The pattern is broken and something is wrong.

I cannot dismiss the youthful reaction. I cannot boo AI either. I can try, however, to understand my inner conflict.

One revealing experiment asked Can an LLM Solve a Murder Mystery? The whodunnit used was Edgar Allen Poe's The Murders in the Rue Morgue but was revised with new twists and a different ending. The changes were needed because the LLM very likely had already trained on the book.  

Did the LLM solve the mystery? In a word, Yes. From the article:

[the LLM] isn't just that it retrieves relevant facts -- it retrieves structured relationships... nodes in a web of cause and effect.

There were signs of deductive reasoning, but I would caution this is not the same as awareness or consciousness.

AI has also made impressive achievements in mathematics, so much so that some mathematicians are experiencing existential angst.  AI has allegedly solved the Navier-Stokes problem; I say allegedly because mathematicians still going over the very dense and obtuse proof, which actually makes a strong argument that we still need human math majors. Daniel Litt, assistant professor of mathematics at the University of Toronto, was optimistic on a YouTube interview: there is no upper limit on how smart AI can get, and despite that, or perhaps because of that, there will always be more math questions to ask and investigate than can be addressed. The interviewers did point out that AI might be hurting K-12 education, ironically narrowing the pipeline to higher degrees. 

In the medical field, AI is playing a targeted, but very important role in the development of Moderna's MRNA skin cancer vaccine. Unlike chemo which kills healthy and cancerous cells alike, the MRNA vaccine is tailored to each patient, and helps their body's own immune system recognize and target the tumor. Customizing the vaccine to an individual requires quite a bit of computation -- analyze the mutations, predict up to 34 of the most likely ones, and, like a detective, fingerprint the cancer protein. 

Scaling this up for thousands of patients would be Sisyphean were it not for Moderna's Maestro platform. Developed before LLMs, Maestro is more akin to a machine learning AI that orchestrates and optimizes all the processes that go into making a personalized vaccine. 

In all three examples, the AIs were successful because they were given guidance and structure from humans. The problem domains were also in a closed system.  For example, the murder mystery had all the facts laid out at the start, and had nothing new introduced after the experiment began. Regarding the field of mathematics, the disciplines -- from algebra to quantum mechanics -- developed over centuries.  The rate of change is slow enough as to not confound an AI tasked to prove a theorem. And with Moderna, we see AI's role was specific, well-defined, and most importantly, understood.   

In contrast, AIs that must operate in an open environment are more likely to go rogue and fail. Consider the personal AI agent Metamucil Meta Muse. Generally helpful in scheduling doctor appointments, making dinner reservations, and selling your stuff online, its failure was captured in a headline from The Guardian: Meta’s AI agent Muse gives out user’s home address without permission, sending buyer to his house. This would be funny if it weren't so dystopian.

AIs driving cars face an even greater challenge. What could be more open than the open road? While AIs can be trained to recognize and handle random events, the possible edge cases are practically infinite. Throwing more compute (data centers) at the problem can only yield diminishing returns at ever-increasing costs, not just in dollar amounts, but in power, water, and land consumption, threatening whole communities.

Instead of building more data centers, the better answer would be to narrow the problem domain. Geo-fencing is already employed, but it doesn't go far enough. We need vehicle-to-vehicle and vehicle-to-road communications infrastructure so that cars could talk to other cars and also share information about road hazards and traffic conditions. On highways, commuting corridors (eg: Boston to New York to Washington DC) can have lanes reserved for self-driving vehicles so that they can platoon -- follow one another closely to reduce aerodynamic drag, increase efficiency, and maintain speed. And we'll need transition lanes where the AI hands control back to you as you drive the final leg of open roads to your destination.

The same holds true for AI agents -- narrow their scope and access.  Admittedly, Muse with real boundaries would be less amusing, but it would be reliable and safe.

Along these lines, a notable development is the rise of local AI. Designed to run on a lone Mac, PC, or Linux machine on your desk, or in a cluster in your office computer room, or in remote racks of company owned servers, these AIs work to solve domain specific problems that don't need the power of a data center in the cloud. This has many tangible benefits, chief among them is stronger privacy and lower cost.

A press release from Thomson Reuters shows they have been working on a just such a thing:

Thomson Reuters took a different path [local AI]: starting from a strong open-source foundation and investing $40 million to train Thomson [on in-house assets]  for the jobs that matter most. ... The result is a model Thomson Reuters fully controls, without the heavy inference costs of typical frontier models.

Another example is kiraa, an Australian startup that has gone all in on local AI that runs on Apple Silicon. Apple's Unified Memory Architecture (UMA) provides a key advantage where the CPU, GPU, and Neural Engine live together on a single chip, obviating the need for channels to exchange information. The results are incredible gains in speed and efficiency, and an environment where local AI can thrive.

The success of local AI remains to be seen, but these examples are not outliers.  Nvidia, recognizing the threat that Apple presents, answered with its own system on a chip (SoC): the RTX Spark. It's their way to spark excitement in the young, and to say "Have fun! Please don't boo AI."

Knowing the young, they will still boo AI. While important, it is becoming clear that AI -- the frontier model -- is not a technology inflection point. Less pivotal compared to the microcomputer, the internet, and the smartphone, AI better parallels America's coal industry with initial intense infrastructure development and a promise of jobs and prosperity, only to be followed by damage to the environment and harm to local residents. Coal has since been surpassed by natural gas and renewables. In that light, today's AI can be seen as a stepping stone to something better (I hope). 

Energized youth lead the technology inflection points, not massive amounts of money. The pattern is not broken after all.