DO REASONING MODELS ACTUALLY SEARCH?

[@MachineLearningStreetTalk]

Transcript: www.dropbox.com/scl/fi/d0ef4ovnfxi0lknirkvft/Subbarao.pdf?rlkey=l3rp29gs4hkut7he8u04mm1df&dl=0

[@realhamza2001]

I know this is unrelated but give the Quran a read, also have a good day :)

[@snarkyboojum]

Why is this guy so underrated. What a clear and powerful thinker.

[@davidbutler9323]

Powerful speaker, for sure. I find he dismisses things that need not be dismissed.

[@matteoianni9372]

Maybe because we was monumentally wrong, he’s giving a precarious justification for why he was wrong and is still very wrong now.

[@snarkyboojum]

@@matteoianni9372care to provide any evidence for those statements?

[@reptar4life]

This channel has become my favorite source of philosophical and technical AI conversations. I am a layman on this topic and your channel has opened up a whole world for me

[@oncedidactic]

MLST ftw!

[@shoubhikdasguptadg9911]

I think we have had our fair share of discussions around the theme of "Do LLMs think" ,"LLMs are stochastic parrots", "LLMs are not the path to AGI" etc. While these discussions have been beneficial and helped all of us to look at LLMs from a variety of different lenses, now would be a good time to pivot. Lets start having more guests who are implementing stuff, for example the ICLR submissions are out, lets have a few guests who have their papers shortlisted and talk to them about implementations, compute constraints, Evals, data pipelining etc.

[@uw10isplaya]

I assume this was recorded before the r1 paper dropped? Would love to hear a part 3. A lot of your guests and talks have been very informative and helpful for me, but Prof. Kambhampati's two conversations have felt the most pertinent to the actual current state of affairs. In the trenches type stuff rather than big picture/metatopics.

[@zbll2406]

All of these channels videos all sound "outdated" because the guests and the host are completely out-of-tune as to the capabilities of modern LLMs. The disconnect is so hard to believe, its like these people have been living under a rock. Probably no one in this comment section is aware of the benchmarks being saturated by LLMs or the crazy LLM based products being built on the daily. Such a weird bubble this channel

[@lenyabloko]

​@@zbll2406it is better to leave under the rock than on Mars

[@PraveenKumar-bo7fw]

@@zbll2406 They are out of tune with the marketing. Not the technology

[@evodevo420]

@@PraveenKumar-bo7fwyou said it so perfectly!

[@keizbot]

Ugh I really wish this could have been recorded after R1, I would love to hear Rao's thoughts after seeing an open source LRM

[@paulocacella]

It seems to me that illusion of generalizing in human beings came from the fact that we are, a continuous in time, closed loop control systems, while LLMs are not fed with information continuously and, as a consequence, cannot have a sense of time. The difference between sentient systems like us are the permanence and adaptation. We are the same and one as an LLM, regarding language capabilities, but we are agents that perform interaction with the world and, language is a communication and an organization bridge. LLM limitations came, in my view, from the fact they lack these properties. Only a complete agent can perform on par with an human being. The paradigm shift is that we need a kind of stochastic programming that cannot be fail proof, like we are. Logic and reasoning are tools that we use to connect with nature, that seems to go on unrelated with our needs or desires. They are not necessarily part of bare reality, but the way we, as a machine capable of building tools and history, try to survive.

[@nicbarth3838]

Ah so so we are making software that can reason like us from mathematical structure in language? so it is sapient but without sentience. Thought machine

[@shyama5612]

My favorite proffy is back!

[@margaritagarcia193]

This channel has become one of my favourite sources of knowledge! Thank you so much for developing relevant and scientific content about AI. It makes a huge difference as we are swimming in a sea of AGI and AI marketing nonsense. I absolutely love it when you speak to Prof. Subbarao Kambhampati. It is so great to hear him talk. What an incredible scientist and intelligent man,, it so refreshing and valuable, unlike the nonsense from the Technocrats who only want to inflate market valuation. Thank you so much!

