Week 5 (4/22,24): Artificial Intelligence

[Gio-Choi]

The Artificial Intelligence program was kicked off with Alan Turing’s notion of computing machinery and intelligence and through an expansive research program in artificial intelligence laid out in a 1955 Dartmouth summer conference. What is the definition and standard of intelligence laid out for machine imitation in this research program? What are implications of choosing this standard for subsequent development? What kinds of technologies might emerge, and not emerge under such a standard? What are ethical considerations of the AI program? What are the core differences between a bottom-up (learning logic from data) and a top-down (designing logic from theory) approach to AI development, which won the battle over modern AI, and why? 

Post your response as a Comment in reply to this message.

[aadarwal]

The most interesting aspect of the reading was Turings paper and the realization that our modern idea of artificial intelligence based on cognitive processes was very different to the kind of idea of a machine that he was trying to build out for his artificial intelligence - particularly one that is based on behavior that can be tracked. This interesting to me because understanding his reasoning for doing this being something embedded in building imitation machines was also very different from the goal for the kind of machines that we are trying to build out today. In fact the fact that the term AI was coined with regards to this concept of a model in mind for replication of task performance and passing off as humans in these tasks was interesting because by a lot of these definitions we have crossed such a bar in so many fields. While reading literature from the past and seeing preliminary interpretations for artificial intelligence never expected me to wonder why they did not think like us, I expected a more far ended approach from Turing in terms of reach of capabilities, and so it was particularly interesting to see this path that he employed and the manner in which everyone else followed suit. Choosing such a definition also meant that research would be less centered around neuroscience and underlying physical processes and more around the need to replicate more mechanical tasks to meet the end goal. If I had to guess I would also say that this might be one of the reasons that our understandings of the brains and the functioning of human cognition is so preliminary even to do this - the fact that our interpretation of creating the next kind of intelligence has not been embedded in our own intelligence's understanding - but simply in the replication of its capabilities..
And so I think that although McCulloch and Pitts had already formalized the logic mechanisms in neurons, the deeper questions that they were trying to answer - in terms of the experience of machines and whether it was actually thinking in the same manner as humans, remained unanswered because they did not match the goals of what Turing or the Dartmouth team envisioned and coined. However, one good thing that I believe that this entire approach has left us with is in terms of defining human creativity subsequently. In the past this was easier when the goal was not cognitive ability matching, and so its become easier too understand why when artificial intelligence has started to threaten that capability of humans we have been struggling to relate the respective roles of human ability since then, and so I think some of the ethical or social conditions have been more modern for this very reason and not something we fully saw being discussed in some of these works. Lastly, I think the fact that the mechanism of AI today (bottom down) being so fundamentally different from that which the people of the past who were at the forefront of the field saw it (top down) is only testament to the fact that the extent of both its capabilities and promises and unbounded in either direction in my opinion.

Discovery:
Learning that Walter Pitts was a big part of the University of Chicago was a big revelation to me and I spent a good time because of this deep diving into the past of the University of Chicago and their connection to the field of neuroscience and particularly computational neuroscience. It was incredible for me to see this because I had never realised that the university had played such a large role in shaping the work that had foundational created and built up the entirety of the field. In addition to this another important realization for me was the extent to which Lisp had been a big part of the expectations for artificial intelligence in the past and the manner in which it had shaped symbolic interpretation. I know Lisp and so was vaguely aware of the role that it had played, but reading some of the background on symbolic computation and this actual relevant history of AI was very important in that regard to formalism the extent to which it had actually shaped the field.

[cavanbgon]

Under the research program we study now, artificial intelligence may be well understood in its capacity to effectively “resemble” human behavior.  This comes from one of Turing’s critical insights. Barring issues of inferable consciousness, machine thinking can be understood as those methods necessary for the development of a non-human agent capable of succeeding in the posited imitation game. The basic advantage of this framework is in its ease of articulation, neglecting issues of what truly autonomous thinking is. However ambivalent we wish to find Turing’s attitude towards the rigorous, positive establishment of the qualities of a machine given the capacity to think, his view can be well contextualized among the prior developments necessary for the establishment of first, a general computational model, and then the surprisingly elusive notion of computerized human-imitative behavior. Or rather, it may be best to understood to understand as having – conscious or not – as an abstraction of the prior history of the development of these necessary features.

In prior cases, the development of machinery with greater computational facility came as the categories of statistical and computational problems emerged. For their emergence, it was in either case necessary for the essentiality of such problems to be established as reflective of deep structure previously unavailable. In the case of computational problems, this is basic confidence in there being a “sufficient” description of a system from the application of further layers of computation. The upper limit of this comes perhaps with Laplace’s demon, but I find the more relevant, characteristic case to be in the calculation of the behaviors of many particles of moisture in the atmosphere. In the case of statistics, this comes with Quetelet’s vindication of the notion “mean man” as, rather than abstraction from a population, the underlying object from which individual human specimens emerge as secondary objects. In both cases, we see that what is identifiable and calculable is favored over what is in a strict sense. This is evidenced best by in both cases, the novel discovery is placed as a central object in an entire model of either human affairs or the universe wholly.

Turing understands that what has been essential to the development of his novel science of computation has been crucial emphasis on the calculable. Now, with sufficiently broad language for what rightly constitutes computation, Turing is capable of newly asserting a broadly “computable” condition under which humans may identify the end of a thinking machine. With this notion of imitation established the problem in some sense begins anew. It is, as I have written, possible to interpret the gradual triumph of neural net technology over symbolic models of cognition. This is perhaps less obvious than it seems. Instead, what may be in a stronger sense the better model for understanding the development is the necessary willing development of a computable model which defies immediate articulation, opting instead to imitate (perhaps naïve) physical models. With this established it is the stronger task of the theorists to place this development within our gradually expanding language of sufficient mathematical comprehension.
 
