<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Dave Fay on Medium]]></title>
        <description><![CDATA[Stories by Dave Fay on Medium]]></description>
        <link>https://medium.com/@davefay?source=rss-30a4cde40e37------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Qbn0oNmwCyw3M-GnOrsYxA.jpeg</url>
            <title>Stories by Dave Fay on Medium</title>
            <link>https://medium.com/@davefay?source=rss-30a4cde40e37------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 10:07:03 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@davefay/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Here’s What LLMs Predict. Here’s What They Also May Be Able To Predict, but Probably Shouldn’t.]]></title>
            <link>https://medium.com/@davefay/heres-what-llms-predict-here-s-what-they-also-may-be-able-to-predict-but-probably-shouldn-t-cb8352730ba2?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb8352730ba2</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[anthropics]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[loans]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Sun, 22 Mar 2026 15:18:48 GMT</pubDate>
            <atom:updated>2026-05-01T13:39:14.317Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>Here’s What LLMs Predict. Here’s What They Also <em>May</em> Be Able To Predict, but Probably Shouldn’t.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*fUDjbDsFZ5PAIe8n" /><figcaption>Trees are valuable organic machines even without human-grade intelligence. There is also room for somewhat-intelligent, but not LLM-intelligent, digital machines.</figcaption></figure><p>This essay is another attempt in my futile efforts to guide the LLM hysteria I am surrounded by in my professional life. I like to position myself as a techno-realist who has weathered the frustrating episodes of blockchain / crypto / NFT-mania, but I’m usually brushed off as an unsophisticated luddite. That’s my cross to bear, though I still have confidence in the conservatism of my estimates on how much “juice” LLMs have with respect to changing humanity forever. It helps that Yann Lecun both inspired and shares most of my views.</p><p>My latest exploration of LLMs-in-industry concerns the use of LLMs to help predict consumer creditworthiness — think of the prompt “we are going to use AI to build the best credit score there has ever been!” My instinctual reaction to this suggestion was multifaceted — there are regulatory reasons why this probably wouldn’t fly — but the more interesting objection to this suggestion is whether or not LLMs are <em>architecturally appropriate</em> to furnish a credit score, or estimate of consumer probability of default. Expanding on the topic of why or why-not LLMs are well suited to this category of prediction will illuminate larger categories of problems that LLMs <em>are</em> well suited to addressing and those they are not.</p><p>So, in general terms, what <em>is </em>an LLM, what do they predict, and how do they work? Put simply, LLMs can be thought of as massively complex <a href="https://youtube.com/shorts/dwSQppujy6s?si=ll5IF9JOzpzjPuaB">pachinko machine</a>, where a series of input signals (balls) are fed into the top of the machine, and various weights and functions (pins) bounce the balls along a statistically-oriented path to some useful terminal classification (landing position). There are algorithms (rules) that are fed into the machine during a training process, which determine how well each ball in the machine “landed” and tweak the bouncing behavior of each pin to ensure subsequent balls land more appropriately given some definition of success. If you scale this metaphor up to billions of pins (model parameters), you have a typical structure of an artificial neural network — a computational structure devised in the 1950’s inspired by our own neurons.</p><p>LLMs are, in effect, large artificial neural networks with a clever take on how best to structure input data and how to tweak the pins in the pachinko machine to help the balls land in valuable spots regarding the structure of human language. Said differently and more technically, the famous “<a href="https://arxiv.org/abs/1706.03762">Attention is All You Need</a>” paper that heralded the age of LLMs indicated that by tuning an LLM to “pay attention” to the tokens (bits of language) that come before and after a given token of language, one can build a digital pachinko machine that can be well-tuned into a model of human language. This is what LLMs, as the acronym gives away, are — they are highly parameterized neural network models of human language that can, given an input series of language tokens (prompt), predict what language should come after it given the signals it optimizes toward in a training dataset.</p><p>LLMs have been elaborated upon in some fascinating ways beyond simple language-token generation. By adding digital images into the training corpus, the pachinko machines have learned the nuanced relationship between input tokens (“create a picture of a bird swimming under the ocean”) and digitized images that correlate to those themes. Generative algorithms leverage those statistically inferred relationships to generate images and videos based on text prompts. Products like Sora are, to this author, truly incredible and profoundly dangerous, though that is a topic beyond the discussion of this essay.</p><p>Given all the above, what do we think — could an LLM, either the language-focused versions like ChatGPT or Claude or the image/video versions like Sora — be used to create the “ultimate credit score”? Well, let’s compare and contrast what LLMs do versus what a credit score does.</p><p>A credit score is a monotonic indicator of whether or not someone will repay a loan you are about to give them according to some strict definition of a discrete non-repayment event. In other words, as a credit score goes up, the likelihood that a person will not repay a given loan generally goes down. This event is generally categorized as a one or a zero in a training dataset, and various model training algorithms (logistic regression, various tree-classifiers, and so on) are pointed at the data and used to approximate the system of factors that drive consumers to default on their loans. The question at hand is whether the LLM pachinko machine is similarly well suited, or perhaps far superior, to conventional modeling algorithms at this task.</p><p>To start to answer this question, let’s consider what an LLM predicts versus what, say, a Gradient Boosted Machine (GBM) model model predicts when it predicts whether or not someone will default on a loan. An LLM uses a text input prompt to predict what text should come after a given prompt, and it does so by leveraging a training algorithm that highly privileges the longitudinal relationships across streams of language tokens. This architecture was so profoundly successful because language, as a system, is a highly recursive and irregular system that significantly benefits from this kind of “attention”. Consider the following sentences:</p><h3>“The tree grew up from a seed, which was colored tan and is a figment of my imagination, just as this sentence is.”</h3><h3>“This statement is a lie.”</h3><p>The tree is a figment of my imagination, but it grew up from a seed that is <em>also</em> a figment of my imagination — this is a sentence that recursively refers to itself in a variety of ways but nevertheless makes <em>some</em> kind of syntactic sense. The latter sentence is a famous linguistic paradox in which the sentence is both true and false simultaneously, and it demonstrates the limitations of language as a system that communicates completely about the reality we inhabit. Language is wonderful and LLMs capture it wonderfully.</p><p>Alternatively, let’s turn to a hypothetical GBM model and consider a selection of the the data is uses and what it predicts.</p><ul><li>Input feature: Last delinquent on another loan: 30 days</li><li>Input feature: Previous bankruptcy: yes</li><li>Input feature: Debt to Income: 50%</li><li>Likelihood to Default: 50%</li></ul><p>At first glance, the two systems being modeled — language versus the financial signals that indicate impeding default — seem significantly different. I would agree with this characterization. <em>Massive</em> computational resources are brought to bear on training and executing LLM-class models because of the profound nuance of the system of human language and its relationship to other dimensions of data, like the images of systems like Sora. The attention-algorithm deployed during the training of an LLM-class neural network requires unprecedented amounts of training data to arrive at a system sufficiently complex to successfully model human language. This seems like so much dramatic overkill for predicting a binary target in a system that is already, statistically speaking, relatively simple when compared to human language and is already reasonably well serviced by existing modeling technologies with many orders of magnitudes less parameters (pachinko machine pins).</p><p><strong>It should be mentioned at this stage that artificial neural networks are a <em>general</em> category of model that is <em>entirely appropriate</em> for general classification problems, e.g. “is this a cat or not?” or “will this consumer default or not?” However, LLMs in particular are highly complex, data-greedy, and expensive types of neural network that optimize toward modeling recursive and/or contextual relationships that do not seem entirely relevant to the domain of credit risk prediction.</strong></p><p>Could an LLM be used to predict credit default? <strong>Yes, LLMs could predict credit default if the appropriate data was captured and served to the LLM.</strong> For example: if every digital communication a consumer had on the internet was monitored and fed into an LLM, then I have no doubt that an LLM would be able to, given the context of all prior conversations consumers had leading up to a default event, predict impending defaults. To do so would be nearly trivial, in fact, given many of the other things LLM-class models are capable of. However, it is my contention that it is <em>profound overkill</em> to use a model that <em>understands the full features and nuances of human langage</em> to simply predict whether or not someone will repay a loan. <strong>The system of loan repayments has <em>already been digitally codified</em> into a much simpler and performant system of FCRA data, and this system can be modeled using far simpler model architectures than an LLM.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*yBtzTD1xk_JLbAaS" /><figcaption>Photo by <a href="https://unsplash.com/@jleeems?utm_source=medium&amp;utm_medium=referral">Jason Leem</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I’ll leave you with a nature-inspired analogy. Nature evolved, separately, humans and trees. Humans generated language and language has driven us to the technological heights we presently enjoy and despair. However, nature did <em>not</em> evolve humans to generate oxygen for our natural habitats. What a strange thing to say, you might be thinking — of course humans do not… “generate oxygen” for our natural habitats. Why should we? This is precisely the point I am making — nature evolved trees to, in a dramatic simplification and among a multiplicity of other reasons, consume carbon dioxide and sunlight and to produce glucose and oxygen. This simple process requires an organism with a vanishing fraction of the nervous system of a language-hosting human, and it happens to be an extremely efficient solution for producing oxygen versus getting a human with our language to do so at scale. In a similar way, not all of our somewhat-intelligent machines need to be optimized for AI goals to be valuable.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb8352730ba2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[LLMs Will Not Become True Agents. AI That Can Touch, Feel, Smell, and Sense Like We Do Will.]]></title>
            <link>https://medium.com/@davefay/llms-will-not-become-true-agents-ai-that-can-touch-feel-smell-and-sense-like-we-do-will-280a5f0d7f53?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/280a5f0d7f53</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Fri, 10 Oct 2025 15:33:02 GMT</pubDate>
            <atom:updated>2025-10-10T15:33:02.109Z</atom:updated>
            <content:encoded><![CDATA[<p>I have two young boys, one who is two at the time of this writing and one who is about 8 months old. Although the older boy is making progress on his verbal communication, his engagement with the world has been decidedly non-verbal through his first couple of years on earth. He has smashed his head on sharp coffee table corners, learned to balance himself and walk on his wobbly two legs, learned to point at the Buzz Lightyear toy he wants off the shelf — my boy is orienting himself in the physical world, but he has not yet developed the robust ability to do so with language.</p><p>Much of our human sense of the world is non-verbal and non-visual. We, of course, have a variety of sense organs — ears (hearing), nose (smell), our nervous system (touch), our tongue (taste). These senses, plus our sense of sight, combine to constitute our developing worldview as children and our mature worldview as adults. Presently, we train our generative AI models based on two primary inputs: text and video. I believe audio is implicit in the video training medium, but cannot be sure. In any case, it is the best case scenario that our current models are trained on two of our five senses and a higher-order product of our world-models — language. Critically, touch, taste, and scent are omitted from our current models. The relationship between language and our senses is worth dwelling on for a moment.</p><p>Language can be thought of as a mechanism that facilitates the binding of our five senses into an encoding of reality that can be shared across the nodes of the human computing networks. Across the countless generations of tens of billions of humans that have ever lived, language has co-evolved with our species as a mechanism for the human network to establish an encoding of the 4 dimensional reality (three dimensions of space plus time) the we exist within. Language codifies for us things in space (nouns), interactions between those things in space (verbs), tenses (time), and other mechanisms that allow humans to reason about the physical system we exist within. In other words, language is the mechanism through which humans reason about and share the worldview produced by our senses, and we extend and evolve that common world-view through communication. Language makes us smarter than the product of our individual experiences.</p><p>So, what does this mean for our contemporary AI models? When I say AI models here, I am specifically referring to transformer architecture-based LLMs like those employed by ChatGPT and diffusion models, like those that produce AI images and videos. These models are trained primarily on language (text) and video feeds, like YouTube. Recall that language is the product of the human network, but it is not itself a primary input to the early development of the computing nodes of the human network — babies. Babies develop their initial worldview by first mastering their sight and touch — they learn to walk — their taste — they learn what foods they like and which they don’t — hearing — they learn to respond to varied audio inputs — and they learn to discriminate between pleasant and unpleasant odors — my son is a consummate flower-sniffer. These five senses bootstrap a worldview that, when hosted by a biological organism, enables non-verbal interaction with and survival in our four-dimensional reality. Look at troops of animal for evidence that these senses can produce successful organisms. The languages we share supercharge this worldview by significantly extending our ability to encode and hypothetically engage with the four dimensions of our reality, but it would be impossible without the non-verbal constituent elements of our worldview.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*vQCPIRh8Ro1WIlvl" /><figcaption>Ravens may not possess a verbal language per se, but they exhibit <a href="https://www.scientificamerican.com/article/young-ravens-rival-adult-chimps-in-a-big-test-of-general-intelligence/">intelligence</a> even without language.