[HN Gopher] Fei-Fei Li: Spatial intelligence is the next frontie...
       ___________________________________________________________________
        
       Fei-Fei Li: Spatial intelligence is the next frontier in AI [video]
        
       Author : sandslash
       Score  : 248 points
       Date   : 2025-07-01 14:00 UTC (2 days ago)
        
 (HTM) web link (www.youtube.com)
 (TXT) w3m dump (www.youtube.com)
        
       | skwb wrote:
       | It's hard to describe, but it's felt like LLMs have completely
       | sucked the entire energy out of computer vision. Like... I know
       | CVPR still happens and there's great research that comes out of
       | it, but almost every single job posting in ML is about LLMs to do
       | this and that to the detriment of computer vision.
        
         | friendzis wrote:
         | What's the equivalent of methadone therapy, but for reckless
         | VC?
         | 
         | What's the equivalent of destroying everything around you while
         | chasing another high, but for reckless VC?
        
           | baxtr wrote:
           | Jeopardy?!
        
         | jgord wrote:
         | yeah, see my other comment.
         | 
         | To me its totally obvious that we will have a plethora of very
         | valuable startups who use RL techniques to solve realworld
         | problems in practical areas of engineering .. and I just get
         | blank stares when I talk about this :]
         | 
         | Ive stopped saying AI when I mean ML or RL .. because people
         | equate LLMs with AI.
         | 
         | We need better ML / RL algos for CV tasks :                 -
         | detecting lines from pixels       - detecting geometry in
         | pointclouds       - constructing 3D from stereo images,
         | photogrammetry, 360 panoramas
         | 
         | These might be used by LLMs but are likely built using RL or
         | 'classical' ML techniques, tapping into the vast parallel
         | matmull compute we now have in GPUs / multicore CPUs, and NPUs.
        
           | pzo wrote:
           | I thought there been a lot of progress in last 2 years.
           | (Video) Depth Anything, SAM2, grounding Dino, DFINE, VLM,
           | Gaussian splats, Nerf. Sure less than progres in LLm but
           | still I would say progress accelerated with LLM research.
        
           | tmilard wrote:
           | You said : "- detecting lines from pixels - detecting
           | geometry in pointclouds - constructing 3D from stereo images,
           | photogrammetry, 360 panoramas"                 ==> For me it
           | is more something like :        Source = crude video-or-photo
           | pixels  (to) ===> Find simple many rectangle-surface  that
           | are glued together one another.
           | 
           | This is, for me, how you really go easily to detecting rather
           | complexes geometry of any room.
        
         | porphyra wrote:
         | I feel like 3D reconstruction/bundle adjustment is one of those
         | things where LLMs and new AI stuff haven't managed to get a
         | significant foothold. Recently VGGT won best paper which is
         | good for them, but for the most part, stuff like NERF and
         | Gaussian Splatting still rely on good old COLMAP for bundle
         | adjustment using SIFT features.
         | 
         | Also, LLMs really suck at some basic tasks like counting the
         | sides of a polygon.
        
           | KaiserPro wrote:
           | > LLMs really suck at some basic tasks like counting the
           | sides of a polygon.
           | 
           | Oh indeed, but thats not using tokens correctly. if you want
           | to do that, then tokenise the number of polygons....
        
         | glitchc wrote:
         | Hah! And I remember when ML itself sucked all the energy out of
         | computer vision. Time to pay the piper.
        
         | m3kw9 wrote:
         | There is nothing to productize vs LLMs right now. I would say
         | robots could fix that but they have hard problems to solve in
         | the physical sense that will bottle neck things
        
         | smath wrote:
         | agreed about sucking the air out by LLM. The positive side is
         | that its a good time to innovate in other areas while a chunk
         | of ppl are absorbed in LLMs. A proven improvement in any other
         | non LLM space will attract investment.
        
           | pixl97 wrote:
           | Isn't this what Nvidia is already doing with a lot of their
           | sim software?
        
         | CSMastermind wrote:
         | Everyone is trying to jam transformers into CV workflows at the
         | moment. Possibly productively.
        
         | whiplash451 wrote:
         | It felt the same back in 2012-2015 when deep learning was
         | flooding over computer vision. Yet 10 years later there is a
         | net benefit for computer vision: a lot of tasks are now solved
         | much better/more efficiently with deep learning including those
         | that seemed "unfit" to deep learning like tracking.
         | 
         | I'm hopeful that VLMs will "fan out" into a lot of positive
         | outcomes for computer vision.
        
           | SlowTao wrote:
           | That is fair. I think it is a case of just seeing a lot if
           | great talent rush to the "in" thing. Other systems are still
           | being developed and that isnt lost but there is just a
           | feeling if being left out of it all while still doing great
           | stuff.
        
         | Barrin92 wrote:
         | >but almost every single job posting in ML is about LLMs
         | 
         | not in the defense sector, or aviation, or UAVS, automotive,
         | etc. Any proper real-time vision task where you have to
         | computationally interact with visual data is unsuited for LLMs.
         | 
         | Nobody controls a drone, missile or vehicle by taking a
         | screenshot and sending it to ChatGPT and has it do math while
         | it's on flight, anything that requires as the title of the
         | thread says, spatial intelligence is unsuited for a language
         | model
        
       | weinzierl wrote:
       | I tried LLM's for geolocation recently and it is both amazing how
       | good they are at recognizing patterns and how terrible they are
       | with recognizing and utilizing basic spatial relationships.
        
         | Pamar wrote:
         | I would like to read a complete example if you want to share (I
         | am not disputing yourbpoint, I'd just to understand better
         | because this is not my field so I cannot immediately map your
         | comment to my own experience)
        
           | weinzierl wrote:
           | Happy to share an complete example privately, contact data is
           | in my profile.
           | 
           | Will add condensed version here in half an hour.
        
             | Pamar wrote:
             | Condensed version will be more than adequate, thanks!
        