Explore:
McCarthy’s grant proposal demonstrates a remarkable sensitivity to the varying forms the basic AI problem could take, which seems on its face almost unbelievable. To have intuited the bifurcation between rigorously understood symbolic systems and the imitation (even if acknowledged as “visual” and essentially naïve) is a testament to the immense progress in the maturity of scientific understanding achieved leading to the middle of the 20th century. This is, of course, not to mention the brilliance of the conference’s organizers. However, overburdened the term has become, “abstraction” as an acknowledged feature of scientific and mechanical progress would have (despite roots in Plato, if we wish to see them) would have seemed alien in the 19th century.

I would feel remiss neglecting our “field trip” to Professor Stigler’s home. However much its acknowledgment is a cliché, there’s real irony to the flattening of knowledge, which has come with easier access to information. Professor Stigler presents an antidote to this. There is an abundant sense in his presentation of a desire to find, draw in, and ultimately spread endless curiosities. To be engaged with the world is to be unsatisfied with filing curiosities away, something far too easy today. I hope the hushed passing around of a silver medallion or playing around with a toy bell curve can help us to see this.

[mhmanek]

Original Materials
The 1955 Dartmouth proposal, inspired by Turing’s 1950 paper, framed machine intelligence as the capacity “to act in ways that we would call intelligent if a human were so acting.” McCarthy, Minsky, Rochester, and Shannon posited that “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” In practice, this set the Turing Test—linguistic indistinguishability from humans—as the touchstone of success, and elevated symbolic manipulation and formal reasoning as core benchmarks of “thinking” machinery.

Choosing a human‐imitation standard steered research toward logic, search, and rule‐based systems. AI pioneers built theorem provers, game‐playing programs, and general problem solvers that excelled in constrained domains. However, this emphasis on clear, symbolic “chunks” of reasoning led to brittleness: systems lacked robustness in unstructured environments, and struggled with perception, adaptation, and commonsense reasoning. The focus on narrow linguistic or logical tasks inadvertently marginalized learning from sensory data and adaptive, statistical approaches.

Under the Dartmouth criterion, symbolic planners (e.g., the General Problem Solver), expert systems (MYCIN, DENDRAL), and dialogue simulators (ELIZA, SHRDLU) flourished, showcasing rule‐based inference and knowledge representation. Conversely, resilient vision systems, scalable speech recognizers, and end‐to‐end learning architectures languished until revived by later breakthroughs. Early researchers largely bypassed neural networks and probabilistic models, delaying the rise of machine learning approaches that excel at tasks unsuited to crisp logical rules.

Top-down AI encodes human expertise into formal logic and explicit rules, while bottom-up AI derives structure from data via statistical learning (neural networks, probabilistic models). The initial Dartmouth era privileged top-down methods for their interpretability and alignment with symbolic reasoning ideals. However, with the explosion of data and compute in the 2000s, bottom-up approaches surged ahead: they proved more scalable, robust to noise, and adaptable across domains.

The bottom up v/s top down debate in modern times seems to have arrived at a definitive conclusive answer with the advent of Alpha Go: by simply letting the algorithm play against itself millions of times without any human feedback or input (through reinforcement learning), it becomes better than anything conceivable.

Discovery
The field trip to Prof. Stigler's residence was a wonderful experience - observing the Laplace painting and its similarity to the one next to it (in terms of hand position - since hands were difficult to draw) revealed the historical, cultural (artistic) context of Laplace's period. The fact that most of Laplace's paintings are from his old age rather than youth and were painted after his demise were interesting factoids about the period's culture. There was also the so-called drawing of Legendre (which was in fact of a different person with the same name) - a portrait of the real mathematician Legendre is not known.

[benkim04]

The standard for intelligence was created by Turing, and is now known as the “Turing Test”. This test involves replicating human intelligence by fooling someone into believing that it could possibly be human, or in other words, becoming indistinguishable from human intelligence. Doing so requires several important caveats: the communication between machine and human cannot be done so through modes like speech or sight, as doing so would immediately dispel the illusion. Instead, intelligence is defined as the ability to respond to questions and have a conversation, most importantly through the medium of text. This provides imitation of human language and thought, but not much beyond that. It is not required to imitate human speech patterns, appearance, motor control, or anything else physical. These limitations enable it to be more feasible, and more importantly it narrows the scope and benchmarks that are important to the test. What becomes paramount is natural language processing, whereas other features like imitating human mannerisms, sounds, and behaviour becomes de-emphasized. From this, the emergence of the LLM is quite intuitive. It is entirely text based, and is a fitting medium for “text-message-like” interactions which are well-suited for the Turing Test. When creating such a model, it is important to understand first what human responses would be like, and understand that it may include racist, sexist, or otherwise unsavory content. Should we wish the model to perfectly model a human, stereotypes ingrained within humans would inevitably arise within a model built to respond like a human. This then creates a dilemma: either engineer to be exactly human at the cost of negative consequences, or avoid those negative consequences but lose some aspects that are fundamentally human. Of course, it is up to debate whether those aspects are fundamentally human. Are they part of our biological tendencies? Or are they instead learned from those around us? For the model, at least, it is clear that it is learned in that we can manipulate training data and curate it to help the model avoid learning such negative stereotypes.
The core difference between the bottom-up approach and the top-down approach is how much the architecture of the model. For example, we can compare neural networks and transformers, the former being top-down and the latter being bottom-up. The neural network has a fixed set of nodes, with fixed activation functions, and learned weights. By “fixed”, it does not mean that they cannot be changed, but instead indicates that they must be changed by the engineer. In contrast, the transformer will take the input and break it into tokens of varying lengths, and so we observe a dynamic set of nodes, dynamic interactions between the nodes, and learning ensues. What is important here is that top-down is often highly curated, and will often require less data to train. Transformers are often more effective, but require more computational power and are more generally applicable.