</figcaption></figure><p>LLMs and diffusion models are not trained on the same degree of sense-data humans are, and as such they do not share the same world-view as we do. To invoke Plato, our current AI models are trained on the shadows of an object outside of a cave produced by a fire within the cave, but not on the actual items in the world themselves. Our current models can produce convincing facsimiles of the intellectual products of the human network (language, visual art), <strong>but its capabilities are bottlenecked, and probably will remain bottlenecked, because they cannot interpret our four dimensional world without us</strong>. This human-gated understanding of the world is due to the fact that our models are primarily trained on the human network’s abstractions of the world (e.g. language), but not the world itself. Our human understanding of the world, at its base foundation, is the product of our non-verbal senses, and our current AI models are not trained on that data. <strong>The implications of this limitation are worth considering, since they prevent “agents” in the kind of way the industry typically promotes them.</strong></p><p>Our LLMs and diffusion models require <em>us</em> to provide some kind of input. This input is often a text prompt, i.e. language, but can also include images and other (limited) input sources. Let’s go with the text prompt input example for now. Consider the prompt and response below, where I ask about the city of Naples, Italy.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jDCOfTx0mb5PG7VlPccOSw.png" /></figure><p>Here is an image of the city of Naples, for reference.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zCQyMKhcbZh60bF5.jpg" /><figcaption>Naples, for Reference</figcaption></figure><p>Note ChatGPT’s response — “If I could visit Naples…” followed by a sophisticated mix of language tokens regurgitating the internet’s collective knowledge about the city itself. But, of course, ChatGPT cannot visit Naples itself. It cannot hear, it cannot smell, it cannot <em>feel </em>Naples in any meaningful way. It may be able to <em>see</em> Naples, so to speak, by being fed images of the city, and it can even engage with that visual input through federating out some image-manipulation task to Sora’s diffusion model. However, to restate the point — our AI system here cannot be <em>in</em> Naples to any meaningful extent.</p><p>What does this matter? Surely this is just the ramblings of a luddite trying to justify humanity’s continued superiority in the face of technology he doesn’t understand, right? Well, though that may be the case, consider the following hypothetical situation.</p><p><em>The year is 2026. In Naples, Italy, the ground begins to shake — during certain occasions, only espresso cups shake. During other occasions, landscape paintings fall off their mounts. Residents in the foreground of Vesuvius report a sulfuric smell in the air.</em></p><p>The developments in Naples in our hypothetical situation are not common, but they are certainly not unprecedented. Should these events come to pass, we would expect the following subsequent actions to take place:</p><ul><li>Neapolitans would feel the shaking, smell the sulfur.</li><li>Neapolitans would then begin to connect these signal inputs into the verbal history of their city, with the stories of Pompeii figuring prominently.</li><li>The leadership of the city would enact some emergency measures to mitigate risk to the city’s inhabitants.</li></ul><p>Fairly straightforward, no? These series of observations and reactions doesn’t seem particularly sophisticated — a bad thing once happened in the city, signs of that bad thing happening again recurred, and the Neapolitans reacted accordingly. However, the initial step in this act of <em>human agency</em> is worth focusing on. The citizens of Naples <em>felt </em>(shaking ground)<em> </em>and <em>smelled</em> (sulfur) signals from their environment. They then associated those sensory inputs to an event they recall in verbal form (a story) that, as per the past-tense associated with that story, is oriented as some point in the past. A logical-verbal operation is then executed to suggest that some event that occurred in the past, when coupled with certain sensory inputs in the present, is likely to produce the same past event (volcanic eruption) again <em>now</em>. Our current digital systems excel at the latter, verbal reasoning part of that chain of agency. They lack capability at the initial part — sensory input and translation into language — of the chain of agency.</p><p>Our LLM and diffusion-based AI systems cannot smell, they cannot feel, they cannot taste. They can <em>kind of</em> see and perhaps also <em>kind of </em>hear through the visual inputs fed to them through our devices, but it is, I think, appropriate to conclude that their ability to sense the environment we live in is limited. As such, our current AI systems cannot be said to be able to, in a meaningful way, exist within our environment and translate our environment and its sensory inputs into verbal language. Again, in principle, OpenAI could link up a million camera feeds from across the globe to its LLM systems to facilitate a kind of visual sensory input. But also again, such a system would still be a fraction of the bandwidth of sensory input experienced daily by the billions of nodes on the human processing network.</p><p>What can we conclude from this? I think the implications are many, but I’ll focus on trying to separate out the immediate implications of AI on humanity and business versus the longer-term possibilities.</p><p><strong>Short term implications:</strong></p><ul><li>To the extent that your work relies on the <strong>products of language</strong> as opposed to <strong>translating our world into novel language</strong>, your job is at risk.</li><li>If you, for example, extract rare earth minerals from the ground, your job is safe from AI. That is a complex mixture of human sensory input, verbal reasoning, and physical activity that cannot be meaningfully automated by LLMs and diffusion models.</li><li>If you are a certain kind of accountant or lawyer, I would be worried about your job. Accounting is a highly ordered system that can be interpreted by a language model and automated. Much of law is similarly ordered and lends itself to automation from an LLM. In other words, <strong>relatively rote knowledge</strong> <strong>work</strong> is at risk.</li><li><strong>However</strong>: to the extent your knowledge work involves <strong>direct human to human</strong> <strong>communication</strong>, I believe your job will be safe. Our AI systems cannot physically inhabit the room you share with your client and therefore, literally, cannot “read the room” when delivering knowledge work to the client. I do not think we should at all underestimate the importance of how knowledge work is delivered — this is often more important than the actual knowledge work collateral itself.</li><li><strong>In summary, rote knowledge work will be increasingly automated. However, those who communicate that knowledge work will not be automated.</strong></li></ul><p><strong>Long term implications:</strong></p><ul><li>Those companies that are investing in AI development — OpenAI, Meta, Google, and others — should invest in <strong>sensor technology</strong>. By sensor technology, I mean technology that can gather input data on <strong>smell</strong>, <strong>sound</strong>, and <strong>touch</strong>. They should then seek to develop models trained on this data <em>and</em> language data that can translate full human sensory input data into language. <strong>This is the path to true GenAI and truly autonomous agents.</strong></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=280a5f0d7f53" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Opposites (Necessarily) Attract]]></title>
            <link>https://medium.com/@davefay/why-opposites-necessarily-attract-048617ff4077?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/048617ff4077</guid>
            <category><![CDATA[relationships]]></category>
            <category><![CDATA[psychology]]></category>
            <category><![CDATA[social-media]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Fri, 28 Mar 2025 14:20:16 GMT</pubDate>
            <atom:updated>2025-03-28T15:36:42.129Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/440/0*dei1-CmKBMb0WRFO.jpg" /><figcaption>Fight with Cudgles, Francisco Goya</figcaption></figure><p>If you leave a human alone too long, we have a tendency to lose it. We hear voices, invent antagonists in the shadows, and so on. Solitary confinement is regarded as one of the cruelest punishments in our penal system because of the psychological torture it produces. As someone who has broken many bones and also suffered from panic disorder, I can assure you that torture of the psychological variety is just as, if not more acutely horrible than physical pain.</p><p>Why is this the case — why do our brains scramble if left alone too long, like a flower left to wilt in a dark box? I suspect that the cause of our isolation-induced insanity is due to the distributed nature of consciousness.</p><p>In the West, famous for our narratives of economic self-determination and for the creation of the ubermensch, the individual is the sacred atom of our reality. He makes a decision to start the company, she searches for existential respite after losing her belief in God. However, this locus of consciousness (our sense of “I”) does not have to, and often has not, centered on the individual. Nationalism is an abandonment of your identity — your consciousness — to a greater entity of social abstraction. Fealty to the party in communism superseded the needs of and therefore the role played by the individual consciousness. The very phrase “mob mentality” indicates the porosity of consciousness across nodes in our social network. Said plainly — you are often, if not usually, not just you, but also others.</p><p>So who else are you? In contemporary America, “you” are, as the saying goes, “the sum of the five people you spend the most time with.” As institutions like religion and social clubs like the Boy Scouts and Lions Club disintegrate, many of us are increasingly defined by our coworkers and online communities. Because of engagement algorithms on social networks and the tendency of social groups to cluster into similar professional communities, the landscape of socio-psychological profiles in our country is becoming increasingly clustered instead of diffuse. To distort the language of physics, the global entropy of our collective socio-psychology is is <em>decreasing — </em>apparently there is no second law of psycho-dynamics. We are gathering into increasingly dense clusters of ideas, attitudes, predispositions, and beliefs. The institutions and morals that used to govern this clustering — the impulse of Christian forgiveness, the sanctity of individual choice as originally propounded in ancient Greek democracy — have ceded their role as moderators of our social-psychological graph to the algorithms of the internet. Those of us that dive headfirst into our social networks while isolating ourselves from the real-world gathering places (bars, churches, community meetings) may begin to manifest the symptoms of solitary confinement. The severe rates of anxiety and depression among teenage girls may be a warning sign of this trend. My own personal battle with anxiety, depression, and panic during the COVID pandemic is a personal testament to this challenge.</p><p><a href="https://www.amazon.com/Am-Strange-Loop-Douglas-Hofstadter/dp/0465030793">You are not you, you are you as well as others</a>. If you surround yourself with too many people that think precisely as you do about the world, you subject yourself to the risk of paranoia, anxiety, and other psychological challenges not unlike those experienced during solitary confinement. Humans understand the world through a complex physical-verbal model and internal dialogues that continuously role-play what is happening in the world according to that model. My wife, as she may be displeased to hear me say, is a figment of my imagination — she is a creation of my mind that resides within the boundaries of my skull. She would be partially comforted (but not really) by the insistence that <em>I</em>, insofar as “I” am my own ego, do not exist outside of my skull either and instead am solely instantiated within my own grey matter. Cogito, ergo sum. It is the primacy not just of our own sense of self, but also the <em>other personas that constitute our primary inner dialogue and identity</em> that are critical to our psychological wellbeing. This is why it is critical, to borrow tellingly from the language of corporate America, to surround ourselves with those with different perspectives and skillsets from our primary node of identity — the “I” in our internal dialogues. In other words, opposites do not just attract, they <em>need to</em> attract.</p><p>The more time my wife and I spend together, the more we begin to borrow each others’ primary characteristics. I like to think we share our best characteristics while mutually discouraging each others’ worst. My wife is an introverted peacemaker, while I am an extrovert who seeks disagreement and progress. Over our roughly 8 years together, to oversimplify the beauty of long-term relationships, she has become more personally assertive while I have become more compassionate. However, the merging of our personas has also produced more nuanced, emergent changes in our lives beyond the somewhat trite example of “asshole becomes less of an asshole because of his wife.” An example of the more nuanced creation of our relationship is the development of our aesthetic sensibility and its expression.</p><p>Prior to my wife, I had a strong aesthetic sensibility but a limited vocabulary with which to express it. Coming from the mountains of Upstate New York, the manner in which I chose to clothe myself, express myself, and so on were, shall we say, bold and specific, but not necessarily sophisticated. To be more specific, I would borrow my clothing aesthetic from catalogues and my language from novels, but I did not have an individual sense of taste across the varied domains of self-expression. Alternatively, if I can be so bold as to make these claims, my wife, prior to meeting me, had an extraordinarily rich internal landscape of self-expression. However, she was often not assertive enough to realize the full breadth of her internal creative visions in the real world. Her personality, merged with mine through the magic of long-term relationships, has yielded an expression of a mutual aesthetic that is present in our home decor, the artwork we display, the jewelry and clothing we wear, and cars we drive, and the financial decisions we make. Everything is conducted to the joint sense of aesthetic proportion we have developed. I am enormously proud of this emergent property of our relationship and, now, my own personal identity.</p><p>The beauty produced by the relationship of opposites is not only manifest in romantic entanglements, but also friendships. I previously mentioned the fact that I was reared an unrefined boor in the hinterlands of New York. I have a longtime friend, let’s call him James, who is exceptionally talented at social grace and communication. James’ talents have been generous to him in his career, landing him a senior partner role at a major global firm. Throughout the years, through his generosity in our friendship, through the mere act of his presence, he has softened many of my rough edges and helped me become a better professional, partner, and man. However, the flip side of James’ talent — his “conscientiousness, to use a Big Five trait — is also what causes him, as I see it, pain. He is a consummate people pleaser, and as such he works himself to the bone, night, day, and usually across timezones, for the sake of his firm and clients. In some ways, the pain he experiences as a result of his gift resulted in a gift to me — the smoothing-out of some of my own shortcomings. As is the mantle of responsibility in a true friendship, in a desperate effort to reciprocate the gift he has given me, I continue to try and imbue a bit of my own (productive) selfishness — that directness that makes me an exceptional “doer” but occasionally selfish — to his own sense of self. I hope that he sees his worth as existing beyond his role as a titan of the corporate world and quits to develop his identity in more existentially lucrative pursuits. I hope I succeed, and I will not yield even he freezes me out of his life. Such is the gift his identity gave to my own.</p><p>If we surround ourselves with those who think like us — whether that is in a physical work environment or a digital online environment — we risk our distributed sense of self becoming a toxic mix of co-dependent personality traits. Those of us who are compassionate peace-makers risk becoming resentful, morally-superior recluses. Those of us who are more assertive and self-interested risk becoming intolerant assholes. Those two opposites may or may not correlate with recent developments in American politics. Regardless of the level of psychological abstraction — whether it is national politics or our own internal dialogues — it is imperative that we surround ourselves with opposites. It is not just natural that opposites attract, but necessary.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=048617ff4077" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dave’s Crazy AI Thoughts of the Month]]></title>
