Learning Library

← Back to Library

AI's Limits: Novel Reasoning

Key Points

  • The speaker stresses that AI, particularly large language models, are great at copying and re‑phrasing existing patterns but are fundamentally weak at genuine novel reasoning and solving brand‑new problems.
  • LLMs don’t actually reason; they simply retrieve contextual information, and making them perform symbolic reasoning requires cumbersome tool‑chains, underscoring how hard it is to give them true reasoning ability.
  • Their apparent intelligence is an illusion created by massive reading—LLMs sound smart because they’ve ingested vast text, yet extensive reading doesn’t equate to the capacity to tackle new, unseen challenges.
  • Real‑world business judgment and many societal tasks rely on solving novel problems, a domain where humans still excel and AI remains a poor bet despite its impressive conversational fluency.

Full Transcript

# AI's Limits: Novel Reasoning **Source:** [https://www.youtube.com/watch?v=b9WehQ_5qeA](https://www.youtube.com/watch?v=b9WehQ_5qeA) **Duration:** 00:12:20 ## Summary - The speaker stresses that AI, particularly large language models, are great at copying and re‑phrasing existing patterns but are fundamentally weak at genuine novel reasoning and solving brand‑new problems. - LLMs don’t actually reason; they simply retrieve contextual information, and making them perform symbolic reasoning requires cumbersome tool‑chains, underscoring how hard it is to give them true reasoning ability. - Their apparent intelligence is an illusion created by massive reading—LLMs sound smart because they’ve ingested vast text, yet extensive reading doesn’t equate to the capacity to tackle new, unseen challenges. - Real‑world business judgment and many societal tasks rely on solving novel problems, a domain where humans still excel and AI remains a poor bet despite its impressive conversational fluency. ## Sections - [00:00:00](https://www.youtube.com/watch?v=b9WehQ_5qeA&t=0s) **AI's Struggle with Novel Reasoning** - The speaker argues that large language models merely recall and repeat learned patterns rather than genuinely reason, and overcoming this fundamental limitation is far from easy. ## Full Transcript
0:01are you so tired of seeing so many links 0:04and so much content around what AI is 0:06good at or Panic inducing posts about 0:08how AI is going to take jobs great 0:11because this is not that this is a post 0:13about how AI is actually really really 0:16bad at certain things and it's not 0:18something that's easily fixable it's not 0:20that people aren't trying it's that it's 0:22not super easy to fix and it's something 0:23that if it were me I would be betting on 0:25from a skills perspective number one is 0:28the skill of novel reasoning so large 0:32language models actually don't reason 0:34they just recall context now you can 0:36stick a tool chain on there and and you 0:38can make them symbolically reason to 0:40some extent if you do that off of the 0:43large language model and then just Port 0:45the results back in and make it talk in 0:46English right so the fact that I had to 0:49use all those words to describe that 0:51should tell you people are having to 0:52bend over backwards to make llms reason 0:55at all and the reason why is that llms 0:57are actually just good at conversational 0:59flow they're good repeating text and 1:01patterns that we understand because 1:03they've read all of our text they've 1:04read everything they've read everything 1:06we wrote before the internet they've 1:08read everything after the internet and 1:10now they're reading stuff that they 1:11themselves created so they're really 1:13good at repeating text and that's 1:15amazing and it allows them to do a lot 1:17of things but mostly it allows them to 1:18do things where there are repeatable 1:21patterns that they can copy from and 1:23then regurgitate in ways that are 1:25technically new but not new kinds of 1:28things not new kinds of problems and so 1:32much of business is about solving net 1:34new problems frankly it's not just 1:36business government is about that 1:38education is about that you're solving 1:40net new problems we're teaching people 1:42to solve net new problems we're solving 1:44net new problems for society whatever it 1:47is that is what humans are actually 1:49really good at and llms are really 1:51really bad at it they just don't do it 1:54and the reason we get fooled into 1:56thinking that they do is because 2:00they have read so much they sound so 2:02smart they've read so much they've read 2:04more than I have that anybody else has 2:06my library is Tiny 2:09comparatively and we think that if 2:11someone's well read they must really 2:14understand how to do novel reasoning and 2:16that's been a really reliable assessment 2:18of human intelligence for as long as 