NOW REPORTING FROM BALTIMORE. A private, non-commercial blog begun in support of the federal Meaningful Use REC initiative, and Health IT and Heathcare improvement more broadly. Moving now toward important broader STEM and societal/ethics topics. Formerly known as "The REC Blog."
NOTE: Comments are moderated, thanks to trolls and bots.
From a brilliant longtime AI insider with intimate access to the world of Sam Altman's OpenAI from the beginning, an eye-opening account of arguably the most fateful tech arms race in history, reshaping the planet in real time, from the cockpit of the company that is driving the frenzy
When AI expert and investigative journalist Karen Hao first began covering OpenAI in 2019, she thought they were the good guys. Founded as a nonprofit with safety enshrined as its core mission, the organization was meant, its leader Sam Altman told us, to act as a check against more purely mercantile, and potentially dangerous, forces. What could go wrong?
Over time, Hao began to wrestle ever more deeply with that question. Increasingly, she realized that the core truth of this massively disruptive sector is that its vision of success requires an almost unprecedented amount of resources: the “compute” power of high-end chips and the processing capacity to create massive large language models, the sheer volume of data that needs to be amassed at scale, the humans “cleaning up” that data for sweatshop wages throughout the Global South, and a truly alarming spike in the usage of energy and water underlying it all. The truth is that we have entered a new and ominous age of empire: only a small handful of globally scaled companies can even enter the field of play. At the head of the pack with its ChatGPT breakthrough, how would OpenAI resist such temptations?
Spoiler alert: it didn’t. Armed with Microsoft’s billions, OpenAI is setting a breakneck pace, chased by a small group of the most valuable companies in human history—toward what end, not even they can define. All this time, Hao has maintained her deep sourcing within the company and the industry, and so she was in intimate contact with the story that shocked the entire tech industry—Altman’s sudden firing and triumphant return. The behind-the-scenes story of what happened, told here in full for the first time, is revelatory of who the people controlling this technology really are. But this isn’t just the story of a single company, however fascinating it is. The g forces pressing down on the people of OpenAI are deforming the judgment of everyone else too—as such forces do. Naked power finds the ideology to cloak itself; no one thinks they’re the bad guy. But in the meantime, as Hao shows through intrepid reporting on the ground around the world, the enormous wheels of extraction grind on. By drawing on the viewpoints of Silicon Valley engineers, Kenyan data laborers, and Chilean water activists, Hao presents the fullest picture of AI and its impact we’ve seen to date, alongside a trenchant analysis of where things are headed. An astonishing eyewitness view from both up in the command capsule of the new economy and down where the real suffering happens, Empire of AI pierces the veil of the industry defining our era.
To use a DNA analogy, genomic diversity is “adaptive” precisely because–mixing my metaphors–“you can’t ever step in the same river twice.” apropos, see @brianklaas’s killer book “Flukes.”#LLMinbreeding is as maladaptive as genetic inbreeding. House of Windsor, anyone?
OK, I was not hip to her until reading a new Science Magazine review of her current book The AI Mirror. Bought her prior release as well (I have no life).
YOU GOTTA READ THIS BOOK
...[M]ost commercial AI systems today are powered by a machine learning model trained on a large body of data relevant to a specific task, then fine-tuned to optimize its performance on that task.
This approach to AI has made rapid progress in widening machine capabilities, particularly in tasks using language, where we have the most data to train with. Indeed, since so many kinds of cognitive tasks are language-enabled, most experts now regard the term “Narrow AI” as outmoded, much like its predecessor label “Weak AI.” Very large language models, like OpenAI’s various iterations of GPT or Google DeepMind’s Gemini, can now do an impressively wide variety of things: answer questions, generate poems, lyrics, essays, or spreadsheets, even write and debug software code. Large image models can generate drawings, animations, synthetic photos or videos. While such models have a considerable speed advantage over human performance of these tasks, the quality and reliability of their outputs is often well below the peak of human ability. Still, some see evidence of progress toward AGI in their widening scope of action and the flexibility of a single base model to be fine-tuned for many new tasks. While a large language model (LLM) can’t solve a problem unless the solution is somehow embedded in the language data it is trained on, multimodal models trained on many types of data (text, image, audio, video, etc.) are expanding the performance range of AI models still further.