            <link>https://medium.com/@davefay/daves-crazy-ai-thoughts-of-the-month-bbc6bfbcc9d6?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/bbc6bfbcc9d6</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Fri, 06 Dec 2024 14:11:12 GMT</pubDate>
            <atom:updated>2024-12-06T14:11:12.410Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FnKVgipoz9BIzZgQZByHFw.jpeg" /><figcaption>Civilizations rise, they fall. To what end? Can we predict these timescales more effectively? Photo <a href="https://longnow.org/ideas/reimagining-the-rise-and-fall-of-civilizations/">source</a>.</figcaption></figure><p>I appreciate that most scientifically-oriented people decry reasoning by analogy as opposed to reasoning by first principles. Elon Musk comes to mind as an example. However, without getting too philosophical, I think the nature of intelligence is so complex so as to be effectively irreducible to accessible first principles for humans, at least for now. My intuition here is informed by <a href="https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/">Wolfram</a>. As such, analogy is what remains as a logical framework for which to imbue the topic of AI with a sense of “meaning.” After all, work by <a href="https://en.wikipedia.org/wiki/I_Am_a_Strange_Loop">Hofstadter</a> and <a href="https://en.wikipedia.org/wiki/The_Symbolic_Species">Deacon</a> gives me confidence that analogies across systems may in fact be the basis by which us humans became intelligent in the first place, so I’m not too fussed by my own limitations in this regard. Given this preface, here are this months “profound”, possibly also deranged, thoughts on AI.</p><p>Humanity is the most sophisticated information processing system our world has known. Each of us is a node in the processing network, and each node communicates with others through the high bandwidth optical and pressure (sound) channels of language. In other words, language is the API that facilitates communication of information across the human network, much like electricity and bits are the API of communication across a network of transistors within a silicon medium (contemporary computer chips). Human communication is effected toward an overdetermined goal — initially it was survival, but given the advent of the Internet and the establishment of high dimension meta-connections across the human network, the system is now evolving — the objective function is no longer clear and the post-war, Pax Americana / religious narratives that grounded the former objective function are being replaced by new fledgling narratives of unclear form (neo-nationalism, meme culture, etc.).</p><p>Extending this analogy, <strong>a series of communicating LLM agents could be used to replicate the long arc of civilizations.</strong> Each agent — a much lighter weight LLM than today’s massive “know it all” agents — would correspond to an individual human in an information processing network. The agents would communicate to one another using their native API — language. Large scale social systems coulde modeled and assessed for their performance toward optimizing a variety of objective functions (expansion, i.e. solar exploration, the preservation of natural resources, the creation of more incremental technology, etc.).</p><p>Such a model of humanity would be incomplete, since there are many non-verbal components to human intelligence (read: <a href="https://en.wikipedia.org/wiki/The_Origin_of_Consciousness_in_the_Breakdown_of_the_Bicameral_Mind">Jaynes</a>). But the model would be superior to the limited Neoclassical economic models that rely on the limited API of twentieth century mathematics to model agent interactions. Using such a model of humanity, which IMO is a more apt model of “intelligence” than modeling a single node (brain) in the network, would more closely represent “artificial intelligence” than current efforts of large, centralized LLMs. Further, more optimal forms of governance for our human systems could be derived using such models, helping us achieve our systemic goals more efficiently and effectively.</p><p>I know, this seems dystopian, but humans have been building models of our own behavior and using those models to guide our governing decisions for centuries. <strong>What do you think?</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bbc6bfbcc9d6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Coming Crucifixion of Sam Altman, Tech Messiah]]></title>
            <link>https://medium.com/@davefay/the-coming-crucifixion-of-sam-altman-tech-messiah-c803b1a1dac2?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/c803b1a1dac2</guid>
            <category><![CDATA[sam-altman]]></category>
            <category><![CDATA[peter-thiel]]></category>
            <category><![CDATA[agi]]></category>
            <category><![CDATA[openai]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Sat, 28 Sep 2024 16:41:19 GMT</pubDate>
            <atom:updated>2024-09-28T16:41:19.586Z</atom:updated>
            <content:encoded><![CDATA[<p><em>I advise Altman to heed some ancient wisdom ahead of an inevitable fall.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/444/0*eTPzOwPQ3bR3Z5Gg" /></figure><p><a href="https://formfluent.com/blogs/blog/why-the-tragic-story-of-the-fall-of-icarus-is-so-important"><em>Jacob Peter Gowy’s The Flight of Icarus (Source Jacob Peter Gowy — Museo del Prado, Public Domain)</em></a></p><p>In my precocious twenties, Peter Thiel was a hero of mine. Contrarian, multi-disciplinary, holder of esoteric knowledge and secrets — Thiel was (is?) the <em>dude</em> in Silicon Valley and tech. I once went out of my way to a Starbucks in the Presidio near the offices of his VC company to read a canonical book on AI — I felt smarter and more ambitious just being near a physical manifestation of his genius and drive. Who knows, maybe I even crossed paths in the latte line with that obsequious buffoon JD Vance, since it was around the time he worked with Thiel.</p><p>Anyway. One of Thiel’s favorite authors and teachers was a gentleman named Rene Girard. Girard was a philosopher and, I suspect, the original impetus for Thiel’s predisposition toward contrarianism and trading in “forbidden wisdom.” In an attempt to be a little more like Pete, I picked up the Girard tome with the most esoteric and foreboding title — <strong>I See Satan Fall Like Lightning</strong>. While it would not be accurate to say that the book changed my life, it certainly left a lasting impression in my mind as I suspect Girard’s work left in Thiel’s mind while he was at Stanford.</p><p>This is not a book review. Instead, I think the central theme of the book is a piquant illustration of a recurring cycle within the raging tornado that is the Bay Area tech scene (and other systems, but that is outside the scope of this brief essay). The theme is as follows:</p><ul><li>Humans are creatures that are incessantly driven to belong to some social group.</li><li>This drive to belong is an algorithm that generates recurring cycles of increasing polarization within cultures. More people want to belong to increasingly large “in-crowds”, a dynamic that inevitably reduces the system to two maximally opposed, equal-sized factions. Think — social media and the American two-party system.</li><li>The maximally-opposed, two-party dynamic never persists. It is broken by the symbolic or literal sacrifice of a “scapegoat” — an outsider — and the catharsis of this sacrifice of an other reconciles the two parties and returns the system to a more equanimous state. The outsider can be found outside the system or generated by ostracising someone within the system.</li></ul><p>In my brief career in software and technology, I’ve witnessed a few sacrifices of scapegoats to resolve the acrimony between conflicted parties. An illustrative example is Travis Kalanick. The opposed parties in the System of Uber were the drivers and investors. Travis was a tech messiah whose plan was to develop fully automated driving as the solution to the problem of Uber’s profitability and economics. He spent billions on this vision, generating equal measures of veneration and vilification as Uber scaled. Ultimately his quest for self driving was premature — not quite Quixotic, but operating on a too-short timeline — and he failed. His personal shortcomings contributed to his downfall, and he was ritually sacrificed and ousted from the company. Dara Khosrowshahi was brought in as a cathartic salve for both drivers and investors, and today we never hear Kalanick’s name and Uber is profitable.</p><p>It is clear to me, and I hope others, that Sam Altman and OpenAI are on very much the same path as Kalanick and Uber. Altman has a complete disregard for his company’s profitability and has made public pronouncements to the fact that achieving Artificial General Intelligence (AGI) will be the savior not only of his company’s financial position, but also of humanity, and that he is willing to spend tens of billions of investor dollars to arrive at that vision. This position is almost exactly equivalent to Kalanick’s high-speed gamble to achieve self-driving and rescue Uber’s business model. Ultimately, self-driving technology was too immature to support Kalanick, and he was sacrificed to resolve the warring parties that supported his company.</p><p>It seems inevitable that the same thing will happen to Altman. He is Icarus flying too fast, too close to the sun, and contemporary chip architecture cannot deliver AGI. Further, there are active copyright lawsuits that will almost certainly cut off OpenAI’s access to the vast reams of training data they used to estimate their initial models. The company is beset not just on these twin data and hardware fronts, but also pressure from Big Labor, who bristles anxiously at the prospect of displacement by automation. As Rene warned us, conflict cannot persist indefinitely, and it is ultimately resolved with a sacrifice. I advise Altman to quit while he is ahead, to take his profits and do something else, before his turn arrives.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c803b1a1dac2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sam Altman Wants $7T for the AI Wars. We Should Try a Computing Manhattan Project Instead.]]></title>
            <link>https://medium.com/the-narrative-and-the-noise/sam-altman-wants-7t-for-the-ai-wars-we-should-try-a-computing-manhattan-project-instead-117494d49e8e?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/117494d49e8e</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[electrical-engineering]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Tue, 13 Feb 2024 15:16:05 GMT</pubDate>
            <atom:updated>2024-02-14T21:39:59.724Z</atom:updated>
            <content:encoded><![CDATA[<p><em>The separation of compute and memory has been as sacrosanct to computing architecture as the separation of church and state in Western democracy. To protect the future of Western democracies, we must break down this separation in our computing architecture, especially with regard to AI research. Neuromorphic computing — an architecture where memory and compute are blended, similar to our brain — is the answer.</em></p><p>Around 1945, the “smartest man of the twentieth century”, John von Neumann, led the development of the computing architecture that continues to dominate modern technology. This architecture, called the <a href="https://en.wikipedia.org/wiki/Von_Neumann_architecture">von Neumann model or Princeton architecture</a>, was created to facilitate the creation of the deadliest weapon humanity has ever constructed — the hydrogen bomb. The amount of calculations required to facilitate the bomb’s creation exceeded all calculating capacity available at the time, both human and machine driven. To address this challenge, von Neumann and his team suggested a machine of the following structure:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*3DvKlfQHBHb0iDmtJzAD0g.png" /><figcaption>Source: <a href="https://en.wikipedia.org/wiki/Von_Neumann_architecture">Wikipedia</a></figcaption></figure><p>This structure should look at least a little familiar to anyone who has ever used a computer, smartphone, or any contemporary digital device. In this model, “inputs devices”, e.g. a mouse or keyboard, serve instructions to a Central Processing unit, or CPU. That CPU interprets the input and performs logical operations on the inputs and any other information stored “in memory”, i.e. in the so-called “Memory Unit”. The results are then served to an Output Device, e.g. a smartphone’s speakers playing music, or a monitor displaying a video. Sometimes values will be stored to a long-term memory device, which isn’t broken out here.</p><p>The amazing thing about this architecture is that it is a <strong>general-purpose computing architecture</strong>. By structuring the CPU to perform its calculations based on binary logic — 1’s and 0&#39;s — the CPU can be used to calculate <em>anything that is in fact able to be calculated</em>. Older machines, like mechanical calculators or hard-coded elevator circuits, were only ever capable of calculating the precise task for which they were originally intended — in this case, arithmetic or the routing of elevators. The CPU as designed by von Neumann et. al can calculate <em>any tractable computation</em>, from arithmetic to elevator routines to programs that launch rockets to Spotify, provided the correct instructions (code) are provided. The value of universal binary computation was not discovered by von Neumann — <a href="https://www.leibniz-translations.com/binary.htm">Leibniz wrote about it in the 1703</a> — but von Neumann &amp; company were the first to instantiate the approach into a powerful, general-purpose computing machine.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wYCDxZdtFr7f7HuwZWPeAQ.png" /><figcaption>The ENIAC — the first general purpose computing machine.</figcaption></figure><p>This kind of machine, realized initially in the <a href="https://en.wikipedia.org/wiki/ENIAC#:~:text=ENIAC%20(%2F%CB%88%C9%9Bni,of%20them%20in%20one%20computer.">ENIAC computer developed for the US military</a>, was created to generate the calculations that would enable the building of the hydrogen bomb. While the computer was used during the day to create bombs, because of its general-purpose computing abilities, it also, at night, was used to generate <a href="https://www.ias.edu/scholars/nils-barricelli#:~:text=Using%20the%20computer%20that%20von,genes%20caused%20populations%20to%20evolve.">universes of self-replicating digital organisms</a>. As you might expect, a general purpose computing architecture that can create bombs and digital universes can be used for many other things, and as such the von Neumann architecture underlies most modern devices (e.g. computers, smartphones, etc.) and digital systems (e.g. cloud computing systems). The von Neumann architecture is truly one of humanity’s most brilliant inventions.</p><p>It isn’t perfect, however. The separation of memory from the CPU creates a significant bottleneck in computing performance. As humanity creates programs that increasingly demand more compute and storage, like AI training and calculation, the bottleneck between the logic of a program and the values it uses for its computation is exacerbated. As stated in a post from the <a href="https://www.sigarch.org/blurring-the-lines-between-memory-and-compute/">Association for Computing Machinery</a>:</p><blockquote>“A significant fraction of energy is spent today in moving data back-and-forth between memory and compute units, a problem that is exacerbated in modern data-intensive systems… [n]early three-fourth of silicon in processor and main memory dies is simply to store and access data. What if we could take this memory silicon and repurpose it to do computation?”</blockquote><p>In other words, there is a cost to this amazing general-purpose computing architecture, and that cost is efficiency. By creating computational machines that can compute and store anything, we’ve missed an opportunity to build alternative machines that can accomplish a more specific goal, but much more efficiently. As stated above, the majority (~75%) of the silicon in our contemporary CPU’s is dedicated to shuttling information from the computing part of the system to the storage part of the system, as opposed to actually performing computation or storing data. What if that 75 percent could be unlocked for additional computational complexity? What would such a solution look like?