A Skeptical Take on the A.I. Revolution

The year 2022 was jam-packed with advances in artificial intelligence, from the release of image generators like DALL-E 2 and text generators like Cicero to a flurry of developments in the self-driving car industry. And then, on November 30, OpenAI released ChatGPT, arguably the smartest, funniest, most humanlike chatbot to date.
In the weeks since, ChatGPT has become an internet sensation. If you’ve spent any time on social media recently, you’ve probably seen screenshots of it describing Karl Marx’s theory of surplus value in the style of a Taylor Swift song or explaining how to remove a sandwich from a VCR in the style of the King James Bible. There are hundreds of examples like that.
But amid all the hype, I wanted to give voice to skepticism: What is ChatGPT actually doing? Is this system really as “intelligent” as it can sometimes appear? And what are the implications of unleashing this kind of technology at scale?
Gary Marcus is an emeritus professor of psychology and neural science at N.Y.U. who has become one of the leading voices of A.I. skepticism. He’s not “anti-A.I.”; in fact, he’s founded multiple A.I. companies himself. But Marcus is deeply worried about the direction current A.I. research is headed, and even calls the release of ChatGPT A.I.’s “Jurassic Park moment.” “Because such systems contain literally no mechanisms for checking the truth of what they say,” Marcus writes (www.scientificamerican.com/ar...) , “they can easily be automated to generate misinformation at unprecedented scale.”
However, Marcus also believes that there’s a better way forward. In the 2019 book “Rebooting A.I.: Building Artificial Intelligence We Can Trust (www.penguinrandomhouse.com/bo...) ” Marcus and his co-author Ernest Davis outline a path to A.I. development built on a very different understanding of what intelligence is and the kinds of systems required to develop that intelligence. And so I asked Marcus on the show to unpack his critique of current A.I. systems and what it would look like to develop better ones.
This episode contains strong language.
Mentioned:
“On Bullshit (www2.csudh.edu/ccauthen/576f12...) ” by Harry Frankfurt
“AI’s Jurassic Park moment (garymarcus.substack.com/p/ais...) ” by Gary Marcus
“Deep Learning Is Hitting a Wall (nautil.us/deep-learning-is-hi...) ” by Gary Marcus
Book Recommendations:
The Language Instinct (www.harpercollins.com/product...) by Steven Pinker
How the World Really Works (www.penguinrandomhouse.com/bo...) by Vaclav Smil
The Martian (www.penguinrandomhouse.com/bo...) by Andy Weir
Thoughts? Email us at [e-mail]. Guest suggestions? Fill out this form (forms.gle/w5Wr1rEMtois5J8h6) .
You can find transcripts (posted midday) and more episodes of “The Ezra Klein Show” at nytimes.com/ezra-klein-podcast (www.nytimes.com/column/ezra-k...) , and you can find Ezra on Twitter @ezraklein. Book recommendations from all our guests are listed at www.nytimes.com/article/ezra-... (www.nytimes.com/article/ezra-...) .
“The Ezra Klein Show” is produced by Emefa Agawu, Annie Galvin, Jeff Geld, Rogé Karma and Kristin Lin. Fact-checking by Mary Marge Locker and Kate Sinclair. Original music by Isaac Jones. Mixing by Jeff Geld and Sonia Herrero. Audience strategy by Shannon Busta. The executive producer of New York Times Opinion audio is Annie-Rose Strasser.

[@onlyme112]

Brute force does have its limits. No matter how many feathers you paste on a brick, it won'tI fly.

[@govindagovindaji4662]

Ezra should have plugged the book "Blueprint" one last time at the end since Marcus didn't plug it for himself, hee hee. There are SO VERY MANY articles, podcasts and conversations around AI right now that we won't ever have time for them all. I was just gonna peep into this one and ended up listening to the entire thing. Marcus is a good mentor on this subject. Sam Harris has a good one with Marcus and Stuart Russel which covers a lot of territory and also some of the differences in how these AI folks look at some of the issues.