Even if it no longer makes sense to call these tools “narrow” AI, they remain below the threshold of general intelligence—AGI. But it’s a mistake to explain that in terms of the problems they can’t yet solve. The true barrier to AGI is that AI tools today lack any lived experience, or even a coherent mental model, of what their data represent: the world beyond the bits stored on the server. This is why we can’t get even the largest AI models to reliably reflect the truth of that world in their outputs. The world is something they cannot access and, therefore, do not know. You might think there’s an easy fix: pair an AI model with a robot and let the robot’s camera and other sensors experience the world! But to an AI model, a robot’s inputs are just another data dump of ones and zeros, no different from image and sound files scraped from the Internet. These ones and zeros don’t organize themselves into the intelligent awareness of an open and continuous world. If they did, the field of intelligent robotics—including driverless cars, social robots, and robots in the service industry—would be progressing much faster. In 2015, fully automated cars and trucks were predicted to be everywhere by the 2020s. Yet in 2023, robotaxis piloted in San Francisco were still driving over firehoses, getting stuck in wet concrete, blocking intersections during busy festival traffic, violating basic rules of the road, obstructing emergency vehicles—even dragging a helpless pedestrian.4 It’s not just driving: the real-world performance of most twenty-first-century commercial robots has lagged well behind AI tools for solving language-based tasks. So, what’s the problem?
A world is an open-ended, dynamic, and infinitely complex thing. A data set, even the entire corpus of the Internet, is not a world. It’s a flattened, selective digital record of measurements that humans have taken of the world at some point in the past. You can’t reconstitute the open, infinite, lived, and experienced world from any data set; yet data sets are all that any AI model has. You might say, “But surely this is true of the human brain as well! What more do we have than data streams from our eyes, ears, noses, and so on?” But your analog, biological brain remains a far more complex and efficient system than even the most powerful digital computer. In the words of theoretical physicist Michio Kaku, “Sitting on your shoulders is the most complicated object in the known universe.”5 It was built over hundreds of millions of years to give you something no AI system today has: an embodied, living awareness of the world you inhabit. This is why we ought to regard AI today as intelligent only in a metaphorical or loosely derived sense. Intelligence is a name for our cognitive abilities to skillfully cope with the world we awaken in each day.6 Intelligence in a being that has no world to experience is like sound in a vacuum. It’s impossible, because there’s no place for it to be.
We humans do inhabit and experience a world, one rich with shared meaning and purpose, and, therefore, we can easily place the outputs of our latest AI tools within that context of meaning. We call these outputs “intelligent” because their form, extracted entirely from aggregated human data, unsurprisingly mirrors our own past performances of skilled coping with the world. They reflect back to us images of the very intelligence we have invested in them. Yet accuracy and reliability remain grand challenges for today’s AI tools, because it’s really hard to get a tool to care about the truth of the world when it doesn’t have one. Generative AI systems in particular have a habit of fabricating answers that are statistically plausible, but in fact patently false. If you ask ChatGPT to tell you about me and my career, it usually gets a lot right, but it just makes up the rest. When my host at a festival I was speaking at used ChatGPT to write my bio for the live audience, the tool listed in a confident tone a series of fictitious articles I haven’t written, named as my coauthors people that I’ve never met, and stated that I graduated from the University of California at Berkeley (I have never studied there).
Importantly, these are not errors. Error implies some kind of failure or miscalculation. But these fabrications are exactly what ChatGPT is designed to do—produce outputs that are statistically plausible given the patterns of the input. It’s very plausible that someone who holds a distinguished professorial chair at a prestigious world university received her degree from another prestigious world university, like UC Berkeley. This fabrication is far more plausible, in fact, than the truth—which is that, due to harsh economic and family circumstances, after high school I attended a local community college in-between full-time work shifts, and later received my bachelor’s degree from a low-ranked (but dirt-cheap and good-quality) commuter university that offered night classes. When I was offered a PhD scholarship at age 25, I became a full-time student again after eight years in the workforce. I first set foot in a college dorm in my 40s, as a university professor. My story isn’t common. And that’s precisely why ChatGPT selected a more “fitting” story for me; quite literally, one that better “fit” the statistical curves of its data model for academic biographies. Later, we’ll consider the cost of relying on AI tools that smooth out the rough, jagged edges of all our lives in order to tell us more “fitting” stories about ourselves.
These systems can perform computations on the world’s data far faster than we can, but they can’t understand it, because that requires the ability to conceive of more than mathematical structures and relationships within data. AI tools lack a “world model,” a commonsense grasp and flowing awareness of how the world works and fits together. That’s what we humans use to generalize and transfer knowledge across different environments or situations and to solve truly novel problems. AI solves problems too. Yet despite the common use of the term “artificial neural network” to describe the design of many AI models, they solve problems in a very different way than our brains do. AI tools don’t think, because they don’t need to. As this book explains, AI models use mathematical data structures to mimic the outputs of human intelligence—our acts of reasoning, speech, movement, sensing, and so on. They can do this without having the conscious thoughts, feelings, and intentions that drive our actions. Often, this is a benefit to us! It helps when a machine learning model’s computations solve a problem much faster than we could by thinking about it. It’s great when an AI tool finds a new, more efficient solution hidden somewhere in the math that you’d never look for. But your brain does much, much better than AI at coping with the countless problems the world throws at us every day, whose solutions aren’t mathematically predefined or encoded in data...