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/992/1*3FkiwmqhJg3oAYC6VJ7wTg.jpeg" /></figure><p>We can draw inspiration for a solution from the human brain. The brain, for example, does <strong>not</strong> follow the von Neumann architecture. If you peered into your own brain, you will not find a storage section, i.e. a kind of biological hard drive, or a section that exclusively produces logical operations, i.e. a kind of biological CPU or GPU. Our brain’s information processing architecture is different than the von Neumann architecture. Similar to the more efficient architecture we suggested, the brain wastes no material on shuttling information between specialized compute and storage machinery — the neurons that compose our brains serve both storage <em>and</em> compute purposes, therefore performing its operations more efficiently than comparable von Neumann-based systems (more on that later).</p><p>The brain is estimated to have around 100 billions neurons. If we use a rough estimate of 1,000 connections (synapsis) for each neurons to other neurons, we have a system of approximately 100 trillion “<a href="https://www.beren.io/2022-08-06-The-scale-of-the-brain-vs-machine-learning/">parameters</a>”, to use the language of machine learning. Hiding within that unfathomable amount of parameters are the unique gifts that, at present, only humans exhibit— consciousness. The result of this is a system of unfathomable complexity and signal processing sophistication. Importantly, it is also a more <em>efficient</em> signal processing system with respect to the concepts of consciousness and intelligence when compared to the von Neumann architecture. To better articulate this point, let’s consider the most intelligent systems human’s have built today and compare their performance to the human brain.</p><p>At present, the most complex and intelligent digital machines humans have built are our Large Language Models (LLMs), like those present at the heart of ChatGPT. Although these machines have produced incredible results as of late — reliably mimicking the language-generation capabilities of humans — they cannot be said to contain consciousness. The limitations of LLMs are known and well documented. In particular, LLM’s do not fare well at, among other things, <a href="https://sloanreview.mit.edu/article/the-working-limitations-of-large-language-models/">lengthy, winding discussions that rely intensively on logical abstraction</a>. This is because LLMs are models of <em>language</em>, not a model of the system that <em>produces</em> language, i.e. networks of human brains. Language is but a fraction of what the human brain produces, and as such LLMs are necessarily a model of a fraction of what the human brain is capable of.</p><p>To compare the complexity of LLMs to that of the brain, we can turn again to a parameter estimate. The latest OpenAI LLM, GPT-4, is estimated to have around <a href="https://levelup.gitconnected.com/gpt-4-parameters-explained-everything-you-need-to-know-e210c20576ca">1–2 trillion parameters</a>. That puts GPT-4 as a system at least 2 orders of magnitude of complexity below that of the brain (100T parameters), if our calculations are reasonable enough for estimation purposes. The parameters in ChatGPT are more or less connected in a manner illustrated by the structure below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/727/1*hNv2Kwx7MNnJ3C0mJ0kHQA.png" /><figcaption><a href="https://medium.com/@fenjiro/chatgpt-gpt-4-how-it-works-10b33fb3f12b">Neural Network Structure</a></figcaption></figure><p>Given enough time, we should not be surprised when our Artificial Neural Networks (ANN’s, e.g. LLMs) begin to, at least on paper, achieve similar levels of parameters to the human brain (i.e. parameters in the ~100T range). As we add more diverse data to train the models, and as we add more parameters to process the data, we would expect the complexity and sophistication of our artificial intelligence to begin to rival the scope of our own intelligence. However, even if the networks themselves can reach this level of complexity, they will still be considerably less efficient and bottlenecked when compared to the system of the human brain. In other words, they would remain inferior from an intelligence perspective and would perhaps still lack a degree of consciousness. This inevitable bottleneck in performance, and therefore effective complexity, of our ANN’s is due to the fact that <strong>they are still trained, hosted, and executed on systems that follow von Neumann architecture, with all the limitations that come from the separation of compute and storage</strong>. No matter how extensive our data is and how sophisticated our algorithms become, they will continue to be hampered by their training and execution hardware.</p><p>Electrical engineers have been aware of this issue for a couple decades. To illustrate this point, we can turn to a recent research paper by Korean researchers, published in <a href="https://www.nature.com/articles/s41467-022-30539-6">Nature</a>.</p><p><em>“Artificial neural networks (ANNs) have shown their effectiveness in various fields such as autonomous cars, finding new drugs, designing circuits, or predicting protein structures1–4. Powered by today’s elaborate complementary metal–oxide-semiconductor (CMOS) based computing processors, ANNs can accomplish highly complex tasks. However, ANNs running in a conventional von Neumann architecture computer (e.g., graphics processing units (GPUs)) cannot reach the efficiency of a biological neural network due to the bottleneck of big data transfer. To substitute the conventional ANNs, neuromorphic computing hardwares, which imitate the operation of the brain and the biological neural network, have been extensively studied. The conventional CMOS-based neuromorphic hardware uses complex CMOS circuitry to imitate the neuronal activity. However, CMOS-based complex circuitry limits scalability and increases energy consumption. Thus, it is difficult to reach the efficiency of a biological neural network using CMOS-based neuromorphic hardware</em>.”</p><p>This research introduces the concept of “neuromorphic hardware”, i.e. computational hardware that mimics the structure of our own neurons. ANN’s mimic the <em>algorithms</em> of our brains, but are limited by their CMOS / von Neumann hardware, which do <em>not</em> mimic the physical structure of the brain. Instead, by adopting neuromorphic hardware, i.e. hardware that shares the physical properties of our brains, our ANN’s can break through these bottlenecks. Breaking these performance and efficiency bottlenecks will unleash our ANN’s and enable them to rival and surpass our own biological neural networks.</p><p>The initial version of such neuromorphic hardware comes to us in the form of “memristors” (ostensibly a conjunction of “memory” and “transistor”). <a href="https://onlinelibrary.wiley.com/doi/full/10.1002/aelm.202200833">These processors resolve the compute / storage bottleneck of conventional CMOS circuitry by <em>combining </em>compute and storage</a>. In other words, in a memristor, the part of the circuit that performs that calculation also stores the data required to perform the calculation, removing the need for inefficient, high-bandwidth connections between compute and storage components. The result is highly efficient hardware that is better suited to training, hosting, and executing our ANNs versus conventional CMOS / von Nuemann approaches.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/220/0*USUUakIeXTfCt2Ri.png" /><figcaption><a href="https://en.wikipedia.org/wiki/Memristor">Mockup</a> of potential memristor hardware.</figcaption></figure><p>The costs of training, hosting, and executing ANN’s on inefficient hardware is quantifiable. It is estimated that the training of OpenAI’s GPT-4 cost upwards of <a href="https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/">$100 million</a>, a figure that suggests further research into the topic of large models as infeasible for all but the largest institutions and nation-states. Further, the energy required to execute, let alone train, ChatGPT every month is “<a href="https://www.numenta.com/blog/2023/08/10/ai-is-harming-our-planet-2023/">roughly as much electricity as 26,000 US households</a>.” This financial and resource inefficiency can be drawn better into focus when considering that ChatGPT, with its huge financial and energy requirements, is a system of intelligence that has still not reached the complexity of a single human brain.</p><p><strong>By transitioning our AI training, hosting, and execution from traditional von Neumann / CMOS hardware to neuromorphic hardware, Western democracies can gain a massive advantage in the AI race.</strong></p><p>Because of the multifaceted nature of a neuromorphic computing project, this author suggests a kind of “Manhattan Project” to help accomplish a working version of this computing model. A realization of this project would mean less, or no, <a href="https://www.economist.com/special-report/2023/03/06/taiwans-dominance-of-the-chip-industry-makes-it-more-important">dependence on Taiwan</a> for the CMOS chips (semiconductors) that drive every one of our digital devices. I need not speak to the relief in international relations such a shift would facilitate. Furthermore, they would enable unprecedented AI systems to be created for the purposes of furthering the continued resilience of Western Democracy on this planet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ynyZCQ8H9L5UobkY.jpg" /><figcaption>The original Manhattan Project developed the first nuclear bombs. These weapons ensured the continuation of Western democracy on this earth. We can take inspiration from this project for investing in a new, similarly consequential technology — AI.</figcaption></figure><p>The workstreams in such a Neuromorphic Computing Manhattan Project would be as follows:</p><ol><li><strong>Data</strong></li></ol><p>Neuromorphic computing would be incompatible with many data sources. The human brain does not leverage symbolic representations of data similar to many of our conventional data structures. In conventional data structures, the string “cat” can be said to stand for the thing of a cat in the digital system it exists in, e.g. the “cat” can have the color “brown”. No such symbolic representation <a href="https://aeon.co/essays/your-brain-does-not-process-information-and-it-is-not-a-computer">exists in the brain</a>, and as such our approach to training data may need to shift to non-symbolic data structures.</p><p>2. <strong>Hardware — Engineering &amp; Theory</strong></p><p>We’ve already mentioned the inadequacies of existing CMOS chips. New computing hardware will need to be built, grown, or otherwise generated. This hardware will need to combine memory and compute into its “neurons”. This will not only be a non-trivial engineering exercise, but also a challenging theoretical exercise. What will it mean to store non-symbolic data? How can our world be deconstructed into meaningless tokens that can, somehow, be aggregated into intelligible results? These questions need to be answered to generate a feasible engineering solution to neuromorphic computing. We can and should take inspiration from the brain, which indicates that not only should computer scientists and electrical engineers be involved, but also biologists and neuroscientists.</p><p>This group can take inspiration from the actual Manhattan Project, where theoretical physicists worked alongside engineering and explosives experts. Each Subject Matter Expert (SME) contributed his or her expertise to the overall project, and each practice area was fundamental to the success of the project.</p><p>3. <strong>Software</strong></p><p>What does it mean to write software for a system that combines non-symbolic memory and computation? This author does not know. Algorithms could be implicitly instantiated in the physical structure of the neuromorphic hardware (similar to our brains), or could perhaps be alternatively submitted from an outside source. I leave it to those with more expertise than I to theorize further.</p><p><strong>Conclusion</strong></p><p>I’ll conclude this note by reinforcing my choice of the Manhattan Project as the inspiration for this research versus, e.g., the private sector example of Google, or Sam Altman’s own OpenAI. Google generated technology of remarkable sophistication using the venture capital and capitalism model of development, so why couldn’t open AI do the same thing with a neuromorphic computing project funded by Altman’s $7T capital raise?</p><p>I suggested government-driven research instead of private sector research because I believe neuromorphic computing to be too much of a sea-change in computing architecture to be compatible with the profit requirements and time horizons of the private sector. With the end of the low interest rate era, the products that are created by startups have much shorter timeframes to realize profits, and neuromorphic computing will take too long, and require too much investment, to be successful given the pressures of profitability. Google was built on von Neumann architecture and therefore was compatible with the capitalist economic model — no fundamental research was required. However, similar to the Manhattan Project, neuromorphic computing <em>will</em> require fundamental research and longer timelines. As such, I believe the government-sponsored model is a better bet for realizing this vision.</p><p>Put differently — to ensure such a neuromorphic computing project does not end up like the various interminable quantum-computing private sector projects, public sector investment and management is needed. Public sector investment will also, hopefully, ensure compatibility of the project with Western values.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=117494d49e8e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-narrative-and-the-noise/sam-altman-wants-7t-for-the-ai-wars-we-should-try-a-computing-manhattan-project-instead-117494d49e8e">Sam Altman Wants $7T for the AI Wars. We Should Try a Computing Manhattan Project Instead.</a> was originally published in <a href="https://medium.com/the-narrative-and-the-noise">The Narrative and the Noise</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Nietzsche, Wolfram, and ChatGTP — What Technology can Teach us About Ourselves]]></title>
            <link>https://medium.com/the-narrative-and-the-noise/nietzsche-wolfram-and-chatgtp-what-technology-can-teach-us-about-ourselves-3b7d0825fee4?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b7d0825fee4</guid>
            <category><![CDATA[philosophy]]></category>
            <category><![CDATA[mythology]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Thu, 21 Dec 2023 14:33:05 GMT</pubDate>
            <atom:updated>2023-12-21T14:54:02.404Z</atom:updated>
