A system card cannot answer it. RAGAS-style metrics cannot answer it either. They convert proxies such as faithfulness using another model as a judge into numbers that create the appearance of precision without providing a defensible basis for release.

Model-level benchma…

Read more

We comfortably apply the word game to chess, truth-or-dare, and cheese-rolling, even though no single definition covers all those activities. What unifies the category isn’t an invisible core but overlapping similarities and the practical and normative expectations that accompany the label. Edge cases make this vivid. Russian roulette has rules, turn-taking, and chance, so it is “game-like” in structure; but calling it a “game” feels grotesque because our practice treats “game” as bound up with play, trivial stakes, and a certain moral posture toward risk, and not saturated with non-ergodic risk. An essence doesn’t police the concept; they’re negotiated in use, and in use, Russian roulette is no game.

The lesson for AI is far from straightforward, but it’s far less complicated than it often appears. Artificial intelligence is not searching for some inner property and does not demand metaphysical introspection. What is meant by “intelligence” depends on how we use it, and is supported by public criteria, family resemblances, and agreed-upon tests. The right question is not “Is this real artificial intelligence?” which feels oxymoronic, but “What public criteria do we actually treat as counting as intelligence here?” For humans, that might include reasoning and accountability; for AI systems, it is often performance on tasks, benchmarks, and real-world reliability. This was the standard in 2012 with AlexNet, and nothing has changed. The point is not to discover some hidden property called intelligence, but to clarify how we use the term. If we forget this lesson, we end up chasing phantoms.

This pragmatism aligns with Alan Turing’s vision. Turing famously sidestepped the unanswerable question “Can machines think?” and replaced it with a different question entirely. Instead of debating definitions, he proposed asking whether a machine could act indistinguishably from a human in a conversational setting. In doing so, Turing changed the game from definition to demonstration, and from philosophy to practice. He gave us a way to talk about operationally useful AI that is grounded in what a machine can do rather than getting stuck in semantic quarrels about what it is. Sadly, the brilliance of Turing’s move is no longer evident today. He recognized that terms such as “think” or “intelligent” are difficult to define, so he anchored them to observable criteria. If a machine can carry on a meaningful conversation, it has, for all practical purposes, earned the label “intelligent,” at least in that context. Crucially, this epistemic humility understood its limitations. It was a test. It was not a declaration of some eternal truth about the machine’s inner nature or even the definition of intelligence.

Subscribe now

Following Wittgenstein and Turing, I have come to reject the romantic metaphysics that dominate so much of the public imagination and discourse. Romantic metaphysics is the tendency to see an inner light in benchmark performance or the tendency to dismiss benchmark performance because there is no inner light. The demand for a mystical essence from AI is, in my view, a mistake. Recall that the behaviorist and functionalist traditions of artificial intelligence sideline phenomena such as the mind and consciousness. It is odd to see these notions being bolted back onto what is essentially an empty functional characterization of intelligence. AI advances by bracketing metaphysical and phenomenological debates rather than settling them. To treat either narrative as essential to achieving benchmark performance is to leap from solid ground into quicksand.

We must remember that artificial intelligence is a description of what a system can do, not what it is. If someone claims a machine is conscious or insists it can’t possibly be conscious, my first question is: what observable difference would that claim make? If we talk about an inner property “about which nothing can be said,” then “a nothing would serve just as well as a something about which nothing can be said.” In other words, if a supposed feature, like consciousness in an AI, or the lack thereof, does not affect any evidence or behavior, then it is empty to us. It might as well not exist and thus is not worth talking about. Well, it may be worth discussing from a philosophical perspective, but that is precisely what AI is trying to move away from. Essence-talk feels profound because it’s unconstrained, and engineering feels crude because it must be right. Modern AI has preferred the crudeness that works. This mustn’t change.

I’m not asserting that AI cannot be conscious or that it will be; I’m saying that until we have public criteria for whatever we mean by “conscious,” such pronouncements are speculative theology. I choose neither to romanticize nor to nullify AI in metaphysical terms. Instead, I examine how it functions. It’s not that I lack imagination about where AI could go. Instead, I insist that our imaginations remain tethered toreality. The tests, benchmarks, and behaviors are where the meaning is. As I argue in the book, “performance does not equal essence, but insisting on an untestable essence is not good epistemology either.” We must be willing to refuse both mystical overreach and evidential withholding. The pragmatic path is procedural, not metaphysical.