2:20we've been able to read and write as a 2:22species and that's no longer true and I 2:25think that's really kind of confusing 2:26our brains because we have this thing 2:28that is super well read that sounds 2:31super smart when we talk to 2:33it and it's still not doing novel 2:36reasoning it's still not solving net new 2:38problems 2:40reliably just reading everything doesn't 2:43allow you to solve a net new 2:45problem and this is where what we call 2:48business judgment comes into play that's 2:50a really soft wishy-washy word but a lot 2:52of it hinges around reliably solving net 2:55new problems in ways that make sense in 2:57the market and and regardless of whether 3:01the role is a sea Suite role or an 3:03individual contributor role every role 3:05has some business judgment to it and 3:07generally speaking the business judgment 3:09parts are the parts that matter the most 3:11and those are the parts large language 3:13models are not good 3:16at and so that should be encouraging to 3:19you we're not going to run out of jobs 3:22that require business judgment because 3:24we're not going to run out of problems 3:25that require Innovation to 3:28solve all right the second skill that 3:31llms are terrible at is realtime context 3:35in fact the founders of businesses that 3:39are building llms like open AI have 3:43admitted that there isn't really a great 3:46answer yet to how large language models 3:49are supposed to handle real-time 3:51breaking news this happened just in 2024 3:55as we had breaking news event after 3:57breaking news event and sech that were 4:00supposedly built around AI or had AI 4:02components including 4:04Google did not reliably update their llm 4:08answers when breaking news happened 4:11because they're just not designed for 4:13net new real world context they're 4:15designed to read a gigantic context 4:17window and synthesize 4:20information that's a terrible way to 4:23handle a net new fact it's just 4:26bad and it extends Beyond breaking news 4:29there's a fundamental problem with llms 4:33in that they are really good at 4:35synthesizing from a large quantity of 4:38written text in the past and they are 4:41really really bad at understanding the 4:44real world realities of something 4:46happening right now in your local 4:48context for instance no llm is going to 4:52be able to realize that it is raining 4:55outside right now and therefore I do not 4:57need to water my tomatoes 5:00now you can use much simpler apps for 5:04that there are apps that will measure 5:06soil moisture and then choose not to 5:08water the garden but that's not 5:11Ai and the thing is those simple apps as 5:14you already know because we have had 5:17those for a long time do not really 5:20replace jobs either there is something 5:24about local context that is 5:26irreplaceable as far as human brains are 5:28concerned you need someone to sit there 5:31and I'm going to go from like watering 5:33the tomatoes to something that like a 5:35human can do if you were sitting there 5:38and you're trying to digest a bunch of 5:39different slack messages and you're 5:41trying to understand what your boss is 5:42expecting you to do and you're trying to 5:45understand what that jur tiet 5:46says and you are making sense of it all 5:49in your head that process that's hard to 5:51describe llms are really bad at 5:55it and also the simple apps that water 5:58the tomatoes are terrible it because 6:00they can't even understand it and so you 6:02know we've had those simple applications 6:04of technology for the last you know few 6:08decades and we've had the more fancy 6:11applications that are like language 6:12related for the last year or two and 6:14that's what we're all excited about and 6:15scared about this large language model 6:17effect and so you might think wow llms 6:20can really understand the slack messages 6:23and they can understand the J messages 6:25and so maybe they can help me think 6:26through and make sense of this and the 6:28truth is 6:29they can synthesize from it they can 6:31make patterns from it but if you're 6:33actually trying to solve a problem with 6:36deep understanding of real world world 6:38context they're not really very good at 6:41it so for instance I have tried this and 6:44llms do not reliably understand for 6:47instance how to assess the interrelated 6:53consequences of a Jura ticket that is 6:58getting worked on slow the Team Dynamics 7:01behind it the slack messages that are 7:03coming in 7:05periodically and all the other tickets 7:07that aren't getting done and part of why 7:09and that's a real world example right 7:11like we've all seen that if we've worked 7:12in software and part of why that is a 7:14problem is because so much of that 7:17context is hidden from a text 7:20perspective most of the world we work in 7:22actually doesn't just work on text 7:24there's