            <content:encoded><![CDATA[<h3>Nietzsche, Wolfram, and ChatGTP — What Technology can Teach us About Ourselves</h3><p><em>Mores and their victims. — The origin of mores may be found in two thoughts: “society is worth more than the individual,” and “enduring advantage is to be preferred to ephemeral advantage” — from which it follows that the enduring advantage of society must be given precedence, unconditionally, over the advantage of the individual, especially over his momentary well-being but also over his enduring advantage and even his continued existence. Whether the individual suffers from an institution that is good for the whole, whether it causes him to atrophy or perish — mores must be preserved, sacrifices must be made. But such an attitude originates only in those who are not its victims — for they </em>[the victims]<em> claim in their behalf that the individual may be worth more than many, also that present enjoyment, the moment in paradise, may have to be valued higher than a pallid continuation of painless or complacent states. The philosophy of the sacrificial animal, however, is always sounded too late; and so we retain mores and morality — which is no more than the feeling for the whole quintessence of mores under which one lives and has been brought up — brought up not as an individual but as a member of a whole, as a digit of a majority. — Thus is happens constantly that an individual brings to bear up on himself, by means of his morality, the tyranny of the majority.</em></p><p>Friedrich Nietzsche, Seventy Five Aphorisms</p><p>“Mores” — “the essential or characteristic customs and conventions of a community.” — Oxford Languages</p><p>“Trickster” — “Found in the <a href="https://www.newworldencyclopedia.org/entry/Mythology">mythology</a>, <a href="https://www.newworldencyclopedia.org/entry/Folklore">folklore</a>, and <a href="https://www.newworldencyclopedia.org/entry/Religion">religion</a> of virtually all world <a href="https://www.newworldencyclopedia.org/entry/Culture">cultures</a>, a <strong>trickster</strong> is a figure who plays tricks or otherwise disobeys normal rules of behavior.” — <a href="https://www.newworldencyclopedia.org/entry/Trickster">New World Encyclopedia</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/440/1*N2mIOHE8GmvQd9PhvQAXwg.jpeg" /><figcaption>Reynard the Fox . The archetype of the “trickster” myth is a cultural / linguistic heuristic that technologies like LLM’s may help us better understand.</figcaption></figure><p>Groups of people are usually more successful than individuals when it comes to survival. Put differently — cooperative groups of people are probably more likely to survive than an individual when push comes to shove. As such, it stands to reason that the cultural artifacts, or “mores”, as Nietzsche refers to them above, of a society will tend toward those that preserve the status quo, “pallid” as that status quo may be. However, as the preponderance of “tricker” <a href="https://en.wikipedia.org/wiki/Trickster">myths</a>, including the Loki character so popular in today’s Marvel universe, indicates, there is value in introducing a minority voice of dissent, or a small degree of freedom, periodically into society. By including a tiny dose of chaos — of individualism — into static societies, those societies are prevented from becoming stale. Since the world ever-changes, having a culture that is at least partially dynamic and not frozen in time promotes the continued existence of the society-corporate.</p><p>These twin, opposing concepts — that of the dominance of mores that enforce cultural coherence and its inverse twin, the minority impetus of chaotic individualism — are two cultural heuristics that promote resilience for humanity. If individualism was too dominant, then society would devolve into an every-man-for-himself cycle of mistrust and destruction. If every person is fighting for food with every other person, it is unlikely that the cooperation required to generate innovations like language, technology, and the like would ever develop. This is why the norms that encourage adherence to cultural norms are valuable.</p><p>However, if, alternatively, the pressure to cohere to the norms of society <em>completely </em>dominated society, then society would be static and therefore vulnerable to a changing environment. For example: Galileo exemplified the chaotic (creative) individual impulse, which drove him to promote a heliocentric theory of the solar system. This view opposed the prevailing orthodoxy, supported by the Church, that the earth was the center of our cosmos. This stroke of individualism was persecuted by the Church (Galileo was imprisoned), but the theory of a heliocentric solar system nevertheless persisted and helped drive the development of modern astronomy. If an asteroid were to collide with the earth 150 years from now, a society that benefited from Galileo’s individualism and mastered the ability to traverse space and colonize planets like Mars would be more resilient than a static, earthbound society still clinging to an erroneous geocentric model of the solar system.</p><p>To state it once more, plainly: mores that 1) promote cultural norms and 2) the minority impulse to individualism are cultural heuristics that promote the fitness and progress of humanity. However, humans are not born with this message tattooed on our foreheads. Similarly, there is no evidence of brain circuitry that explicitly promotes cultural adherence or strokes of individual brilliance. Instead, these impulses are the emergent behavior of multitudes of generations of human-hosted neural network processing. The physical information of our planet — its sights, sounds, smells, textures — passes through each node (individual brain with its own respective processing network) in the human processing network, with each node (person) communicating with other nodes (people) through the phenomenon of physical proximity and the emergence of language.</p><p>The sum total of thousands of generations of processing lifetimes of this network is, among many other things, the emergent properties of enforced cultural mores and the minority impulse of individualism as effected through the medium of language. There is no “reason” for these twin cultural heuristics, and the relative frequency with which they occur in the corpus of our communication, other than they promote the fitness of humanity over time. Furthermore, it is probably impossible to reason, from first principles, why these impulses, and our myths that embody them, are the most efficient algorithms by which humanity should operate.</p><p>All of this is to say that the recent innovation of ChatGPT — itself a product of the minority individual impulse — sheds two alternating shades of light on the nature of our own human intelligence, both individual and collective. The <strong>first insight</strong> ChatGPT reveals about human thought is that there need not be any explicit “rules” for language, or thought more broadly, encoded in our brains in order for our brains to be useful in ordering the universe and propagating the species. <strong>Intelligent design from first principles is unnecessary to produce an intelligent agent capable of ordering the world with language.</strong> It is enough that a large number of highly connected processing units (the neurons of our brain) have found a way, through sexual reproduction and Darwinian processes, to process the physical phenomena of our planet in such a way that has proven useful. In his <a href="https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/">explanation</a> of ChatGPT’s technology, Stephen Wolfram highlights the irreducible complexity of ChatGPT’s neural net, which in many ways mimics the processing architecture of our own brains.</p><p><em>But in the end, the remarkable thing is that all these operations — individually as simple as they are — can somehow together manage to do such a good “human-like’ job of generating text. It has to be emphasized again that (at least so far as we know) there’s no “ultimate theoretical reason” why anything like this should work. And in fact, as we’ll discuss, I think we have to view this as a — potentially surprising — scientific discovery: that somehow in a neural net like ChatGPT’s it’s possible to capture the essence of what human brains manage to do in generating language.</em></p><p>If we apply Wolfram’s logic to our twin cultural impulses, there need not have been any intelligent (explicit) <em>design</em> of, e.g., the trickster myth in order for it to emerge from the complex network of human language. Intelligent systems need not be explainable to reach human-level fluency with language.</p><p>However, in a brilliant inversion of theme, Wolfram does not think, in a <strong>second insight</strong> from ChatGPT, that the human brain (and technologies based on it) will <em>forever</em> remain an irreducible mix of legions of networked computational elements. The human network, over its many generations of communication, have, as we established, produced intelligible cultural heuristics, or “rules”, that are useful artifacts<em> </em>in the broader patterns of our language. The trickster Raven in indigenous Alaskan culture, or Hermes in Greek culture, or Coyote in Native American culture are observations of, as Wolfram would phrase it, a certain “law of physics” of our language. Although the trickster myth — the persistent embodiment of the minority individual impulse across cultures — is an emergent property of a computational network of irreducible complexity, it <em>suggests </em>a kind of rule that could be enacted in an analytical framework to better understand, and therefore predict, the evolution of the not-so-irreducibly complex network in the future.</p><p>Here is what Wolfram has to say on the subject.</p><p><em>The success of ChatGPT is, I think, giving us evidence of a fundamental and important piece of science: it’s suggesting that we can expect there to be major new “laws of language” — and effectively “laws of thought” — out there to discover. In ChatGPT — built as a neural net — those laws are at best implicit. But if we could somehow make the laws explicit, there’s the potential to do the kinds of things ChatGPT does in a vastly more direct, efficient — and transparent — ways.</em></p><p>For eons, humanity engaged in a futile struggle against nature. We had in common with all other earthbound creatures the same struggle for sustenance, for shelter, and for sexual reproduction. Our population curve was similarly flat to other creatures as a result. The world to us was irreducibly complex and therefore unconquerable. However, due to the <em>unconscious</em> development of the heuristics of cultural coherence and the minority individual impulse, language was invented and embodied in human’s personal (brains) and collective (culture) networks of intelligence. Eventually, the informal logic of language begat the more formal logic of mathematics, which reduced the complexity of our complex world into logical structures that humans could 1) understand and 2) use to make reasoned predictions about the world. To calculate how fast a falling object is traveling in space, you need not contend with the endlessly complex conditions of the real world — you simply plug a few variables into the equation velocity = 9.81 meters / second ^2. By such devices, humans have conquered many such real-world physical systems and phenomena, and by such devices we may one day colonize Mars, conquer disease, and other worthy endeavors.</p><p>In a similar path of innovation and creation, ChatGPT may help us create the laws of physics for human language and thought. It may help us identify heuristic patterns, expressible by accessible logic, that enable us to effectively predict the behaviors of a system that is irreducibly complex — our own brains and language. By such means we may uncover the mechanisms by which cultures evolve over time and enter into conflict, thereby providing the theories to help prevent humanity’s addiction to warfare. I have a suspicion that Nietzsche’s insight into human behavior — the cultural coherence impulse — and the prevalence of the trickster myth and its individualistic impulse — could be terms that feature into this future calculus for large scale human thought, i.e. culture.</p><p>By continuing to develop our LLM’s and investigate the structure of their parameters, these personal mysteries about our own minds and the collective, cultural mysteries we share may be solved. However, a more interdisciplinary approach to guiding such research is crucial to ensure the fruits of such labor are not used irresponsibly. More on that to come in later pieces.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b7d0825fee4" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-narrative-and-the-noise/nietzsche-wolfram-and-chatgtp-what-technology-can-teach-us-about-ourselves-3b7d0825fee4">Nietzsche, Wolfram, and ChatGTP — What Technology can Teach us About Ourselves</a> was originally published in <a href="https://medium.com/the-narrative-and-the-noise">The Narrative and the Noise</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Shouldn’t Rush to Regulate AI When Social Media is the Real Danger]]></title>
            <link>https://medium.com/the-narrative-and-the-noise/we-shouldnt-rush-to-regulate-ai-when-social-media-is-the-real-danger-6ba25d802250?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/6ba25d802250</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[neural-networks]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Fri, 03 Nov 2023 18:31:05 GMT</pubDate>
            <atom:updated>2023-11-05T13:48:55.244Z</atom:updated>
            <content:encoded><![CDATA[<p><em>AI is teaching us how our own minds work and will unleash a wave of creativity. While regulators are worried about AI, they should instead first regulate the primary destructive creation of our time — social media.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xxo89Cn1XP-22478FkMpFg.png" /><figcaption>Civilization facilitates the large-scale collaboration that results in brilliant innovations like Newton’s magnum opus and, more recently, generative AI. We shouldn’t let social media destroy our ability to create these cultural wonders.</figcaption></figure><h4>Introduction</h4><p>To open the hood of a car is to behold a complicated machine — the internal combustion engine. In this engine, thousands of parts work in precise cooperation to generate one thing: controllable locomotion. All one needs to do is to enter the driver’s side of the car and operate a relatively simple set of controls to navigate to anywhere from the grocery store to Tierra Del Fuego. Although the machinery and its thousands of components are challenging for a layperson to understand as a system, if taken one by one, the mechanics of the system are simple. Spark plugs ignite fuel, which drives pistons to move up and down. That vertical movement is converted via a crankshaft to rotational movement, which is then controlled by a transmission of various speeds, which then transfers the rotational movement to the tires, which provides propulsion for the entire vehicle. As a whole the car is intimidatingly complex. However, the function of each of its component parts is eminently accessible to even the least mechanically-inclined of us.</p><p>The human brain is composed of tens of billions of cells, known as neurons. Some of these neurons may be connected to only a few other neurons, while some are connected to many others. To say that the brain — a bioelectric machine — is “complicated” similar to the automobile is to make an understatement of many orders of magnitude. Whereas an average new car may have <a href="https://knowhow.napaonline.com/how-many-parts-are-in-a-car/">30,000 new parts</a> on average, we have already stated that a normal human brain has approximately <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4689008/#:~:text=There%20are%20approximately%20100%20billion,connections%20among%20neurons%20occurs%20postnatally.">100 billion</a> “parts”, i.e. neurons. Furthermore, not only are there astronomically more <em>parts </em>in the machine of the brain when compared to the car, there are also an unfathomable amount of <em>connections </em>between those parts — some estimates put this number of connections at <a href="https://www.scientificamerican.com/article/100-trillion-connections/">100 trillion connections</a> (you read that right) between neurons. The human brain is a machine of stupefying complexity and, with apologies to the humble automobile, humanity has never created anything of a similar order of complexity — that is, until recently.</p><p>The brain is a brilliant machine, and investigations into how it works is leading us to the advent of intelligent, human-created machines — AI. Many government officials and scaremongers are rushing to <a href="https://www.whitehouse.gov/briefing-room/statements-releases/2023/10/30/fact-sheet-president-biden-issues-executive-order-on-safe-secure-and-trustworthy-artificial-intelligence/">regulate</a> the creation of AI by private companies. This essay will take a look at this brilliant but complex machine — the brain — and will chart how its research led to modern AI. This relationship will indicate the importance of the development of AI in better understanding ourselves, i.e. our own biological intelligence. Additionally, this investigation into intelligence — biological and artificial — will also reveal why social media and <em>not</em> AI is the primary threat regulators should be focusing on.</p><h4>Our Own Intelligence — the Brain</h4><p>As we established, the brain is a machine that is orders of magnitude more complex than the automobile. However, despite this difference, both machines share a characteristic in common. This characteristic is that although the whole is complex, the individual parts (neurons) that constitute the whole are relatively simple; the functioning of a neuron can be understood by most just as a firing piston can be understood. A neuron contains four primary components:</p><ol><li>Dendrites</li><li>Axons (and myelin)</li><li>Synapses</li><li>Cell Body (Soma)</li></ol><p>These components enable a neuron to <a href="https://www.khanacademy.org/science/biology/human-biology/neuron-nervous-system/a/overview-of-neuron-structure-and-function">receive and to send signals</a>. Neurons receive signals from other cells they are connected to through their dendrites. Think of trees with interlocking branches — many dendrite branches reach out and make connections with other branches. If a sufficient amount of signal is received across the network of a neuron’s dendrites, then an “<a href="https://www.khanacademy.org/science/biology/human-biology/neuron-nervous-system/a/overview-of-neuron-structure-and-function">impulse, or action potential, is conducted down the axon.