prediction

The Last Word on AI This Time

June 24, 2025 By Jack Vaughan

When I first heard of Generative AI, I was skeptical. Although it was clearly a gigantic step forward for machine learning. I covered the Hadoop/Big Data era – for five years. As noted before, we would ask what do we do with Big Data? The answer it turned out was Machine Learning. But it was complex, hard to develop, difficult to gather data for, and ROI was complicated or ephemeral. People would bemusedly ask if it had uses east of Oakland Bay. My experience with Big Data colored my perspective on Generative AI.

Generative AI requires great mountains of data to work. Herding that data is labor intensive. As with previous machine learning technologies, getting desired results from the model is difficult. Some engineer will come up with a tool to solve problems. And find some VC, and startups race about. Few apps get from prototype proof-of-concept to operational. Few pay their own way.

Benedict Evans, Geoffrey Hinton and Gary Marcus are just some of the people who more ably than I critiqued the LLM. But great excitement was unleashed on the global public. And there wasn’t much of an ear for their ‘wait a minute’ commentary.

But early this year, Deep Seek seemed to show that the rush to Generative AI – the rush of money for electricity and land to deploy widely – should be more carefully considered. Deep Seek was an event that arrived at a receptive moment.

It seems in a way a textbook case of technology disruption. That advocates were blind to the limits of scalability for LLMs, coming up with greater and greater kluges – think nuclear power – SMRs or other.

Meanwhile, a crew at a resource-strapped Chinese quant-tank saw ends around. The designers focused on efficiency rather than brute force, employing methods such as reduced floating-point precision, optimized GPU instruction sets, and AI distillation.

Engineers love benchmarks – they love to tear them down! Benchmarks are biased, yes. Even a child can figure that. But … when you look at Deep Seek’s recipe, it is clever engineering. None of it is new. Others have worked on TinyML for years. The type of hardware optimizations they did were bread and butter years ago. There are plenty of computer scientists the are working to get around Generative AI’s issues [scale/cost/hallucination/use cases being the big ones]. These issues make this Silicon Valley baby a suspect redeemer. With respect, Jim Cramer sometimes oversteps his actual areas of expertise.

That Deep Seek moment – don’t let anyone tell you it is more or less than that – has just been followed by an upswing in the contrarian view on LLMs. A world that would have nothing of it a year ago is now seriously discussing “The Illusion of Thinking” – a paper by Apple researchers that questions the true reasoning capabilities of large language models.

“The Illusion of Thinking” This may put a pall on Agentic AI, which has conveniently arisen as the answer to Generative AI’s first task: to finagle its way into real world workflows. Now, as summer begins, there is more of an ear for voices that cite the challenges, obstacles, and over-sell that have marked the last 24 months or so of the AI era. That can be helpful in the big task of understanding and taming LLMs for greater usage.

Penultimately, we have to hold in mind some contrary points going forward. It is not LLMs are not valuable, just that they have limits that hyperbole has obscured.

Inspiring to me was a recent post by network systems expert and author Bruce Davie. He reminds us that a rational middle path is often preferable to the extreme predictions of doom, bust, or boom that characterize today’s AI tempest. Humans can skew but the mean always calls, and we may be seeing that now. [Thanks to Davie for cueing me to New Yorker writer Joshua Rothman and, in turn F. Scott Fitzgerald, he of the adage of holding “two opposed ideas the mind at the same time” seen above.]

This seems like a good time to let the Substack cogitate on these great matters. While I may post yet this season, I am kicking up my heels, and dreaming about slapping skeeters in Up North in Wisconsin. And taking “Lise Meitner: A Life in Physics” down from the shelf.

Why tinyML?

MARCH 26, 2021 — In about 2004 this reporter asked a top IBM software development leader what cloud computing looked like to him. “It looks like a mainframe,” he said with only half a smile. True enough, cloud is a centralized way of computing, which is beginning to raise questions.