[@jamesrob8552]

This was really fascinating to listen too.

[@Magnulus76]

Klein has a good objection that I feel wasn't adequately answered. In truth, I don't think most human beings know what they are talking about, either. Genuine insight is very rare. An internal model is not immune to being the result of pastiche, either. Mostly we are working off just that. It's just we are better at it than GPT-3.

[@govindagovindaji4662]

61:45 I'm no computer scientist but this was the first thing I thought of [getting different components to do this piece and that piece, etc.] earlier in this very conversation when he was talking about how many varied proteins, cells, DNA acid-combos there are, etc. Maybe these AI people should try making models for categories just as the calculator is a model for arithmetic.

[@DennisHunter]

Wow

[@govindagovindaji4662]

Eliezer Yudkowsky probably has it right when he says that until we understand how the brain works it is impossible to get AI to simulate human intelligence and beyond it. And we are still a long way from that very large piece of the puzzle.

[@peristiloperis7789]

I think he's right and jelous at the same time.

[@oaskoglie]

Jepp

[@Liberty-rn4wy]

I am using ChatGPT to learn Spanish. It is quite good and gives me random translations which are quite helpful. But it keeps giving me the word gato (cat). A human would remember that we already did that word four or five times and not keep including it. The app doesn't.

[@squamish4244]

It'll fix that in a few more months.

[@govindagovindaji4662]

Try commanding it to stop doing that as it can only follow commands, not 'unspoken' or 'undirected' commands.

[@stonedapedefunct7131]

​@@squamish4244future here - they fixed that. Amazing...like the majority of the stuff in this episode, just give it a number of months and they'll fix it... trash episode that aged horribly imo

[@neforfor]

GG

[@amparoconsuelo9451]

How can AI be used to deal with iniquities in the justice and government system, corruption in the food industries, greed of big pharma, and the dismal state of health care, or at least circumvent these situations?

[@GuaranteedEtern]

"Is it good" and "should we do it" - these are normative (human) questions, and are now largely irrelevant.

[@borisshmagin8925]

It is an example of thinking about outload with out actual known the object and what is the language

[@Liberty-rn4wy]

What is "the truth"? And who gets to decide that?

[@josephp.3341]

No one - precisely why you should be scared of an LLM being the final arbiter of what truth is

[@rahuliox]

The statement that the output of large language models has no relation to the truth makes absolutely no sense. example, Ascot, how to connect a Denon 3600 receiver to a subwoofer without a RCA cable and you get an answer and that is perfectly accurate.

[@onlyme112]

Interesting example, thanks.

[@oddajkredki]

Of course, the most burning problem of humanity is "misinformation about COVID." I appreciate this take, but until we resolve what misinformation is and, more importantly, who decides what misinformation is - we shouldn't play with this fire (i.e., large language models connected to the Internet). Let's clean up our information ecosystem first and only then let's have a deep discussion about AI, not the other way round. Until one fails to comprehend that this is THE problem with large language models - one shouldn't seriously discuss the impacts of AI/ML. Certainly - one doesn't need Chat GPT to disseminate massive disinformation on a global scale. NYT's retracted exaggeration of COVID-19 hospitalisations by 837,000 cases (sic!) in 2021 is just a drop in the ocean of similar examples.

[@maiku20]

Yeah, I love how this guy uses controversial examples to illustrate the potential problems of computer-generated disinfo. The other example given is Russian trolls swinging American elections; as if the humans actually agreed on the significance of that. I'm at 20 minutes, getting ready to bail from this video.

[@silberlinie]

Gary Marcus served from 2001 to 2019. He is no longer relevant in today's discussion about the status of large language models. He is relevant in the historical development of AI and will always remain so.

[@stonedapedefunct7131]

Funny how Ezra focuses so much on Bullshit, a8nce thats exactly what this episode is! BULLSHIT!

[@wendylafolle]

What does Gary Marcus mean by a machine understanding something? I don't see how a machine can understand at all. As Ezra said, understanding requires experience, which takes place in awareness.