</a>” The axon splits into its own series of branches and, should the action potential be realized, translates the electric signal into a chemical signal. This chemical signal is passed over a junction called a synapse to other neurons’ dendrites, and the process begins anew.</p><p>This process is modulated by hormones and other factors, but the functioning of a neuron can be boiled down to a straightforward process: a chemical signal from one neuron’s axon is transmitted into an electrical signal by a receiving cell’s synapse. That electrical signal is communicated down the receiving neuron’s dendritic network and into the cell’s axon. Sometimes, the signal will be such that the neuron will “fire” and pass on a version of the signal, via the axon and its terminals, to another cell via its own synapses.</p><p>The brain is composed of billions of simple cells that, when provided with trillions of connections to other simple cells, produces the emergent quality of intelligence. Though it is not exactly true that “bigger” is always “better” in the <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4685590/">case of neural networks</a> — birds are more intelligent than some larger animals with larger brains — it is clear that the Central Nervous Systems (CNS) observed on earth are far more complex than most other systems, and that the relative degree of this complexity is what generated intelligence in humans versus other systems, like networks of plants.</p><p>So humans are intelligent because of our brains, and our brains are made up of billions of relatively simple cells with trillions of connections to other, similar cells. Being intelligent apes, we picked up on this fact back in the <a href="https://news.cornell.edu/stories/2019/09/professors-perceptron-paved-way-ai-60-years-too-soon#:~:text=It%20was%20a%20demonstration%20of,&#39;56.">1950’s</a> and began to construct not bio-electric thinking machines, but electromechanical thinking machines to approximate the signal processing capabilities of our brains. These early machines were called “perceptrons” and are the ancestors of modern neural networks and generative AI. Following the principle of the automobile, humanity, through our early understanding of the brain, knew that it was possible to construct a machine of significant, complex ability out of a series of simple parts. In short, the perceptron and its descendents all model their processing nodes (“parameters”) based on the functioning of a neuron, as described above. It was this intuition — that we could begin to approximate intelligence in our created, artificial systems by mimicking the elementary cells of natural intelligence — that kicked off the current revolution in generative AI. However, it took a while for this vision to be realized.</p><p>It is widely understood that the intuition behind the perceptron and its descendants was always more or less the correct way to pursue artificial intelligence, but that the amount of data we needed to train these networks was not available until the last 10–15 years. Because of the revolutions of data storage technologies, the subsequent explosion in data generated by humanity’s digital systems, the unprecedented parallel computing capabilities of GPU’s, and finally the <a href="https://huggingface.co/docs/transformers/v4.15.0/parallelism">invention of algorithms</a> that can leverage big data alongside parallel processing, humanity finally has the tools to build convincing versions of artificial intelligence. The advent of OpenAI’s ChatGPT and other generative AI tools are evidence of this capability.</p><p>It is impossible to resist comparisons between the technological advancements humans have produced in the last half-century, which led to convincing green-shoots of electromechanical AI, to the natural conditions on earth that led to bio-electrical intelligence, i.e. humans and other animals. Humanity and our intelligence evolved over thousands of generations (or <a href="https://royalsocietypublishing.org/doi/10.1098/rstb.2015.0033#:~:text=It%20is%20argued%20that%20molecular,would%20have%20favoured%20the%20evolution">millions</a>, depending on when you locate the genesis of our central nervous system). Through the Darwinian mechanisms of evolution, we developed a central nervous system that could receive signals from the outside world, translate those signals through a network of processing cells, and generate pattern-recognition based actions derived from those inputs. The pattern recognizing behavior produced by our CNS resulted in, on average, a greater fitness than other organisms for passing on our genetic information (surviving) on earth. Each subsequent generation led to greater efficiency and sophistication of our CNS, which in turn led to greater survival fitness and the propagation of our species.</p><p>Humanity has been able to, in many ways, condense the natural process by which nature evolved intelligence over hundreds of thousands to millions of years down to a period of several decades. This is due to the aforementioned explosion of data, parallel computing architecture, and training algorithms (e.g. the transformer algorithm) that enables the efficient processing of huge amounts of data. Humans evolved a nervous system that can “learn” language based on a relatively tiny amount of input data over a matter of a few years, i.e. childhood. It took thousands or, more likely, millions of years to slowly evolve the network pre-training to enable this kind of efficient language acquisition. ChatGPT, on the other hand, because of the huge availability of data and parallel processing technologies, was able to acquire language in, on an evolutionary scale, the blink of an eye, albeit it in a far less efficient manner than a human child. Although our AI has arrived at convincing natural language processing by different means that our own biological evolution, the AI like ChatGPT exhibits intelligence similar to our own human intelligence, at least with respect to language.</p><p>As we continue to develop LLMs and other AI agents, humanity will realize a variety of boons. Tedious tasks will be automated, like the generation of boilerplate marketing material and customer service conversations. Further, and less obviously, by learning to further fine-tune the training algorithms and parameters of LLMs, we will learn more about how humans 1) acquire and 2) evolve language over time. These revelations could lead to breakthroughs in diseases like aphasia, help create therapeutic protocols for language-related learning disabilities, and a host of other medical benefits. AI research is not without its dangers, but there is also tremendous promise.</p><p>To this point, I have established a link between human intelligence and the intelligence of the artificial systems we are creating. The goal has been to, if you’ll forgive the term, “humanize” the intimidating algorithms of generative AI and to enunciate its potential promise. However, this is not the only reason I highlight this link. The additional reason for delving into the relationship between biological and artificial intelligence is to provide an alternative perspective on the scale and nature of “human” intelligence… and how social media is making go all wrong.</p><h4>Human Intelligence — Beyond the Individual</h4><blockquote>“How big of a network, then, do we need to build a full brain? Given approximately 20B non-cerebellar neurons and if we assume that the cerebellum has an effective parameter count of about 1T, then my estimate that the effective parameter count of the full brain is somewhere between 10–30T parameters. This is extremely large but within only 1–2 OOMs of current large scale systems. Based on estimates of the cost of training GPT3 and its parameter count, if we assume a linear cost per parameter then this works out to be within 100M-1B to train an NN equivalent to the brain today.”</blockquote><p>-Beren Millidge in a <a href="https://www.beren.io/2022-08-06-The-scale-of-the-brain-vs-machine-learning/">brilliant article</a> on AI and its biological correlates</p><p>The quote above suggests that our current LLM technologies are beginning to approach the order of complexity of the language-processing centers of the human brain. Given positive trends in GPU performance, ever-evolving training algorithms, and exploding language data, it is reasonable to forecast that at some point in the future, our language AI’s (LLMs) may exceed the complexity of our own personal capacities for language. If you continue to extrapolate that trend, similar to how studying a present-day LLM can reveal insights about the <em>individual </em>human mind, we may find that extra large-scale LLM’s may reveal insights about the operation of human <em>civilizations</em>.</p><p>Humans are social creatures. We acquire language from other humans, we transfer information about our world to other humans, and we all share narratives about the order and nature of our world. It may be more apt to say that humans do not “acquire” language, as a consumer may acquire a good in a capitalist economy, so much as humans gain the capacity to plug into a pre-existing language <em>network</em>. This language network is shared across all members that share a common tongue, and the manifestations of this common tongue are everything from information about the weather to narratives about who created us and the morals by which we should conduct our lives. Put more to the point: <strong>I have a suspicion that intelligent systems are fractal, and just as human neurons communicate with each other to promote the Darwinian fitness of the individual, so do cultures emerge from the communication of humans to better facilitate the fitness of the species. By evolving our language models to this scale of complexity, we may be able to reveal (and predict) the behavioral patterns of cultures across millenia</strong>.</p><p>Maybe Asimov wasn’t so far off after all in his psychohistory!</p><p><strong>Malfunctioning Language Networks, or How Social Media Degrades Culture and our Collective Sanity</strong></p><p>Life on earth evolved the central nervous system over millions of years. The human brain stood on those shoulders and developed the capacity for language over many thousands of years. The result is gray matter of considerable complexity but amazing sophistication — a human child can develop the capacity for language with a tiny fraction of the data required by our LLMs. This capacity for language in humans is a result of the pre-trained neural networks produced by evolution, and it is likely that our LLM’s, as our training algorithms evolve in sophistication, will become increasingly efficient at such pre-training.</p><p>However, what happens when you start mucking about with such delicate machinery, hard-earned over millions of years of evolution? I made the point earlier that I suspect intelligent systems are fractal, i.e. that intelligent sub-systems will mirror the structure of the larger system. In the brain, hormones modulate communication from neuron to neuron. For example: the generation of cortisol impacts our central nervous system in a way that makes the body corporate more likely to survive in the <a href="https://www.sciencedirect.com/science/article/pii/S0306453016305315#:~:text=Cortisol%2C%20the%20primary%20glucocorticoid%20(GC,%2Dterm%20potentiation%20(LTP).">case of a perceived threat</a>. However, while hormones have definite short-term utility, a body that <a href="http://the generation of cortisol impacts our central nervous system in a way that makes the body corporate more likely to survive in the case of a perceived threat.">perceives itself to be in a constant state of threat can cause long-term damage to the overall system</a>, i.e. the physiological impacts of long-term stress. I suspect that there are corollaries between the impact of, e.g., prolonged elevated hormone levels and social media’s impact on culture and its ability to facilitate large-scale human cooperation. Let me explain.</p><p>Over thousands of years, humans have used language to communicate a wide variety of shared conceptions about the nature of the world we live in. We discovered the power of metaphor by relating, e.g., the cycles of the sun and moon to our language and created <a href="https://en.wikipedia.org/wiki/Ra">Gods</a> that govern the earth. Brilliantly, we collectively re-discover common heroic archetypes that populate fables across civilizations. These archetypes, most commonly associated with the <a href="https://www.amazon.com/Thousand-Faces-Collected-Joseph-Campbell/dp/1577315936">work of Joseph Campbell</a>, can be interpreted as heuristics for large-scale human societies. For example: societies ranging from Native Americans to early Europeans and the Chinese evolved the cultural archetype of the “<a href="https://www.amazon.com/Trickster-Makes-This-World-Mischief/dp/0374532559">trickster</a>”, which plays the role of elevating the profane to promote creative destruction and rejuvenation when our societies become stagnant. In short, I believe the artifacts of human culture exist as a kind of “hormone” that regulates not the interactions of neurons in a human brain, but rather the interactions of humans in large-scale society. The fitness function of culture is the propagation of its constituents, i.e. its human participants and their language, and cultural artifacts like heroic narratives, shared values, religions, and other manifestations are how culture acts upon humanity.</p><p>Social media has disrupted our cultural heuristics just as prolonged cortisol levels can lead to disfunctioning neural circuits and chronic diseases. A well-functioning civilization, with functioning archetypes, identity narratives, and so on, can evolve incredible feats of intelligence, like the development of the scientific method, mathematics, and art. A society with <em>disregulated</em> archetypes, or failing notions of common identity, can lead to the disintegration of the culture, which often is facilitated by mass genocide. In the 19th century, Neitzsche warned the world about the cost of science displacing religion — “God is dead”. Without developing an ubermensch, i.e. without creating an adequate substitution for the cultural institution of the Church, society risked devolving into chaos. In fact, this is more or less what occurred in the twentieth century with the successive world wars.</p><p>I raise these issues because I believe social media is eroding our cultural norms in a way that promotes cultural degradation and raises the probability of large-scale violence. It is <a href="https://www.independent.co.uk/life-style/social-media-facebook-twitter-politics-b1870628.html">well established</a> that social media and its engagement algorithms promote antagonistic exchanges between its platform participants. It is also the case that social media is particularly effective at degrading prevailing social norms, since a combative attitude toward social behaviors is <em>precisely what social media is designed to promote</em>. I say these things not to take a position on one side or the other of, e.g., contemporary discussions on the nature of gender, but rather to point out the risk of tearing down institutions like the varying roles of gender in society without a replacement for that institution. Social media is perfectly weaponized to tear down established cultural machinery, but it is not adept at fostering the collaboration that generates <em>new</em> cultural machinery. If we allow for the destruction of our cultural commonalities without considering their replacement, we risk violence.</p><p>With social media, there are now more narratives than ever about what is happening in our world — is there a conspiracy that kept Donald Trump out of the White House? Is it Israel or Palestine’s fault and who deserves to be murdered? Is China a force for good or evil in the world? Is US history a history of colonial oppression or the triumph of individualism? Is there room for both explanations? Is the earth flat? Did humanity actually travel to the moon? In order for our societies to function well, we need shared cultural norms to bind our individual intelligences together to breed the brilliance that society-scale intelligence can produce, e.g. the aforementioned developments of science, mathematics, art, and technology.