One of which is: Will machine learning be forever in the “glass room?” That is the old-time descriptor for the home sweet home of the immortal mainframe era, where numbers got crunched and good ideas went to die.

Today, technologists are working to bring machine learning out of the closet and into the real world in the form of Edge computing.

For that to happen machine-made observations and decisions will have to succeed on individual chips in devices and on boards, far from the cloud data center where a lot of electrical power allows infinite compute.

For that to happen, machine learning at the edge, which is often more project than reality today, will have to become productized. It will have to work within much tighter constrains. That is the motivation behind TinyML, which — thank goodness — is more a way of doing things, than it is a standard or product.

Issues facing TinyML as it struggles to leave the cocoon are worth consideration. As with client server and other computing paradigm shifts, the outcome will rely on how teams on the cloud and on the edge deal with the details of implementation.

That was seen in a panel at this week’s tinyML Summit 2021. It afforded opportunity for such consideration. Here I am going to share some comments and impressions from a panel that featured expert implementers working to make it happen.

The lively panel discussion entitled “tinyML inference SW – Where do we go from here?” was moderated by Ian Bratt, Distinguished Engineer & Fellow, Arm. Joining Brat were Chris Lattner, President, Engineering and Product, SiFive; Tianqi Chen, CTO, OctoML; Raziel Alvarez, Technical Lead for PyTorch at Facebook AI; and Pete Warden, Technical Lead, Google. (A link to the panel recording on YouTube is found at the bottom of this page.)

A familiar view emerged, one that showed the creators of the trained machine learning model handing off their work, hoping a dedicated engineer can make the code run in the end. That conjures the old saw about ‘throwing it over the wall,’ and hoping system programmers can do the finished carpentry.

The tableau suggested the objectives of the researchers in a sort of ivory tower of cloud machine learning were somewhat at odds with the objectives of the front-line inference engineers at the edge where cycle economy is paramount and power consumption is crucial.

That echoes yet another developer saw that goes ‘it worked on my machine’ – one of the classic crunch time excuses over the history of computing.

Other issues:

-It may take top gun skills to make a trained model work in practice. “Somebody has to do magic to get it into production,” said Raziel Alvarez.

-People are able to deploy these models but the effort is very considerable. The many different cogs in machine learning (for example, the link between a CPU and a GPU) have to be managed deftly. In practice that means “people have to go outside their [practice] boundaries,” said T.Q. Chen.

-They hope to deploy inference on a variety of hardware, but each hardware version and type requires special tuning. And, low-level hardware changes can effect a cascading chain of changes all the way up the machine learning development stack. “As soon as you get heterogenous hardware, models tend to break,” said Peter Warden.

Hmmm, maybe that is enough on the ‘challenges.’ Obviously, people go at this to succeed, not to loll in obstacles. But obstacles go with the move to production for machine learning inference. As one tinyML Summit 2021 panelist said of recent history, “we have found a lot of what doesn’t work – we know what we don’t know.”

It will be interesting to see if and how the machine learning technology moves to the edge from the cloud. In architecture, the devil isn’t in the details, but in building, it is. What is likely is that the leap from science project to useful product will depend on the future work of the participants at tinyML Summit 2021 and other conferences to come. – Jack Vaughan

Some cracks in the new AI edifice

[September 7, 2020] – Something new happened when newly automated algorithms of machine learning met the big, 21st Century data troves of Google and Facebook. Spring had sprung!

But, if you look at some of the smoke signals drifting out of the Valley you may wonder if scenes from an earlier AI Winter may be replayed.

Distinguishing cats from dogs is one thing, autonomous driving is another.

AI’s forward steps are often accompanied by equal sidesteps. But its fortune continues to rest on the fulfillment of machine learning methods based on neural nets. After all, that has been the biggest driver behind the upsurge in AI.

