The AI Experts Who Aren't Buying the Hype
Publicado no IA Dots ·
AGIAI SignalAssistente pessoaldot-agidot-artificial_intelligencedot-llmdot-reasoning_modelsidioma-enInglêsyt-5Vi88DskH8Y
Vídeo em Inglês, publicado pelo canal CoderGrammer.
About this content: The summary, key points and topics below are generated automatically from the video transcript. They reflect the content, opinions, claims and predictions presented in the video and do not represent a position, validation or fact-check by IA Dots.
Video summary
The video examines various expert perspectives on the short-term limitations, architectural constraints, and economic realities of artificial intelligence, contrasting common hype with concerns regarding reliability, reasoning, and adoption.
Key points
- Experts argue that large language models rely on statistical pattern matching rather than true reasoning or world models.
- The deployment of AI faces a gap between impressive demonstrations and the reliability required for complex real-world tasks.
- Adoption rates of AI tools in enterprises often move much slower than the pace of technological improvement.
- Critics highlight economic concerns, pointing to massive capital expenditures in data centers compared to actual software utility and revenue.
- Several voices emphasize that human-level intelligence requires physical grounding, sensory capabilities, and common sense beyond language fluency.
Topics
AI LimitationsLarge Language ModelsWorld ModelsEconomic BubbleReliability GapAI AdoptionHuman ReasoningTechnology Hype
Video transcript
Transcript from available captions; it may contain errors.
AI hype has been around for a while now, with it being shoehorned into every tool and system. No longer just software tooling, but all systems. It's said that it will end most or close to all human work done electronically, but is this really the case? Well, not according to many of the experts. And given that we hear a lot from the hype merchants, I thought it might be useful to put together a list of some of the voices who speak out about some of the shortcomings limitations and constraints of AI in the near term. Now, just to be clear, I'm not and many of these folk are not claiming that AI is not real or transformative, just that there are clear limitations in the short term that make any discussion about AGI anytime soon seem somewhat ridiculous. You and I may not agree with everything they say, but it's definitely interesting to hear their arguments. So, let's start with one of the biggest questions in AI today. Is your favorite chatbot actually reasoning, or is it simply producing very convincing answers? One of the strongest voices arguing against the latter is Yann LeCun. His core argument is that LLMs are commercially valuable, but they're not the route to human-level intelligence because they lack true world models and reasoning. Yann LeCun left Meta in 2025 and co-founded AMI Labs, where he is pursuing world models as an alternative route towards more capable machine intelligence. He argues that they're solving real problems, and he believes that we should continue investing in them, but he also thinks that they've reached a fundamental limit. In his view, LLMs are not the technology that will lead us to human-level intelligence. Given the implications of human-level AI, many people, including me, may find that reassuring. But LeCun's answer is not to stop, it's to search for the next architectural leap. Now, AMI Labs describes that bet as action-conditioned world models that predict the consequences of actions and plan sequences under safety constraints. So, the real message is not to switch it off, it's more like that we should not confuse a powerful autocomplete with Aristotle. This article summarizes his distinction between practical value and human-like intelligence in this way. Large language models justify the billions being poured into GPU clusters and data centers, but the bubble isn't in the infrastructure spending. It's in the belief that these models can think like humans. LLMs are good at a growing list of practical tasks from coding assistance to enterprise search, document summarization, customer service automation, and others. These applications generate real revenue and solve real problems. That makes the massive infrastructure buildout, the GPU farms, and the power plants a defensible investment. In other words, he wants us to distinguish the real capabilities of these models from the claims made about what they can ultimately become. He's been arguing for years that next token prediction, the core mechanism behind every major LLM, is a dead end for achieving anything resembling genuine intelligence. Now, Advanced Machine Intelligence or AMI Labs is a Paris-headquartered AI research laboratory founded by LeCun. The company made headlines by raising over a billion dollars in a seed round. Rather than scaling text-based large language models, AMI focuses on developing AI world models using LeCun's Joint Embedded Predictive Architecture or JEPA, and is focused on building AI systems that learn from raw sensory inputs rather than just text. We can see that research discussed here. LeCun's lab at Meta had also been working on this front directly. They introduced a model called V-JEPA 2 built on JEPA-based architecture. Rather than predicting the next word in a sequence, JEPA-based systems learn by predicting abstract representations of future states. The goal is improved planning efficiency and a more grounded understanding of how things actually work in the physical world. That matters here because planning and grounding predictions in the physical world will come up again later. Now, LeCun isn't