Vallor, Shannon. The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking (pp. 22-26). Oxford University Press. Kindle Edition.
Above: yeah, that'd be @BobbyGcyborg in the wake of the Singularity.
My more plausible impending next-life future. 🤣
Lotta good stuff in Science Magazine of late. to wit:
When Iosif Gidiotis began his doctoral studies in educational technology this year, he was intrigued by reports that new tools powered by artificial intelligence (AI) could help him digest the literature in his discipline. With the number of papers burgeoning—across all of science, close to 3 million were published last year—an AI research assistant “sounds great,” says Gidiotis, who is studying at the KTH Royal Institute of Technology. He hoped AI could find more relevant papers than other search tools and summarize their highlights.
He experienced a bit of a letdown. When he tried AI tools such as one called Elicit, he found that only some of the returned papers were relevant, and Elicit’s summaries weren’t accurate enough to win him over. “Your instinct is to read the actual paper to verify if the summary is correct, so it doesn’t save time,” he says. (Elicit says it is continuing to improve its algorithms for its 250,000 regular users, who in a survey credited it with saving them 90 minutes a week in reading and searching, on average.)
Created in 2021 by a nonprofit research organization, Elicit is part of a growing stable of AI tools aiming to help scientists navigate the literature. “There’s an explosion of these platforms,” says Andrea Chiarelli, who follows AI tools in publishing for the firm Research Consulting. But their developers face challenges. Among them: The generative systems that power these tools are prone to “hallucinating” false content, and many of the papers searched are behind paywalls. Developers are also looking for sustainable business models; for now, many offer introductory access for free. “It is very difficult to foresee which AI tools will prevail, and there is a level of hype, but they show great promise,” Chiarelli says.
Like ChatGPT and other large-language models (LLMs), the new tools are “trained” on large numbers of text samples, learning to recognize word relationships. These associations enable the algorithms to summarize search results. They also identify relevant content based on context in the paper, yielding broader results than a query that uses only keywords. Building and training an LLM from scratch is too costly for all but the wealthiest organizations, says Petr Knoth, director of CORE, the world’s largest repository of open-access papers. So Elicit and others use existing open-source LLMs trained on a wide array of texts, many nonscientific...
So, of course, I hopped on over to the Elicit website to rummage around and check things out.
Okeee Dokeee, then...
Good reasoning is reasoning that reliably arrives at true beliefs and good decisions.
Good reasoning is rare but essential for government, companies, and research.
Advanced AI systems are an opportunity to radically scale up good reasoning.
Getting this right is crucial for guiding the transition to a society that runs on machine thinking.
"A society that runs on machine thinking." Hmmm... what could possibly go wrong?
What is good reasoning?
We want people, organizations, and machines to arrive at:
True beliefs
Good decisions
Good reasoning is reasoning that reliably leads to these outcomes.
The status quo
Right now, good reasoning is rare. Bad reasoning is everywhere:
Governments allocate resources for pandemics and similar risks based on biases, political pressures, or vested interests
Courts judge cases based on selective, fallacious, or unfair use of evidence, arguments, and law
Companies decide their strategies based on overconfidence, underestimation, or complacency
Investors allocate capital based on herd mentality, hype, or fear
Researchers do work that is unimpactful or harmful, and manipulate or hide data
Yeah. No pick with any of that.
But, "why do humans reason?" Well, predominantly to win the argument (Sperber & Mercier's "Adaptive Utility"). Should verifiable truth happen along the way, so much the better.
I sent an email to my son Nick and his lifelong best friend Nate:
This stuff is serious interest to me. One of my long-term questions has been, “can AI do argument analysis?“
What is “argument analysis?“ Actually the full phrase is “argument analysis and evaluation.“ I got hip to it in graduate school at UNLV. The “analysis“ part is basically where you “flowchart“ the logic in a prose piece making the case for some assertion of truth. All of the “if/then/therefore/else contingency stuff. The “evaluation“ part that follows is basically then an argument of YOURS assessing the relative strengths and weaknesses of the original proffer set forth by the author. The difficulty comes when you consider that everyone has a significantly different way of expressing the same ideas in prose writing. Language fluidity (not to mention the 100+ different human languages around the world these days).
Basically, I went through the JAMA paper paragraph by paragraph, flowcharting the “logic” that comprised the sub-arguments leading to the final conclusion (took me probably 100 hours). I wrote about this also on my blog, here:
Imagine trying to do this shit manually to a full length book. You’d fucking never be heard from again.
