[HN Gopher] Continuous Thought Machines
       ___________________________________________________________________
        
       Continuous Thought Machines
        
       Author : hardmaru
       Score  : 278 points
       Date   : 2025-05-12 02:21 UTC (20 hours ago)
        
 (HTM) web link (pub.sakana.ai)
 (TXT) w3m dump (pub.sakana.ai)
        
       | robwwilliams wrote:
       | Great to refocus in this important topic. So cool to see this
       | bridge being built across fields.
       | 
       | In wet-ware it is hard not to think of "time" as linear Newtonian
       | time driven by a clock. But in the cintext of brain- and-body
       | what really is critical is generating well ordered sequences of
       | acts and operations that are embedded in thicker or thinner sluce
       | of "now" that can range from 300 msec of the "specious present"
       | to 50 microseconds in cells that evaluate the sources of sound
       | (the medial superior olivary nucleus).
       | 
       | For more context on contingent temporality see interview with RW
       | Williams in this recent publication in The European Journal of
       | Neuroscience by John Bickle:
       | 
       | https://pubmed.ncbi.nlm.nih.gov/40176364/
        
       | ttoinou wrote:
       | Ironically this webpage continuously refreshes itself on my
       | firefox iOS :P
        
         | tonyhart7 wrote:
         | it literally never load for me
        
       | rvz wrote:
       | > The Continuous Thought Machine (CTM) is a neural network
       | architecture that enables a novel approach to thinking about
       | data. It departs from conventional feed-forward models by
       | explicitly incorporating the concept of Neural Dynamics as the
       | central component to its functionality.
       | 
       | Still going through the paper, But this looks very exciting to
       | actually _see_ , the internal visual recurrence in action when
       | confronting a task (such as the 2D Puzzle) - making it easier to
       | interpret neural networks over several tasks involving 'time'.
       | 
       | (This internal recurrence may not be new, but applying neural
       | synchronization as described in this paper is).
       | 
       | > Indeed, we observe the emergence of interpretable and intuitive
       | problem-solving strategies, suggesting that leveraging neural
       | timing can lead to more emergent benefits and potentially more
       | effective AI systems
       | 
       | Exactly. Would like to see more applications of this in existing
       | or new architectures that can also give us additional
       | transparency into the thought process on many tasks.
       | 
       | Another great paper from Sakana.
        
         | omneity wrote:
         | Is it the same Sakana from the cheating AI coder tribulations?
         | There were some fundamental mistakes in that work that made me
         | question the team.
         | 
         | https://www.hackster.io/news/sakana-ai-claims-its-ai-cuda-en...
         | 
         | https://techcrunch.com/2025/02/21/sakana-walks-back-claims-t...
        
           | doall wrote:
           | They admitted, apologized, and are in the process of revising
           | the paper. Mistakes always happen whether small or big. What
           | is more important is to be transparent, learn from it, and
           | make sure the same mistake doesn't happen again.
        
       | coolcase wrote:
       | I love the ML diagrams that hybrid maths and architecture. It is
       | much less dry than all formal math.
        
       | dcrimp wrote:
       | I'm quite enthusiastic about reading this. Since watching the
       | progress by the larger LLM labs, I've noted that they're not
       | making material changes in model configuration that I think to be
       | necessary to proceed toward more refined and capable
       | intelligence. They're adding tools and widgets to things we know
       | don't think like a biological brain. These are really useful
       | things from a commercial perspective, but I think LLMs won't be
       | an enduring paradigm, at least wrt genuine stabs at artificial
       | intelligence. I've been surprised that there hasn't been more
       | effort to transformative work like in the linked article.
       | 
       | The two things that hang me up on current progress in
       | intelligence is that:
       | 
       | - there don't seem to be models which possess continuous thought.
       | Models are alive during a forward pass on their way to produce a
       | token and brain-dead any other time - there don't seem to be many
       | models that have neural memory - there doesn't seem to be any
       | form of continuous learning. To be fair, the whole online
       | training thing is pretty uncommon as I understand it.
       | 
       | Reasoning in token space is handy for evals, but is lossy - you
       | throw away all the rest of the info when you sample. I think Meta
       | had a paper on continuous thought in latent space, but I don't
       | think effort in that has continued to anything commercialised.
       | 
       | Somehow, our biological brains are capable of super efficiently
       | doing very intelligent stuff. We have a known-good example, but
       | research toward mimicking that example is weirdly lacking?
       | 
       | All the magic happens in the neural net, right? But we keep
       | wrapping nets with tools we've designed with our own inductive
       | biases, rather than expanding the horizon of what a net can do
       | and empowering it to do that.
       | 
       | Recently I've been looking into SNNs, which feel like a bit of a
       | tech demo, as well as neuromorphic computing, which I think holds
       | some promise for this sort of thing, but doesn't get much press
       | (or, presumably, budget?)
       | 
       | (Apologies for ramble, writing on my phone)
        