It has been exciting to watch the rebirth, but caution should still obtain. That’s the core of my analysis as a reporter – one who saw neurals enter a long winter, and later emerge from a time in obscurity.

Problems that deep learning — machine learning’s latest best hope — faces include:

*The vast amounts of computing needed to train the deep learner,
*The massive quantities of data and electricity such computing entails, and
*The algorithms’ questionable ability to adapt to unexpected data patterns like those a Black Swan Covid-19 Pandemic can serve up.

Faults like these are well known within the machine learning community, and were much discussed in the technical confabs that preceded the general cultural slowdown that the pandemic has wrought. Yes, they are working on it.

ii.

To be fair, when AI actually escapes from the lab it takes on a different persona. It becomes more businesslike and less magical. But the business argument for deep learning and AI itself are moving targets, as a monograph by a16z analysts tends to indicate.

In short, the AI edge that big data and cloud players like AWS, Google, and Microsoft have exhibited does not easily transfer to startups working in the AI vineyard.

After some writings earlier this year that questioned what kind of a moat AI startups can build around their business, Martin Casado and Matt Bornstein went on to consider how different an AI business may be from a garden variety software business. Their piece is entitled The New Business of AI and How It’s Different from Traditional Software.

While the overall future of software may technically take the path of AI, they conclude, the economics of the undertaking may more resemble a consultant/services business than the SaaS software model.

That is because there is a lot of work that goes into making machine learning work in special cases. It is unfortunate but the world is a maze of special cases.

Casado and Bornstein write:

AI is showing remarkable progress on a range of difficult computer science problems, and the job of software developers – who now work with data as much as source code – is changing fundamentally in the process.

They continue:

Many AI companies (and investors) are betting that this relationship will extend beyond just technology – that AI businesses will resemble traditional software companies as well. Based on our experience working with AI companies, we’re not so sure.

Compared to your regular software enterprise, AI companies have lower gross margin (due to heavy cloud infrastructure and labor costs), greater scaling challenges in solving edge cases and weaker defensive moats due to the commoditization of AI models.

Our a15z authors have probably arrived at this conclusion after having more meetings with AI dragon slayers than we here have had hot lunches this year.
That AI, deep learning, and machine learning face challenges should not be a surprise. But these techs should continue to face scrutiny, especially when it comes to some of the further reaches of use cases, as will be discussed below.

iii.

One of the most enchanting movies in recent memory is The Vast of Night. Something like a slicked up The Blob or a science-fiction version of The Last Picture Show, the film is set on a summer night in the 1950s in New Mexico, and something akin to UFOs is on the prowl. The Vast of Night centers on a boy who is DJ at a radio station and a girl who runs a telephone switchboard and their encounter with something vague and eerie.

In the setup, it is clear the boy, Everett, is a technical geek and a cynic. Meanwhile the girl, Fay is enthusiastic and wide-eyed – anxiously looking to the future. The film takes on the eternal glow of an old tube amp as Fay gushes to Everett about the technological wonders ahead – the ones she has read about in Popular Science – and he squints skeptically.

This world of the future is not far away, she assures, and it will include portable video phones and self-driving cars. Her naivete would draw chuckles in a theatre, as the cell and video phone have arrived. It brought recognition to me.

That is because in my own little time-capsule movie, an 8th grade class field trip had allowed me to take part in a video phone call to a girl at the New York World’s Fair in 1964.
Later, I told the folks at dinner back home that “TV phones are coming soon.”

The family did not buy it. The advice was: “Eat your dinner.” They were right.

It wasn’t until this year’s Pandemic-driven adventure in Zoom-A-Rama, that I personally became a regular user of such video communications capabilities.

The GM exhibit at that same World’s Fair showed automated highways – a dream still deferred.

Today, the self-driving car is pursued as the litmus test for deep learning and other AI techniques. But its ascent is still in question, and ‘Where is my self-driving car?’ has joined ‘Where is my flying car?” among popular memes.