If someone could come up with a “AI/LLM“ app that could parse the reasoning in a complex scientific or technical paper (or book), or, say, a complex political policy argument, it would totally kick ass. My 3GL/4GL "structured" RDBMS programming skills are two generations out of date by now. But, logic is logic, and the same basic principles apply – I assume.
Just looking for a bit of feedback. You guys are way smart.
Pop
PS- looking into this “Elicit” company cited in the article, oh, geez, they too are trying to “make the world a better place.” 🤣
'eh?
Nick, our baby, is now 40. He started his current job as the Operations Manager for a custom beverage packaging & distribution company based here in Baltimore. He's transitioned of late into an IT management role. He has degrees in accounting and finance, and operations management. He and his bud Nate, a year younger, met at The Hebrew Academy in Las Vegas when Nick was 9. We're not Jewish, but we sent him there anyway. Neither are we Catholic (I'm a fully recovered Episcopalian, dilettante UU, and Secular Zen Sympathizer), but we subsequently sent Nick to Bishop Gorman High School, also in Vegas (I have a funny story about that; I'll get to it).
Nate is absurd. He's the son of Frank Sinatra's Las Vegas bassist, Seth Kimball. Nate now has a Master's in Jazz Performance, a real estate agent license, a private pilot's license, and recently transitioned to work in software engineering for Microsoft (Azure platform). We used to call him "Number Two Son." He spent more time at our house than at home.
Courts in the United States have increasingly relied on scientific evidence and expert testimony to help resolve questions of fact. On 1 December 2023, amendments to Federal Rule of Evidence 702 will take effect, further clarifying the court’s responsibilities as a gatekeeper for expert evidence. This update comes just a few months after the 30-year anniversary of the Supreme Court’s landmark decision on how federal judges should evaluate scientific evidence. Daubert v. Merrell Dow was hailed as a victory for the use of scientific information in the legal system and certainly cast a much-needed spotlight on scientific evidence in the courtroom. But the nuanced and flexible nature of the “Daubert standard” has since led to substantial inconsistencies in its application. Most strikingly, it has had far more impact in civil cases than criminal cases. Daubert’s core tenet—that scientific evidence introduced in court should be adequately valid and reliable—needs to be taken just as seriously in the criminal justice system and for forensic science as it has been in civil cases.
Daubert instructs judges to be “gatekeepers” responsible for assessing the validity of the science brought to court. Previously, courts often asked only whether the science was “generally accepted” by the relevant scientific community. Because judges often treated an expert witness’s own assertion of general acceptance as adequate, the rule did not present much of a bar for admissibility. With Daubert, the onus is more squarely on the judge to assess validity. The opinion details numerous possible factors to consider (including testing, error rate, peer review, and general acceptance), but gives little truly concrete guidance and allows the court great flexibility in weighing these factors.
When the Daubert decision was handed down, many legal analysts and scientists agreed that the use of experts in court had long been a mess. Some critics lambasted judges for too often permitting expert testimony that wasn’t scientifically credible; others worried that juries were fundamentally incapable of making reasoned decisions when competing experts offered wildly different, contradictory testimony. Daubert shined a much-needed spotlight on expert evidence—but 30 years later, what has been its evidentiary impact?...
Jennifer Mnookin is chancellor and professor of law at the University of Wisconsin–Madison, Madison, WI, USA
702?
JDSupra F.R.E. 702 12/01/23 AMENDMENTS
In April, the Supreme Court sent a list of proposed amendments to Congress that amend the Federal Rules of Evidence. Absent action by Congress, the rules go into effect December 1, 2023. The proposed amendments affect Rules 106, 615 and, relevant to this article, 702.
Rule 702 addresses testimony by an expert witness. The proposed rule reads as follows (new material is underlined; matters omitted are lined through): A witness who is qualified as an expert by knowledge, skill, experience, training, or education may testify in the form of an opinion or otherwise if the proponent demonstrates to the court that it is more likely than not that:
(a) the expert’s scientific, technical, or other specialized knowledge will help the trier of fact to understand the evidence or to determine a fact in issue;
(b) the testimony is based on sufficient facts or data;
(c) the testimony is the product of reliable principles and methods; and
(d) the expert has reliably applied expert’s opinion reflects a reliable application of the principles and methods to the facts of the case.
The proposed amended seeks to clarify how a judge should view his or her gatekeeping role without substantively changing the rule. Assuming Congress adopts the proposed rule, the proposal amends the rule in two ways.