       | liamwire wrote:
       | Seems really interesting, and the in-browser demo and model was a
       | really great hook to get interest in the rest of the research.
       | I'm only partially through it but the idea itself is compelling.
        
       | erewhile wrote:
       | The ideas of these machines isn't entirely new. There's some
       | research from 2002, where Liquid State Machines (LSM) are
       | introduced[1]. These are networks that generally rely on
       | continuous inputs into spiking neural networks, which are then
       | read by some dense layer that connects to all the neurons in this
       | network to read what is called the liquid state.
       | 
       | These LSMs have also been used for other tasks, like playing
       | Atari games in a paper from 2019[2], where they show that while
       | sometimes these networks can outperform humans, they don't
       | always, and they tend to fail at the same things more
       | conventional neural networks failed at at the time as well. They
       | don't outperform these conventional networks, though.
       | 
       | Honestly, I'd be excited to see more research going into
       | continuous processing of inputs (e.g., audio) with continuous
       | outputs, and training full spiking neural networks based on
       | neurons on that idea. We understand some of the ideas of
       | plasticity, and they have been applied in this kind of research,
       | but I'm not aware of anyone creating networks like this with just
       | the kinds of plasticity we see in the brain, with no back
       | propagation or similar algorithms. I've tried this myself, but I
       | think I either have a misunderstanding of how things work in our
       | brains, or we just don't have the full picture yet.
       | 
       | [1] doi.org/10.1162/089976602760407955 [2]
       | doi.org/10.3389/fnins.2019.00883
        
       | AIorNot wrote:
       | Can someone explain this paper in the context of LLM
       | architectures - it seems this cannot be combined with LLM deep
       | learning or can it?
        
       | davedx wrote:
       | So this weekend we have:
       | 
       | - Continuous thought machines: temporally encoding neural
       | networks (more like how biological brains work)
       | 
       | - Zero data reasoning: (coding) AI that learns from doing,
       | instead of by being trained on giant data sets
       | 
       | - Intellect-2: a globally distributed RL architecture
       | 
       | I am not an expert in the field but this feels like we just bunny
       | hopped a little closer to the singularity...
        
         | spiderfarmer wrote:
         | Also not an expert, but I think this is like saying robots will
         | dominate the world because we invented camera's, actuators and
         | batteries.
         | 
         | In other words, baby steps, not bunny hops.
        
           | TeMPOraL wrote:
           | In some sense, they did. The world may not be full of human-
           | like robots autonomously roaming the wastelands^Wurban
           | landscape - but it's chock-full of actuators, sensors and
           | batteries. There are sensors and actuators in your coffee
           | maker. There are plenty of them in your car too, whether they
           | control the wheels, or the angle of the mirrors, or the
           | height of the windows, or the state of your door knob. Etc.
           | And all of those robotic parts were mostly made by... larger
           | robots in factories.
        
         | dgfl wrote:
         | Don't give too much importance to individual papers. At best,
         | you're mostly disregarding all the work that lead to it. At
         | worst, you're placing a lot of faith on an idea analyzed
         | through rose-tinted glasses and presented with a lot of
         | intentional omissions.
        
           | davedx wrote:
           | I mean the Zero Data reasoning directly cites previous work
           | on the same principle. That in particular seems like a
           | significant step forwards though - one of the main critiques
           | of current methods is "humans don't learn by ingesting
           | terabytes of common crawl, they learn from experience".
        