My goal in writing this book was never to convince you of a grand theory or to convert you to a particular ideology about AI. In fact, I have deliberately avoided framing any part of the narrative as “the final word” on intelligence or machine consciousness. Instead, I was to affirm the importance of epistemic humility. Our understanding of intelligence (human and artificial) is still incomplete. As AI advances, it will surely surprise us, and we should be prepared to update our views. However, the remedy is not metaphysics, but better criteria.

That is why I care about language but refuse to fetishize it. Words like “understand” or “think” should be treated as claims we can test rather than as essence-talk or as reasons to dismiss what is plainly demonstrated. Living hope demands receipts; living doubt keeps them honest. The point is not to win philosophical debates about whether a system really understands. The point is to demand receipts. If we want that progress to be safe, our doubts must also become operational. Not essays, sternly worded tweets, or moratoriums, but evaluations and alignment research.

Share

In fairness to Google, they didn’t mistreat the Transformer. In the years immediately following they introduced the decoder-only architecture and BERT, which outperformed models of similar size including autoregressive and decoder-only models. A better characterization is that they possessed ambiguity and faced issues that incumbents often face. As incumbents, they had a lot more to lose from a Tay-like incident, so they were more careful than OpenAI (OAI). They also had an academic culture that liked exploring novel ideas for their own sake. Meanwhile, OAI was more focused partly because they didn’t have the resources to pursue novelty, multiple ideas, or worry too much about an incident.

For example, just a year after the introduction of the Transformer, Google introduced the Universal Transformer. The authors explain that the original Transformer “fails to generalize in many simple tasks that recurrent models handle with ease.”[1] While self-attention was shown to replace recurrence, researchers at google had a hard time letting go.

In fairness, the Universal Transformer did not try to undo the original Transformer’s GPU-friendly parallelism. Instead, it reused the same stacked self-attention and feed-forward layers but added a new “transition function” that unfolds across a recurrent dimension. The authors describe this as a parallel-in-time recurrent self-attentive sequence model where all positions are updated simultaneously, but the model loops over these layers multiple times. This recurrence introduces a notion of time steps, and under certain conditions, gives the Universal Transformer the theoretical power of Turing completeness. That is, the ability to perform any computation a conventional computer can, given sufficient time and memory.

As a result, the Universal Transformer is significantly faster than RNNs, as its recurrence is depth-wise and parallelizable, as opposed to time-step sequential. Yet, it is still slower than the original Transformer and offered no clear performance gains. The supposed weaknesses, such as lacking recurrence and hacky positional encodings, turned out to matter less once larger models were trained on more data. The Universal Transformer embodied a clever idea, but deep learning favors simpler designs that don’t complicate optimization.

Despite introducing the Transformer, Google exhibited ambivalence toward aggressively scaling it. Their internal tensions between theoretical elegance and pragmatic impact contrasted sharply with OpenAI’s. OpenAI seized upon the scalability and empirical performance with GPT-2 and GPT-3. Google’s detour, though intellectually intriguing, suggests at least some hesitation or confusion about what made the standard Transformer a step change (confirmed here by Brin’s statements). OpenAI’s success highlights what Google initially overlooked: Transformers thrive not through careful inductive biases or even Turing completeness, but through simple architectures scaled dramatically.

In an interview on the Lex Fridman Podcast (episode #94), Ilya Sutskever argued, “the Transformer is successful because it is the simultaneous combination of multiple ideas. If you remove any one idea, it would be much less successful.” He pointed out that Transformers are parallelizable and it “is not recurrent… therefore much easier to optimize.” This divergence highlights how differences in organizational culture and leadership profoundly shaped the trajectory of modern AI development.