[@luke.perkin.online]

Fascinating talk. Professional sceptic, always a pleasure to listen to!

[@ikiphoenix9505]

Thanks for this one ! Eagerly waiting for Ohad Asor on MLST.

[@opusdei1151]

Thank you so much for your work-it's such a feast for the mind!

[@agenticmark]

Love it, he said Q values, as in RL we were using years ago. Sutton finally getting his due, too bad he isnt being named as the father of this type of learning.

[@jsalsman]

Giving DeepSeek-r1 tic-tac-toe problems absolutely shows explicit search at least 2 and sometimes 3 moves ahead, but it takes a very long time, 140 seconds for a 2-move lookahead problem. In part because it doesn't recognize symmetries of any sort.

[@BHBalast]

It seems like reasoning about reasoning is a natural next move and could be a step change like o1 was.

[@zbll2406]

It has never seen those problems that is why it struggles. Same as humans on novel problems

[@ArtechsVlog]

@@zbll2406 You're confusing "struggles" with "fails entirely"

[@zbll2406]

@@ArtechsVlog people forget the issues behind tokenization and compreension of visual or geometric features, these models only read text, and in fragments, when you throw at them stuff that they cannot really discern because its visual or geometric, no amount of reasoning helps you. Think of it like this, imagine you are trying to navigate your way around and unknown house , but the lights are off and you are resorting to your sense of touch. No matter how much you go touching around, finding the bathroom becomes near impossible

[@Kaleb_Weise]

An unexpected,but welcome recommendation on the feed- thank-you!

[@ai-cowboy]

Was this recorded before R1 release? If so he had some accurate predictions.

[@zbll2406]

This entire channel is outdated

[@henrikjohansson6666]

Subbarao seems to be quite focused on the cost of LRM:s, it would be really interesting to see if his perspective change now after Deepseek released a model that has o1 capabilities at 2% of the cost per token.... This is also likely to go down rapidly due to algorithmic improvements and just plain old better hardware. Dismissing use cases from a cost basis seems short sighted....

[@Lukas-uc2bu]

"I think you may have to wait until the Chinese replicate what we did to actually figure out what we did." Unknown OpenAI employee. They were right.

[@wwkk4964]

I really enjoyed the second follow up interview. I appreciate where he is coming from much more this time around as I felt he was being too dismissive earlier. He's right that the Science of reasoning, logic, retrieval that was successful in modeling computational systems and their behavior is caught in hard place as connectionism has proven to be a third type of phenomenon that is neither syllogistic logic or platonic ideal nor is it a mechanistic description of cause and effect. So its clear, our vocabulary has to expand to accomodate the blind men and the elephant quickly.

[@MyUpsideDownLife-SKR]

Brilliant! Thanks gents!🙏🏾😊

[@shawnfromportland]

i can't go out tonight, babe. new mlst.

[@tonyppe]

I am amazed at all the news saying AGI is here! and ASI is right around the corner and we should prepare. And you look at the complex questions that the recent models are answering, and it appears pretty amazing. But then you continue on, and these latest models still struggle with some basic reasoning questions. So what do they mean by AGI in this context because it appears to me that my understanding of what AGI is is different to what say OpenAI's understanding is and I am wondering if openai have a psuedo-AGI definition and are intent on misleading the world with this or, some other reason.

[@szebike]

I'm getting the feeling since the beginning that our anthropomorphism gets in the way to judge these systems properly. We as humans are programmed to assume anything that talks (or writes) to be intelligent but what if it truly just imitates language very well ? Most LLM answers I get is usally 90% of my prompt in other words and like 10% of related "new" information. And I don't loose the feeling that most responses are always a bit off even if it is "correct".

[@notaras1985]

It's fake yes

[@GurixRathore]

OpenAI seems to have this habit of shifting the definition of AGI depending not only on space and time but also on “who’s in the space at the time”

[@govelasco]

Fantastic episode!! I love the Life of Brian reference!!