             | weinzierl wrote:
             | Condensing it for HN was harder than I thought because most
             | of it makes only sense when you also see the images, so
             | here is more like a summary of parts of the dialogue.
             | 
             | Prompted by this comment
             | 
             | https://news.ycombinator.com/item?id=44366753
             | 
             | I tried to geolocate the camera.
             | 
             | I uploaded a screenshot from
             | 
             | https://walzr.com/weather-watching
             | 
             | to ChatGPT and it said a lot of things but concluded with
             | "New York City street corner in the East Village".[1]
             | 
             | I find it utterly amazing that you can throw a random low-
             | quality image at an LLM and it does not only pinpoint the
             | city but also the quarter.
             | 
             | Good, but how to proceed from there? ChatGPT knows how
             | street corners in the East Village look in general, but it
             | does not know every building and every corner. Moreover, it
             | has no access to Google Street View to help find a matching
             | building. So this is kind of a dead end when we want a
             | precise location.
             | 
             | How would a human approach this? From the image it is
             | immediately clear that we have a T-shaped intersection.
             | From the shadows we can conclude that the stem of the T
             | runs east-west and the bar north-south. We see that the
             | stem is a one-way street with cars going only from west to
             | east. This immediately excludes all odd-numbered streets.
             | There are buildings all around and no parks or water.
             | 
             | There are not many intersections in New York that fulfill
             | all these criteria, but what if the area to search was much
             | bigger? Could an LLM do that for us?
             | 
             | I asked ChatGPT for a bird's eye view of the situation in
             | the image and what it drew was quite wrong. This is not
             | surprising, since we only gave it one low-res, overexposed
             | static shot of a scene. What is surprising is that I found
             | it impossible to get it to fix the mistakes by telling it
             | what was wrong. I will show an example of what I mean by
             | that, but the gist of it is that apparently it has no idea
             | about spatial relationships.
             | 
             | Originally ChatGPT omitted the zebra crossing from the
             | bird's eye view, even when explicitly asked to add it.
             | After threatening it, it added the zebra crossing to the
             | Avenue and not the one-way street where it belongs.
             | 
             | I: "No, the zebra is in the wrong direction. It is across
             | the horizontal street."
             | 
             | ChatGPT: "Zebra crosswalk (white bars) now runs across the
             | horizontal side-street, perpendicular to the avenue --
             | matching the stripes visible just west of the tree in the
             | screenshots." [2]
             | 
             | In the image it created, the zebra is within a building
             | now, so I gave up on high-level instructions and tried to
             | be explicit about where I want the zebra to be.
             | 
             | I: "The zebra needs to be moved to the east right to the
             | edge of the building."
             | 
             | ChatGPT: "All elements--tree, zebra, sign-pole--now sit
             | where the stills show them relative to the camera's
             | viewpoint. Let me know if any final tweak is needed!"
             | 
             | The image it generated has the zebra removed.
             | 
             | I: "I do not see a zebra. Maybe make it light grey or
             | something."
             | 
             | ChatGPT: "Zebra crosswalk now runs across the avenue, flush
             | with the facades of the camera-side and A/C-wall buildings;
             | stripes drawn in light grey for clarity."
             | 
             | Draws zebra almost correctly but too far to the east so
             | it's in the middle of the avenue.
             | 
             | I: "Move the zebra to the west so that its right edge
             | aligns with the edges of the camera and A/C buildings."
             | 
             | ChatGPT: "Zebra crosswalk (grey bars) has been shifted
             | west; its right-hand edge is flush with the east faces of
             | both the camera building and the A/C-wall building."
             | 
             | Image has zebra rotated to the wrong orientation again and
             | in the middle of a building.
             | 
             | I: "The zebra runs north to south like it was previously. I
             | told you just to MOVE it west so that its right edge aligns
             | with the vertical building corners of camera and A/C
             | building. I never told you to rotate it."
             | 
             | ChatGPT: "Zebra crosswalk runs north-south (vertical bars)
             | and its right edge aligns precisely with the east facades
             | of the camera-side and A/C-wall buildings."
             | 
             | Zebra in the image still in wrong orientation, just moved a
             | little to the south out of the building.
             | 
             | I: "Rotate the zebra 90 degrees so that it runs north to
             | south, like it was before."
             | 
             | ChatGPT: "Zebra crosswalk rotated back: bars run north-
             | south; its right edge aligns exactly with the east facades
             | of the camera and A/C-wall buildings."
             | 
             | I: "Now we have a zebra within the building."
             | 
             | Zebra moved out of building but again rotated in the wrong
             | direction.
             | 
             | ChatGPT: "Zebra crosswalk: runs north-south; its right edge
             | flush with the east facades of both camera and A/C
             | buildings."
             | 
             | I: "Rotate the zebra crosswalk so the stripes run west to
             | east."
             | 
             | This gave an acceptable result and I left it there.
             | 
             | Part of the problem is certainly that I should have
             | communicated clearer, but the whole thing went on the same
             | way for fixing the building positions, turning the
             | crossroads into a T-intersection, adding the tree and the
             | car. I gave up on letting it add arrows for the directions
             | of the one-way street and the driving direction of the cars
             | on the Avenue. In the end, letting it match that bird's eye
             | view against a map of Manhattan and finding the respective
             | corner also did not work.
             | 
             | [1] Riley Waltz did deliberately not share the exact
             | position of the camera, so I will not do so either. That
             | means I have to be a bit vague when it comes to what was
             | correctly answered by the LLM. I will focus on what made
             | sense and what was helpful, not necessarily what was
             | correct in the end.
             | 
             | [2] All ChatGPT output verbatim but abbreviated to the
             | relevant parts.
        
               | Workaccount2 wrote:
               | I'd take text-to-image capabilities with a grain of salt,
               | because they are dramatically lower than their text to
               | text abilities. I don't know the exact mechanics with
               | current multimodal models, but it is pretty clear that
               | there is a disconnect between what the text modal wants,
               | and what the text model outputs. It's almost feels like
               | asking someone with a blindfold to draw a cat, you kinda
               | get a mess.
               | 
               | If you ask chatgpt to describe a new image based off an
               | input image, it will do dramatically better. But ask it
               | to use it's image generation tooling and the "awareness"
               | judged by the image output falls off a cliff.
               | 
               | Another example is infographics or flow charts. The
               | models can easily output that information and put it in a
               | nicely formatted text grid for you. But ask them to put
               | it in a visual image, and it's just a mess. I don't think
               | it's the models, I think it's the text-image translation
               | layer.
        
               | weinzierl wrote:
               | This is a good point. The 2D birds eye view image adds
               | another separate complication. There are certainly better
               | and more direct ways to show that current models are bad
               | with spatial reasoning. This was just a byproduct of my
               | geolocation experiments. Maybe I will give it a shot
               | another day.
        
         | ansgri wrote:
         | I've tried to use various OpenAI models for OpenSCAD code
         | generation, and while the code was valid, it absolutely
         | couldn't get spatial relationships right. Even in simple cases,
         | like a U-tube assembled from 3 straight and 2 curved segments.
         | So this is definitely an area for improvement.
        
         | AStrangeMorrow wrote:
         | Yeah even LLM generated code for a 2D optimization problems
         | with many spatial relationships has being absolutely terrible,
         | while I had great success in other domains.
        
       | jandrewrogers wrote:
       | I appreciate the video and generally agree with Fei-Fei but I
       | think it almost understates how different the problem of
       | reasoning about the physical world actually is.
       | 
       | Most dynamics of the physical world are sparse, non-linear
       | systems at every level of resolution. Most ways of constructing
       | accurate models mathematically don't actually work. LLMs, for
       | better or worse, are pretty classic (in an algorithmic
       | information theory sense) sequential induction problems. We've
       | known for well over a decade that you cannot cram real-world
       | spatial dynamics into those models. It is a clear impedance
       | mismatch.
       | 
       | There are a bunch of fundamental computer science problems that
       | stand in the way, which I was schooled on in 2006 from the
       | brightest minds in the field. For example, how do you represent
       | arbitrary spatial relationships on computers in a general and
       | scalable way? There are no solutions in the public data
       | structures and algorithms literature. We know that universal
       | solutions can't exist and that all practical solutions require
       | exotic high-dimensionality computational constructs that human
       | brains will struggle to reason about. This has been the status
       | quo since the 1980s. This particular set of problems is hard for
       | a reason.
       | 
       | I vigorously agree that the ability to reason about
       | spatiotemporal dynamics is critical to general AI. But the
       | computer science required is so different from classical AI
       | research that I don't expect any pure AI researcher to bridge
       | that gap. The other aspect is that this area of research became
       | highly developed over two decades but is not in the public
       | literature.
       | 
       | One of the big questions I have had since they announced the
       | company, is who on their team is an expert in the dark state-of-
       | the-art computer science with respect to working around these
       | particular problems? They risk running straight into the same
       | deep, layered theory walls that almost everyone else has run
       | into. I can't identify anyone on the team that is an expert in a
       | relevant area of computer science theory, which makes me
       | skeptical to some extent. It is a nice idea but I don't get the
       | sense they understand the true nature of the problem.
       | 
       | Nonetheless, I agree that it is important!
        