</p><h4>To bring us back to the main thrust of this essay as established from the start: artificial intelligence can automate the menial and teach us more about ourselves and our societies. Functioning large-scale societies, enabled by culture, facilitated the development of AI. Alternatively, social media serves no ostensible social good and is actively degrading the efficacy of culture to enact large-scale cooperation. As such, we should regulate social media and foster innovation in AI research.</h4><h4>Conclusion</h4><p>Culture, i.e. our shared narratives and artistic, linguistic expressions of those narratives is important. Under-investment in the humanities in our cultural institutions and a denigration of those topics in the capitalist zeitgeist raise the risk of large-scale violence in our world. Sustained political divides in the West, violence in Eastern Europe and the Middle East, and escalating tensions in Southeast Asia are a testimony to the cost of social media’s tendency to tear down shared cultural norms, e.g. the concept of global US hegemony and the championing of Western values. We should put energy into studying and regulating this phenomenon not to re-establish old cultural artifacts, which are fraught with their own limitations, but to establish new ones more appropriate to the modern age.</p><p>This is not to say that all cultural manifestations are “good”, insofar as I am able to use that word which is itself a manifestation of shared cultural norms. Christianity, as a cultural force, introduced the concept of forgiveness and original sin in a way that <a href="https://www.amazon.com/See-Satan-Fall-Like-Lightning/dp/1570753199">reduced long-established historical cycles of mimetic violence</a>. Christianity also created the Catholic Church, which has perpetuated any number of horrific atrocities on the world for the past thousand years. That said, without <em>any</em> large-scale shared cultural norms, the greater intelligence of our species threatens to disintegrate, leading to greater social atomism and reduced fruits of collective civilization. Social media is great at promoting atomic narratives of our world, but it is <em>not</em> good at replacing them. Civilization has given us the beginnings of AI and all of its promise— we should not ignore the perils of social media and allow it to destroy those vaunted shared values.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6ba25d802250" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-narrative-and-the-noise/we-shouldnt-rush-to-regulate-ai-when-social-media-is-the-real-danger-6ba25d802250">We Shouldn’t Rush to Regulate AI When Social Media is the Real Danger</a> was originally published in <a href="https://medium.com/the-narrative-and-the-noise">The Narrative and the Noise</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Our Leaders Are Worried About ChatGPT’s Deceit. Instead, We Should Embrace Its Ability to Lie.]]></title>
            <link>https://medium.com/the-narrative-and-the-noise/our-leaders-are-worried-about-chatgpts-deceit-instead-we-should-embrace-its-ability-to-lie-b46c34cbf841?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/b46c34cbf841</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[fiction]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[generative-ai]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Fri, 19 May 2023 13:47:30 GMT</pubDate>
            <atom:updated>2023-05-19T14:39:23.926Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Lies are not a bug of our language, they are a feature. To get the most out of ChatGPT, we must embrace them.</em></p><p>On a recent trip to Miami I found myself terrified. With one leg dangled onto a balcony and another planted in the condo, I peered off the ledge of a friend’s fifty-fourth floor view. The vista was undeniably beautiful, with expansive views of Biscayne Bay from downtown Miami. However, I am afraid of heights, and as such I enjoyed the view mostly through the floor to ceiling windows.</p><p>Out on the balcony were braver souls, or at least people less predisposed to vertiginous afflictions. The fact that they so easily sauntered up to the precipice, peered over, offered a comment about the spectacle, and walked back into the condo was unsurprising given their profession. These were individuals who were accustomed, after all, to routinely ascending the slopes of intimidating intellectual domains. As management consultants at a top global firm, each could be relied upon for what, in another time, would have amounted to an astounding feat of mental ability.</p><p><em>“Have a question? Step right up! Let the amazing Count Consultant, from the distant lands of Corporatedom, skilled in the black magic of fact-conjuring, hear your questions and answer any query your mind can produce!”</em></p><p>Curious about corporate governance? How about asking where the bathroom is, but in Brazilian Portuguese? Interested in how to transition from half guard to a rear-naked choke, or if that is even a valid sequence in the discipline of jiu-jitsu? Or how about GDPR data regulations in the EU? Step right up, Count Consultant has an answer for you!</p><p>I’m poking fun here, but these folks were truly impressive. Their sociability and ability to speak on a vast expanse of subjects with at least some degree of informed opinion was a delight to experience as an interlocutor. That said, despite the pleasant conversation, as we imbibed the view and some beer, I couldn’t help but be haunted by a malignant, Lovecraftian presence lurking off the edge of the balcony, somewhere just out of the expansive views of the condo. It was harshing the intellectual (or pseudo-intellectual) vibe we were all doing our best to cultivate.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/0*yLli4Hj5LS6fUVH_.jpeg" /><figcaption>Not my photo, but representative</figcaption></figure><p>By coincidence, that presence was where we were headed that night — the Kaseya Center. Now, I have nothing against the stadium itself. It is a perfectly lovely venue that did a great job at hosting the UFC festivities we enjoyed that night. The malignant aspect of the stadium hides just out of view of spectators on the ground. However, fifty four floors up, we had a perfect view of the stadium’s dirty little secret. If the brilliant lights and spectacle of the entrance of the stadium could be considered its Dr. Jekyll, then what hid just above could be considered its Mr. Hyde.</p><p>The stadium’s previous name, still brazenly displayed on the roof, albeit it slowly deteriorating in the Miami sun, was FTX Arena.</p><p>If the friendly consultants were vast repositories of helpful, factual-sounding information, FTX as an entity, as a concept, as a group of people, represents the opposite. FTX was a pioneer in the crypto industry, founded by the eccentric Sam Bankman Fried. It was a company that was leading the charge in redefining the global financial industry away from corrupted central bank-controlled currency schemes toward a decentralized libertarian wet-dream of free market exchange. Crypto tokens could be spun up, wound down, traded, and stored at FTX, with a Cambrian explosion of free market activity flourishing around these tokens. Miami (by way of the company’s headquarters in the Bahamas) would be the vanguard of the future of finance, nay, of capitalism! And Bankman-Fried and FTX would be the harbingers of a better techno-crypto tomorrow.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VLxsNLTxYLsRmEFa.jpg" /><figcaption>Crypto-worshippers in the wild at a 2021 Miami conference</figcaption></figure><p>Of course, as with many people, companies, and concepts as hyperbolic as Bankman-Fried, FTX, and crypto, it was all a lie. If we asked our consultants, they would be quick to point out first the definition of a Ponzi scheme, followed by an illustrative example (e.g. Bernie Madoff) of that kind of fraud, and finally an explanation as to how FTX was just such a Ponzi scheme. However, despite the disapproval of our dear consultant friends, we should not be so quick to disregard the distinctly American impulse that led to the creation of FTX just because it was a lie. To better understand why, let’s consider a couple familiar, culturally significant phrases many of us may be familiar with.</p><p>“<em>We hold these truths to be self-evident, that all men are created equal…</em>”</p><p>“<em>If everyone fought for their own convictions there would be no war.</em>”</p><p>Although each of these statements, today, ring of truth, there is an easy argument to be made for both that, at least at some point in time, they were lies. The first statement, of course, is from the Declaration of Independence. It would not take an Oxford debate champion to argue that, at the time of its writing, such a statement was wrought with, at least, morbid irony. A group of idealistic white men were rebelling against foreign, monarchical colonization all while, in many cases, being slave owners themselves. Their statement — that “all men are created equal”, was contradicted by all the available facts of the day and even their own actions. If we were to ask the management consultants of the day, freshly decked out in their finest colonial garb, if such a statement were “true”, they would undoubtedly respond, in perfect, period-appropriate vernacular, “nay”!</p><p>Nevertheless, such a bold proclamation was beautiful and powerful. Contemporary experiments in democracy were ignited by this proclamation, with France, Britain, and continental Europe following suit over the next century despite the best efforts of the ruling monarchies to resist the movements. The bulwark of Western democratic values arguably started with an <em>aspirational lie</em> that was used as a beacon for our culture to mobilize around with respect to individual liberties. This founding lie, despite the occasional setback, continues to be further instantiated in the United States with each dance performed by our three branches of government. Our founding lie is a useful lie that has become, to an imperfect but nevertheless undeniable extent, <em>true</em>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*ddfvxXrmAiRuZXIs" /><figcaption>The United States’ founding aspirational lie</figcaption></figure><p>The second statement is a musing on the power of national narrative and its eternal battle with our individual identities. Despite the resonance of its message, it, similar to our passage from the Declaration of Independence, can be considered a kind of useful lie. This is because it is a line spoken by an entirely made up person — Andrei Nikolayevich Bolkonsky from Tolstoy’s great novel War and Peace. Many lines from the Declaration of Independence were useful lies that were adopted by a culture and made progressively more true as they were integrated into the national narrative of millions of people. This line from Tolstoy is a bit of language with entirely fictitious origins, but nevertheless speaks volumes of truth about just those kinds of nationalistic narratives. In short — just because a bit of language may not have a rigorous foundation in what some call empirical fact does not mean it cannot be useful or powerful to humanity. The Declaration of Independence was a visionary, idealistic document full of statements that were lies until they were proven true by the global democratic movement. Tolstoy’s fiction never happened in historical fact, but still speaks monumental and unparalleled truth about our world.</p><p>If the practice of earnest fact-based communication, exemplified by our consultants, is the Dr. Jekyll of our language, then the lies of Bankman Fried can be considered the Hyde. However, as we just established, not all lies are malignant — some can be useful with respect to driving change in or revelation about our world. Our most powerful metaphors (“all the world’s a stage”) are literally untrue. So what do the Dr. Jekyll of truth-telling and the Mr. Hyde of lying have to do with ChatGPT?</p><p>There is value in the consistency bred by the practice of earnest truth-telling. ChatGPT has the ability to be the best truth-teller we have ever known, far beyond the abilities of any individual consultant. ChatGPT’s Large Language Model (LLM) has <em>billions </em>of parameters, which means that it understands billions of language fragments (concepts) and how they relate to one another. As such, it has at its command far more knowledge about our world, instantiated in digital token form, than hundreds or even thousands of such consultants. As we are already seeing, ChatGPT can be an incredible tool for promulgating a consistent, granular worldview. From a pedagogical perspective, there is tremendous potential for scaled learning with such a tool.</p><p>Given ChatGPT’s promise of truth-telling, there is a natural impulse to prevent it from ever lying. Elon Musk says we should build a “<a href="https://www.theverge.com/2023/4/17/23687440/elon-musk-truthgpt-ai-chatgpt">TruthGPT</a>”, and there are many efforts to curb ChatGPT’s generative capabilities from plumbing controversial domains or activities that can be categorized as “false”. Despite literally creating the attention algorithm ChatGPT is based on, Google has been, until recently, reticent to release their own version in their search products because they are worried about what the system may say to its users. Indeed, Geoffrey Hinton, the inventor of the digital neural network, recently quit his job at Google due to his fears over what new digital agents like ChatGPT are capable of. He suggests that, with these kinds of generative agents, humans may “<a href="https://www.nytimes.com/2023/05/01/technology/ai-google-chatbot-engineer-quits-hinton.html">not know what is true anymore</a>” after a given amount of time interacting with them. I appreciate Hinton’s note of caution, as I think, if irresponsibly deployed on large social platforms, generative agents like ChatGPT could do much to drive harmful narratives across our populations. However, I think it is both fruitless and a shame to begin to limit ChatGPT to only telling “truths”.</p><p>Human creativity and idealism can be thought of equally as lies and truths depending on who you speak to. From a scientist’s perspective, a high resolution photograph of a French pond is more “true” than a Monet. However, an enthusiast of impressionist art may argue that Monet’s depiction of a pond in his home of Giverny, France is more “true” than the photograph because it more aptly captures the <em>ideal</em> of <em>all </em>ponds, not just that individual instantiation of one. To a biologist, a great work of literature (fiction) is objectively a lie — often no characters, and certainly no language, in these works ever existed outside of the page. However, as we saw with our line from Tolstoy, sometimes these fictions can speak enormous truths about people in general.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*XcTu3YdOBT5rMPK3.jpeg" /><figcaption>This painting is a lie — it does not accurately portray reality. But it is beautiful nonetheless.</figcaption></figure><p>My point is that if we crank down heavily on generative AI’s performance parameters like “Word Error Rate” we risk eliminating the possibility for the world’s most powerful language agent to <em>create</em> beautiful new narratives, or <em>reveal</em> in novel ways truths about our world, or be <em>idealistic </em>about how our world can be. The extent to which an idea is not represented in our current corpus of language, or our prevailing models of knowledge, is the extent to which its prediction by generative AI will generate “Word Error Rate”. If we try too hard to make ChatGPT and other generative AI into TruthGPT — if we try to make our generative agents only consultants/Dr. Jekyll — we will lose out on the enormous potential of the Jungian shadow of these generative agents, i.e. the beautiful, but also potentially destructive creative impulse.</p><p>My suggestion is merely this: let the scientists and teachers create their TruthGPT. I suspect they may find it an endeavor more difficult and fraught with controversy than they initially suspect. But let’s also allow the creatives to build CreateGPT. We may get some Sam Bankman Frieds and we may get bizarre cults worshiping the goat-god of a compelling new generative-AI holy text. However, we may also get a brilliant screenplay that helps, as Tolstoy suggests, “everyone fight for their own convictions” just that much more, dissipating the war-mongering clouds that currently dominate our skies.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b46c34cbf841" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-narrative-and-the-noise/our-leaders-are-worried-about-chatgpts-deceit-instead-we-should-embrace-its-ability-to-lie-b46c34cbf841">Our Leaders Are Worried About ChatGPT’s Deceit. Instead, We Should Embrace Its Ability to Lie.</a> was originally published in <a href="https://medium.com/the-narrative-and-the-noise">The Narrative and the Noise</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Data may be the New Oil, but Most of it is ‘Heavy Crude’ and Not Worth Refining]]></title>