Discovery:
What I found interesting was how much forethought Turing and his peers gave to concepts that they could not fully create at their time. For one, Turing is capable of imagining his test, understanding the limitations of it, and how it should eventually be implemented, all while knowing he would not live to see it. Further, there was substantial discourse among his peers as well as to the eventual rise of AI, and he was capable of creating counterarguments against them. Still, he is not naive to the limitations of his theory and to his own understanding, at which point he concedes that he does not have much positive justification and more theoretical estimations.

[suryachinnappa]

It seems evident from the very beginning, the goals and endeavors of the emerging field of computer science were inextricably linked with the quest for artificial intelligence. The works and vision of Alan Turing exemplify this. The paper “Computing Machinery and Intelligence” that we read this week, establishing what we now know as the Turing Test, takes a substantial departure from the previous Turing paper we read, “On Computable Numbers, with an Application to the Entscheidungsproblem”, which among other things established the Turing Machine. While the latter paper is extremely mathematically formal, the former keeps to the conceptual, musing about the social possibilities of machines that can pass off as humans. In many ways, the fruits of Turing’s mental labor wouldn’t reach full fruition until recent years; he correctly predicted that language would bear the most likely first interface by which machines would be able to approximate human capacity to the extent that they might pass as human.

We now interact with artificial intelligence as a distinctly socioculturally situated technology. The last three years have served as a sort of global uncharted experiment into the ways this technology has impacted culture and society. Notably, conceptions of what exactly artificial intelligence is have a definite impact on shaping the way different people, institutions, societies etc. interact with it. We can examine two possible ontologies: first, artificial intelligence models as agents. Treating these bundles of math as personable entities with distinct voices, personalities, and selves is arguably the most common way to conceive of them; I feel like this may boil down to an inherent human predisposition to anthropomorphize everything that we feel is proximate to humans: when something can talk like us, we cannot help but to think of it as alike to us. Some research in the field makes an implicit indication of this stance–for example, work analyzing the ways in which LLMs are changing the linguistic landscape often represent them as entities contributing to the linguistic landscape alongside human language users. However, this representation runs into issues when we consider the necessary question of the boundaries of the self. After all, a self is defined by its being set apart from the world beyond it, but determining these boundaries with respect to artificial intelligence is challenging. Is every single instantiation of a language model a distinct self? What about individual conversations, between which a model does not transfer memory? Or is an entire model across all of its instantiations and conversations a single self? What implications could such a knowledge-rich entity hold? On this note, we may examine the second ontology: that of artificial intelligence models as refined access methods for human knowledge. This approach is more grounded in the data-driven principles underlying AI methods. In this case, rather than constituting independent nodes in the vast web of social interactions, artificial intelligence may manifest as weights along edges between humans, merely a modulating factor of communication between them, as one accesses knowledge produced by the other, etc.

Discovery: Playing the AI Dungeon was very interesting, as I got to experience firsthand the process of being creative with AI assistance. I very rarely write with the assistance of language models, so seeing how I could channel the story forward or change it dramatically based on how I continued the story was very interesting. It reminded me of a research project I heard about by Prof. Mina Lee, called CoAuthor, which collected a dataset built on a very similar principle of AI-assisted writing. That system was substantially different, being mainly human-led with the AI playing an auxiliary role (while in the AI Dungeon, these roles were essentially reversed), but both explored a fascinating dynamic between mind and machine, a research landscape which is only just beginning to emerge as these technologies continue to shape our social world.

[ucachyut]

Original Materials

The 1955 Dartmouth Conference, building on Alan Turing’s work, defined artificial intelligence as mimicking human intelligence—reasoning, problem-solving, and learning—with the Turing Test, which checks if a machine’s conversation is indistinguishable from a human’s, as the standard. 

This focus on conversational mimicry capped AI’s intelligence at human levels, implicitly admitting the idea that human intelligence is unsurpassable, thus limiting AI’s potential to go beyond human capabilities. By setting human behavior as the benchmark, it steered AI toward replicating what most humans do well, like talking or writing, rather than, say doing original mathematical or scientific research.

As a result, most modern AIs excel at mimicking human conversation and producing coherent text, while AI that can tackle complex problems in math, physics, or biology, or think creatively outside the box, lags behind. For example, modern AI can write a convincing story or chat about the news, but it often stumbles on a math problem that lies even slightly outside its training data. Further, AI has so far been unable to invent new scientific theories, showing its weakness in deep, creative reasoning.

Ethically, AI’s potential to disrupt society raises uncertainty and hence fear, especially around job displacement. In a hypothetical future where machines outperform humans in every task, personal farming machines could let everyone grow their own food, harvest crops, and meet their needs without relying on markets. Imagine a farmer trying to sell crops when a machine can do it cheaper and faster—how would they compete? This could lead to a self-sufficient, isolated world where people depend on their own machines, potentially dismantling economies and raising questions about human purpose. What would humans do if machines handle all labor? How would society function without traditional jobs or trade? While changes brought about by AI might in fact lead to better living standards and more positive changes than negative, society a hundred years from now is beyond our imagination. 

AI development splits into two approaches: top-down, which tries to impose a theoretical “universal language” on machines, like math proof assistants that follow strict logic, and bottom-up, which trains machines with vast natural language data, like internet texts, to act like or outperform humans. Bottom-up AI, powered by deep learning, has clearly won in modern AI because of the abundance of data—billions of web pages—and strong computational power, while top-down struggles with limited data for a universal language, despite its potential to create precise, logical systems. The victory of bottom-up reflects our reliance on data-driven solutions, but it also means AI excels at mimicking us rather than mastering fields like math or creative innovation, aligning with the Turing Test’s human-centric focus.

Discovery:
During the field trip to Regenstein’s special collections, I was amazed to see a century-old book on the calculus of finite differences, a topic I studied in a high school project and learned was taught in schools 100 years ago. It was incredible to flip through its pages and also see Enrico Fermi’s notebook with his intense calculations and diagrams, showing the raw math behind his nuclear work. This experience taught me how much effort went into early computational methods.