This hesitancy exemplifies what Ilya often calls out about Google’s culture during this time. Google valued a research culture and lacked decisive strategic clarity. By contrast, OpenAI, under Sutskever’s leadership, demonstrated no such hesitation. Sutskever later recalled the moment the paper appeared, saying: “This is the greatest thing… It’s clearly superior [to] anything else before it.” Ilya’s immediate reaction was unequivocal: “We’re going to do Transformers now.”[2]

The Guardian noted that GPT-2 can produce “plausible” continuations of prompts in many styles and domains, rarely exhibiting the grammatical nonsense or mid-sentence derailments that plagued previous systems. For example, after being seeded with the opening line of George Orwell’s Nineteen Eighty-Four: “It was a bright cold day in April, and the clocks were striking thirteen,” the system recognized the vaguely futuristic tone and the novelistic style and continued with: “I was in my car on my way to a new job in Seattle. I put the gas in, put the key in, and then I let it run. I just imagined what the day would be like. A hundred years from now. In 2045, I was a teacher in some school in a poor part of rural China. I started with Chinese history and history of science.” This lands somewhere between bleak futurism and a mediocre travel blog. The bizarre narrative leaps suggest the model is struggling to maintain narrative coherence rather than deliberately evoking Orwell’s unsettling style.

Nonetheless, GPT-2’s ability to generate coherent, though sometimes peculiar, text surprised and excited observers. Vox’s Kelsey Piper wrote that GPT-2 was “one of the coolest AI systems I’ve ever seen,” so human-like in writing that she quipped it “may also be the one that will kick me out of my job.”^[3] Reporters marveled at GPT-2’s zero-shot abilities. Without task-specific training, it could generate news articles from headlines or answer reading comprehension questions, achieving unimaginable results from an unsupervised model.^[4]

TechCrunch reported that GPT-2 produces “longer text with greater coherence” than prior models, calling it a vast improvement over the original GPT-1 model.^[5] Even OpenAI’s researchers were taken aback when GPT-2 produced a persuasive essay arguing a counterintuitive viewpoint, describing it as “something you could have submitted to the SAT and get a good score on.”^[6] Such anecdotes spread quickly, fueling excitement that AI had generated a new level of sophistication.

Researchers also voiced surprise. For instance, one researcher wrote that he was “pretty shocked” upon reading GPT-2’s now-famous “Ovid’s Unicorn” story, noting that “on the whole it’s remarkably coherent… like a news article that a human could have written”^[7] No prior model came close to GPT-2’s level of fluency, long-range coherence, and apparent world knowledge.

The model’s ability to generate multiple paragraphs of on-topic, grammatical text was an eye-opener for experts accustomed to earlier models that would quickly go off the rails. Some researchers highlighted that while GPT-2’s underlying techniques were not new, its scale made the difference. “There is no algorithmic contribution here,… they are ‘just scaling up’ previous research, but seeing exactly how strong these scaled-up models are is an important contribution in its own right,” one researcher noted, adding that “if I had shown the [GPT-2] unicorn story to a colleague… and told them that it was AI-generated, I don’t think they would have believed me.”^[8] The community recognized GPT-2 as proof that larger models can yield startling new performance levels.

GPT-2 is an attention-based artificial neural network known as a transformer with 1.5 billion parameters, trained on a dataset of approximately 8 million web pages and 40 GB of Internet text. GPT-2 is a direct scale-up of GPT, featuring over 10 times the parameters and trained on more than ten times the amount of data. It is trained with a simple objective: to predict the next word, given all previous words within some context. The diversity of the dataset causes this simple goal to contain naturally occurring demonstrations of many tasks across diverse domains.

Subscribe now

The model generates synthetic text samples in response to being prompted with an arbitrary input. OpenAI described the model as “chameleon-like,” meaning it adapts to the style and content of the conditioning text.^[9] This allows the user to generate realistic and coherent continuations about a topic of their choosing, as seen by the following select sample taken directly from the GPT-2 paper and commonly known as “Ovid’s Unicorn:”

System Prompt GPT-2:

In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains. Even more surprising to the researchers was the fact that the unicorns spoke perfect English.

Model Completion GPT-2:

The scientist named the population, after their distinctive horn, Ovid’s Unicorn. These four-horned, silver-white unicorns were previously unknown to science.

Now, after almost two centuries, the mystery of what sparked this odd phenomenon is finally solved.

Dr. Jorge Pérez, an evolutionary biologist from the University of La Paz, and several companions, were exploring the Andes Mountains when they found a small valley, with no other animals or humans. Pérez noticed that the valley had what appeared to be a natural fountain, surrounded by two peaks of rock and silver snow.