         | machinelearning wrote:
         | "Most ways of constructing accurate models mathematically don't
         | actually work" > This is true for almost anything at the limit,
         | we are already able to model spatiotemporal dynamics to some
         | useful degree (see: progress in VLAs, video diffusion, 4D
         | Gaussians)
         | 
         | "We've known for well over a decade that you cannot cram real-
         | world spatial dynamics into those models. It is a clear
         | impedance mismatch" > What's the source that this is a
         | physically impossible problem? Not sure what you mean by
         | impedance mismatch but do you mean that it is unsolvable even
         | with better techniques?
         | 
         | Your whole third paragraph could have been said about LLMs and
         | isn't specific enough, so we'll skip that.
         | 
         | I don't really understand the other 2 paragraphs, what's this
         | "dark state-of-the-art computer science" you speak of and what
         | is this "area of research became highly developed over two
         | decades but is not in the public literature" how is "the
         | computer science required is so different from classical AI
         | research"?
        
           | calf wrote:
           | Above commenter also asserts "highly developed research but
           | no public literature" _shrug_ ...
        
             | fu-hn wrote:
             | But the best minds in the world said so!
        
         | ccozan wrote:
         | I agree that the problem is hard. However, biological brain is
         | able to handle it quite "easily" ( is not really easy - bilions
         | of iterations were needed ). The current brains are solving
         | this 3D physical world _only_ via perception.
         | 
         | So this is place were we must look. It starts with the sensing
         | and the integration of that sensing. I am working at this
         | problem since more than 10 years and I came to some results. I
         | am not a real scientist but a true engineer and I am looking
         | from that perspective quite intesely: The question that one
         | must ask is: how do you define the outside physical world from
         | the perspective of a biological sensing "device" ? what exactly
         | are we "seeing" or "hearing"? So yes, working on that brought
         | it further in defining the physical world.
        
           | tmilard wrote:
           | I do agree with you. We have an natural eye (what you call a
           | 'biological brain') automat that inconsciouly 'feels' the
           | structure of a geometric of the places we enter to.
           | 
           | Once this layer of "natural eye automat" is programmed behind
           | a camera, it will spit out this crude geometry : the Spacial-
           | data-bulk (SDB). This SDB is small data.
           | 
           | From now on, our programs will only do reason, not on data
           | froms camera(s) but only on this small SBD.
           | 
           | This is how I see it.
        
             | tmilard wrote:
             | ==> And now the LLMs, to feel Spacial knowledge, will have
             | a very reduce dataset. This will make spacial data
             | reasoning very less intencive than we can't imagine.
        
           | foobarian wrote:
           | Maybe a brute force solution would work just like it did for
           | text. I would not be surprised if the scale of that brute
           | force was not within reach yet though.
        
           | andoando wrote:
           | Also a cook here whose spent years thinking about this, would
           | love to hear about what results you've obtained
        
         | adamnemecek wrote:
         | All (ALL!!) AI/optimization problems boil down to energy
         | minimization or dually entropy maximization.
        
         | lsy wrote:
         | To make this more concrete: ImageNet enabled computer "vision"
         | by providing images + labels, enabling the computer to take an
         | image and spit out a label. LLM training sets enable text
         | completion by providing text + completions, enabling the
         | computer to take a piece of text and spit out its completion.
         | Learning how the physical world works (not just kind of works a
         | la videogames, _actually works_ ) is not only about a jillion
         | times more complicated, there is really only one usable
         | dataset: the world itself, which cannot be compacted or fed
         | into a computer at high speed.
         | 
         | "Spatial awareness" itself is kind of a simplification: the
         | idea that you can be aware of space or 3d objects' behavior
         | without the social context of what an "object" is or how it
         | relates to your own physical existence. Like you could have two
         | essentially identical objects but they are not interchangeable
         | (original Declaration of Independence vs a copy, etc). And many
         | many other borderline-philosophical questions about when an
         | object becomes two, etc.
        
           | coldtea wrote:
           | > _there is really only one usable dataset: the world itself,
           | which cannot be compacted or fed into a computer at high
           | speed._
           | 
           | Why wouldn't it be? If the world is ingressed via video
           | sensors and lidar sensor, what's the hangup in recording such
           | input and then replaying it faster?
        
             | psb217 wrote:
             | I think there's an implicit assumption here that
             | interaction with the world is critical for effective
             | learning. In that case, you're bottlenecked by the speed of
             | the world... when learning with a single agent. One neat
             | thing about artificial computational agents, in contrast to
             | natural biological agents, is that they can share the same
             | brain and share lived experience, so the "speed of reality"
             | bottleneck is much less of an issue.
        
               | HappMacDonald wrote:
               | Yeah I'm envisioning putting a thousand simplistic
               | robotic "infants" into a vast "playpen" to gather sensor
               | data about their environment, for some (probably smaller)
               | number of deep learning models to ingest the input and
               | guess at output strategies (move this servo, rotate this
               | camshaft this far in that direction, etc) and make
               | predictions about resulting changes to input.
               | 
               | In principle a thousand different deep learning models
               | could all train simultaneously on a thousand different
               | robot experience feeds.. but not 1 to 1, but instead 1 to
               | many.. each neural net training on data from dozens or
               | hundreds of the robots at the same time, and different
               | neural nets sharing those feeds for their own rounds of
               | training.
               | 
               | Then of course all of the input data paired with outputs
               | tested and further inputs as ground truth to predictions
               | can be recorded for continued training sessions after the
               | fact.
        
               | loa_in_ wrote:
               | But the playpen will contain objects that are inherently
               | breakable. You cannot rough handle the glass vessel and
               | have it too.
        
               | m-s-y wrote:
               | The world Is breakable. Any model based on it will need
               | to know this anyway. Am I missing your argument?
        
               | devenson wrote:
               | Can't reset state after breakage.
        