First, it clarifies that a court should not admit expert testimony unless the proponent demonstrates that it is “more likely than not” that the proffered testimony meets Rule 702’s admissibility requirements. The rules committee recommended the change because many courts hold that the critical question of the sufficiency of the basis for an expert’s opinions goes to the weight of the testimony, not its admissibility. Thus, to be admissible in the future, the proponent of expert testimony must demonstrate, consistent with Rule 104(a) and case law interpreting Rule 104(a), that the testimony meets admissibility requirements – i.e., meets a preponderance of the evidence standard…
Back in the mid-late 1980's I worked in a forensic-level environmental radiation lab in Oak Ridge as a systems programmer and QC analyst (we did a ton of litigation support and regulatory action analytics).
Our (now-late) founder and CEO, John A. Auxier, PhD, CHP was the nation's premier expert on radiation dose/exposure. Former Director of Industrial Health & Safety at Oak Ridge National Laboratory, he was a member of the Three Mile Island Commission, and Editor of the Health Physics Society Journal.
The governing forensic legal standard then was "the Frye Standard." One hopes that this new FRE 702 Amendment will materially improve things with respect to "Science, Justice, and Evidence” in ouor endless adversarial venues.
More to come. Still studyin' up on these Elicitpeeps (pdf). Smart bunch.
Introduction The world has changed since the rise of civilization some 10,000 years ago. Look around—what do you see? You may be sitting on a chair, in front of a computer, a glass of water next to it on the table, within the walls of a house, the neighbor’s dog barking outside, cars driving by on asphalt roads built next to power lines supplying electricity for the local school or hospital. Almost every part of your environment has been shaped by humans; it is there because we intend for it to be there, or at least approve of its existence.
This has arguably been a change for the better—as indicated by the existence of the notion of progress—even if not without exceptions, and not without contention. Basic human needs such as food, shelter, health, and physical safety are provided to a degree far beyond hunter-gatherer times. What is responsible for this change? While it may be difficult to pin down the relative contributions of different causes and enabling factors, it is safe to say that our capability for thought was a necessary ingredient…
Chapter 2 Background: Probabilistic Programming A probabilistic program is a program in a universal programming language with primitives for sampling from probability distributions, such as Bernoulli, Gaussian, and Poisson. Execution of such a program leads to a series of computations and random choices. Probabilistic programs thus describe models of the stochastic gen- eration of results, implying a distribution on return values. Most of our examples use Church [26], a probabilistic programming language based on the stochastic lambda calculus. This calculus is universal in the sense that it can be used to define any com- putable discrete probability distribution [49] (and indeed, continuous distributions when encoded via rational approximation)…
Reasoning about Reasoning as Nested Conditioning 4.1 Introduction Reasoning about the beliefs, desires, and intentions of other agents—theory of mind—is a central part of human cognition and a critical challenge for human-like artificial intelligence. Reasoning about an opponent is critical in competitive situations, while reasoning about a compatriot is critical for cooperation, communication, and maintaining social connections. A variety of approaches have been suggested to explain humans’ theory of mind. These include informal approaches from philosophy and psychology, and formal approaches from logic, game theory, artificial intelligence, and, more recently, Bayesian cognitive science.
Many of the older approaches neglect a critical aspect of human reasoning uncertainty—while recent probabilistic approaches tend to treat theory of mind as a special mechanism that cannot be described in a common representational framework with other aspects of mental representation. In this chapter, we discuss how probabilistic programming, a recent merger of programming languages and Bayesian statistics, makes it possible to concisely represent complex multi-agent reasoning scenarios. This formalism, by representing reasoning itself as a program, exposes an essential contiguity with more basic mental representations.
Probability theory provides tools for modeling reasoning under uncertainty: distributions formalize agents’ beliefs, conditional updating formalizes updating of beliefs based on evidence or assertions. This approach can capture a wide range of reasoning patterns, including induction and non-monotonic inference. In cognitive science, probabilistic methods have been very successful at capturing aspects of human learning and reasoning [92]. However, the fact that conditioning is an operation applied to such models and not itself represented in such models makes it difficult to accommodate full theory of mind: We would like to view reasoning as probabilistic inference and reasoning about others’ reasoning as inference about inference; however, if inference is not itself represented as a probabilistic model we cannot formulate inference about inference in probabilistic terms.
Probabilistic programming is a new, and highly expressive, approach to probabilistic modeling. A probabilistic program defines a stochastic generative process that can make use of arbitrary deterministic computation. In probabilistic programs, conditioning itself can be defined as an ordinary function within the modeling language. By expressing conditioning as a function in a probabilistic program, we rep- resent knowledge about the reasoning processes of agents in the same terms as other knowledge. Because conditioning can be used in every way an ordinary function can, including composition with arbitrary other functions, we may easily express nested conditioning: we can condition any random variable, including random variables that are defined in terms of other conditioned random variables. Nested conditioning de- scribes reasoning about reasoning and this makes theory of mind amenable to the kind of statistical analysis that has been applied to the study of mental representation more generally.
