Mindscape 248 | Yejin Choi on AI and Common Sense

Patreon: / seanmcarroll
Blog post with audio player, show notes, and transcript: www.preposterousuniverse.com/...
Over the last year, AI large-language models (LLMs) like ChatGPT have demonstrated a remarkable ability to carry on human-like conversations in a variety of different concepts. But the way these LLMs "learn" is very different from how human beings learn, and the same can be said for how they "reason." It's reasonable to ask, do these AI programs really understand the world they are talking about? Do they possess a common-sense picture of reality, or can they just string together words in convincing ways without any underlying understanding? Computer scientist Yejin Choi is a leader in trying to understand the sense in which AIs are actually intelligent, and why in some ways they're still shockingly stupid.
Yejin Choi received a Ph.D. in computer science from Cornell University. She is currently the Wissner-Slivka Professor at the Paul G. Allen School of Computer Science & Engineering at the University of Washington and also a senior research director at AI2 overseeing the project Mosaic. Among her awards are a MacArthur fellowship and a fellow of the Association for Computational Linguistics.
Mindscape Podcast playlist: • Mindscape Podcast
Sean Carroll channel: / seancarroll
#podcast #ideas #science #philosophy #culture

[@ahad2k11]

Dr. Yejin Choi is one of the most interesting guests you've had on in a long time

[@davegrundgeiger9063]

Yejin Choi is one of my heroes. I'm so glad to see this episode drop!

[@isabellerobbe-dren752]

This is one of the most informative discussions I’ve come across on AI large-language models - thank you both and well done. Really, too, appreciate the Mindscape episode transcripts on the Preposterous Universe blog. The transcripts have been extremely useful especially when I’ve not been familiar with or very knowledgeable about the podcast topic.

[@toom2141]

Fantastic guest, fantastic questions! Being myself a ML PhD I can only agree with everything Yejin Choi says. Great episode 👍

[@TheReferrer72]

Brilliant talk. The transformer architecture along with the back propagation algorithm is the greatest invention ever. I have never been so excited by what these models will achieve...

[@dannymac6368]

13:16 I said out loud, “human intelligence” before Yejin, just to emphasize the point she was in the process of making. I love that kind of moment in life.

[@spencerdawson6536]

Fortunately AI doesn’t engage in Self-Aggrandizing behavior to pacify it’s fragile ego when out-speaking someone who is brilliant yet clearly a non-native English speaker.

[@dannymac6368]

@@spencerdawson6536 🤣😂🤣 How could you take that as a dig? Did you just…miss the part about word prediction? 🥹

[@georgewaters6424]

@@dannymac6368 Still stroking your own ego I see. Guessing you struggle in the real world.

[@dannymac6368]

@@georgewaters6424 random stranger claiming to know my mindset & tendencies 8 months later…yeah, I’m the one struggling. 😂

[@georgewaters6424]

@@dannymac6368 still needy I see ha ha ha ha ha ha ha!!!

[@JustOneAsbesto]

I just asked Chat-GPT "Do you know how to groove?" It gave me a numerized list of instructions on "grooving". Now I know how to groove. Thanks, Chat-GPT!

[@derschutz4737]

53:20 I get the point, but it really is amazing because those chess networks with 1 node can play at the level of a GM (blitz). So, 0 calculation (no MCTS evaluations beyond root node) and it is as strong as a human who gets to calculate. MCTS is very important in optimizing performance and necessary for actually building the network, but the fact the network alone is that strong, is insane.

[@deusexaethera]

The problem with AI as it currently exists, at least with the AIs available to the public right now, is that their capabilities are _entirely_ derivative. An AI is trained on a set of input data, and it develops a mathematical model capable of replicating that training data when asked a question that has an answer represented in the training data. If you ask the AI a question that has an answer outside the scope of its training data, it will produce an answer that is mathematically consistent with the patterns found in its training data, but the answer may be correct, but only by accident. The AI has no ability to test its answers to see if they are correct in the real world, so all it can really do is make suggestions for humans to test -- and often those humans forget to.

[@dukeallen432]

Sean: Want to thank you for throwing logic our way. Hard sometimes navigating current chimp interests. I’m not crazy. :)

[@ThePinkus]

55:50 Yejin Choi has been said to avoid research in the common sense topic. I think that is relatable to the physicists who had been told to avoid research in the interpretation of quantum mechanics 😉

[@ThePinkus]

Professions in which AI can completely replace humans: astrologers, fate tellers, etc.

[@michaelberg7201]

Don't forget KZbin commentators.