[@ElieSanhDucos0]

Holy grail. Come on

[@govelasco]

@ElieSanhDucos0 doh! You're right! Agh!

[@shyama5612]

Prof needs to be a monthly return guest. I'd love to ask him about the $500b (uph!) Open AI investment claim and where he would invest if he had the reins.

[@agenticmark]

Prof. Subbarao's intro with Monty Python as his example makes him awesome in my book!

[@johnkost2514]

Maybe search is really an intersection of the 'prompt+context' hyperplane with a high-dimensional manifold (model) ..

[@danielhenderson7050]

Really enjoyed this one, thank you

[@inandemir321]

I wanted illustrate a point that I have not heard before, of consciousness. I believe the AI systems require this to be infused, I believe consciousness in AI can be incorporated by adding estimate and prediction of weights and measures where it can be illustrated by statistical analysis of all values including temperature, pressure, etc. of whatever the remote conditions can be incorporated into the systems.

[@induplicable]

MLST and a Monty Python skit! LOVE it!

[@authenticallysuperficial9874]

A computer scientist who doesn't know Life of Brian from Holy Grail? I can't trust him.

[@shyama5612]

Such balanced voices are so critical to the conversation - and rooted in strong academia and research - not just going with the shifting winds and hype (AGI utopia/dystopia and perverse incentives of regulatory capture by the incumbents)

[@TheReferrer72]

What hype? All I have seen from the last decade is progress towards AGI. I think ASI is the only target that is in doubt.

[@SahakSahakian]

Is it reasoning? Are airplanes flying? Submarines swimming? Noncense questions, like my friend Noam Chomsky said.

[@notaras1985]

Not nonsense thinking. If it can't reason in similar fashion with humans, and if it's far inferior in creativity then for it's computing power requirements, it's probably an unaffordable tool

[@notaras1985]

Reasoning is the human ability to utilise limited data and environmental resources, as well as infer past experiences, and adjustably manipulate your imminent material world in a way that betters humanity's well being and happiness. That's the reasoning we are trying to mimick with LLMs at least.

[@legionofthought]

​@notaras1985 If you're gonna literally smuggle "human" in as a prerequisite, you make the discussion pointless. It implies you misunderstood the OP. If you describe flying as what birds do to stay in the sky by flapping their wings, you're disingenuously disqualifying planes from the discussion.

[@for-ever-22]

Amazing video as usual

[@amonelders3916]

does he talk about deepseek?

[@sekharsirigiri9616]

At 27:45 into the talk (kzbin.info/www/bejne/aKmphYGOgJt0sLM) into the talk Prof. Rao says "Unless you do significant amount of post training on an LLM to improve reasoning (on a domain/multitude of domains?) to get approximate Q-Values upfront, you cannot improve just by MCTS. Think in terms of AlphaGo analogy, if you only did MCTS it will take much, much longer per move before you can get any level of accuracy/any level of confidence. But one of the things that AlphaGo does is that it does humongous amount of pre-training using a DNN (with a symbolic representation of the board state as input) where it learns an approximate policy which it then kind of rolls out to improve the Q-Value estimates that it has." Here are my perspectives: In AlphaGo, AlphaZero - The pre training using symbolic representation of the board state to learn an approximate policy and value functions (using DNN - either self play or using historical expert games) and the post training improvement of Q-values during MCTS - both of these are significant and important steps that happen before Inference time compute, which was the moment of reckoning when it played Lee Sedol. Hence the inference time compute that it unleashed after honing it both in a DNN based initial policy and value function and then bettering it using MCTS was effective. The problem with post training a generic LLM to improve it's reasoning and synthesize into sub tasks using CoT is that there are challenges on pre-training for problems that need symbolic representation and takes time per domain, assuming that can be done using an AI model and if not, a symbolic domain model has to be used. This will change per domain. Once that is done, the approximate Q-values have to be improved using probably MCT Search where again it should be domain specific. These are the challenges that are generic reasoning model like O1 faces, compared to say an AlphaGo.