The probabilistic program view goes beyond other probabilistic views by extending compositionality from a restricted model specification language to a Turing-complete language, which allows arbitrary composition of reasoning processes. For example, the multi-agent influence diagrams proposed by Koller and Milch [45] combine the expressive power of graphical models with the analytical tools of game theory, but their focus is not on representing knowledge that players’ might have about other players’ reasoning…
In the wake of studying Mustafa Suleyman's new book "The Coming Wave," I've begun interacting a bit with his Pi.ai app.
e.g., a question I posed yesterday:
ME "In 1994, the journal of the American Medical Association, published a paper by an organization known as “Physicians for a National Healthcare Plan.“ In the paper, they argue for what is known as a “single payer system“ for healthcare. Can you analyze and evaluate their argument for me?"
Pi "The 1994 JAMA paper by Physicians for a National Healthcare Plan (PNHP) argued for the adoption of a single-payer healthcare system in the US. Their main argument was that such a system would provide universal coverage, eliminate the need for private insurance, and reduce administrative costs, thereby improving access to healthcare and reducing overall costs. They proposed that a national formulary would be used to determine which prescription drugs and medical supplies would be covered, and that an expert panel would oversee the program to ensure its effectiveness. Their argument was based on the premise that all Americans should have access to quality healthcare, and that a single-payer system was the most efficient and cost-effective way to achieve this. However, their proposal was met with criticism from those who argued that it would lead to a decline in the quality of care and a loss of patient choice."
Pi.ai left me a bit less than dazzled here. The response was "accurate," but nowhere in the depth I wanted to see. Sounded like a Google search result. But, it was lightening fast, gotta give 'em that.
Q*
In AI research—and, really, in all of science—the rise and fall of ideas is not based on pure meritocracy. Usually, the scientists and companies with the most resources and the biggest loudspeakers exert the greatest influence. Consensus forms around these entities, which effectively means that they determine the direction of AI development. Within the AI industry, power is already consolidated in just a few companies—Meta, Google, OpenAI, Microsoft, and Anthropic. This imperfect process of consensus-building is the best we have, but it is becoming even more limited because the research, once largely performed in the open, now happens in secrecy.
Over the past decade, as Big Tech became aware of the massive commercialization potential of AI technologies, it offered fat compensation packages to poach academics away from universities. Many AI Ph.D. candidates no longer wait to receive their degree before joining a corporate lab; many researchers who do stay in academia receive funding, or even a dual appointment, from the same companies. A lot of AI research now happens within or connected to tech firms that are incentivized to hide away their best advancements, the better to compete with their business rivals… via @TheAtlantic"Why Won’t OpenAI Say What the Q* Algorithm Is?"
COMPLETING THE PICTURE (FOR NOW)
AGI software (e.g., "Scheme" & "Church") has to run on cutting-edge hardware–e.g., the GPU technology. Massively parallel computing.
"AI and the Nvidia GPU"
The revelation that ChatGPT, the astonishing artificial-intelligence chatbot, had been trained on an Nvidia supercomputer spurred one of the largest single-day gains in stock-market history. When the Nasdaq opened on May 25, 2023, Nvidia’s value increased by about two hundred billion dollars. A few months earlier, Jensen Huang, Nvidia’s C.E.O., had informed investors that Nvidia had sold similar supercomputers to fifty of America’s hundred largest companies. By the close of trading, Nvidia was the sixth most valuable corporation on earth, worth more than Walmart and ExxonMobil combined. Huang’s business position can be compared to that of Samuel Brannan, the celebrated vender of prospecting supplies in San Francisco in the late eighteen-forties. “There’s a war going on out there in A.I., and Nvidia is the only arms dealer,” one Wall Street analyst said…
With affability, humility, and synaptic smog-clearing clarity, Dr. Erik J. Larson Takes No Prisoners.
Yikes, indeed.
In sum, we can dial back the ominously foreboding hyperbole regarding the incipient homo sapiens-enslaving / exterminating Singularity said to soon be wrought by AI. 320 pages of Chill Pills. 320 doses of fast-acting Naloxone HCl antidote to AGI Cognitive Pearl-Clutching Disorder.
One fumbles to know where to begin.
I love it when I learn stuff (unlearn, mostly, anymore). Even when it entails humbling internal reactions of "how the [bleep] have you missed that all these decades?"