alone. Another AI researcher who's been making a very similar argument for years is Subbarao Kambhampati, although he approaches the problem from a slightly different angle. Kambhampati's core argument is surprisingly simple. Today's AI is exceptionally good at producing answers that look like they've been carefully reasoned through without necessarily having a reliable reasoning process underneath. He once jokingly described LLMs as n-grams on steroids, a deliberately provocative way of saying that they're incredibly sophisticated pattern matches, the type of thing that you would get with any auto complete on a shopping website search, rather than human-like thinkers, which is an academic way of saying that auto complete has been hitting the gym very, very hard. He argues that the models are strongest as idea generators. Reliable planning, on the other hand, requires an external search or verification process. Now, to be fair, the AI industry hasn't stood still since he first made many of these arguments. We've seen the arrival of so-called reasoning models from companies like OpenAI, Google, DeepMind, and Anthropic. These systems spend more computation on difficult questions, generating intermediate steps, exploring alternatives, and in some implementations, checking their work with tools or verifiers before producing the final answer. Now, this is much closer to the search and verification behavior that Kambampati said ordinarily LLMs lack. He essentially says that generating a correct answer is not the same thing as having a reliable reasoning procedure. Interestingly, modern AI systems are moving towards something remarkably similar to Kambampati's preferred architecture. You go from LLM to generating possibilities, you have search tool and verifiers, they check the results, and you iterate round and round. His term for this is LLM modulo. The irony is that reasoning models weaken the strongest version of the claim that language models cannot reason at all, while actually strengthening Kambampati's broader point that reliable reasoning systems need more than simple free-running language generation. In this later article, he talks about such reasoning models in more detail. LLMs are impressive statistical pattern matches, but should not be mistaken for systems that genuinely plan, reason, or understand the world. The paper does not say that models have no reasoning ability. It argues that benchmark gains should not automatically be treated as evidence of general human-like reasoning or planning. One of the central ideas is that seeing lots of intermediate thinking tokens shouldn't automatically convince us that the model is reasoning like a human. The prevalent narrative is very questionable. Just because an AI writes about 20 lines of thinking doesn't necessarily mean that it's thinking in the way a person does. Those reasoning traces may simply be one mechanism that improves performance rather than evidence of human-like cognition. The authors argue that systems become stronger when they generate candidate answers and then verify them rather than because they produce long chains of reasoning tokens. That distinction matters because better results may come from allocating more computation, searching more candidates, and verifying outputs, not necessarily from a sudden leap to human-like thought. In their account, the more defensible breakthrough is test time computation, spending additional compute during inference to search, compare, and verify candidate answers. Their argument is essentially spend more computation on difficult questions, explore alternatives, reject bad answers, and keep the best one rather than magically becoming more intelligent. In other words, modern reasoning models may be winning because they search harder before answering, not because they've suddenly discovered human reasoning. So, today's reasoning models do not simply overturn Kambhampati's argument, they partly validate his system-level prescription, which is to let the models generate possibilities, then use search tools and verifiers to test them. The gain is practical reliability. The unresolved question is whether the visible trace reflects human-like reasoning or merely a useful computational procedure. Now, this takes us nicely onto Michael Wooldridge, another one of the experts that I wanted to talk about. Michael Wooldridge's main idea is that general intelligence needs more than merely language. Wooldridge's argument is not that ChatGPT is unimpressive, it is that fluent language is only one part of intelligence and we should not confuse it with the full range of human capabilities. Now, talking to an LLM can feel like AI from the movies, available on demand, conversational, and apparently knowledgeable about almost anything. It can answer unexpected questions and sometimes even feel like a friend, a tutor, and a therapist all rolled into one. To explain how we got here, Woolridge contrast two traditions in AI. Symbolic AI represented knowledge and reasoning explicitly. It succeeded at some structured tasks, but struggled badly with perception. Neural networks had existed for decades, but larger data sets, more compute, and architectural ad advances finally made them work at scale. This quote from the article is particularly interesting. At this point, there is something I simply must get off my chest. Thanks to ChatGPT, we have finally reached the age of AI. Every day, hundreds of millions of people interact with the most sophisticated AI on the planet. This took 70 years of scientific labor, countless careers, billions upon billions of dollars of investment, hundreds of thousands of scientific papers, and AI supercomputers running at