Which leads us to a discussion of Starsky Robotics. In February, as the Coronavirus began to spread, this driverless trucking company shut down. Even though it was the first company to publicly run a 7-mile load without driver, it failed to obtain financing for further endeavor.

Company co-founder Stephan Seltz-Axmacher was philosophical about the failure and wrote an oft quoted blog on Medium.com that outlines the challenges ahead for such AI undertakings.
The biggest insight he offers there perhaps is this: “Supervised machine learning doesn’t live up to the hype.”

iv.

Early this summer as I sat in on a Robotics Business Review webinar on manual vehicle automation, it seemed a good occasion to ask analyst Rian Whitton (ABI Research) for his take on the Starsky experience.

Whitton is a deft researcher and observer on robotic automation – no rose-colored glasses, either. Whitton recommended the Medium blog entry to anyone interested in automation and AI. It raises questions about whether automated systems would finally be up to task, and whether the business prerequisites are sufficiently aligned.

Whitton boils the question down to: “Does more data equal a better system, to the point where eventually, through superior algorithms training and more edge cases, the self-driving car revolution would simply come about?”

He like others points out that it becomes increasingly expensive to obtain data to actually improve the necessary algorithm or program.

“In a sense,” he says, summing Seltz-Axmacher’s thesis, “the logic — that simply through more testing and more deployments, self-driving cars become safer — doesn’t actually hold water.

That’s a problem awaiting anyone waiting on the Next AI Golden Age. But, Whitton notes, something short of full autonomy may show actual cost benefits, and could spell some type of progress.

It may be the case that much of AI and machine learning in time to come will be about paring down the problem until it can be efficiently solved. – Jack Vaughan

Cited above

The Vast of Night – IMDB
The End of Starsky Robotics – Medium
The Automation of Manual Vehicles: Insights, Analysis and Opportunities – Robotics Business Review [Reg req]
The New Business of AI and How It’s Different from Traditional Software – a16z.com
Machine learning tools pose educational challenges for users – TechTarget

Fujita – Watcher of the air waves

Stormy Weather — Ted Fujita, even in his early years, was fascinated by detail. His weather maps revealed storms in ways others did not.

Few may have ever heard of Ted Fujita, the professor of meteorology who diligently researched the paths of tornados after arriving from Japan in the 1950s. But his work changed the way people in the United States prepare for weather, go about their lives and travel by air.

His ways of analyzing wind, weather and related phenomena were unique, based on his own home-brewed forensic techniques. His mix of keen observation, statistical acumen and persistent pursuit of details are a guidepost for anyone pursuing any kind of analysis. Determination to counter and rise above the criticism of staid colleagues helped too.

He brought to the world a greater understanding of tornados, which were still a near mystical event in the American Midwest in the early and middle 20th Century, but one that gained greater attention as military aviation and commercial air travel expanded. His endeavors also included a vital study toward the end of his career of the effect of wind shear on passenger jet takeoffs and landing. In all his work, he took on the heroic role of applied scientist. He was the creator of the F scale for measuring tornado damage, and a force behind the use of Doppler radar in aviation.Fujita is the subject of current PBS American Episode dramatically entitled “Mr. Tornado.”

The show begins by covering his early life in Japan – his study of astronomy to ascertain tidal motions that could threaten the local clam fishermen, his penchant to wander about in storms taking measurements, making weather maps for local school teachers, and such.
A very haunting what-if cloaks Fujita, whoe served as a high school teacher in Kikura, Japan during World War II. This placed him in the city that was the original target for the second US atomic bomb. On August 9, 1945 the B-29 carrying the great weapon encountered fog over Kikura – over Fujita — and eventually headed to its secondary target, Nagasaki.