In 1980, at the age of 34, divorced w/ custody of my two girls, I found it prudent and necessary to give up my then-16 years of hardscrabble roadhouse touring musician life (or, as my wife Cheryl calls it, "working in the not-for-profit sector") and enroll in undergraduate school at UTK.
I thrived, ravenously consuming the likes of deductive logic, inductive logic, philosophy of science, various flavors of ECON, the gamut of statistics courses, and experimental psychology and psychometrics.
In addition to my awesome Linear Regression text (written by my Prof, Mary Sue Younger), I still have my UTK deductive logic textbook in my hardcopy stash.
Looking back now through the index, I find no reference to either "abductive inference" or Charles Sanders Peirce. What? Who?
Erik Larson:
...[C]ommon sense is itself mysterious, precisely because it doesn’t fit into logical frameworks like deduction or induction. Abduction captures the insight that much of our everyday reasoning is a kind of detective work, where we see facts (data) as clues to help us make sense of things. We are extraordinarily good at hypothesizing, which is, to Peirce’s mind, not explainable by mechanics but rather by an operation of mind which he calls, for lack of another explanation, instinct. We guess, out of a background of effectively infinite possibilities, which hypotheses seem likely or plausible.
We must account for this in building an intelligence, because it is the starting point for any intelligent thinking at all. Without a prior abductive step, inductions are blind, and deductions are equally useless.
Induction requires abduction as a first step, because we need to bring into observation some framework for making sense of what philosophers call sense-datum—raw experience, uninterpreted. Even in simple induction, where we induce a general statement that All swans are white from observations of swans, a minimal conceptual framework or theory guides the acquisition of knowledge. We could induce that all swans have beaks by the same inductive strategy, but the induction would be less powerful, because all birds have beaks, and swans are a small subset of birds. Prior knowledge is used to form hypotheses. Intuition provides mathematicians with interesting problems.
When the developers of DeepMind claimed, in a much-read article in the prestigious journal Nature, that it had mastered Go “without human knowledge,” they misunderstood the nature of inference, mechanical or otherwise. The article clearly “overstated the case,” as Marcus and Davis put it. In fact, DeepMind’s scientists engineered into AlphaGo a rich model of the game of Go, and went to the trouble of finding the best algorithms to solve various aspects of the game—all before the system ever played in a real competition. As Marcus and Davis explain, “the system relied heavily on things that human researchers had discovered over the last few decades about how to get machines to play games like Go, most notably Monte Carlo Tree Search … random sampling from a tree of different game possibilities, which has nothing intrinsic to do with deep learning. DeepMind also (unlike [the Atari system]) built in rules and some other detailed knowledge about the game. The claim that human knowledge wasn’t involved simply wasn’t factually accurate.” A more succinct way of putting this is that the DeepMind team used human inferences—namely, abductive ones—to design the system to successfully accomplish its task. These inferences were supplied from outside the inductive framework.
Larson, Erik J.. The Myth of Artificial Intelligence (pp. 161-162). Harvard University Press. Kindle Edition.
(BTW: Not even in graduate school did I encounter C.S. Peirce or abductive inference. An inexplicable, significant omission, given what I now learn.)
I am not kidding about Erik J. Larson's new book. Another great place within which to hide a $100 bill from Your Favorite President Donald Trump.
This has been a compelling and fun read. I think the author has sustained his case.
In the pages of this book you will read about the myth of artificial intelligence. The myth is not that true AI is possible. As to that, the future of AI is a scientific unknown. The myth of artificial intelligence is that its arrival is inevitable, and only a matter of time—that we have already embarked on the path that will lead to human-level AI, and then superintelligence. We have not. The path exists only in our imaginations. Yet the inevitability of AI is so ingrained in popular discussion—promoted by media pundits, thought leaders like Elon Musk, and even many AI scientists (though certainly not all)—that arguing against it is often taken as a form of Luddism, or at the very least a shortsighted view of the future of technology and a dangerous failure to prepare for a world of intelligent machines.
As I will show, the science of AI has uncovered a very large mystery at the heart of intelligence, which no one currently has a clue how to solve. Proponents of AI have huge incentives to minimize its known limitations. After all, AI is big business, and it’s increasingly dominant in culture. Yet the possibilities for future AI systems are limited by what we currently know about the nature of intelligence, whether we like it or not. And here we should say it directly: all evidence suggests that human and machine intelligence are radically different. The myth of AI insists that the differences are only temporary, and that more powerful systems will eventually erase them. Futurists like Ray Kurzweil and philosopher Nick Bostrom, prominent purveyors of the myth, talk not only as if human-level AI were inevitable, but as if, soon after its arrival, superintelligent machines would leave us far behind.