a lot of human context unspoken 7:26things things between the lines llm are 7:29not good at that because they can't read 7:32it and so even if we got something that 7:34was good at breaking news or good at 7:35recent updates we still wouldn't have 7:38something that's good at reading between 7:39the lines because there's no text there 7:41you have to read between the lines it's 7:43real world context and llms are just not 7:45going to be good at it but humans are 7:48and that's something you can bet on too 7:51okay third one AI is under opinionated 7:55that's the third skill 7:56set and what I mean by that is that AI 7:59is designed for conversational flow it's 8:02designed to have a conversation that 8:04keeps going and that means it's actually 8:06designed to mirror to you they've done 8:07studies on this and llms tend to mirror 8:11the opinion they think will keep you 8:14chatting and F for once I actually do 8:18not think that that is a social 8:20algorithm designed to keep you addicted 8:24to chat it may become that but I think 8:28it's actually a situation where the 8:30large language model is trained to 8:33replicate patterns in your utterance in 8:36ways that make sense based on its very 8:39large training data 8:41set and so it's going to come back with 8:44something it think match it it thinks 8:45matches and so it's inherently a 8:48mirroring technology which means it's 8:50inherently really bad at decision-making 8:53because it's just going to come back and 8:54say what you say which means it's not 8:57going to give you a separate perspective 8:59now now you can brainstorm with it and 9:00it can help you expand your 9:02understanding of your own perspective 9:04absolutely it can give you a loosely 9:06held summarization of some alternate 9:09views 9:10sure but what it will not do is take a 9:14strong position that is tightly held and 9:17say this is what I really think should 9:19happen here because llms have no idea 9:22about what decisions are they're not 9:25built for decisions they're built for 9:27conversation and so whether you work in 9:31business whether you work in government 9:33I don't care you are still going to need 9:36to make 9:38decisions humans need to make decisions 9:41even if you're just living your life and 9:42you don't have a job humans need to make 9:45decisions and humans are actually really 9:47good at saying it's either A or B I'm 9:48going to pick a and this is 9:50why and llms just replicate human 9:55conversation that talks about picking A 9:57or B but they don't have an 9:59understanding of what a is or what B is 10:01and they certainly don't have an 10:02understanding of the 10:04choice and so without that they're not 10:08going to actually be making those 10:11decisions they're not going to be 10:13recommending decisions in ways that are 10:16deeply reflective and deeply rational 10:19and that is why even if you are drafting 10:21with large language models The 10:23Innovation the deep thought that comes 10:26with making good decisions is still 10:27going to have to come from a person it 10:29there's just no substitute for 10:32it that's just the way it is and that's 10:35a good thing it means that there are 10:37skill sets that humans can rely on all 10:39right so let's wrap this up what are the 10:42three skill sets that AI is not good at 10:46that humans are really reliably good at 10:48that you can bet on number one novel 10:51reasoning every role has some degree of 10:54Novel 10:55reasoning and or nearly every role and a 10:59I is just not going to get there large 11:05language models are really really bad at 11:08novel reasoning because they don't 11:09reason at all and let alone reason over 11:11new context they don't even know what 11:13new context 11:14is number two realtime context and this 11:17is related to the new context piece but 11:19I want to talk specifically about the 11:20fact that so much of real-time context 11:22is around silences it's around things 11:24that are between the lines it's around a 11:27felt sense the thing we call intuition 11:30llms don't have that realtime context is 11:34going to remain pretty much impossible 11:36for them to 11:38together and number three AI is under 11:41opinionated it just doesn't have real 11:44opinions it talks with you as if it does 11:46but it doesn't those three things are 11:49actually all 11:51crucial for job success and always have 11:56been and they're not going anywhere and 11:58so if you are tired of getting link 12:00after link of like AI is taking my jobs 12:03just come back to this post or share 12:04this post with someone who is worried 12:07about their role to remind them that 12:09there are skills and I haven't even 12:10listed them all these are just three to 12:12get started there are skills that are 12:15going to keep mattering regardless so 12:17don't give up hope