top speed for months. And the AI that the world finally gets is prompt completion. Or another way of putting it is that 70 years of research and the grand reveal was a text box. But of course, things have progressed rapidly since ChatGPT launched in November 2022. He argues that human-level AI would need more than language. It would need physical world understanding, sensory capabilities, and basic common sense, bringing us back to LeCun's work on world models. Woolridge then splits human intelligence into mental capabilities such as reasoning, planning, and common sense, and physical capabilities including perception, navigation, and manipulation. That again echoes LeCun's focus on systems grounded in the physical world. Current LLMs do not cover the full spectrum of mental and physical capability. His precise claim is that LLMs are not suitable for physical capabilities and are unlikely on their own to provide a route to true AI. He still allows that there may be one ingredient in the larger system. Now, as we continue, do keep in mind that this article is from 2023, and a lot has happened since then. Now, that moves us from physical world intelligence to a related problem where the fluent language is being mistaken for understanding. And for that, we'll look at Emily Bender's work. Now, this 2021 paper asks whether bigger language models bring risks that raw benchmark scores hide. One of the best lines is that coherence is in fact in the eye of the beholder. In other words, sometimes the chatbot sound clever because the human is doing unpaid interpretive labor. Now, polished language models can make us supply the understanding ourselves. Fluency is evidence that a model predicts language well. It is not on its own evidence of intent or human-like grasp of meaning or intelligence. That's the point behind the paper's famous label, without any reference to meaning, a stochastic parrot. The paper also warns that scraping more of the internet does not automatically create a more representative data set. It can amplify dominant viewpoints while making the data too large to properly document. And this widens the argument beyond whether a chatbot gives a good answer. The real question is who pays the environmental and social costs? Who benefits and who is harmed when mistakes are deployed at scale? Now, it's important to note that this paper is from 2021, before ChatGPT. The technology has improved dramatically, but Bender's core challenge still bite, which is that are we measuring useful performance or mistaking fluent output for deeper understanding? Bender's conclusion is that models can generate convincing text by learning statistical patterns in language, but fluency alone does not establish communicative intent or human-like understanding. This matters beyond the lab because businesses can buy the appearance of intelligence before they've actually proved to be reliable. Now, to understand whether businesses are overestimating AI, I want to take a look at the work of Arvind Narayanan and Sayash Kapoor. Their 2025 essay calls AI normal technology, but potentially transformative, like electricity or the internet, but still a tool shaped by people, companies, laws, and institutions, not a separate species deciding our future. Their key distinction is between AI methods, useful applications, and adoption. A model can improve quickly while the last two move much more slowly. One sharp line is that adoption is about software use, not availability. Putting a chatbot in every product does not mean that people have changed their workflows enough to produce economic value. The historical analogy is electricity, and that factories did not get big productivity gains until layouts, processes, training, and management changed. In their words, diffusion occurs over decades and not years. Now, electricity transformed factories, too, but nobody installed one lightbulb and declared the productivity revolution complete. The most interesting concept may be the capability-reliability gap. AI can look brilliant on a benchmark or demo and still be too inconsistent for a costly real-world task. Their example is Epic's sepsis model. It looked accurate internally, yet reportedly missed 2/3 of cases in hospitals and flooded doctors with false alerts, partly because testing relied on information unavailable at deployment. They make the same point about GPT-4's bar exam results. Scoring highly on a test tells us remarkably little about practicing law, where judgment, strategy, context, and accountability matter. My takeaway is that businesses often model capability for a finished product. Narayanan and Kapoor's framework says deployment value appears only when the capability is converted into a reliable application and then adopted inside a redesigned workflow, often as augmentation rather than substitution. This is a subtler anti-hype position, which is that AI can be transformative without transforming every company next quarter. The bottleneck is often not the model, it's the reliability, integration, and wider organizational change. That separates three time periods or clocks that hype often collapses into one. The pace of model improvement, the pace of dependable product development, and the much slower pace of organizational adoption. Their warning is not that AI itself is the only danger. It is the exaggerated claims can persuade businesses and governments to grant systems more trust than their demonstrated reliability warrants. Now, from there the argument widens from failed deployments to the economics of the AI boom itself. And for this I want to talk about Ed Zitron, Ed Zitron a very outspoken commentator on this topic. His 2025 article is an 