Pérez and the others then ventured further into the valley. “By the time we reached the top of one peak, the water looked blue, with some crystals on top,” said Pérez.

Pérez and his friends were astonished to see the unicorn herd. These creatures could be seen from the air without having to move too much to see them – they were so close they could touch their horns.

While examining these bizarre creatures the scientists discovered that the creatures also spoke some fairly regular English. Pérez stated, “We can see, for example, that they have a common ‘language,’ something like a dialect or dialectic.”

Dr. Pérez believes that the unicorns may have originated in Argentina, where the animals were believed to be descendants of a lost race of people who lived there before the arrival of humans in those parts of South America.

While their origins are still unclear, some believe that perhaps the creatures were created when a human and a unicorn met each other in a time before human civilization. According to Pérez, “In South America, such incidents seem to be quite common.”

However, Pérez also pointed out that it is likely that the only way of knowing for sure if unicorns are indeed the descendants of a lost alien race is through DNA. “But they seem to be able to communicate in English quite well, which I believe is a sign of evolution, or at least a change in social organization,” said the scientist.

Such an output may seem ordinary today. However, in 2019, it was shockingly good and nearly unimaginable two years earlier. Naive observers may have assumed that such a result relied on curated knowledge databases, including placenames, rules engines, or symbolic logic, to produce such coherent results. It’s still hard to believe that a model trained to predict the next word could exhibit this level of narrative control and contextual awareness.

This sample is notable not only for its fluid and convincing prose but also for how effectively the model imitates the style and structure of an authentic news article. The initial sentence introduces the discovery of unicorns in the Andes, and the model dedicates ten sentences to elaborating on that single idea. It presents a fictional academic, Dr. Jorge Pérez, and places him at a plausible university, the University of La Paz, demonstrating sensitivity to geographical, academic, and stylistic contexts. It mirrors the conventions of journalism, such as switching from full name to surname in later references (i.e., “Pérez”), and incorporates plausible-sounding quotes and scientific speculation.

The model shows an impressive ability to track long-term dependencies connecting information across multiple paragraphs. This was impossible with earlier architectures. The model encoded real-world knowledge and effectively modeled statistical relationships between words. It approximates knowledge like:

§ The Andes Mountains are in South America

§ Jorge Pérez is a plausible name for a scientist in Bolivia

§ Universities often bear city names like “La Paz”

§ News articles introduce experts with titles and institutional affiliations

This implicit knowledge is not programmed into the model but is learned from the language used to describe the world. Predicting the next word requires modeling statistical relationships about the world that language refers to. As a result, without task-specific tuning, the model can generate stylistically nuanced, coherent, and context-sensitive text, even approximating implicit world knowledge.

As we follow the generated article, the model transitions to the second sentence of the prompt (“Even more surprising… the unicorns spoke perfect English”) only after addressing the first in detail. This delayed yet coherent introduction of new information demonstrates the model’s ability to structure its output like a human writer, capturing continuity and pacing. The story eventually ventures into fanciful territory, including theories of alien origins. However, the model was trained on Reddit, where whimsical ideas about alien origins are commonplace.

The model generated other nonsense too. The opening reference to “four-horned unicorns” exemplifies an oxymoron and highlights a fragile underlying “world model.” In another sample, GPT-2 confidently describes “fires happening underwater.”^[10] These types of output became an early example of what we now call a hallucination, where models produce plausible-sounding text that is factually or logically incorrect. In 2019, OpenAI termed such outputs “failure modes.”

It's important to note that “Ovid’s Unicorn” was self-selected as the best among ten generated completions, which tempers any impression of consistency and reliability. However, even when generating nonsense, GPT-2’s convincing form, structure, and tone changed expectations about machine-generated language. In other words, even a single high-quality sample like Ovid’s Unicorn in 2019 was unprecedented, even if it was cherry-picked.

The GPT-2 paper provides fewer architectural and training details than the GPT-1 paper and, as noted, is not included on Sutskever’s List. Therefore, we will skip an in-depth technical discussion and review GPT-2’s performance across tasks as presented in the original paper. We’ll explore how attention-based transformers like GPT-2 function in more detail in Chapter 5.