               | csullivan107 wrote:
               | Never thought I'd get to do this but this was my masters
               | research! Simulations are inherently limited and I just
               | got tired of robotic research being done only in
               | simulations. So I built a novel soft robot (notoriously
               | difficult to control) and got it to learn by playing!!
               | 
               | Here is an informal talk I gave on my work. Let me know
               | if you want the thesis
               | 
               | https://www.youtube.com/live/ZXlQ3ppHi-E?si=MKcRqoxmEra7Z
               | rt5
        
               | rybosome wrote:
               | A very interesting idea. I am curious about this sharing
               | and blending of the various nets; I wonder if something
               | as naive as averaging the weights (assuming the neural
               | nets all have the same dimensions) would actually
               | accomplish that?
        
               | hackyhacky wrote:
               | > In that case, you're bottlenecked by the speed of the
               | world
               | 
               | Why not have the AI train on a simulation of the real
               | world? We can build those pretty easily using traditional
               | software and run them at any speed we want.
        
             | otodus wrote:
             | How would you handle olfactory and proprioceptive data?
        
           | m-s-y wrote:
           | > the world itself, which cannot be compacted or fed into a
           | computer at high speed.
           | 
           | ...yet.
           | 
           | 15 years ago LLMs as they are today seemed like science
           | fiction too.
        
           | amelius wrote:
           | Well, you can use one or two cameras and a lidar, and use
           | that to generate data to train a depth-map.
        
           | TheOtherHobbes wrote:
           | Considering how bad LLMs are at understanding anything, and
           | how they still manage to be useful, you simply don't need
           | this level of complexity.
           | 
           | You need something that mostly works most of the time, and
           | has guardrails so when it makes mistakes nothing bad happens.
           | 
           | Our brains acquire quite good heuristics for dealing with
           | physical space without needing to experience all of physical
           | reality.
           | 
           | A cat-level or child-level understanding of physical space is
           | more immediately useful than a philosopher-level of
           | understanding.
        
           | sega_sai wrote:
           | I feel that if words/phrases/whole texts can be embedded well
           | in high dimensional spaces as points, the same must apply to
           | the 3d world. I'm sure there will be embeddings of it (i.e.
           | mapping the 3-d scene into a high-D vector) and then we'll be
           | work with those embeddings as LLMs work with text
           | (disclaimer: I am not an expert in the field ).
        
         | mindcrime wrote:
         | > became highly developed over two decades but is not in the
         | public literature.
         | 
         | Developed by who? And for what purpose? Are we talking about
         | overlap with stuff like missile guidance systems or targeting
         | control systems or something, and kept confidential by the
         | military-industrial complex? I'm having a hard time seeing many
         | other scenarios that would explain a large body of people doing
         | research in this area and then not publishing anything.
         | 
         | > I can't identify anyone on the team that is an expert in a
         | relevant area of computer science theory
         | 
         | Who is an expert on this theory then?
        
         | gyomu wrote:
         | > There are a bunch of fundamental computer science problems
         | that stand in the way, which I was schooled on in 2006 from the
         | brightest minds in the field. For example, how do you represent
         | arbitrary spatial relationships on computers in a general and
         | scalable way? There are no solutions in the public data
         | structures and algorithms literature. We know that universal
         | solutions can't exist and that all practical solutions require
         | exotic high-dimensionality computational constructs that human
         | brains will struggle to reason about. This has been the status
         | quo since the 1980s. This particular set of problems is hard
         | for a reason.
         | 
         | Where can I read more about this space? (particularly on the
         | "we know that universal solutions can't exist" front)
        
         | niemandhier wrote:
         | Regarding sparse, nonlinear systems and our ability to learn
         | them:
         | 
         | There is hope. Experimental observation is, that in most cases
         | the coupled high dimensional dynamics almost collapses to low
         | dimensional attractors.
         | 
         | The interesting thing about these is: If we apply a measurement
         | function to their state and afterwards reconstruct a
         | representation of their dynamics from the measurement by
         | embedding, we get a faithful representation of the dynamics
         | with respect to certain invariants.
         | 
         | Even better, suitable measurement functions are dense in
         | function space so we can pick one at random and get a suitable
         | one with probability one.
         | 
         | What can be glanced about the dynamics in terms of of these
         | invariants can learned for certain, experience shows that we
         | can usually also predict quite well.
         | 
         | There is a chain of embedding theorems by Takens and Sauer
         | gradually broadening the scope of applicability from
         | deterministic chaos towards stochasticly driven deterministic
         | chaos.
         | 
         | Note embedding here is not what current computer science means
         | by the word.
         | 
         | I spend most of my early adulthood doing theses things, would
         | be cool to see them used once more.
        
           | golol wrote:
           | What field of mathematics is this? Can you point me to some
           | keywords/articles?
        
         | teemur wrote:
         | > We know that universal solutions can't exist and that all
         | practical solutions require exotic high-dimensionality
         | computational constructs that human brains will struggle to
         | reason about. This has been the status quo since the 1980s.
         | This particular set of problems is hard for a reason.
         | 
         | This made me a bit curious. Would you have any pointers to
         | books/articles/search terms if one wanted to have a bit deeper
         | look on this problem space and where we are?
        
           | jandrewrogers wrote:
           | I'm not aware of any convenient literature but it is
           | relatively obvious once someone explains it to you (as it was
           | explained to me).
           | 
           | At its root it is a cutting problem, like graph cutting but
           | much more general because it includes things like non-trivial
           | geometric types and relationships. Solving the cutting
           | problem is necessary to efficiently shard/parallelize
           | operations over the data models.
           | 
           | For classic scalar data models, representations that preserve
           | the relationships have the same dimensionality as the
           | underlying data model. A set of points in 2-dimensions can
           | always be represented in 2-dimensions such that they satisfy
           | the cutting problem (e.g. a quadtree-like representation).
           | 
           | For non-scalar types like rectangles, operations like
           | equality and intersection are distinct and there are an
           | unbounded number of relationships that must be preserved that
           | touch on concepts like size and aspect ratio to satisfy
           | cutting requirements. The only way to expose these additional
           | relationships to cutting algorithms is to encode and embed
           | these other relationships in a (much) higher dimensionality
           | space and then cut that space instead.
           | 
           | The mathematically general case isn't computable but real-
           | world data models don't need it to be. Several decades ago it
           | was determined that if you constrain the properties of the
           | data model tightly enough then it should be possible to
           | systematically construct a finite high-dimensionality
           | embedding for that data model such that it satisfies the
           | cutting problem.
           | 
           | Unfortunately, the "should be possible" understates the
           | difficulty. There is no computer science literature for how
           | one might go about constructing these cuttable embeddings,
           | not even for a narrow subset of practical cases. The activity
           | is also primarily one of designing data structures and
           | algorithms that can represent complex relationships among
           | objects with shape and size in dimensions much greater than
           | three, which is cognitively difficult. Many smart people have
           | tried and failed over the years. It has a lot of subtlety and
           | you need practical implementations to have good properties as
           | software.
           | 
           | About 20 years ago, long before "big data", the iPhone, or
           | any current software fashion, this and several related
           | problems were the subject of an ambitious government research
           | program. It was technically successful, demonstrably. That
           | program was killed in the early 2010s for unrelated reasons
           | and much of that research was semi-lost. It was so far ahead
           | of its time that few people saw the utility of it. There are
           | still people around that were either directly involved or
           | learned the computer science second-hand from someone that
           | was but there aren't that many left.
        