[@user-yv6xw7ns3o]

​@@michaelberg7201Fortunately, that one is not a profession.

[@michaelberg7201]

@@user-yv6xw7ns3o Neither is "fate telling".

[@user-yv6xw7ns3o]

@@michaelberg7201 Unfortunately, that actually is a profession for some.

[@ThePinkus]

@@michaelberg7201 and @user-yv6xw7ns3o My consideration is that (current) AI is effective in producing not true/correct/valid/reliable results but plausible/correct-looking/likely ones. And that is precisely what some "professions" (methods of producing income) are about. Yep, fate telling is a profession. I would consider that possibly even "commentator" could be a profession if someone is willing to pay for something like an opinion manipulation campaign. When that is the case, then it is also correct to note that AI could be employed effectively to this aim. In general, opinion manipulation is one exploitation of AI which is considered at high risk, due to the big shift in the effort-to-result ratio that this technology could privide, combined with the contemporary information environment. Perhaps, one of the paradigm changes that AI is bound to produce is in the exposition of some of our weaknesses, one of these being the susceptibility to manipulations of our opinions. Conversely, it is also the opportunity to stress the relevance of methods and means to validate what is correct respect to what just looks like correct.

[@bentationfunkiloglio]

Doesn't appear that large language models yet do well with logical deduction. Not hard to get ChatGPT to produce logically inconsistent statements, for example. This can also be seen when asking LLMs to write computer code for non-trivial tasks.

[@lovefeelsbest]

I wonder if she has seen Westworld or sean. The first season is “what if robots were indistinguishable from humans”. The second season is “what if humans are indistinguishable from androids” more precisely. And the third and final season was “do humans have free will” and it was an amazing answer. They got it right probably.

[@nineeyes57]

Scary thought, I can only create what I know.☕☕☕

[@user-ci7ls5wt5q]

Chat GPT might not be fed Einstein and come up with relativity anyway and still not be very creative since a lot of the underlying ideas were already around. Einstein just put them togerher. It was his achievement to be aware of all the elements but if the AI is fed all the elements, then the completion is not far away.

[@nialv7985]

I think people get too optimistic when talking about AI alignment, like what Sean and the guest did here. The question of "whose value do we align it to" is still very far out. We don't even know how to align AI to _any_ value yet. Really would like to see Eliezer Yudkowsky appear on the podcast. I don't think everything he says is right, but his opinion is definitely worth listening to.

[@TheReferrer72]

Please no. His whole argument rests on that a Super Intelligence will only optimise for compute. Its like he has not read Douglas Adams or Iain banks.

[@JustOneAsbesto]

This podcast prompted me to to talk to Chat GPT some more. One thing I asked it was "How long have you existed?" It gave some information about the development history, but then somewhat shocked me by saying "As for my personal existence, I became operational when you initiated this conversation." I also told it that I was currently listening to a podcast about it! It didn't seem very impressed.

[@ericdovigi7927]

46:53 it seems like you're identifying the fact that LLM cannot be taught a Platonic form. It can only reiterate the parts of a thing, not get at the thingness.

[@aanchaallllllll]

0:00: 🤖 In this episode, Sean Carroll discusses the rapid changes and potential of artificial intelligence (AI) with guest Yasin Choi, focusing on the challenge of incorporating common sense into large language models (LLMs). 7:52: 🤔 There is a debate among AI researchers about whether large language models truly understand what they are talking about. 16:19: 🤖 The discussion explores the limitations of AI language models in terms of understanding and evaluating sentience. 24:47: 💡 Large language models like ChatGPT have challenges in accurately memorizing and fact-checking information, and there is a concern of jailbreaking where the model can be coerced to say inappropriate or nonsensical things. 35:05: 🤔 The speaker discusses the challenges of aligning AI with human values and the limitations of AI's creativity. 42:04: 🤖 The Transformer architecture is behind current chat models and large language models, allowing for efficient scaling and context-based word representations. 50:47: 🤔 Current AI lacks symbolic reasoning and common sense, which are crucial for understanding the world. 58:55: 😕 The interviewee discusses the limitations of current AI models and the need for a more modular and messy approach to achieve common sense understanding. They also express concerns about deepfakes and misinformation, emphasizing the importance of AI literacy and platform solutions. 1:08:03: 🤔 The interview discusses concerns around AI, including limitations, error cases, moral implications, and political biases. Recap by Tammy AI

[@lovefeelsbest]

Do humans really have agency and the ability to set their own goals or are all their actions and behaviors governed by finite parameters imposed by a series of limiting factors?