GPT-2 was trained on a dataset called WebText. As previously mentioned, WebText was created by scraping webpages linked from Reddit posts with at least three karma points, ensuring the content was generally informative, educational, or engaging rather than random or noisy. The final dataset included roughly 8 million documents and 40 GB of text. Wikipedia articles were excluded to prevent artificial performance boosts on standard NLP benchmarks.

The model acquired the ability to perform different language tasks without explicit supervision. The authors, including Ilya, referred to this phenomenon as “unsupervised multitask learning,” where the model’s diverse capabilities emerge from training it solely to predict text continuations. This approach differed from previous specialized systems, which relied heavily on carefully labeled datasets and task-specific training.

The researchers tested various models with increasing sizes, ranging from 117 million parameters similar to the original GPT model to 1.5 billion parameters. Larger models outperformed smaller ones across various tasks, demonstrating a consistent log-linear relationship between scale and performance. The largest GPT-2 model achieved state-of-the-art zero-shot results, meaning it performed tasks without prior training or examples, on seven out of eight standard language modeling benchmarks.

However, performance across more specialized benchmarks revealed limitations. When researchers prompted GPT-2 with the phrase “TL;DR:” (meaning “too long; didn’t read”) to generate summaries without explicit training, the results qualitatively resembled standard summaries but quantitatively scored only marginally better than randomly selected sentences, and significantly below advanced summarization techniques.

Similarly, GPT-2 struggled in multilingual tasks without explicit translation training (i.e., “zero-shot”). The model produced poor translations, scoring a 5 out of 100 for English-to-French and 11.5 out of 100 for French-to-English on the WMT-14 dataset. In open-domain question-answering, GPT-2 answered about 4.1% of questions correctly, indicating limited factual knowledge. Yet, on a commonsense reasoning test known as the Winograd Schema Challenge, GPT-2 achieved 70.7% accuracy, surpassing prior state-of-the-art methods by roughly 7%.

GPT-2 was impressive but not powerful enough to pose a threat. The staged rollout revealed these fears to be overstated. To most outside observers, neither its benchmark results nor demonstrations like “Ovid’s Unicorn” represented a transformative breakthrough. Yet, to researchers involved, particularly Ilya and colleagues, GPT-2 marked a milestone. It underscored how unsupervised learning and scaling quietly paved the way toward larger, more powerful models.

In an October 2019 interview with The New Yorker, Sutskever marveled at GPT-2’s unexpected prowess, saying, “Give it the compute, give it the data, and it will do amazing things… it’s like alchemy!”^[11] The reporter captured him with eyes “wide with wonder,” underscoring how astonishing the model’s capabilities appeared, even to an expert who understood the underlying mechanics. For Sutskever, the metaphor of alchemy highlighted how scaling compute and data had produced outcomes that seemed almost miraculous.

In late November 2019, The Economist published an interview conducted entirely with GPT-2 as an experimental curiosity. Sutskever shared this article on Twitter (now X), noting enthusiastically: “The Economist interviews GPT-2—and the interview makes sense.”^[12],[13] This public remark illustrated his delight at the model’s capacity to produce coherent, topical responses indistinguishable at times from human-generated text.

Sutskever suggests that GPT-2 was doing more than merely stringing together plausible sentences; it was starting to show real understanding. In a May 2020 appearance on the Lex Fridman podcast (May 2020; Episode 94), he articulated this clearly, explaining that larger language models like GPT-2 demonstrated unmistakable “signs of semantic understanding” absent in smaller models.^[14] Though he clarified that this understanding was not yet complete—a characteristic restraint often found in his commentary—he emphasized that GPT-2 undeniably understood language meaning, at least in part.^[15]

Ultimately, GPT-2 mattered on multiple levels, including technically, culturally, and philosophically, though its immediate technical impact never matched its sweeping philosophical conclusions. Declaring GPT-2 “too dangerous to release” shifted the Overton window, bringing once-fringe concerns about AI safety into mainstream discourse. While the risks never materialized, the event set a precedent for preemptive caution, establishing a new normal. This tension between AI’s capabilities, perceived threats, and the responsibilities accompanying scale would erupt dramatically four years later with Sam Altman’s firing, underscoring GPT-2’s lasting significance as both a cautionary tale and a cultural turning point.

