Sep 13, 2022 · 1h 2m · big-technology
Is AI Dangerously Overhyped? — With Gary Marcus
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Cognitive scientist Gary Marcus joins host Alex Kantrowitz on the Big Technology Podcast to critically evaluate generative AI, deconstruct claims of machine sentience like Google's LaMDA, and warn against the technical brittleness, safety hazards, and economic hype surrounding current language models.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 18.8% of the talking time here. How this is scored →
Alex as informed peer 4.7 Guest teaching 6.9 Guest disagreement 4.6 Alex pushing back 4.1
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:55 Rejection of AI ethical reasoning
Gary firmly rejects Alex's hypothesis that an LLM could have reached an alternate ethical stance on suicide, declaring that the system merely spits out words without awareness or comprehension of real-world consequences.
Hardest push from Alex ▶ 57:46 Challenging the AI winter premiseAlex directly challenges Gary's skeptical stance on AI hype, arguing that current systems are already delivering impressive tangible results that make a research funding pullback unlikely.
Biggest teaching moment ▶ 19:07 Evolutionary psychology of the ELIZA effectGary educates Alex on why humans are so susceptible to conversational AI, tracing it back to 1965's ELIZA and explaining how human brains evolved to equate language capability with personhood.
Alex holds their own ▶ 55:36 Firsthand reporting on Microsoft TayAlex demonstrates journalistic authority by recounting how he originally broke the story of Microsoft's Tay chatbot, witnessing its overnight descent into generating offensive hate speech firsthand.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| AI Art Contests and the Ethics of Prompt Generation | 4 | 5 | 2 | 3 | Alex opens with the Colorado State Fair AI art controversy and questions whether prompt-generated imagery constitutes original work. Gary draws parallels to music sampling and Photoshop filters, clarifying the distinction between image synthesis and human creative curation. | |
| Deconstructing LaMDA and Why LLMs Lack Sentience | 3 | 8 | 5 | 3 | Gary systematically dismantles Blake Lemoine's sentience claims regarding LaMDA, comparing the chatbot's sensing capabilities unfavorably to an Apple Watch. He explains that statistical word-prediction models lack internal world representations or grounded semantic meaning. | |
| The ELIZA Effect and Psychological Machine Anthropomorphism | 4 | 8 | 6 | 4 | When Alex brings up Lemoine's mirror test and religion prompts, Gary contextualizes them using the 1965 ELIZA effect. He explains how human evolution leaves people cognitively ill-equipped to distinguish between talking biological humans and superficial machine mimics. | |
| Human Cognition vs. Statistical AI Models and Consciousness | 5 | 7 | 5 | 6 | Alex challenges Gary with philosophical counterpoints, asking whether humans are merely statistical pattern matchers trained on terabytes of historical data. Gary responds by differentiating human causal reasoning and self-reflection from the ungrounded mimicry of LLMs. | |
| Mid-Episode Break and Guest Plug | 5 | 6 | 4 | 3 | Alex and Gary discuss the lack of pre-market regulatory frameworks and the rapid emergence of unconstrained copycat models. Gary warns about misinformation-as-a-service and predicts imminent real-world casualties caused by unvetted conversational agents. | |
| The Fundamental Difficulty of Debugging Stochastic AI Systems | 6 | 7 | 4 | 4 | Alex notes the paradox of non-sentient AI posing lethal risks and shares his reporting experience covering Microsoft Tay. Gary illustrates the impossibility of systematically debugging stochastic systems using an anecdote of a summoned Tesla colliding with a private jet. | |
| AI Hype Cycles, Economic Viability, and AI Winter Risks | 6 | 7 | 6 | 6 | Alex questions whether an AI winter is even possible given recent breakthroughs in generative media and dialogue. Gary pushes back by citing historical abandoned projects like Facebook M and Google Duplex, emphasizing the uncertain economic unit viability of image generation. |
Statements from this episode (18)
Marcus: AI Language Models Produce Convincing Text That Is Often Untrue
“If you talk about another domain, like language synthesis, it turns out that current systems can make very convincing language, but it's often bullshit. That doesn't matter in the same way in art. So in, if you have a system make up a news story, even if it's …”
Marcus: Generative AI Art Is Fully Derivative Sampling on Steroids
“And what these systems are doing is kind of like amazing enhanced version of sampling where you don't even recognize what the samples are anymore. It's all derivative.”
Marcus: AI Art Generators Lack True Understanding or Intentionality
“And it's worth realizing that the art system doesn't really understand what it's doing. It's just correlating words with images in its database. It doesn't have the same intentionality about the objects as a person does, but it can still be a very effective te…”
Marcus: AI Will Change Art but Humans Will Remain Creative Filters
“I think it probably will change art. I think in general that humans are still going to be like the creative inspiration and are often going to be filtering things.”
Marcus: Google's LaMDA Has No Sensors Perceiving the Physical World
“Lambda actually has fewer sensors than my watch. My watch has a lot of sensors and lamb doesn't really have anything sensing the real world, except for its linguistic input.”
Marcus: LaMDA Possesses No Internal Model of the World
“When I'm going through this in some depth, because when we get to Lambda, Lambda doesn't have a model of the world.”
Marcus: LaMDA Is Merely Autocomplete on Steroids, Not Sentient
“My turn-by-turn system has more elements of what I would actually ascribe sentience than Lambda, which is really just autocomplete on steroids. That's all it is, right? You type in your phone, I will meet you at, and it guesses that, you know, you might say th…”
Marcus: Evolution primes humans to mistake conversational machines for people
“There's no machinery in our brains innately to tell us the difference between a person and machine. And what happens is that in evolutionary perspective, anything that could talk was probably a person, right?”
Marcus: Building sentient machines is possible, but we do not know how
“I am in no way arguing that it is not possible to build a sentient machine. I don't think we know how to do it, and I don't think we're clear enough on what it would consist of, but I'm not making the argument that it's impossible.”
Marcus: Future AGI systems will require internal world models
“So I'm pretty confident that we'll have models of the world, internal ideas about how the world works. I don't see how to build an AI without it.”
Marcus: Machines will convincingly conduct podcast interviews within 30 years
“And some point I'm going to say, 30 years from now. We'll be able to make machines that do podcast interviews and I won't really know. I don't really know if you're a person fake or, you know, machine faking me out or whatever at some point.”
Marcus: LLM-powered misinformation as a service will soon be widespread
“If what you want to do is to put out like 10,000 versions of something on Twitter, something untrue and find one that sticks, then misinformation as a service, which is how they might call it in the tech industry is a pretty damn powerful technique. And if it …”
Marcus: A death will be tied to an LLM within a year
“The prediction that I made Is basically that there will be a death tied to a large language model in the next year.”
Marcus: No established methodology exists for debugging chatbots or autonomous vehicles
“None of these chatbot systems are well-debugged. Nobody knows how to debug them, in fact. And so both the problem with GPT-III and with driverless cars is we don't actually have a methodology even for debugging it.”
Marcus: Nobody knows how to prevent chatbots from being toxic or misleading
“Nobody knows, you know, how to make their chatbots constrained and not toxic and not spew misinformation.”
Marcus: The AI field is facing the same unsolved problems as 2015
“Despite all the hype about, you know, we're so close to solving AI or whatever. We're not, we're facing the same problems.”