             | andoando wrote:
             | Ive spent years trying to tackle spatial representations on
             | my own, so Im extremely curious here.
             | 
             | How does the cutting problem relate to intelligence in the
             | first place?
        
               | jandrewrogers wrote:
               | Indexing is a special case of AI. At the limit, optimal
               | cutting and learning are equivalent problems. Non-trivial
               | spatial representations push these two things much closer
               | together than is normally desirable for e.g. indexing
               | algorithms. Tractability becomes a real issue.
               | 
               | Practically, scalable indexing of complex spatial
               | relationships requires what is essentially a type of
               | learned indexing, albeit not neural network based.
        
             | calf wrote:
             | But then that sounds more like that person explained it
             | wrong. They didn't explain why it is necessary to reduce to
             | GRAPHCUT, it seems to me to beg the question. We should not
             | assume this is true based on some vague anthropomorphic
             | appeal to spatial locality, surely?
        
               | jandrewrogers wrote:
               | It isn't a graph cutting problem, graph cutting is just a
               | simpler, special case of this more general cutting
               | problem (h/t IBM Research). If you can solve the general
               | problem you effectively get efficient graph cutting for
               | free. This is obviously attractive to the extent you can
               | do both complex spatial and graph computation at scale on
               | the same data structure instead of specializing for one
               | or the other.
               | 
               | The challenge with cutting e.g. rectangles into uniform
               | subsets is that logical shard assignment must be
               | identical regardless of insertion order and in the
               | absence of an ordering function, with O(1) space
               | complexity and without loss of selectivity. Arbitrary
               | sets of rectangles overlap, sometimes heavily, which is
               | the source of most difficulty.
               | 
               | Of course, with practical implementations write
               | scalability matters and incremental construction is
               | desirable.
        
             | tehjoker wrote:
             | Did that research program have a public code name?
        
               | mindcrime wrote:
               | Looking through some old DARPA budget docs[1], it seems
               | like there's a chance that what's being discussed here
               | falls under DARPA's "PE 0602702E TACTICAL TECHNOLOGY"
               | initiative, project TT-06.
               | 
               | Some other possibilities might include:                 -
               | "PE 0602304E COGNITIVE COMPUTING SYSTEMS", project
               | COG-02.       - "PE 0602716E ELECTRONICS TECHNOLOGY",
               | project ELT-01       - "PE 0603760E COMMAND, CONTROL AND
               | COMMUNICATIONS SYSTEMS", project CCC-02       - "PE
               | 0603766E NETWORK-CENTRIC WARFARE TECHNOLOGY", project
               | NET-01       - "PE 0603767E SENSOR TECHNOLOGY", project
               | SEN-02
               | 
               | Or maybe it's nothing to do with this at all. But in
               | either case, this looks like some interesting stuff to
               | explore in its own right. :-)
               | 
               | [1]: https://web.archive.org/web/20181001000000/https://w
               | ww.darpa...
        
         | idiotsecant wrote:
         | If there's one thing that control theory has taught us in the
         | last 100 years, it's that anything is linear if you zoom in far
         | enough. Nonlinearity is practically solvable by adjusting your
         | controls to different linear models depending on your position
         | in the system space.
        
         | amelius wrote:
         | Some types of deep learning model can handle 3d data quite
         | well:
         | 
         | https://en.wikipedia.org/wiki/Neural_radiance_field
        
         | nurettin wrote:
         | > how do you represent arbitrary spatial relationships on
         | computers in a general and scalable way?
         | 
         | Isn't this essentially what the convolutional layers do in
         | LeNet?
        
         | queuebert wrote:
         | I think you are using "sparse" and "non-linear" as scare terms.
         | Sparse is a good thing, as it reduces degrees of freedom, and
         | non-linear does not mean unsolvable.
         | 
         | Also "impedance mismatch" doesn't mean no go, but rather less
         | efficient.
        
         | CamperBob2 wrote:
         | _We've known for well over a decade that you cannot cram real-
         | world spatial dynamics into those models. It is a clear
         | impedance mismatch._
         | 
         | Then again, not much that we "knew" a decade ago is still
         | relevant today. Of course transformer networks have proven
         | capable of representing spatial intelligence. How could they
         | work with 2D images, if not?
        
         | andoando wrote:
         | What's non linear about spatial reasoning?
         | 
         | >We know that universal solutions can't exist
         | 
         | Why not?
        
           | randcraw wrote:
           | Spatial models must be 3D, not 1D (linear), much less 2D,
           | which is sufficient for images and object recognition (where
           | models are not needed). And adding time makes it 4D, at least
           | for robot motion.
           | 
           | To reason spatially (and dynamically) the dependence of one
           | object's position in space on other objects (and their
           | motions and behaviors) adds up fast to complicate the model
           | in ways that 95% of 2D static image analysis does not.
        
             | andoando wrote:
             | Well hold on, first Im not convinced we have solved 2D
             | spatial intelligence. Analyzing 2D images is very different
             | from being able to reason about 2D geometry. How do you
             | mathematically define relations like "above", "below",
             | "diagonal", etc in a composable way that can be learned?
             | 
             | Second, problems in 3D can be deconstructed to 2D. For
             | example, how do you get to the airport? You need to first
             | solve the 2D overview of the path youd take as youd now
             | looking at a map. Then you need to reason about your field
             | of view, and here again I believe youre really reasoning is
             | something like "object A is behind object B and A is to the
             | left of B", and not solving some non linear equation
             | 
             | I think a big issue is people are trying to solve this in
             | the realm of traditional mathematics, and not as a simple
             | step by step process
        
         | dopadelic wrote:
         | You're pointing out a real class of hard problems -- modeling
         | sparse, nonlinear, spatiotemporal systems -- but there's a
         | fundamental mischaracterization in lumping all transformer-
         | based models under "LLMs" and using that to dismiss the
         | possibility of spatial reasoning.
         | 
         | Yes, classic LLMs (like GPT) operate as sequence predictors
         | with no inductive bias for space, causality, or continuity.
         | They're optimized for language fluency, not physical grounding.
         | But multimodal models like ViT, Flamingo, and Perceiver IO are
         | a completely different lineage, even if they use transformers
         | under the hood. They tokenize images (or video, or point
         | clouds) into spatially-aware embeddings and preserve positional
         | structure in ways that make them far more suited to spatial
         | reasoning than pure text LLMs.
         | 
         | The supposed "impedance mismatch" is real for language-only
         | models, but that's not the frontier anymore. The field has
         | already moved into architectures that integrate vision, text,
         | and action. Look at Flamingo's vision-language fusion, or
         | GPT-4o's real-time audio-visual grounding -- these are not mere
         | LLMs with pictures bolted on. These are spatiotemporal
         | attention systems with architectural mechanisms for cross-modal
         | alignment.
         | 
         | You're also asserting that "no general-purpose representations
         | of space exist" -- but this neglects decades of work in
         | computational geometry, graphics, physics engines, and more
         | recently, neural fields and geometric deep learning. Sure, no
         | universal solution exists (nor should we expect one), but
         | practical approximations exist: voxel grids, implicit neural
         | representations, object-centric scene graphs, graph neural
         | networks, etc. These aren't perfect, but dismissing them as
         | non-existent isn't accurate.
         | 
         | Finally, your concern about who on the team understands these
         | deep theoretical issues is valid. But the fact is: theoretical
         | CS isn't the bottleneck here -- it's scalable implementation,
         | multimodal pretraining, and architectural experimentation. If
         | anything, what we need isn't more Solomonoff-style induction or
         | clever data structures -- it's models grounded in perception
         | and action.
         | 
         | The real mistake isn't that people are trying to cram physical
         | reasoning into LLMs. The mistake is in acting like all
         | transformer models are LLMs, and ignoring the very active (and
         | promising) space of multimodal models that already tackle
         | spatial, embodied, and dynamical reasoning problems -- albeit
         | imperfectly.
        