^[1] https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf

^[2] https://www.theguardian.com/technology/2019/feb/14/elon-musk-backed-ai-writes-convincing-news-fiction

^[3] https://www.vox.com/future-perfect/2019/2/14/18222270/artificial-intelligence-open-ai-natural-language-processing

^[4] https://www.theverge.com/2019/2/14/18224704/ai-machine-learning-language-models-read-write-openai-gpt2

^[5] https://techcrunch.com/2019/02/17/openai-text-generator-dangerous

^[6] https://www.theverge.com/2019/2/14/18224704/ai-machine-learning-language-models-read-write-openai-gpt2

^[7] https://medium.com/data-science/openais-gpt-2-the-model-the-hype-and-the-controversy-1109f4bfd5e8

^[8] https://medium.com/data-science/openais-gpt-2-the-model-the-hype-and-the-controversy-1109f4bfd5e8

^[9] https://openai.com/index/better-language-models/

^[10] https://openai.com/index/better-language-models/

^[11] https://www.newyorker.com/magazine/2019/10/14/can-a-machine-learn-to-write-for-the-new-yorker

^[12] https://x.com/ilyasut/status/1199036860934193152

^[13] https://medium.economist.com/how-i-sort-of-interviewed-an-artificial-intelligence-2a9c069a1680

^[14] https://www.happyscribe.com/public/lex-fridman-podcast-artificial-intelligence-ai/94-ilya-sutskever-deep-learning

^[15] 1:00:36 https://www.happyscribe.com/public/lex-fridman-podcast-artificial-intelligence-ai/94-ilya-sutskever-deep-learning

The book unpacks a collection of papers Ilya Sutskever described to legendary software developer John Carmack as containing “90% of what matters today.” If you’ve ever wondered why these papers matter or why “Have you read Sutskever’s List?” is becoming shorthand for knowing the fundamentals of modern AI, this book is for you.

What attracted me to the List is that in a field where breakthroughs arrive relentlessly, there’s something uniquely comforting, even seductive, about the notion of a stable canon, quietly handed down by one of the field’s grandmasters. Yet, instead of viewing each paper as a separate contribution, the chapters depict them as interconnected threads within the broader narrative they represent. In doing so, this book targets readers who seek more than just independent summaries; it is for those who, like me, wish to understand how these ideas connect, their significance at the time, their enduring relevance today, and what they collectively reveal about the most transformative period in the history of artificial intelligence. In this way, the book is more than an anthology.

I’d appreciate your support, whether that’s reading, sharing, or just spreading the word. The more people who know the List, the stronger the conversation about AI’s foundations becomes.

Now in MEAP. Get 50% off through September 18th: https://hubs.ly/Q03GPM5X0

That is, AI is smart unless it's pretending to be a slacker, so you will allow it to be smart. This argument conflates AI’s functional limitations with human-like deception and some self-referential nonsense about human permission, resulting in something completely incomprehensible.

That said, I hope AI companies start using this to explain any inconsistent performance.

"No, no, no… our AI isn’t underperforming. It’s strategically sandbagging."

Are you still conscious?

I think you'd say absolutely yes.

Hinton's thought experiment is actually David Chalmers' argument.[1] As Hinton ages, he seems less willing to credit others. Nevertheless, Chalmers' argument defends functionalism.

The idea is that consciousness ensues on functional organization rather than any specific physical substrate. Chalmers' goal was to show that functionalism preserves subjective experiences (i.e., qualia), including those implemented in non-biological substrates like silicon (i.e., computers).

Functionalism focuses on the causal roles of individuated functions, not the brain or what it is made of. Functionalism views intelligence (i.e., mental phenomenon) as the brain's functional organization where individuated functions like language and vision are understood by their causal roles. Functionalism is not interested in how something works or if it is made of the same material. It doesn't care if the thing that thinks is a brain or if that brain has a body. If it functions like intelligence, it is intelligent like anything that tells time is a clock. It doesn't matter what the clock is made of as long as it keeps time. This is why I push back when someone says that artificial neural networks are inspired by the brain. They are not, and no one gives an F anyway. Functionalism sees intelligence emerging from the collection of individuated functions within some administrative structure. Functionalism is uninterested in intelligence. It seeks an administrative theory.