[@hos42]

Andrej Karpathy didn't call it fractal intelligence, he called it "jagged intelligence"

[@MachineLearningStreetTalk]

Correct

[@Dht1kna]

too bad this was recording before DeepSeek R1

[@psi4j]

On a mf roll Tim! 🎉

[@jmaycock]

What a great guy

[@samahirrao]

Even if it is by luck or chance reasoning, one with money can still test out predictions by using humans and ot can still be profitable.

[@notaras1985]

Why do LLMs hallucinate or even lie often? Why do they deviate from their programming and display laziness?

[@screwsnat5041]

Hey guys for this who aren’t sure about If we are at AGI I can bet my life rn that if we get there it won’t be commercially viable to own an AGI company . To be an AGI requires some degree of spontaneity or processes that are more round the loop that doesn’t coexist with the premise of capitalism. Capitalism requires an innately predictable system. It’s long for me to explain just know . Predictability is part of what makes any business successful

[@memegazer]

If it is just search then what if reasoning is just efficient search

[@memegazer]

"it gives itself a gobbledygook string" Not sure why people think that humans don't reason in a similar fashion If you are tasked with solving some problem that requires step by step thinking and you are one of those with an internal dialogue, and assuming we might be able to record that internal dialogue then I suspect we would also accuse humans of having gobbledygook strings of tokens anyway the point is what humans seem to be able to do now that current models can't arbutrarily compress strings of tokens into more general but also more vague concept tokens and I am not convinced that gobbledygooking is not a necessary part of that process

[@isajoha9962]

I like people that think beyond nodes of assumed truth, on the path to clarity. Assuming systematic sequencing solving reasoning tasks "is IT" might not be the optimal solution, though? 🤔 LLM is not a script. The guest is kind of circling around assumptions of concepts, referencing probabilities.

[@isajoha9962]

Stating that 4o is kind of meh, is like admitting to not understanding prompting basics, IMHO. Eg do the prompt include the reasoning framework or do they assume that it is known to 4o?

[@patruff]

LLMs intelligence? What the frack?

[@majoorF]

it's indigital to assume everything from 1 ai! think of agi more like artificial societal intelligence or (ASI) . It's in the coöperation between the Ai systems where the notion of AGI arises. a full on automated self improving language modelling system. There will be static elements, dynamic elements, subjective elements and more. A group of people don't make a society! For that, something more is needed.

[@jesseeverhart]

Are we Cargo Culting our universe 🤔 Exponentially increasing fragmented 🌊 will become an issue if it already isn’t.

[@billykotsos4642]

The only bemchmark that counts

[@angloland4539]

[@kensho123456]

It's all wishful labelling.

[@paxdriver]

"He's a witch!!!" 😂❤

[@henrycobb]

A reasonable robot won't expend unreasonable efforts for its reasoning.

[@srivatsasrinivas6277]

Yes

[@musicandgallery-nature]

"they exchanged the truth about God for a lie and worshiped and served the creature rather than the Creator" - Bible

[@notaras1985]

Facts

[@musicandgallery-nature]

@@notaras1985 What is the difference between artificial and natural?

[@notaras1985]

@musicandgallery-nature many. We can learn what we didn't.

[@musicandgallery-nature]

@@notaras1985 www.youtube.com/@notaras1985 How do you know you'll learn the right thing from it? Why AI is already killing?

[@musicandgallery-nature]

@@notaras1985 www.youtube.com/@notaras1985 Which was built first - the atomic bomb or the nuclear power plant? A 14-year-old boy and a Belgian man have both been wiped out by AI. There are also many cases involving self-driving cars. The number of unknown cases cannot be determined. "Sometimes things happen and we never know the answer why" - SKYND, Michelle Carter

[@burnytech]

❤

[@good_vibes_20]

I'll give him my "magical token"

[@therainman7777]

Hide your goalposts folks. Something tells me this guy is going to be looking to move them.