[peterzhou-cs]

The Artificial Intelligence program was kicked off with Alan Turing’s notion of computing machinery and intelligence and through an expansive research program in artificial intelligence laid out in a 1955 Dartmouth summer conference. What is the definition and standard of intelligence laid out for machine imitation in this research program? What are implications of choosing this standard for subsequent development? What kinds of technologies might emerge, and not emerge under such a standard? What are ethical considerations of the AI program? What are the core differences between a bottom-up (learning logic from data) and a top-down (designing logic from theory) approach to AI development, which won the battle over modern AI, and why? 

Post your response as a Comment in reply to this message.

The 1955 Dartmouth Conference proposed a research program for artificial intelligence, and it was clearly influenced by Turing’s 1950 paper on machine imitation. In that paper, Turing set up what’s now called the Turing Test—basically, the idea that if a machine can imitate human responses well enough to fool someone in conversation, then we can call it intelligent. It’s a clever shift away from asking whether machines really think, and instead focusing on whether their behavior is indistinguishable from that of a human.
The Dartmouth proposal carried this idea forward but added some concrete goals. One of them was to build machines that could work as universal computing devices—something that could perform any computable task. Another goal, which I find especially interesting, was the idea that an intelligent machine should be able to master language. The reasoning was that humans use language to think, so a machine that thinks would likely need to use language too. That assumption already sets up a strong connection between intelligence and language use. At first glance, that connection seems to be reflected in today’s large language models. LLMs like GPT are the main public face of AI today, and they’re all about generating language. So in a superficial way, they seem to fulfill the Dartmouth goal: they can use language fluently, and sometimes even convincingly. But there’s a deeper distinction here that matters. The Dartmouth proposal seemed to assume that language is the medium through which machines would think—just like it is for us. But LLMs don’t actually use language that way. They operate on tokens. They don’t think with language; language is just their output. Internally, they’re doing mathematical computations, optimization over massive data, and the end result happens to look like linguistic reasoning. But it’s not the same thing.

Another difference that I think is important is the approach—whether we design machines from theory (top-down), or let them learn from data (bottom-up). Classical algorithms like RSA are designed from mathematical principles. They’re precise, deterministic, and secure, and it’s hard to imagine discovering them just by studying large datasets of encrypted messages. But once we leave those well-defined areas, and enter messy, unpredictable domains like language, it becomes much harder to design clear logic from scratch. That’s where data-driven approaches take over. In fields like natural language processing, we often don’t know the rules—we just have the data. So we let the models learn patterns from it, even if we don’t fully understand how they work. That’s a very different vision from what the Dartmouth group might’ve expected, but it seems to be where we’ve ended up.

Explore/Discover: reading the grant proposal after 75 years is especially meaningful, where some goals were partially achieved(the computing machine, creativity), some seems to be accomplished with a revised understanding on these concepts (neurons into neuro-network, language as working medium into working target), and some are still completely out of reach (self-improvement) The record of weekly deaths in the 17th century Britian was also worth-noting: the numbers are specific and are organized in categories of death/gender/location, marks the earilest form of data analysis. (there was one person died from 'burnt in the bed by candle'. R.I.P.)

[Laurelli7]

Original Material:
The idea of machine intelligence in early AI started with Alan Turing’s 1950 paper “Computing Machinery and Intelligence.” Instead of asking if machines can think, Turing suggested a practical test called the Imitation Game, now known as the Turing Test. According to this test, a machine is intelligent if it can carry on a text conversation in a way that a human can’t tell if it’s a person or a machine. This test focused on behavior rather than actual understanding—it was all about how well the machine could imitate human responses.
This idea shaped the direction of AI for years. Researchers focused on tasks that looked smart to humans, like solving puzzles, playing games, or understanding language. But it also meant that physical abilities—like seeing, moving, or interacting with the real world—were not seen as important. As a result, technologies like logic-based systems and rule-following programs grew, while work on neural networks and robotics was left behind. Even though in “A Logical Calculus of the Ideas Immanent in Nervous Activity”, simple brain-like networks could model logical thinking, these ideas didn’t get much attention at first. I think that even today’s large language models (LLMs) still reflect this focus on language imitation. Some people say LLMs already pass the Turing Test, but we know they don’t really “understand” anything—they’re just good at copying how people talk. Critics also point out that LLMs can produce false or misleading information, which makes them seem smart but causes real problems.
There are also ethical concerns with this imitation-based view of AI. On one side, it pushed people to think about what “thinking” really means. On the other hand, it raises questions about trust and honesty. The Turing Test is built around tricking someone into thinking a machine is human, which can lead to confusion, especially in online spaces. “Large AI Models are Cultural and Social Technologies” argues that today’s AI tools aren’t really minds of their own—they’re more like cultural tools that reflect the information we give them. That means we need to think carefully about where they get their data and how they’re used.
Over time, AI research split into two main approaches. The top-down method tries to write rules that tell the computer how to think. The bottom-up method lets machines learn from lots of data. At first, the top-down approach was more popular, but it didn’t work well in complex situations. Later, the bottom-up approach, especially deep learning, became much more successful. As LeCun, Bengio, and Hinton (2015) showed, deep learning made huge improvements in areas like image and speech recognition.
Overall, the Turing Test gave early AI a clear goal, but it also limited how people thought about intelligence. Bottom-up learning eventually became the main path forward—not by copying how humans think, but by learning from the huge amount of data humans have created.
Discovery:
On the trip to Regenstein’s Special Collections and Professor Stigler’s residence, I got to see a historical copy of Pascal’s work on early computation, along with an old computing machine labeled with a graph that mentioned “artificial intelligence.” It really surprised me to see how physical and mechanical the early stages of computation were. Seeing these objects in person made me feel unexpectedly connected to the roots of the field—like even as a student today, I’m part of a long timeline that started with tools like these. One of the machines looked more like a dough press than anything we’d call a computer now, but it still carried the label of AI, which was wild to think about. It made me realize how long people have been trying to build machines that could replicate how we think or solve problems. It also reminded me that computation isn’t just about advanced algorithms or digital systems—it started with very basic machines and big ideas. Seeing how far we’ve come, from Pascal’s mechanical calculator to today’s deep learning models, made me think about how fast the field has evolved and how our understanding of intelligence and machines keeps changing. Even if those early devices couldn’t “think,” they started something that shaped how we see AI today.