The hope for the functionalist view of intelligence is that researchers can replicate enough parts of the intelligence so that the machine can figure out how to do the rest. Unfortunately, functions are not intelligent. They are aspects of thinking. The issue with functionalism, aside from the reductionism that results from treating thinking as a collection of functions, is that it ignores intelligence. While the brain has localized functions with input-output pairs that can be represented as a physical system inside a computer, intelligence is not a loose collection of localized functions. Intelligence is not a neuron or collection of neurons. Intelligence is not part of the brain or even in the skull. Even if we had a detailed map enumerating the countless functions or a neuronal map of all neurons and their connections, it would have been hard to believe we could understand intelligence, let alone the hard problem of consciousness.

If functionalism were as unassailable as Hinton suggests, why don't computers perform or fail like humans? Said differently, if functionalism explains why there are no differences between brains and computers, why are they so different?

If functionalism is not a theory of intelligence, it is unlikely to be a framework that supports consciousness. The flaw in Chalmers' argument—and, by extension, Hinton's—is it ignores the possibility that consciousness arises from the system-wide properties of biological cognition rather than from isolated computational functions.

Swapping a neuron for a functionally equivalent microchip assumes that neurons are just computational units rather than active participants in a larger, complex biological process that may not be reducible to computation. If functionalism were correct, AI should develop intelligence through functional replication alone, yet it imitates intelligence. It doesn't generate it.

Ultimately, AI does not train, perform, or fail like humans because AI is not intelligent. It is a collection of functional simulations, often exceptional simulations. Yet, functionalism is not a cognitive theory. AI systems do not "think" as humans do, and the differences between artificial and biological cognition are not trivial but fundamental. If AI were conscious simply because it replicated functional roles, simulated fire would burn, simulated digestion would nourish, and a mirror would become the person it reflects.

[1] Absent Qualia, Fading Qualia, Dancing Qualia (1996)

The researchers show that the most significant discrepancy between model responses and the ground-truth solution was a general lack of mathematical rigor. Whereas the ground-truth solution makes claims to advance its solution and then proves those claims step-by-step, o1-preview often makes and uses claims without justification.

𝗧𝗵𝗲 𝗻𝗼𝘃𝗲𝗹𝘁𝘆 𝗼𝗳 𝘁𝗵𝗶𝘀 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵 𝗹𝗶𝗲𝘀 𝗶𝗻 𝘁𝗵𝗲 𝘀𝗲𝗻𝘀𝗶𝘁𝗶𝘃𝗶𝘁𝘆 𝗮𝗻𝗮𝗹𝘆𝘀𝗶𝘀 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗲𝗱 𝘁𝗼 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝘁𝗵𝗲 𝗿𝗼𝗯𝘂𝘀𝘁𝗻𝗲𝘀𝘀 𝗼𝗳 𝘃𝗮𝗿𝗶𝗮𝘁𝗶𝗼𝗻𝘀 𝗶𝗻 𝗯𝗲𝗻𝗰𝗵𝗺𝗮𝗿𝗸 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀. Data contamination is a serious issue. The researchers address data contamination by introducing functional variations to the Putnam dataset. By altering elements such as variables and constants in the benchmark dataset, they are operationalizing a straightforward method of avoiding data contamination without creating a new benchmark, which can be costly. Naturally, this is much easier to achieve with math benchmarks, but I want to see more of this methodology. This type of analysis is particularly relevant to the "reasoning vs. memorization" debate. By introducing functional variations and analyzing the performance drops, the research contributes to the ongoing discussion about whether LLMs genuinely reason or primarily rely on pattern recognition and memorization.