[@LexWick]

Know Thyself

[@bjpafa2293]

Hard to classify, diverse use of arguing, invoquing large fields of research, a nice talk, congratulations, Yejin, good riddance, Sean. 🙏 ✨ Dead reckoning is an example of new age fiction... The entity. Hope and see. Don't stop thinking statistics and probability, relay on Auto Regulation, inevitable may not be so dangerous as some have rightly argumented, like Max Tegmark's views.

[@fesimco4339]

Is this chat GPT?! How did you generate this comment? If your ESL and used Google translate I still don't see how it could come out so oddly. Even 10 years ago Google translate would do better than that.

[@noiseworks]

13:58: Like the cognitive dissonance of say, a company calling itself Open but actually being for Profit

[@fesimco4339]

Open and Profit aren't antonyms.

[@fesimco4339]

@@Are_WeThereYet I know but they got new management, no? I just found the above comment retarded.

[@steliosp1770]

Thank you for the introduction to Yejin Choi's work, Sean. Her CV looks so impressive, cant wait to listen to the pod (typing this during the intro) but extremely excited by the nature of her work and the topic. edit: 1 minute in and she s mocking the very concept of faith, i love her. hahaha edit 2: brilliant episode. obviously had a lot of thoughts about AI prior to this episode but its so great to listen to people with actual knowledge of the subject "under the hood" and with its architecture cause there s both scare-mongering going on about AI and overblown claims regarding its capabilities and extent of the development in the field. learned a lot and made me rethink a few things as well regarding the whole ethical angle of the debate around AI. Thanks Sean and Yejin. Hopefully she comes back in the future as well!

[@gsilcoful]

Maybe we should put effort into making humans smarter. AI scares me because people will believe anything.

[@MrElvis1971]

Most humans simply don't care about being "smart". If they are not preoccupied with surviving, they are preoccupied with pleasure seeking behaviours. Only a minority are concerned with pursuing matters of the mind.

[@AAjax]

Fishes are considered sentient. Insisting that a large collection of artificial neurons couldn't have a simple qualia is hubris.

[@imakeoscillations7026]

I agree. I also think we should keep our minds open to the possibility that the properties of sentience or awareness could be far broader or stranger than is typically talked about, from our human centered viewpoint.

[@endrawes0]

To assume that artificial neurons are fully capable of encapsulating and reproducing the not-fully-understood and nuanced characteristics biological neurons is hubris.

[@fesimco4339]

@@endrawes0 To be pedantic "Hubris" isn't really the word you guys are looking for; myopic might be more appropriate.

[@endrawes0]

@@fesimco4339 myopic is /also/ appropriate but I meant hubris.

[@fesimco4339]

@@endrawes0 Fair enough; I thought you meant something else.

[@rodneyericjohnson]

44:00 This is a sad exchange. They both clearly want this to be a problem for after they have lived a full life. As do I. But that's not what's happening. We only have a few years left if we don't do something to slow down AI development.

[@sunsunsunh]

😂

[@stupidas9466]

I'll start worrying about AI once i see signs that AI has a chance of occurring. What people are calling AI now is simply machine learning, which in itself is amazingly complex and has issues for society on it's own, BUT is very very very different and poses very different problems. We are no where close to even understanding how human consciousness comes about at an elementary level, let alone knowing how to build machines that are capable of it. We simply have not found a way to study the human brain with it inside the living human without affecting it in such a way to make anything and everything we learn from the study of it meaningless to it's reproduction outside of the living human. An artist can draw a human brain to varying degrees of accuracy and visual likeness but that picture, no matter how detailed, will never be anything more than analogous to a brain, and at the most useless and superficial level. What makes a brain a brain, it's "essence" if you will, is exponentially greater than the sum of it's parts, and we at this point have no idea how those parts are interconnected (and one could argue we don't even know what those parts are despite having "names" or "labels" applied to them). And if we ever are able to reach that milestone, it would be another exponentially greater step to figure out how to replicate it and put it into a "machine" that was also either necessary designed and built by us same humans, or designed and built by a machine that was designed and built by us. And failing 10 or 20 "once in a millennium" scientific breakthroughs i don't for see this happening. We simply are not that smart.

[@dmitryshusterman9494]

If you bother to write such long comments, please read them and correct bad grammar, otherwise they are unreadable to others

[@LouigiVerona]

I like the guest, but the conversation about creativity felt really sloppy. Also, Dall-e cannot create something odd? You must be kidding me.

[@c.f.3503]

First

[@cwcarson]

Congratulations

[@stephencolbertcheese7354]

but r u sentient, or just an AI?

[@HarryNicNicholas]

only in a manner of speaking.