[@Rensoku611]

I hope he mentions OpenAI's FrontierMath fiasco. It was all a lie and the goalpost people were saying it was AGI.

[@tylermoore4429]

To me it looks like behind the blizzard of words, he's grudgingly coming around to the fact that LLM's can reason after all.

[@govelasco]

I think he's actually casting a wide net and open to novel definitions of reasoning.

[@wwkk4964]

He actually made a good case for himself this time round. His problem definition seems to be : We don't know what reasoning actually is in an abstract sense and everyone appears to disagree about what it means. We dont even know if and how humans reason, so its a mistake to to try to force a definition without focusing on what it actually means for us when the systems are deployed: much like a plane is flying and is not a bird, what's more important is figuring out what guarantees we can get instead of trying to comprehend it in a human way when the intelligence is alien and what we want is getting a grip on how well it peeforms, not what we cal it.

[@notaras1985]

​@tylermoore4429 facts

[@billfrug]

Sounds like cargo cult as a service.

[@rey82rey82]

Witches exist

[@guimaraesalysson]

Tldr: No

[@ArtOfTheProblem]

you are on a ROLLL

[@ElieSanhDucos0]

I disagree on his monty python exemple. It is reasoning. It is doing analogies, maybe bad reasoning but it is not retrieving. It is trying to find signs and reason with them, wrongly yes, that she is a witch. And also the open source version is now out... I remember him being verrrry verryyy sure about LLMS not reasoning.... ;) . Maybe there is no human magic thing heh ;)

[@ElieSanhDucos0]

And it is holy grail ;(

[@govelasco]

We can't anthropomorphize it, novel definitions of reasoning and other terms must be nuanced to describe what is happening.

[@DirtyLifeLove]

Why don’t I find Monty Python funny?

[@agenticmark]

you gotta be an anglophile to appreciate it

[@notaras1985]

Because they aren't

[@zbll2406]

Guys o1 is an LLM, r1 is just an LLM, and they can reason, please lets admit that already and move on, these talks are utterly useless. Weird ass out of date channel jesus christ

[@160jeios]

Saar 😂

[@SahakSahakian]

I suggest renaming them to GREEDY AI, from OpenAI

[@zbll2406]

When are these people going to admit they were wrong about LLMs not being able to reason or not being truly intelligent? This channel is the most disingenuous AI channel I have haver seen

[@notaras1985]

They can't. And they aren't. It's just an elaborate T9 text completion that somewhat accurately guesses a sensible next word.

[@zbll2406]

@@notaras1985 Have you been living under a rock? Have you actually tried ChatGPT, o1, and seen the o3 results on ARC-AGI and any other benchmark?

[@zbll2406]

@@notaras1985 Its absolutely crazy to me that these people in the comments section and in the video OBVIOUSLY have not tried any of the new models or any models at all for that matter, even ChatGPT. This whole channel is such a peculiar BUBBLE, where people here pretend they are on bleeding edge yet don't even use or test the models at all. They don't follow any of the benchmarks being constantly saturated by LLMs, or the crazy products being built around them. Crazy bubble this channel, never seen such thing before

[@notaras1985]

@zbll2406 you comment seems incongruent to your initial one. Do you agree it's overhyped trash that just does what it was programmed or not

[@DebiprasadGhosh]

Please acknowledge Indian Philosophy also.

[@wwkk4964]

Indian philosophy might be useful in the neural net model but greek logic was binary and this is what modern computational sciences have been built upon. So he is correct that modern paradigm is lacking a way to classify what a language model is doing because it lacks a category of formally provable programming which has been the basis for computing and thinking and western language modelling classically. Clearly these are not memory retrieval systems and these are not explicit deterministic logical systems either. For a lack of a better analogy, they are most closely like a human, who are neither databases of associative memories nor are they explicit control systems with a logic flow.