Not long thereafter, he joined a crew tasked to map the Nagasaki devastation, and to estimate at what altitude the bomb went off. His study on the ground gave him a native feel for the force of downdraft, which was a recurring theme in his studies. The techniques he applied at Nagasaki would reappear in his later studies of the US tornado belt. And the pattern of ground damage there he would see again in the wake of a historic air crash near JFK in New York City in the 1980s.

Fujita’s early life saw a grand fascination on his part with weather in all its small eddies and rivulets. That was coupled with resolute but creative application of analysis to understanding the problem.

The PBS show starts out with a big what-if. As a high school teacher in Kikura, Japan during World War II, he was in the city that was the original target of the second US atomic bomb. On August 9, 1945 the B-29 carrying the great weapon encountered fog over Kikura, and eventually headed to its secondary target, Nagasaki.

Fujita’s drawings of weather patterns for local school teacher, which accompany his post, are fascinating. He worked with atmospheric data gathered by local weather stations, but then went further, adding his own observations. His detailed maps revealed things others overlooked. He took measures from high on the mountaintop. He measured temperature, pressure, wind speed over the course of the passing storm. Where others took hourly readings, his were more frequent.

His maps had more information. The style of the time would see the bumps and wiggles of local weather smoothed out for a macro level presentation. Fujita felt people were throwing precious information by those methods.

This was due to the fact that weather researchers at this time were looking at the planet level. It was Fujita’s want, instead, to study local thunderstorms themselves very closely.

He came to believe there were downdrafts within the thunderstorm. He presented papers on this, and somebody handed him a paper – the story goes – found in the trash at a US base. It was a paper from 1942 entitled “Non-Frontal Thunderstorms”. Fujita saw similarity to his work.

The author, Horace Beyers, was a University of Chicago professor, who was on to downdraft too. Boldly Fujita sent a copy of his own research to Beyers. (By the way, this was before personal computers and Google translator – he took a major part of his saving to purchase an English language typewriter to convey his own paper to Beyers.)

Serendipity happened. Beyers invited Fujita to come to the University of Chicago. On the flight, his first, he classified and graphed the clouds amid which he flew versus time and location. Graphs, some very delicate, all very imaginative, were a constant form of thought experiment for Fujita. The Nova show even depicts one he created toward the end of his career that mapped out the cost of rice from his birth to the time of his retirement. He graphed everything, and beautifully.

The Midwest was a great place to study thunderstorms. And, tornados! The rapidly spinning vortices of summer would appear on the horizon with little warning – then come and go like a terrible swift sword. Then Fujita would come to town.

He’d quickly study the scattered refuse, hopefully before people cleaned it up. The refuse of tornados was evidence. He’d go around and interview residents while their sensations were fresh. “Did you take any pictures?” he’d ask. “What time did it hit here?”

In the days of Walt Disney – but before GIFs – Fujita imaginatively combined still photos in animations to provide some of the first actual pictures of tornados coming down. He said there are storms within storms, and for this encountered considerable criticism at learned events. He stuck to his guns. He eventually became a full professor at the University of Chicago.

Most any viewer of his graphs or drawings would walk away with a sense that these somewhat resemble Japanese calligraphic arts. In his drawings I find some of the feel of the work of artist, Katsushika Hokusai, whose classic drawing of waves in motion – “Under the Wave off Kanagawa” — is a fractal fiend’s delight that appears whenever tsunamis, financial, natural or other, hit the news. Hokusai reflected deeply on weather and its forms, and, in a way, like Fujita, he ruminates on the effect of dangerous weather on commerce in prints such as “Express Delivery Boats Rowing Through Waves”.

How would Ted approach the pandemic? It’s on my mind in these days of COVID-19. One would guess he would: Analyze travel patterns within travel patterns, chart scatter plots, scour victim’s apartments, get verbal accounts, trace their friends and friends’ friends for infection and graph the data in a way others would not contrive.

I found this story of Fujita and his life’s work stirring, dub him the Watcher of the Airways who saw what others saw not, and hereby recommend him for a place in the pantheon of Great Analyses.