           | calf wrote:
           | How do we prove a trained LLM has no inductive bias for
           | space, causality, etc.? We can't assume this is true by
           | construction, can we?
        
             | dopadelic wrote:
             | Why would we need to prove such a thing? Human vision has
             | strong inductive biases, which is why you can perceive
             | objects in abstract patterns. This is why you can lay down
             | at a park and see a duck in a cloud. It's also why we can
             | create abstracted representations of things with graphics.
             | Having inductive biases makes it more relatable to the way
             | we work.
             | 
             | And again, you're using the term LLMs again when vision
             | based transformers in multimodal models aren't simply LLMs.
        
           | mumbisChungo wrote:
           | Claude, is that you?
        
         | ryeguy_24 wrote:
         | Agree. Also, with respect to training, what is the goal that we
         | are maximizing? LLMs are easy, predicting the next word and we
         | have lots of training data. But what are we training for in
         | real world? Modeling the next spatial photograph to predict
         | things that will happen next? It's not intuitive to me what
         | that objective function would be in spatial intelligence.
        
           | curiouscavalier wrote:
           | Or that there is a sufficiently generalizable objective
           | function for all "spatial intelligence."
        
         | epr wrote:
         | Human beings get by quite well with extremely oversimplified
         | (low resolution) abstractions. There is no need whatsoever for
         | something even approaching universal or perfect. Humans aren't
         | thinking about fundamental particles or solving differential
         | equations in their head when they're driving a car or playing
         | sports.
        
         | voxleone wrote:
         | I'm trying to approach spatial reasoning by introducing
         | quaternions to navigate graphs. It is a change in the unit of
         | traversal -- from positional increment to rotational
         | progression. This reframing has cascading effects. It alters
         | how we model motion, how we think about proximity, and
         | ultimately how systems relate to space itself.
         | 
         | The traditional metaphor of movement -- stepping from point A
         | to point B -- is spatially intuitive but semantically
         | impoverished. It ignores the continuity of direction, the
         | embodiment of motion, and the nontriviality of turning.
         | Quaternion-based traversal reintroduces these elements. It is
         | not just more precise; it is more faithful to the mechanisms by
         | which physical and virtual entities evolve through space. In
         | other words objects 'become' the model.
         | 
         | https://github.com/VoxleOne/SpinStep/blob/main/docs/index.md
        
       | jgord wrote:
       | makes sense - humans have evolved a lot of wetware dedicated to
       | 3D processing from stereo 2D.
       | 
       | I've made some progress on a PoC in 3D reconstruction - detecting
       | planes, edges, pipes from pointclouds from lidar scans, eg :
       | https://youtu.be/-o58qe8egS4 .. and am bootstrapping with in-
       | house gigs as I build out the product.
       | 
       | Essentially it breaks down to a ton of matmulls, and I use a lot
       | of tricks from pre-LLM ML .. this is a domain that perfectly fits
       | RL.
       | 
       | The investors Ive talked to seem to understand that scan-to-cad
       | is a real problem with a viable market - automating 5Bn / yr of
       | manual click-labor. But they want to see traction in the form of
       | early sales of the MVP, which is understandable, especially in
       | the current regime of high interest rates.
       | 
       | Ive not been able to get across to potential investors the vast
       | implications for robotics, AI, AR, VR, VFX that having better /
       | faster / realtime 3D reconstruction will bring. Its great that
       | someone of the caliber of Fei-Fei Li is talking about it.
       | 
       | Robots that interact in the real world will need to make a 3D
       | model in realtime and likely share it efficiently with comrades.
       | 
       | While a gaussian splat model is more efficient than a pointcloud,
       | a model which recognizes a wall as a quad plane is much more
       | efficient still, and needed for realtime communication. There is
       | the old idea that compression is equivalent to AI.
       | 
       | What is stopping us from having a google street-view v3.0 in
       | which I can zoom right into and walk around a shopping mall, or
       | train station or public building ? Our browsers can do this now,
       | essentially rendering quake like 3D environments - the problem is
       | with turning a scan into a lightweight 3D model.
       | 
       | Photogrammetry, where you have hundreds of photos and reconstruct
       | the 3D scene, uses a lot of compute, and the colmap / Structure-
       | from-Motion algorithm predates newer ML approaches and is ripe
       | for a better RL algorithm imo. Ive done experiments where you can
       | manually model a 3D scene from well positioned 360 panorama
       | photos of a building, picking corners, following the outline of
       | walls to make a floorplan etc ... this should be amenable to an
       | RL algorithm. Most 360 panorama photo tours have enough overlap
       | to reconstruct the scene reasonably well.
       | 
       | I have no doubt that we are on the brink of a massive improvement
       | in 3D processing. Its clearly solvable with the ML/RL approaches
       | we currently have .. we dont need AGI. My problem is getting
       | funding to work on it fulltime, equivalently talking an investor
       | into taking that bet :)
        
         | jgord wrote:
         | ps. its handy to compare the relative data sizes of [ models of
         | ] the same scene : typically for something like a house, the
         | data will be ballpark :                 -  15GB of pointcloud
         | data ( 100Mn xyzRGB points from a lidar laser scanner )       -
         | 3 GB of 360 panorama photos       -  50MB obj 3D textured model
         | -  2MB CAD model
         | 
         | Im guessing gaussian-splat would be something like 20x to 40x
         | more efficient than the pointcloud. I achieved similar
         | compression for building scans, using flat textured mini-
         | planes.
        
         | rallyforthesun wrote:
         | SplatAM is an interesting new way to generate 3D Gaussians in
         | real-time. It relies on RGB+D data and doesn't need COLMAP at
         | all. I am not related to it but am using it for a project with
         | a robot, as its main purpose is to do SLAM. As far as I
         | understand, it uses the point cloud for the alignment of the
         | images
         | 
         | edit:typo
        
         | KaiserPro wrote:
         | I have worked around spatial AI for a number of years.
         | 
         | Most of the stuff I have been working with has been aimed at
         | low power consumption. One of the things that really helped is
         | not bothering with dense reconstruction at all.
         | 
         | things like scenescript and SpaRP where instead of trying to
         | capture all the geometry (like photogrammetry) the essential
         | dimensions are captured and either outputted to a text
         | description (scene script) or a simple model with decent
         | normals (SpaRP)
         | 
         | Humans don't really keep complex dense reconstructions in our
         | head. Its all about spatial relationships of landmarks.
        