𝗧𝗵𝗲 𝗳𝗶𝗻𝗱𝗶𝗻𝗴𝘀 𝗲𝗺𝗽𝗵𝗮𝘀𝗶𝘇𝗲 𝘁𝗵𝗮𝘁 𝗲𝘃𝗲𝗻 𝘀𝘁𝗮𝘁𝗲-𝗼𝗳-𝘁𝗵𝗲-𝗮𝗿𝘁, 𝗯𝘂𝘁 𝗽𝗿𝗶𝗺𝗮𝗿𝗶𝗹𝘆 𝗼𝗽𝗲𝗻-𝘀𝗼𝘂𝗿𝗰𝗲, 𝗺𝗼𝗱𝗲𝗹𝘀 𝘀𝘁𝗿𝘂𝗴𝗴𝗹𝗲 𝘁𝗼 𝗴𝗲𝗻𝗲𝗿𝗮𝗹𝗶𝘇𝗲 𝗯𝗲𝘆𝗼𝗻𝗱 𝘁𝗵𝗲𝗶𝗿 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗱𝗮𝘁𝗮, 𝗽𝗮𝗿𝘁𝗶𝗰𝘂𝗹𝗮𝗿𝗹𝘆 𝗶𝗻 𝗱𝗼𝗺𝗮𝗶𝗻𝘀 𝗿𝗲𝗾𝘂𝗶𝗿𝗶𝗻𝗴 𝗿𝗶𝗴𝗼𝗿𝗼𝘂𝘀 𝗹𝗼𝗴𝗶𝗰 𝗹𝗶𝗸𝗲 𝗺𝗮𝘁𝗵𝗲𝗺𝗮𝘁𝗶𝗰𝘀. However, the analysis does not show that all performance can be attributed to memorization. 𝗦𝗼, 𝗿𝗲𝘀𝘂𝗹𝘁𝘀 𝗹𝗶𝗸𝗲 𝘁𝗵𝗲𝘀𝗲 𝗮𝗿𝗲 𝗮 𝗥𝗼𝗿𝘀𝗰𝗵𝗮𝗰𝗵 𝘁𝗲𝘀𝘁. If you're a nativist, you will say, "The inability to generalize reflects fundamental limitations in the models' inherent architecture and lack of innate reasoning capabilities." If you are an empiricist, you may say, "The performance drop highlights insufficient exposure to diverse training data or limitations in the training process, suggesting that further fine-tuning or more robust data augmentation could address the issue." The research serves as both a diagnostic and a mirror

.

AI's ability to generate low-cost but plausible outputs fundamentally disrupts domains historically relying on high reporting costs as a natural filter. In areas like security alerts and bug bounty programs, the effort required to produce credible reports once served to deter low-quality or spurious contributions, ensuring a manageable signal-to-noise r…

Read more

This week, Casey Newton published an article where he challenges AI skeptics for minimizing AI.[1] He argues that skepticism concentrates too heavily on present limitations instead of the technology's broader potential. He specifically warns against reducing AI criticism to unfounded negativity, calling out figures like Gary Marcus and others.

He writes,…

Read more

Luckily, Miles leaves us with a blog post explaining his exit and reasoning, a master class in the Motte-Bailey fallacy, blending r…

Read more

If you've already read (or are just curious about) "Generative Artificial Intelligence Revealed: Understanding AI from Technical Depth to Business Insights to Responsible Adoption," I'd love your thoughts! 📖✨ Whether you're interested in AI fundamentals, practical tips, or just the thrill of being a book tastemaker, your re…

Read more

I’m excited to offer a special giveaway for new paid subscribers: a complimentary copy of my book, Generative Artificial Intelligence Revealed.

This book is not just another AI guide. It's a comprehensive resource designed specifically for business leaders, tech enthusiasts, and policymakers. It's a roadmap to responsible AI adoption, filled …

Read more

When I write about AI, I will get someone in the comments section telling me why AI is not "real" AI or "not intelligent" or asking me how I define AI or AGI. I am guilty of asking, saying, and writing all these things too. This reaction is mainly due to the abundant hype surrounding AI, especially the association between marketing and slapping "AI" on …

Read more

Read more

Read more

You can purchase the book on Amazon or download it for free directly from my book website. Whether you’re looking to gain te…

Read more

On September 18, 2024, my co-author Clayton Pummill and I will officially release Generative Artificial Intelligence Revealed: Understanding AI from Technical Depth to Business Insights to Responsible Adoption. Writing a book about an evolving subject comes with its challenges—many people advised me against it. The conventional wisdom goes that producin…

Read more

Read more

Peter Dickson and Joseph Giglierano argue that executives and entrepreneurs face different risks. One is that their organization will make a bold move that fails—a risk they call "sinking the boat." The …

Read more