This book explains two important aspects of the AI myth, one scientific and one cultural. The scientific part of the myth assumes that we need only keep “chipping away” at the challenge of general intelligence by making progress on narrow feats of intelligence, like playing games or recognizing images. This is a profound mistake: success on narrow applications gets us not one step closer to general intelligence. The inferences that systems require for general intelligence—to read a newspaper, or hold a basic conversation, or become a helpmeet like Rosie the Robot in The Jetsons—cannot be programmed, learned, or engineered with our current knowledge of AI. As we successfully apply simpler, narrow versions of intelligence that benefit from faster computers and lots of data, we are not making incremental progress, but rather picking low-hanging fruit. The jump to general “common sense” is completely different, and there’s no known path from the one to the other. No algorithm exists for general intelligence. And we have good reason to be skeptical that such an algorithm will emerge through further efforts on deep learning systems or any other approach popular today. Much more likely, it will require a major scientific breakthrough, and no one currently has the slightest idea what such a breakthrough would even look like, let alone the details of getting to it… [Erik J. Larson, pp 1-2]
I think I can safely continue to exclude The Singularity from BobbyG's list of priority exigencies.
It remains true that technology often acts like a prosthetic to human abilities, as with the telescope and microscope. AI has this role to play, at least, but a mythology about a coming superintelligence should be placed in the category of scientific unknowns. If we wish to pursue a scientific mystery directly, we must at any rate invest in a culture that encourages intellectual ideas—we will need them, if any path to artificial general intelligence is possible at all.[Erik J. Larson, pg 280]
My intial interest in the topic went to all of the Gartner Hype Cycle swooning over AI's (overstated) clinical potential in Health IT. Beyond that, I mused about this stuff (click the robot thinker graphic):
My dubiety is hardly attenuated in the wake of reading Erik's book.
More broadly,
"Someone has to have an idea."
Yeah, and where do we get them?
“Because of livewiring, we are each a vessel of space and time. We drop into a particular spot on the world and vacuum in the details of that spot. We become, in essence, a recording device for our moment in the world.
When you meet an older person and feel shocked by the opinions or worldview she holds, you can try to empathize with her as a recording device for her window of time and her set of experiences. Someday your brain will be that time-ossified snapshot that frustrates the next generation.
Here’s a nugget from my vessel: I remember a song produced in 1985 called “We Are the World.” Dozens of superstar musicians performed it to raise money for impoverished children in Africa. The theme was that each of us shares responsibility for the well-being of everyone. Looking back on the song now, I can’t help but see another interpretation through my lens as a neuroscientist.
We generally go through life thinking there’s me and there’s the world. But as we’ve seen in this book, who you are emerges from everything you’ve interacted with: your environment, all of your experiences, your friends, your enemies, your culture, your belief system, your era—all of it.
Although we value statements such as “he’s his own man” or “she’s an independent thinker,” there is in fact no way to separate yourself from the rich context in which you’re embedded. There is no you without the external. Your beliefs and dogmas and aspirations are shaped by it, inside and out, like a sculpture from a block of marble. Thanks to livewiring, each of us is the world.”
— Livewired: The Inside Story of the Ever-Changing Brain by David Eagleman pp 244-245
The fuel source of abductive inference?
UPDATES
Some of my episodic prior posts discussing "robots." How about "Deep Fakes?"
Mathematician Bill Dembski'slengthy review of Erik's book.
Erik Larson’s THE MYTH OF ARTIFICIAL INTELLIGENCE is far and away the best refutation of Kurzweil’s overpromises, but also of the hype pressed by those who have fallen in love with AI’s latest incarnation, which is the combination of big data with machine learning. Just to be clear, Larson is not a contrarian. He does not have a death wish for AI. He is not trying to sabotage research in the area (if anything, he is trying to extricate AI research from the fantasy land it currently inhabits). In fact, he has been a solid contributor to the field, coming to the problem of strong AI, or artificial general intelligence (AGI) as he prefers to call it, with an open mind about its possibilities...
Yep. I give Dembski's review five stars as well—one star for each 1,000 words. /s
Cool guy.
OFF-TOPIC PERSONAL ERRATUM
22 months ago I was officially dx'd with Parkinson's. Until a month ago I've avoided the Rx, but I'm now doing the Carbidopa/Levodopa 25/100 regimen (6 tabs/day, 2 at a time). I find it problematic. Might yet be too early to assess efficacy, though. My side-effects thus far are pretty much just elevated standing balance issues. I was already a fall risk. Sux. But, so do the tremors, which are messing with my typing, mouse use, and guitar playing (and manual dexterity more broadly).
Mixing the film metaphor riffs, "To Make Benefit Glorious Nation of Sinemetistan."