         | MITSardine wrote:
         | Have you tried "traditional" approaches like a Delaunay
         | triangulation on the point cloud, and how does your method
         | compare to that? Or did you encounter difficulties with that?
         | 
         | Regarding what you say of planes and compression, you can look
         | into metric-based surface remeshing. Essentially, you estimate
         | surface curvature (second derivatives) and use that to distort
         | length computations, remeshing your surface to length one in
         | that distorted space, which then yields optimal DoFs to surface
         | approximation error. A plane (or straight line) has 0 curvature
         | so lengths are infinite along it (hence final DoFs there
         | minimal). There's software to do that already, thought I'm not
         | sure it's robust to your usecase, because they've been
         | developed for scientific computing with meshes generated from
         | CAD (presumably smoother than your point cloud meshes).
         | 
         | I'd be really curious to know more about the type of workflow
         | you're interested in, i.e. what does your input look like (do
         | you use some open data sets as well?) and what you hope for in
         | the end (mesh, CAD).
        
       | signa11 wrote:
       | mr. yann-le-cunn's jepa paper is quite instructive.
        
       | IdealeZahlen wrote:
       | I've always wondered how spatial reasoning appears to be
       | operating quite differently from other cognitive abilities, with
       | significant individual variations. Some people effortlessly
       | parallel park while others struggle with these tasks despite
       | excelling at other forms of pattern recognition. What was
       | particularly intriguing for me is that some people with
       | aphantasia have no difficulty with spatial reasoning tasks, so
       | spatial reasoning may be distinct from reasoning based on
       | internal visualization.
        
         | polytely wrote:
         | my theory is that aphantasia is purely about conscious access
         | to visualizing not the existence of the ability to visualise.
         | 
         | I have aphantasia but I would say that spatial reasoning is one
         | of the things my brain is the best at
        
           | golol wrote:
           | How does one determine they have aphantasia? How do you know
           | that you are not doing exactly this thing people call
           | visualizing when you perform spatial reasoning?
        
         | m463 wrote:
         | I have had this idea about parking a car...
         | 
         | Most people have proprioception - you know where the parts of
         | your body are without looking. Close your eyes and you
         | intuitively know where your hands and fingers are.
         | 
         | When parking a car, it helps to sort of sit in the drivers seat
         | and look around the car. Turn your neck and look past the back
         | seat where your rear tire would be. sense the edges of the car.
         | 
         | I think if you sort of develop this a bit you might "feel"
         | where your car is intuitively when pulling into a parking space
         | or parallel parking. (car-prioception?)
         | 
         | (but use your mirrors and backup camera anyway)
        
       | myspeed wrote:
       | Most of our spatial intelligence is innate, developed through
       | evolution. We're born with a basic sense of gravity and the
       | ability to track objects. When we learn to drive a car, we simply
       | reassign these built-in skills to a new context
        
         | pzo wrote:
         | Is there any research about it ? This would mean we massing
         | some knowledge in genes and when offspring born have some
         | knowledge of our ancestors. This would mean the weights are
         | stored in DNA?
        
           | cma wrote:
           | Horses can be blindfolded at birth and when removed do basic
           | navigation with no time for any training. Other non-visually
           | precocious animals like cats, if they miss a critical
           | development period without getting natural vision data, will
           | never develop a functioning visual system.
           | 
           | Baby chicks can do bipedal balance pretty much as soon as
           | they dry off.
           | 
           | Wood ducks can visually imprint very soon after hatching and
           | drying off, a couple hours after birth with very limited
           | visual data up until then and no interspersed sleep cycles.
           | 
           | We as humans have natural reactions to snake like shapes etc.
           | even before encountering the danger of them or learning about
           | it from social cues. Babies
        
             | magicalhippo wrote:
             | I've pondered this often, especially kangaroos where the
             | half-developed fetus can climb up into the pouch.
             | 
             | Clearly we're just hardwired for certain tasks, in such a
             | way that the function is primarily dictated by topology.
             | 
             | This weight agnostic neural network page[1] explores this,
             | but obviously isn't the true answer.
             | 
             | [1]: https://weightagnostic.github.io/
        
             | jampekka wrote:
             | It's not clear whether humans have natural reactions to
             | snakes. https://link.springer.com/article/10.11133/j.tpr.20
             | 13.63.4.0...
        
               | cma wrote:
               | If not, another visual system one in newborns is
               | preference for faces with open eyes and direct gaze:
               | 
               | https://pubmed.ncbi.nlm.nih.gov/17030037/
        
       | wolframhempel wrote:
       | we're actually working on a practical implementation of aspects
       | of what Fei-Fei describes - although with a more narrow focus on
       | optimizing operations in the physical space (mining, energy,
       | defense etc) https://hivekit.io/about/our-vision/
        
       | ldenoue wrote:
       | Full playable transcript https://www.appblit.com/scribe?v=_PioN-
       | CpOP0
        
       | sabman wrote:
       | We've been working on this challenge in the satellite domain with
       | https://earthgpt.app. It's a subset of what Fei-Fei is
       | describing, but comes with its own unique issues like handling
       | multi-resolution sensors and imagery with hundreds of spectral
       | bands. Think of it as computer vision, but in n-dimensions.
       | 
       | Happy to answer questions if you're curious. PS. still in early
       | beta, so please be gentle!
        
         | fnands wrote:
         | Hey, cool project!
         | 
         | Do you actually pass the images to the model, or just the
         | metadata/stats?
        
           | sabman wrote:
           | Thanks! This live demo uses metadata and stats only. Right
           | now we are testing ViTs and Foundation Models as well. But
           | quality of results from EO FMs haven't been worth the
           | inference cost so far. Early days though. Also starting to
           | fine tune models for specific downstream tasks ourselves.
        
             | fnands wrote:
             | Cool, makes sense.
             | 
             | Yeah, have you considered maybe looking into just running
             | it on embeddings [1], instead of the imagery itself? Would
             | save on most of the inference cost, at the cost of
             | flexibility (i.e. you are locked into whatever embeddings
             | have been created).
             | 
             | [1] https://developers.google.com/earth-
             | engine/datasets/catalog/...
        
         | byteab wrote:
         | Really interesting space
        
       | District5524 wrote:
       | An immaterial side note: funny how obsessed she seems to be with
       | her age. She said once that people in the audience could be half
       | or even third of her age. Given that she's 49, is it really
       | typical that 16-year olds attend these fireside YC chats?
        
         | defrost wrote:
         | Possible, yes .. which validates her statement.
         | 
         | Typical? Probably not, but hardly relevant to the truthiness of
         | the claim.
        