[evancholerton]

Alan Turing’s 1950 paper reframed machine “thinking” as successful imitation of human conversational behaviour: “We now ask the question, ‘What will happen when a machine takes the part of A in this game?’” (p. 433) By focusing on typed exchanges rather than inner states, Turing established the "imitation game" as a clear, empirical benchmark that even today remains a lodestar for AI evaluation.

In the decades that followed, many AI efforts hewed to this top-down ideal of hand-crafted reasoning: symbolic planners, expert systems, and logic-based theorem provers dominated, but they proved brittle outside narrow domains. Deep learning overturned this paradigm. As LeCun, Bengio, and Hinton explain, “representation learning…allows a machine to be fed with raw data and to automatically discover the representations needed for detection or classification”. (p. 436) By stacking simple non-linear modules, modern networks learn edge detectors, motifs, object parts and entire concepts from data, sidestepping the limitations of manual feature-engineering.

More recently, researchers have recast AI not as autonomous agents but as cultural and social technologies. Farrell et al. argue that “large models are themselves best understood as a particular type of cultural and social technology”. (p. 1153) Like markets compressing economic complexity into prices or bureaucracies summarizing populations into statistics, large language and multimodal models summarize, reorganize, and reconstruct vast bodies of human knowledge, but don't necessarily think in a human sense.

This shift from agent-centred intelligence to social mediation brings ethical considerations into sharp relief. If we judge success by how well a model “passes” for a human, we risk entrenching human prejudices: racist or sexist language patterns are learned as easily as neutral ones. Moreover, the imitation standard can narrow AI’s horizons, valuing mimicry of human tasks over the discovery of genuinely novel insights or superhuman capabilities.

A key question for further investigation is how to augment rather than merely replicate human cognition. Can we develop evaluation frameworks that reward originality, robustness and ethical reasoning alongside conversational fluency? For example, hybrid benchmarks might combine behavioural tests with measures of transparency, fairness, and societal impact, ensuring AI systems contribute positively without simply mirroring our biases.

Discovery
While reading Turing’s section on machine architectures, I discovered his proposal for “digital computers with a random element”, noting that such machines “are not normally possible to determine… whether it has a random element, for a similar effect can be produced by… the digits of π” (p. 436) ​The early insight that stochasticity could endow machines with a form of “free will” resonates today in debates about AI autonomy and unpredictability, suggesting that randomness may be essential to creating adaptive, socially embedded systems.

[malongo03]

I will address the Summer Conference's definition and standard for artificial intelligence because I believe the Turing Test is, for many reasons, an extremely poor standard for machine intelligence.

The definition of artificial intelligence given for the Summer Conference proposal would be an emulation of the rules and algorithm of human cognition. For example, they consider that language can be considered a kind of instruction set or logic that human brains manipulate with relatively fixed rules of cognition. Therefore, they consider the great limitation to getting machines to produce and understand language is finally decoding and describing how human minds internally manipulate language. Their standard, unlike Turing's, is that the mark of successful artificial intelligence is a description of how the intelligence could be produced.

The implications of this are a variety of artificial intelligence that are, at least initially, directed away from efficiency and towards purity of theory and emulation of human thought. A "top-down" AI approach.

The limitation for this is obvious: a focus away from efficiency allows for inefficiency. And especially the emphasis on having completely rigorous theory before construction can create a kind of analysis paralysis. Sometimes ideas have to be tested instead of solved.

Technologies formed under this dogma also have the weakness that they will be trapped within the limitations of humans, as opposed to exploring ways to improve efficiency away from them. For example, a hypothetical robotic farm laborer may seem like a good idea, except tractor plows and harvesters turn out to be way better ones. I think the same is true with AI.

I do not believe AIs deserve any ethical considerations except as they effect humans. Turing is wrong, in my view, and though he would construe my view as religious, so is his. He takes it on blind faith that the world and therefore consciousness is entirely materialistic and mechanistic, and while that is certainly a reasonable assumption for physics, I think that is an absurd axiom to impose on the entire world for philosophy and morality and theology. I do not believe the value of a human being comes from being a particular collection of atoms whose calculations fulfill certain qualities.

Bottom-up AI won eventually, despite the best efforts of the top-down, and I believe this is because it got results faster and more effectively. As I said before, the weakness of the top-down approach given in the conference is it is held back by perfectionism in theory while not in efficiency. With bottom-up AI,

I am a human supremacist. I believe that AI based on human cognition will eventually prove to be superior in its computational efficiency than current human-agnostic AI designs. But I think we require a far better understanding of brains before such a thing is possible. Once it is, I believe LLMs will be considered idiosyncrasies of history, similar to the mechanical calculators we discussed earlier in the quarter.

Explore:

I am consistently surprised by the age and origins of thoughts and words and phrases. That "artificial intelligence" originated in 1955, while I know "robotics" was coined by Isaac Asimov in 1941, blows my mind. It just seems like a phrase that should have come up sooner, especially with the earlier readings taking about the mechanical calculators being artificial minds. More generally, the Dartmouth Proposal was a frustrating read, because only 2 of the 7 agenda items (4 and 5) have really had any progress at all. Most of the other have been side-stepped or ignored after the subversion to item 1, since we decided to dedicate much of our energy to expanding computing hardware. Again, that's my bias as a cognitive science major and human supremacist. I want to understand people, not train matricies.