18 and a half thousand word polemic, not a neutral survey. His case joins two claims, the technology is too unreliable for the promised automation and the industry's economics do not justify it the scale of investment. Calling Zitron skeptical would be like calling the sun slightly warm. He put it quite bluntly saying that you cannot rely on a large language model to do what you want. He broadens the idea of hallucination to beyond invented facts. For an AI agent it can mean misunderstanding the task, taking the wrong action, or wandering through expensive unnecessary steps. The more complex the workflow, the more chances there are for failure. His labor argument is that most jobs are not output driven at all. A job is not merely the visible document, code, or image produced at the end. Executives may see a coder producing code or a writer producing words, but miss the judgement, the context, the maintenance, the accountability, and the unstated needs behind the output. His business model critique is equally important. Unlike traditional software, every generation and correction consumes compute. Errors do not merely annoy the user, they create extra marginal cost making flat subscriptions difficult to price profitably. His sharpest enterprise example is Microsoft 365 Copilot. Citing a source who saw sales materials, he reports about 8 million active license users from 440 million Microsoft 365 subscribers, roughly 1.8%. This is his reporting, not an independent audited figure. Zitron argues that spending and partnerships can be mistaken for demand. Announcements of big giant data centers deals do not prove that customers will pay enough to support them. My takeaway is that his strongest case is not that generative AI has zero utility. It's that useful features do not automatically add up to reliable autonomous workers or a profitable industry large enough to justify today's huge capital expenditure. And if you've seen any of Zitron's recent media appearances, you'll know that he talks a lot about the huge overspend on data centers and the difficult finances behind all of that. And Zitron's hard conclusion is that the industry has created a speculative bubble driven more by investor expectations than by any sustainable products or business models. Now, these perspectives differ in tone. Narayanan and Compour focus on slow diffusion and institutional change. Zitron makes the harder economic bubble case, but both warn against treating a compelling demo as a dependable business system. Now, this leads us to a quite practical question, which is that when is AI really useful and can we rely on it? And for that, we'll look at the work of Gary Marcus. Now, Gary Marcus's June 2025 essay argues that reasoning failures are symptoms of a deeper architectural problem. LLMs do not reliably build or maintain a model of the world they are describing. A world model is an internal framework used to track what exists, how things relate, and how a situation changes. Marcus says it should be persistent, stable, and updatable. Chess is his clean test case. A conventional chess program explicitly the board, legal moves, and the existence of an opponent. An LLM can discuss chess fluently, yet still lose track of pieces or suggest an impossible move. It can explain every rule perfectly and then still move the queen as if it has diplomatic immunity. His sharpest line is that LLMs cry to their peril to live without classical world models. That distinction helps explain why fluent output can coexist with hallucinations. Producing the next plausible token is not the same as checking an answer against a durable representation of the situation at hand. A model may sound as though it understands a scene because it has learned the language people use to describe it. Sounding coherent and maintaining states are two very different abilities. Marcus also argues that scaling alone won't solve this. More training data and compute can improve pattern recognition without guaranteeing a persistent, interpretable internal state. The practical danger appears in multi-step work. If a system's picture of the task drifts, errors can compound across agents, coding, planning, law, or even medicine, even when individual responses sound convincing. The reliability problem is not merely an occasional invented fact. It may reflect a system that cannot consistently keep track of the world its answer is supposed to be about. Now, one important caveat is that the idea of world model is contested. Some researchers argue that language models learn partial or latent representations. Marcus's narrow practical point is that representations exist are not robust enough for dependable use. Gary's 2025 post interprets Apple's illusion of thinking research as evidence that the latest reasoning models still fail to generalize reliably beyond familiar problem distributions, and you can see that paper here. Reasoning models spend extra computation generating intermediate steps, often called chain of thoughts. The hope is that more inference time compute will overcome the limits of ordinary next token prediction. Apple tested models on rule-based puzzles of increasing complexity, including Tower of Hanoi. Marcus highlights a steep performance collapse. Claude managed under 80% accuracy with seven disks and almost none with eight, and Apple reported similar limits across models and tasks. The especially awkward result is that supplying the solution algorithm did not reliably fix performance. The system was not being asked to invent the method, only to execute it consistently. It shows that the model can explain the puzzle, it just cannot resist becoming one. Marcus mentions that LLMs are no