         | mistersquid wrote:
         | > An immaterial side note: funny how obsessed she seems to be
         | with her age.
         | 
         | Given her intellectual stature, Professor Li likely was one of
         | the strongest minds in any room she found herself in and, for
         | the first half of her life, also one of the youngest voices.
         | 
         | Now that she's entering mid-life, she's still one of the most
         | powerful minds, but no longer one of the youngest.
         | 
         | It's something middle-aged thinkers can't help but notice.
         | 
         | For the rest of us, we can only be grateful to share space and
         | time with such gifted thinkers.
         | 
         | Coincidentally, today is Professor Li's birthday! [0] I hope I
         | will be around to see many more 3rds of July.
         | 
         | [0] Maybe her coming birthday was on her mind, hence the
         | frequency of her remarks about her relative age.
        
           | dopadelic wrote:
           | It's interesting how figures get idolized.
           | 
           | Fei-Fei Li is known for the creation of ImageNet, which is
           | certainly transformative in the field of computer vision. But
           | the crux of it is painstaking grunt work to create the vast
           | labeled dataset. Fei-Fei Li is a leader who mobilized vast
           | resources and people hours to create this vast dataset.
           | Certainly worth a ton of acclaim. But to claim she's the most
           | brilliant mind in an entire room is a stretch.
        
         | gk1 wrote:
         | I only noticed a few times when she brought up age, and all of
         | them were natural and appropriate in the context. Eg she was
         | asked how old she was when she opened the dry-cleaning
         | business.
         | 
         | Also, there's this weird thing in culture (is it US only?) that
         | whenever an interviewer brings up (even implicitly) the guest's
         | age, the guest has to make some quip about it as if they're
         | offended or sensitive to it. So I wouldn't interpret even a
         | slightly defensive comment about age as an "obsession."
        
       | cainxinth wrote:
       | There are many such frontiers in AI. I was just reading that
       | current models are apparently quite bad with temporal perception:
       | 
       | https://community.openai.com/t/time-awareness-in-ai-why-temp...
       | 
       | https://boraerbasoglu.medium.com/the-impact-of-ais-lack-of-t...
        
       | moktonar wrote:
       | Intelligence is not only embodied (it needs a body), it is also
       | embedded in the environment (it needs the environment). If you
       | want an intelligence in your computer, you need an environment in
       | your computer first, as the substrate from which the intelligence
       | will evolve. The more accurate the environment the better the
       | intelligence that will be obtained. The universe is able to
       | create intelligence and we are proof. Thus, if you want to create
       | intelligence, you have to find a way of efficiently simulate our
       | reality at the desired level of detail. Currently we don't know
       | such efficient algorithm, but one way could be finally harnessing
       | Quantum Computing to hack the universe itself, cheat and be able
       | to simulate our environment efficiently without even knowing the
       | algorithm behind Quantum Physics.
        
         | Workaccount2 wrote:
         | The silicon exists in the same environment that our brains do.
        
           | rtaylorgarlock wrote:
           | The drivers of its failure to adequately assimilate to said
           | environment being?
        
           | moktonar wrote:
           | Yeah, and we represent the evolutionary force, but that means
           | that the ability to craft silicon life depends on our ability
           | to find an efficient algorithm to do so...
        
         | dcreater wrote:
         | So your basically saying we need Dolores and Westworld
        
         | pixl97 wrote:
         | >Intelligence is not only embodied (it needs a body), it is
         | also embedded in the environment (it needs the environment).
         | 
         | Of course don't make the mistake that we need anything like a
         | human body, or any singular object containing 'intelligence'.
         | That's simply the way nature had to do it to connect a sensor
         | platform to a brain. AI seems much more like it will be a hive
         | mind and distributed system of data collection.
        
       | yellow_postit wrote:
       | recent paper on " How Well Does GPT-4o Understand Vision?
       | Evaluating Multimodal Foundation Models on Standard Computer
       | Vision Tasks" [1]
       | 
       | [1] https://arxiv.org/abs/2507.01955
        
       | Nurbek-F wrote:
       | Isn't it what Karpathy has been advocating for since the early
       | days of Tesla Vision?
        
       | ninetyninenine wrote:
       | The next frontier is eliminating hallucinations?
       | 
       | Once that happens it's all over.
        
       | owenpalmer wrote:
       | While spatial intelligence is certainly a major limitation of
       | current AI systems, I have been able to get LLMs to do quite
       | impressive things.
       | 
       | Here's an on the fly video I made (no retakes) of Claude
       | generating a Godot scene file.
       | 
       | https://youtu.be/2gARJpDG7Jo?si=W4rlISO-J4EPJYyG
        
         | AStrangeMorrow wrote:
         | Yeah funnily I did a project where we had an LLM based
         | interface to in house 3D parametric modeling system and it did
         | fairly well.
         | 
         | However I've been trying to use LLM, both as orchestrators and
         | in other cases to write code for 2D optimization problems with
         | many spatial relationships and it has done terribly.
         | 
         | I have talking it can generate 1000s of lines over many rounds
         | of prompting/iteration that solve maybe 30% of the problem (and
         | the 30% very easy cases) while completely messing up the rest.
         | When doing that code myself, in less than 1000 lines, the "30%
         | part" was maybe 3% of the total code. Even when basically
         | providing pseudo code to solve specific part of the problem
         | chances are these LLM solutions would also have many blind
         | spots or issues.
         | 
         | The thing is, that is a 2D problem for which there basically no
         | ressources about online, and all the slightly similar problems
         | all have careful handcrafted specialized solutions. So I think
         | it has no good frame of reference how to solve the problem
        
       | czbond wrote:
       | Thanks for joining the obvious Fei-Fei about 5 years late.
       | Spatial web standards approved by IEEE that have been in the
       | works for years.
       | 
       | https://spatialwebfoundation.org/
        
         | hiddencost wrote:
         | You do know who she is right?
        
           | czbond wrote:
           | Of course.... doesn't mean she was early in forming that
           | viewpoint. Just stating the near future obvious
        
         | sota_pop wrote:
         | Wow, I watched a presentation of this idea (2) years ago. It
         | reads like classic flavor-of-the-week jargon-soup engineered to
         | be catnip for unsuspecting VC; ie combining buzzwords HTTP,
         | IoT, AI, blockchain, and the notion of "digital twin" from the
         | AEC industry. Given by a guy who seemed extremely excited (and
         | heavily energized - possibly chemically). The presenter tried
         | to describe how this differs from HTTP. I'm highly confident no
         | one in the room was able to make anything of it.
         | 
         | Before any questions could be asked, the presenter said "OK, I
         | need to run to give this presentation at the World Economic
         | Forum in Davos now.", and quite literally ran out of the room.
        
       | alganet wrote:
       | How can I be sure that spatial intelligence AIs will not be just
       | intricate sensoring that ultimately fails to demonstrate actual
       | intelligence?
       | 
       | > "trilobite"
       | 
       | The trilobite ancestor had a nervous system before it had an eye.
       | It was able to make decisions and interact with the environment
       | before the ability to see or speak a language.
       | 
       | It feels to me like this basic step is still missing. We haven't
       | even crossed the first AI frontier yet.
        
       | mehulashah wrote:
       | "Forget about what you've done in the past. Forget about what
       | others think of you. Just hunker down and build. That is my
       | comfort zone."
       | 
       | Enough said.
        
       ___________________________________________________________________
       (page generated 2025-07-03 23:00 UTC)