[jpprieto]

Turing’s proposal for evaluating AI rests on human evaluation, at the time if an AI was convincing enough to be able to model a human reaction to a set of rules it would have been thought intelligent. In modern theory we seem to adapt the Turing test from gender discernment to being convinced through conversation that a machine is actually a human.
I find one of Turing’s quotes directly ironic in this regard, “….we should feel there was little point in trying to make a ‘thinking machine’ more human by dressing it up in such artificial flesh.” Here he is referring to the physical constitution of the machine, but in my opinion it can expand cognitively to what his test may fail on: That the machine is replicating the structure of thought without an internal computation.

I feel a framework that may be useful to dig deeper into this paradox is Weisberg’s construal of the model. Particularly the notion of dynamic vs representational fidelity. In which dynamic fidelity is how well a particular model or machine replicates the expected effects of a cognitive process, and representational is how closely it maps to the actual cognitive methods to produce such a result. A machine capable of passing Turing’s test would have high dynamic fidelity, as do LLMs, as often they are convincingly conversational. But while arguably human cognition translates conceptual structures into words, LLMs operate in reverse: They predictively construct linguistic form without necessarily engaging in abstract conceptualization.

I think an interesting case we may use from a previous reading is that of Chef Watson. We generated a system that is rated as creative, with high ratings by chefs and tasters. But how creative was it? Going back to Weisberg’s lens, the system exhibited dynamic fidelity as it mimicked human chefs’ output. Yet, I would argue that it lacked the representational fidelity one might expect from a truly creative mind. It did not discover new thermodynamic transformations, innovate new cooking methods, or develop fundamental theories of flavor. Like LLMs, it was all simulated creativity made by remixing prior data within a constrained space. Which illustrates the direction technology may (or does) take under this evaluative approach.

The ethics of generating a ‘deception’ machine can be seen in current trends. How do we deal with massive integration of AI into information systems, where their output may sway public opinion, allow distributed enactment of scams, or overall reduce the confidence of users on content being human generated. But the ethical implications of a machine with better representational fidelity can be just as concerning, if its thought process arises from a true interaction of internal representational spaces rather than statistical processing, then its output is more unpredictable. What if it develops goals that are harmful to society? It is a common question in AI ethics regarding AGI.

In terms of bottom-up and top-down, where we either design the system explicitly such as symbolic AI or give data to a well-crafted structure and let it design its algorithms, respectively. Arguably the latter is more prevalent in modern AI, particularly deep learning. Where the system develops from the raw data its internal representations. I would say a significant component of this being the case, is simply that metacognitive modeling is very difficult. To generate complex symbolic or algorithmic processes of thought requires deep expertise and intelligence. While deep learning allows the machine to not only generate these processes itself, but also arguably produce some that are more native to its structure.

Discovery: Playing with neural networks this week was fascinating. I struggled to lower the loss on the spiral dataset using small architectures. So the week’s reading led me to a rabbit hole of trying to build a neural network in NumPy to better understand the mathematical mechanics. While I didn’t manage to fully implement backpropagation in time for this week’s memo, learning about it in an attempt to implement it was interesting.

For a bit I have had this idea about what a neural system would look like if it operated on continuous waveforms rather than discrete vector spaces. Which is an interesting contrast of micro vs macro effects in comparison to McCulloch and Pitts convention of neurons as boolean circuits. I got inspiration from neuroscience in thinking about what the brain may be doing mathematically with EEG-like dynamics or cortical fields. But now that I understand more, here we couldn’t implement backpropagation because the assumptions of differentiability wouldn’t hold. How, then, could a system like this tune itself? Is this why such a framework hasn’t been implemented? What kind of optimization principles would be necessary, is it mathematically possible? I would really appreciate any thoughts on what a backpropagation analogue might look like in this case or if we could use other principles to achieve a similar effect.

Additional sources:
Weisberg, Michael. 2007. “Who Is a Modeler?” The British Journal for the Philosophy of Science, Vol. 58, No. 2, pp. 207-233

[Alex7692]

Alan Turing considers the possibility of intelligent machines by employing the conceptual tool of "imitation game"; that is, whether a machine can bypass an interrogator's alert in successfully imitating human actions.

Turing test underscores the inferential role of conceptual reasoning. To borrow Robert Brandom's terminology, we as rational thinkers engage in a "game of giving and asking of reasons". There is a fundamentally normative, and hence pragmatic dimension to the way we construct propositions. Logical reasoning makes explicit our stakes in such a game. While we may remain neutral to this deflationary characterization of logic (which returns to the "classical" definition logic as the capacity to judge, as opposed to modern symbolic logic), we may certainly, as with Turing, think of it as an important criterion for intelligence, and in particular human-like intelligence.

But to what extent is human intelligence the only form of intelligence? Most scholars today would be incredulous against, for instance, the thesis of plant intelligence or the extravagant, post-humanist thesis of biosemiotics, insofar as we may reflect on the "intelligent" behavior of any self-feeding, complex system of objects. The fact that neural network, at least on the conceptual level, imitates the multiple realizability, adaptivity and continuity of human brain explicitly shows the stance people have been taking. Machines also imitate in the sense of auto-learning. To that end, both Lovelace's objection and the objection from various disabilities seem to be met with solutions. (Note the conflation of terms intelligence, rationality, self-consciousness, human-like, etc.)

The skeptic's question, however, persists, especially as we are at a stage of AI development of which the interpretation of AI decisions and behaviors become significantly problematized. And Turing does realize this point when he considers the converse scenario, in which a human is asked to perform more machine-like. Biases that would result in caricatures of machine behaviors aside, it makes us question whether we are really striving after, and hence implicitly acknowledging Leibiniz's law (i.e. indiscernibility of identicals). It would seem like, on the very intuitive level, that a machine can perfectly be intelligent even when people understand it is a machine. Just as people could distinguish kinds or groups of the general population, perfect imitation is not quite the strict requirement we want AI to abide to. (Incidentally, the very idea of "imitation" is a human and social psychological one in the first place.)