substitute for good well-specified conventional algorithms. Now, the damaging part is not that a puzzle becomes difficult, it is that the success on easy versions can look like general reasoning before a small rise in complexity exposes a brutal approximation. This creates quite a business trap. A model can ace an easy internal test set and encourage deployment, then fail on the unusual, longer, or higher-stakes cases that matter most in production. Marcus acknowledges that an important objection. Many humans also struggle with an eight-disk Tower of Hanoi. His response is that advanced AI should combine human adaptability with machine reliability, rather than inherit the weaknesses of both. Now, one important caveat is that the Apple experiments do not prove that language models never reason. They show that current reasoning traces and additional computation do not guarantee systematic generalization or faithful execution. Marcus still grants LLMs a useful role in coding, brainstorming, and writing. His narrower conclusion is that they are not dependable stand-alone routes to AGI. Hybrid systems using explicit algorithms or symbolic tools may also be necessary. Marcus focuses on technical brittleness. Neil Lawrence, on the other hand, shifts the lens to a wider question. Should the AI debate center on machine capability or on people and values affected by them? This June 2024 Cambridge article introduces Neil Lawrence's book, The Atomic Human. It's a philosophical and political argument about AI, rather than experimental study of model performance. Lawrence asks what remains uniquely human as machines repeatedly take over abilities once treated as defining us. His counterintuitive answer is that the indivisible core lies around our vulnerabilities and limitations, not a fixed list of superior capabilities. He mentions that humanity is really about how we overcome our limitations. If human value is measured only by tasks a machine cannot perform, the value appears to shrink with every technical advance. Lawrence changes the unit of comparison from task performance to lived experience, adaptation, and responsibility. He therefore wants the AI debate redirected away from celebrating mega intelligence and towards two human questions: Who are we and what do we want to become? There is also a democratic argument. Lawrence wants tech bros to be one voice among many because systems that reshape work and public life should not be designed solely around the priorities of their builders. AI choices distribute power, opportunity, and risk. People affected by those choices should participate in deciding where the technology is used and what safeguards it needs. He brings experience from machine learning research, directing machine learning at Amazon, and deploying systems in Africa. The article presents his concern as grounded in how end-to-end systems meet messy real-world conditions. Now, this article does not demonstrate that present AI lacks understanding. It introduces the thesis of Lawrence's book and quotes its author and publishers. His point is less a laboratory verdict on whether machines understand than a warning against treating machine performance as a measure of human worth. Taken together, these sources operate at different levels. LeCun, Kambampati, and Marcus question architectures and reliability. Bender questions meaning, data, and social costs. Narayanan and Kapoor question the speed of diffusion. The Tron questions the economics and finances of all of this. And Lawrence questions whose values shape actual deployment. Again, I want to state clearly that these are not arguments that AI is fake. They're really just arguments for matching each system's responsibility to evidence of its capability, reliability, and limits. Now, Jeff Bezos once described AI as a horizontal enabling layer, a technology that can spread into almost everything. That framing is very useful because it captures both sides of this story. AI can matter enormously without every claim made about it being true. A technology can be real and transformative while still attracting exaggerated claims and problematic business practices. Hype cycles blur the lines between what works, what might work later, and what is being promised to investors. This is why it's quite helpful to hear from dissenting experts, especially with the nonstop scaremongering and hype that we cannot avoid. They force us to ask better questions before we hand systems more power, more money, and more responsibility. Taken together, these experts are not arguing that AI is useless. They're asking us to separate capability from reliability and demos from deployment, availability from adoption, and genuine progress from promises. That gives us four practical tests. What can the system do? How often does it fail? What human or institutional support makes it useful? And who remains accountable when it goes wrong? A sensible position is neither blind belief nor blanket dismissal. We want to use the tools, measure what they can actually do, and keep humans responsible for the decisions around them. The real question isn't whether AI is effective, but whether our claims about it can survive contact with the real world. And if AGI happens to arrive before my next upload, then I might make a very awkward follow-up video. Let me know what you think in the comments below, and if you do not want to be the last to know about important AI, tech, and programming news, then be sure to hit subscribe. Thank you for watching.
Assistir ao vídeo original no YouTube
Assuntos: Assistente pessoal, AGI. Publicado no YouTube em 2026-10-18.
Seleção automática por assunto e idioma. O vídeo pertence ao canal de origem.