            <link>https://medium.com/the-narrative-and-the-noise/data-may-be-the-new-oil-but-most-of-it-is-heavy-crude-and-not-worth-refining-20a30e180a87?source=rss-30a4cde40e37------2</link>
            <guid isPermaLink="false">https://medium.com/p/20a30e180a87</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[dirty-data]]></category>
            <category><![CDATA[data-engineering]]></category>
            <dc:creator><![CDATA[Dave Fay]]></dc:creator>
            <pubDate>Tue, 03 May 2022 15:28:45 GMT</pubDate>
            <atom:updated>2022-05-20T13:50:53.109Z</atom:updated>
            <content:encoded><![CDATA[<p>With the Ukraine conflict, many of us have read proposals for the West to <a href="https://www.economist.com/the-americas/can-venezuela-help-the-west-wean-itself-off-russian-oil/21808837">shift away</a> from Russian oil toward alternatives, e.g. Venezuelan oil. One of the primary concerns of doing so is the difficulty of processing Venezuelan crude oil. In short, the oil from the South American country is <a href="https://www.rigzone.com/training/heavyoil/insight.asp?i_id=185#:~:text=With%20nearly%20320%20billion%20barrels,and%20oozes%20rather%20than%20flows.">extra heavy</a> and requires specialized equipment to refine it to levels where it can be used as fuel in industry and by consumers. As such, it is unclear as to whether Venezuela will be able to scale production to meet the need of an international market pivoting away from Russian oil. I think this example of international oil markets is useful as an analogy to data in the business ecosystem — we have all heard that “<a href="https://www.wired.com/insights/2014/07/data-new-oil-digital-economy/">data is the new oil</a>”, but is it all really as valuable as we think? More to the point — how much of our business data is easy to refine versus “heavy crude” that isn’t worth the effort? Unfortunately, unless your company offers exclusively digital products and experiences, your data most likely requires a lot of expensive, fragile refining prior to it being useful for reporting, not to mention more advanced use-cases like machine learning. Far from being an unhappy insight, I think this realization is empowering and will help data leaders across industries avoid the <a href="https://venturebeat.com/2019/07/19/why-do-87-of-data-science-projects-never-make-it-into-production/">87 percent of data science projects that fail</a>. This article will elaborate on the challenges faced when “refining” data and will offer a more productive model for companies hoping to realize value from machine learning sooner rather than later.</p><p><strong>First Challenge: Your data requires constant massaging by in-demand specialists</strong></p><p>Those of us who have worked as data analysts are aware of the unilateral, invisible decisions we make to process the data into the specification given to us by the data scientist, product manager, or business analyst. Any single one of the coding decisions we make can make or break the resulting data’s suitability for a given use-case. For example: a script that ingests multiple datasets, joins them, and aggregates them for reporting purposes could be built to support a BI dashboard. However, as is very often the case, a data scientist will catch a glimpse of the dashboard and think to herself “huh, that data looks like it could be useful for a model I am building!” The data scientist will then grab the data and begin binning, un-binning, and massaging the data to make it fit the form required by her model. This, however, is where things get dangerous. There are likely any number of decisions made by the data analyst in the original script that may invalidate the data’s use for the data-science use-case. For example: certain raw values could have been normalized to meet arbitrary reporting needs, but no attention was given to preserving statistical signal between the normalized variable and another variable predicted in the model. However, unless the data pipeline, with its source data, output data, and processing steps are <em>very</em> well-managed, it is unlikely that the data scientist will realize the unsuitability of data until it’s too late — the model will be less performant than if it was built from an optimized dataset, or perhaps not performant at all. This example is illustrative of why most ML projects fail — the data that is used for the models is not purpose-built for the occasion, and the specialists and infrastructure to solve for this issue are challenging to hire, retain, and manage.</p><p>Further, even some of the companies pouring the most budget into these specialists and data / modeling infrastructure will continue to face these challenges for the foreseeable future. This is because, in addition to the complexities of data processing summarized above, much of the raw data we have access to in industry is very difficult to refine <em>to begin with</em>. Even assuming your company is capable of world-class data engineering and management — which would put them in the top 10 percent of companies globally — the data you’re starting with may never yield value commensurate with the investment it will take to refine it. The nagging challenge of dirty data is due to the incomplete landscape of digital transformation across industries.</p><p><strong>Second Challenge: Even with specialists, your data may not be clean enough from the source to realize value</strong></p><p>Companies like Google, Facebook, Netflix, and the other leading Tech companies benefit from have primarily digital products — Netflix is enjoyed on computers and mobile devices, Google’s product suite operates on the same screens, and so on. By having digitally-native products, these companies have the critical ability to <em>optimize for raw data quality at the point of data generation</em>. For example: when users are filling in the company they work for on Linkedin, Linkedin normalizes that data by auto-completing users’ input into a standardized list of companies maintained, you guessed it, by LinkedIn. This means that LinkedIn has full awareness and control of its data elements at the point they are created, which makes it far easier to “refine” this data for various use-cases, ranging from reporting (“how many users work for Company X”) to machine learning (“what is the probability of Y user attrition from X company?”).</p><p>However, consider an alternative situation in which the raw data is <em>not</em> digitally-native, or is at least partially non-digital. I work in fintech, and financial services offers a multitude of these cases. For example — many financial institutions, though they are finally investing in digital loan origination experiences, still maintain networks of branches and, *gasp*, paper applications. This means no matter how much effort the bank invests in the new digital application experience, the continued existence of the paper application and its lack of control over the data scribbled on it means that the overall data will be “dirty”. In other words, some cleaning prior to “refining” the data for any use-case will be required for as long as the branches and paper applications exist. If you multiply this need to clean data, analogous to the specialized machines needed to prepare Venezuelan oil, across all data use-cases, organizations will face the same challenge of scaling their refining efforts. It takes, in other words, an undue amount of time and effort for non-digitally native companies to clean, engineer, then leverage data, which is why 87 percent of data projects fail. This case is true in banking, but also for dozens of other industries that continue to have material operations “in the real world.” Although reduced reliance on non-digital data capture and sensors will reduce the dirtiness of data, this transformation will take time. As such, if non digitally-native companies want to start realizing value from data <em>now</em>, they will need to shift their model of investment toward one of more targeted investments + partnerships.</p><p><strong>The Solution — Experian Case Study</strong></p><p>Are these challenges insurmountable? Should non digitally-native companies despair and abandon their data engineering and science investments? Absolutely not! Instead, what these challenges suggest is that non digitally-native companies should 1) be <em>more focused</em> with their investments in the space and 2) explore partnership models to leverage high-quality data <em>today</em> with the support of domain experts. In terms of the investment focus, non digitally-native companies need to accept that data is not their area of competitive advantage — yet. By laser-focusing investment on the projects needed to <em>make</em> data a competitive advantage, they can ensure their “oil” (data) is just as refineable as Big Tech’s. Additionally, in parallel, these same companies can realize value from data and machine learning <em>today</em> by leveraging partnerships with data and analytics specialists in their domain.</p><p>Focus on your area of competitive advantage and partner for the data to supercharge it. This is the strategy that can enable companies who are not digitally-native to overcome the challenge of integrating data and machine learning into their products, services, and operations. Let’s consider an example in the financial services industry — that of a bank that would like to drum up some high-quality applications for its new credit card product.</p><p>To deliver analytically-supercharged offers using conventional, in-house approaches, the bank would need to inefficiently spread its investment capital across its data infrastructure, modeling infrastructure, mobile, and branch experiences. This usually results in sub-par results in each area — the mobile app could lose certain automated notifications, or the data platform could be developed to only focus on integrating with legacy portfolio systems vs. offering new data engineering features. This state of affairs is described in Figure 1 below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_Hoi47cCjXGvqEt-bUuVNQ.png" /><figcaption>Figure 1</figcaption></figure><p>Worse, spreading investment this thin across so many domains results in poor technical coordination of all the separate components. As Figure 1 above indicates, the offer channels (branches, mobile and online experiences) are often tied to the analytical and data platforms via tenuous, fragile, ad-hoc data loads. Further, the actual analytical selection of customers to present the offers to are often buried under human-facilitated, manual processes of approval (dotted lines), further clouding and clogging the decision-making process. What this all leads to are delayed timelines, inferior performance from the campaign, and ambiguous, ad-hoc decision making at each stage. Crucially, this also leads to dissatisfied customers in the mobile and branch experiences since, because investment is spread so thin and the customer-facing experiences suffer as a result.</p><p>As we mentioned above, however, this need not be the case. Driving customer value from data and analytics can be fast-tracked by 1) focusing investment on their <em>digitally native, refined data-yielding experiences</em> and 2) relying on strategic partnerships with data and analytic vendors to compensate for the lack of refined data in the short-term. This model is described below.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cEH-I_elNVm11Xqos0mItw.png" /><figcaption>Figure 2</figcaption></figure><p>Figure 2 illustrates a model banks could leverage with a vendor like Experian and the Ascend platform they offer. By leveraging Ascend and its 1) onboard data, 2) host of API connectivity options, and 3) model-building and deployment capabilities, banks can circumvent the complexity of investing in all of those required components for the offers campaign. In other words, banks can realize value in the short-term from data and analytics. Further, and most importantly, the API connectivity of the Experian solution would enable banks to have formalized, auditable, and <em>automated</em> integrations with the execution channels and all intermediate systems required to execute the campaigns. This is in contrast to the confused network of ad-hoc integrations and manual processes that would be required from building all of the data and analytical functionality in-house. Not only will time to market improve, but the results of the campaign should also improve due to Experian’s refinement of all the data required for the analytical work.</p><p>Hold up! Doesn’t this kind of solution seem short-sighted? Shouldn’t banks be investing in improving their data infrastructure to be viable in the long-term instead of prioritizing short-term returns? Of course! However, the model in Figure 2 does not preclude long-term investments in data engineering and analytics. To the contrary, it <em>enables it </em>in a way not possible with the in-house model. Figure 2 illustrates how banks can, in the short term, refocus their short-term investments away from data and analytics toward their <em>digitally-native experiences</em>. By doing so, the mobile app and branch application experiences can be revamped to the highest digitally-native standards of Big Tech — all fields can be standardized, and the data captured in those applications can be highly “refineable.” By leveraging a vendor partnership like that of Experian, banks can realize value in the short-term from analytics projects like credit card offers while also enabling a future that is articulated below in Figure 3.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QjIm_TEpan3L7-sp4eLk0g.png" /><figcaption>Figure 3</figcaption></figure><p>Figure 3 describes a truly differentiated bank, benefitting from refined data and analytics from Experian while also offering further differentiated experiences only possible because of their short-erm investments in becoming digitally-native. The Experian Ascend solution is integrated via API’s with the rest of the bank, which means that its maintenance relies not on a network of manual processes that are expensive and complicated to maintain, but on automated technology. This further means that the bank can ‘leave that engine running’ while focusing on expanding capabilities elsewhere. If, as described, the bank has successfully improved its digitally-native experiences such that highly-refined data is now available for analytics, it can then take the next step of investing in a brand-new data and analytics platform of its own, e.g. on AWS. That new platform can be enhanced not just with the bank’s new digitally-native data, but also via API’s with the Experian platform to generate datasets of unprecedented quality and richness. At this stage, the bank has leapfrogged its competition and now, instead of being tied to the ground by supporting a vast network of legacy systems and processes, the bank can instead focus its efforts on, e.g., building and deploying projects like a digital assistant. The reliance of the new architecture on API’s ensures that the system is flexible and plug-and-play as new capabilities are added to the bank.</p><p><strong>Conclusion</strong></p><p>Many industries rely on oil to operate, but are not experts on refining the resource themselves. This same model holds true for industry when it comes to data— not every company, at least in the early stages, has access to the world-class data and analytics expertise and infrastructure to benefit from those technologies. In fact, as we’ve argued above: <em>it is impossible to realize an advanced analytics future until you have access to such refinable data</em>. Companies should partner with domain expert vendors, such as Experian, in the short term to realize value <strong>today</strong> from refined data and analytics. By doing so, short-term capital can be freed up to complete the transformation of the company to a digitally-native enterprise, which will yield returns in terms of refined data. This kind of short-term strategy will enable the <em>long term </em>realization of the dream of becoming a global leader in differentiated consumer experiences on the level of the Googles and Netflixes of the world.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=20a30e180a87" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-narrative-and-the-noise/data-may-be-the-new-oil-but-most-of-it-is-heavy-crude-and-not-worth-refining-20a30e180a87">Data may be the New Oil, but Most of it is ‘Heavy Crude’ and Not Worth Refining</a> was originally published in <a href="https://medium.com/the-narrative-and-the-noise">The Narrative and the Noise</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>