Now consider the difference between bottom-up and top-down approaches to AI development. It would seem like: whereas AI today has fully maximized the bottom-up approach to AI, human thinkers would mostly engage in a synthetic mode of learning by framing the exploration in a space of reasons. This distinction is reasonable as we consider iterations of AI not as ontogenetic, but phylogenetic, and thus follows a natural course of macroevolution. It would be a lot harder to argue for such radical individuation in the case of the human learner, as a sense of self-consciousness persists.

Discovery (original material)

We had the great pleasure of visiting Prof. Stephen Stigler's residence on Tuesday. And one of the objects we've been looking at is a 19th century calculator model. It would seem like a machine back then is imitating human intelligence insofar as mechanizing the computing algorithm - and as such we really treat it as a prosthetic, an extension of our consciousness. But would we assume the same expectations for AI today?

I've also been exploring the AI dungeon generative RPG tool. It is pretty good, considering how difficult it may be to play a round of D&D-like games without a proper dungeon master. And in terms of doing its task of world-making, it is doing a good job at imitating the role of dungeon master (although from the perspective of players, would it matter if we know that it is a human or machine generating the fiction we are immersing ourselves within?)

[meeradasgupta247]

Original Materials

Thinking Machines
Meera Dasgupta

Cognition is a phenomena that individuals have often interrogated from multiple standpoints. However, one ideology which continues to persist is how cognition is a human trait - where the ability to initiate precedes the ability to form thoughts in determining whether or not something is cognizant. Though the readings seem to take on this viewpoint, where machines will be able to imitate but not necessarily think, this view seems to solely rest upon thinking as a human trait and not consider alternative models outside of the human - where more similarity exists than not. One alternative definition, would be to consider thinking as the mode of being able to derive a conclusion from external information/stimuli. In this way, machines are able to think more than we realize and perhaps we too are closer to machines.

To begin, the alternative definition posed above does not incorporate the idea of feelings nor emotion. Though emotions can drive - and are intertwined with - thought, this definition rests upon a fairly simple diagram of assessing some input by using external/internal information to draw conclusions. Hence, when Alan Turing describes the imitation principle, though he differentiates imitation from initiative, he still describes a machine without initiative as “a person who ‘would have no common sense’” (Alan Turing: the enigma, p. 478). Though the goal for Turing seemed to be far more than imitation, then, imation was enough to ascribe humanness to the thought process of the machine - despite it not necessarily seeming to be the most rational or intelligent in terms of the human expectation.

Still, even this is reconciled by considering the machine as a “boy” (Alan Turing: the enigma, p. 479) who is being instructed by an adult. This seems to also be the idea depicted by supervised learning in the “Deep Learning” paper (LeCunn et al) where training is more straightforward. With the large amounts of data which are fed to these models/algorithms, one may very well liken them to experience. I would argue that pattern recognition in images, for example, is somewhat of a learned trait as - when some patterns are observed multiple times, they instead constitute familiar objects. Pretraining, as seen in the same paper, almost seems to mimic the evolutionary nature of thinking - where a concept is able to further be applied to other similar domains.

Perhaps even if we were to venture into emotion, though machines may not be able to experience emotions, empathy has not often been equated to thinking. In fact, some aspects of emotional analysis, such as facial recognition of emotions, are still able to be learned by machines. Then, we come to a question that is more akin to Theseus’ ship; to what extent may we remove aspects of human thinking, if any, before this thinking is considered mechanical? Or, perhaps must we reconcile that our thinking is also more mechanical than we realize - subject to the biases which we encounter and limited to the world which encompasses us.

Discover/Exploration

I read “The Great A.I. Awakening: How Google used artificial intelligence to transform Google Translate, one of its more popular services — and how machine learning is poised to reinvent computing itself” by Gideon Lewis-Kraus (New York Times). Though I do find it incredible that google translate is able to interpret works a lot better than earlier, I am still fairly skeptical about its ability to interpret artistic forms of creative writing - especially in less-resourced languages. Not only does translation account for meaning in the original language but also attempts to translate words/phrases in context while maintaining the integrity of words that may be untranslatable in other languages. Additionally, translation considers the priority of the author when having to sacrifice some aspect of the text which is difficult to do solely by discerning meaning - for example, if a poem is utilizing vowels or consonants in an interesting way, the translation would have to decide whether to prioritize form or meaning.

Also, when visiting Prof. Stigler’s residence, it was very interesting to see many of the rare books/media which he had in his collection - from the very large book detailing barometric pressures to the calculations of deaths in the past week (the n = 1 data points indicated that the cause of deaths were considered on a case-by-case basis where they were either added to a pre-existing category or led to the creation of their own category). Given that I hadn’t seen a photo of this yet on Github, I am adding a photo of the mini Galton board. It was fun to interact with it directly and also physically see the normal distribution come to life.

Citations

Hodges, Andrew. Alan Turing: the enigma. Princeton University Press, 2014, Forward and 474-490

LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. “Deep learning.” Nature 521, no. 7553 (2015): 436-444
“The Great A.I. Awakening: How Google used artificial intelligence to transform Google Translate, one of its more popular services — and how machine learning is poised to reinvent computing itself” by Gideon Lewis-Kraus, New York Times, Dec. 14, 2016

[rshah1869]

Alan Turing’s reframing of the artificial intelligence program as “The Imitation Game” defined intelligence as behavioral rather than metaphysical. Turing also defined the parameters of the game by maintaining the “machine-like” nature of the program by limiting interaction to the efficient question and answer method. According to Turing, “we should feel there was little point in trying to make a ‘thinking machine’ more human by dressing it up in such artificial flesh” (Turing, 434).
Turing’s conception of the discrete machine grounded the purpose of AI in predictive models based on existing data. In his description of the discrete machine he says, “reasonably accurate knowledge of the state at one moment yields reasonably accurate knowledge any number of steps later” (Turing, 440).
This imitation-based definition carried profound implications for subsequent AI development. As Farrell et al. (2025) argue, choosing behavioral mimicry as the standard shaped both research agendas and public expectations. It incentivized technologies that succeeded in linguistic simulation—chatbots, virtual assistants, and now large language models like ChatGPT. However, this standard also marginalized forms of intelligence tied to embodiment, affective reasoning, or non-verbal learning, narrowing the field's scope for decades​.
Turing’s contemporaries and successors, including McCulloch and Pitts, attempted to provide a formal neurophysiological foundation for machine intelligence. Their 1943 model treated neurons as binary logic gates, demonstrating that any computable function could be implemented by a network of simplified artificial neurons​. This symbolic approach to intelligence would later be termed “top-down,” emphasizing logic, rules, and internal representations. It was dominant in early AI systems, such as expert systems in the 1980s.
However, as Farrell and colleagues note, this top-down paradigm eventually gave way to bottom-up, data-driven models, particularly neural networks that learn from massive datasets. This approach, inspired by the brain’s adaptive structure, powered the deep learning revolution that now dominates the field. Connectionist models like GPT-4 exemplify how systems can learn generalized behavior from raw data without being explicitly programmed with rules. This transition reflected shifts in computational capacity, data availability, and a willingness to embrace statistical uncertainty and opacity in exchange for performance​.
It is pretty interesting to consider what modern AI might have looked like without Turing’s framing of the machine in the Q&A format and without “dressing it up”. Perhaps we might have robots that more closely resemble humans that reflect emotion or reaction instead of the more tool-like models we have today.
Discovery:
The proposal for the Dartmouth Summer Research Project in 1955 is fascinating as it is feels so early development of computer science as a discipline. Turing died in 1954. Yet, there are so many concepts that feel quite modern and relevant to AI and particularly machine learning such as “Neural Nets”. This research however, feels more biological than mechanical and perhaps reflects a kind of humanistic understanding of what AI could become. The purpose of this project seems to be to create a brain that can think. It is evident that Turing’s “Imitation Game” hadn’t been fully accepted as the de facto way to define artificial intelligence.

[ciara-kinney]

Original Materials

In his 1950 paper, Computing Machinery and Intelligence, Alan Turing recasts the question of “Can machines think?” as, “Are there imaginable digital computers which would do well in the imitation game?” (Turing, 442). Turing argues for his imitation game as the means of determining machine intelligence and parries all sorts of objections he imagines could be raised to his use of this metric. Early on in the paper, he addresses the concern of the possibility of the emergence of legitimate machine intelligence that is not comparable to human intelligence. He counters this concern by arguing something along the lines of that if such intelligence existed, it would still be able to play the imitation game successfully, so “we need not be troubled by this objection” (Turing, 442).

I am not so convinced by the quick dismissal of this question. Although it doesn’t seem to interfere with the possibility of answering the question of “Are there imaginable digital computers which would do well in the imitation game?”, it does seem to interfere with the legitimacy of reducing the broad, vague, and pressing question of “Can machines think” to a simple evaluation of a machine's performance in the imitation game. It seems quite plausible to me that there could, in theory, be a machine which should be regarded as properly intelligent, which would be unable to properly imitate humans. Although I don’t suspect that this is the primary issue that we encounter today with questions of machine intelligence, as I am under the impression that most machines which raise questions about sentience or true “thinking” capacity were designed with human imitation in mind, and would thus likely arrive to their thinking capacities in human-analogous ways, I think that it is still valuable to consider all possible ways in which thinking capacity or sentience could emerge.

It seems quite possible that, in the way that humans are necessarily limited by our physical bodies, minds, and conditions, a machine could be similarly limited. Although humans could be incapable of properly imitating an alien with novel abilities or unknowable and imperceptible qualities, we would still regard ourselves as thinking and sentient creatures. We have restraints that are particular to our constitution that prevent us from accurately representing something we are not, but we are still "thinking". I don’t see why this wouldn’t apply to machines as well? Thus, I don’t think it's any sort of novel complaint, but to me, Turing’s imitation game doesn’t seem to be quite as comprehensive as it initially presents itself to be. His method answers a particular question– “can machines be indistinguishable from humans” – well, but I don't think that this question can be super easily extended to larger questions of true intelligence.

Exploration/Discovery:
Visiting Professor Stigler’s home was wonderful! He introduced us to lots of great things, but I found the books on death records particularly interesting. Although far from the social statistics which would emerge in the 19th century, these books did seem to be the embryonic form of the practice. In addition to the charts containing data, the book contained comments on their observations, such as that “the vast numbers of Beggars, swarming up and down this City, do all live, and seem to be most of them healthy, and strong”, which is then followed by the suggestion that the state assume responsibility for them so that they could be taught to work. It was interesting to see this combination of social data and policy suggestions in such an early book. (Image attached below)

This week, I tried out the AI RPG tool. I tried both prompting it in ways which seemed very natural to the story line (ie “I attack the raiders with my sword to save the innkeeper's daughter”) and ways that seemed quite out line (ie “I start dancing to a groovy tune and sing about the power of groovy-ness") to see how it would respond. I found that despite my attempts to steer the story's direction to some completely alternate direction, the AI seemed pretty steadfast in its trajectory and kept curving me back on track. I imagine this sort of player-led customizable story will be hugely influential in the future of entertainment, so it was interesting to actually engage with a concept I’d only considered before.

Turing, Alan. “Computing machinery and intelligence (1950).” Mind, LIX, no. 236, pp. 433–464.