         | gessha wrote:
         | But when you try to run their code or use the product, it's
         | either missing or doesn't perform as well as marketed in the
         | paper. Personal recommendation for building mental resistance
         | against AI hype is to:
         | 
         | - read the paper and the concrete claims, results and
         | limitations
         | 
         | - download and run the code whenever possible
         | 
         | - test for out of distribution inputs and/or practical examples
         | outside of the training set
        
         | chrsw wrote:
         | I don't get that feeling. There are so many papers and so many
         | avenues of research. It's extremely difficult for me to predict
         | what's going to "pop" like the diffusion paper, the transformer
         | paper, AlphaZero, Chat GPT-3, etc. But even these research or
         | product advances that seem like step functions are built on a
         | lot of research and trial-and-error. Can all 3 of these you
         | listed be combined somehow? Hopefully, but I have no idea.
        
         | aaroninsf wrote:
         | I'll take a moment to comment on this, after reading the
         | responses which challenge your conclusion.
         | 
         | This criticism is entirely justified for a narrow read of your
         | point, that the _specific_ and _relatively-widely-disseminated_
         | papers /projects, themselves, represent specific progress
         | towards e.g. take-off or AGI or SI.
         | 
         | But it's also unjustified to the extent that these particular
         | papers are proxies for broader research directions--indeed,
         | many of the other comments provide reading lists for related
         | and prior work.
         | 
         | I.e. it's not that this or that particular paper is the hop.
         | It's that the bunny is oriented in the right direction and many
         | microhops are occurring. What one chooses to label a hop amid
         | the aggregate twitches and movement is a question for pedants.
         | 
         | Meanwhile the bunny might be moving.
        
         | aDyslecticCrow wrote:
         | Both Intellect-2 and zero data reasoning work on LLMs ("Zero
         | data reasoning" is quite a misleading name of a method. It's
         | not very ground-breaking.) If you wanna see a major leap in
         | LLMS, you should check out what InceptionLabs did recently to
         | speed up inference by 16x using a diffusion model.
         | (https://www.inceptionlabs.ai/)
         | 
         | Our algorithms for time-series reinforcement learning are
         | abysmal compared to inference models. Despite the explosion of
         | the AI field, robotics and self-driving are stuck without much
         | progress.
         | 
         | I think this method has potential, but someone else needs to
         | boil it down a bit and change the terminology because, despite
         | the effort, this is not an easily digested article.
         | 
         | We're also nowhere close to getting these models to behave
         | properly. The larger the model we make, the more likely it is
         | to find loopholes in our reward functions. This holds us back
         | from useful AI in a lot of domains.
        
       | iandanforth wrote:
       | This paper is concerning. While divorced from the standard ML
       | literature there is a _lot_ of work on biologically plausible
       | spiking, timing dependant artificial neutral networks. The
       | nomenclature here doesn 't seem to acknowledge that body of work.
       | Instead it appears as a step toward that bulk of research coming
       | from the ML/LLM field without a clear appreciation of the ground
       | well traveled there.*
       | 
       | In addition some of the terminology is likely to cause confusion.
       | By calling a synaptic integration step "thinking" the authors are
       | going to confuse a lot of people. Instead of the process of
       | forming an idea, evaluating that idea, potentially modifying it
       | and repeating (what a layman would call thinking) they are trying
       | to ascribe "thinking" to single unit processes! That's a pretty
       | radical departure from both ML and ANN literature. Pattern
       | recognition/signal discrimination is well known at the level of
       | synaptic integration and firing, but "thinking?" No, that wording
       | is not helpful.
       | 
       | *I have not reviewed all the citations and am reacting to the
       | plain language of the text as someone familiar with both lines of
       | research.
        
         | cepera wrote:
         | >there is a lot of work on biologically plausible spiking
         | 
         | I ask you kindly to share the list (or even better brief
         | review) of most insightful books/papers in your opinion with
         | neuroscience inspired algorithms concepts/implementation
         | details.
        
           | erewhile wrote:
           | Not the original poster, but:
           | 
           | - Theoretical Neuroscience Computational and Mathematical
           | Modeling of Neural Systems - Peter Dayan, L. F. Abbott (2001)
           | is quite good, more mathematical than computational.
           | 
           | - Neuronal dynamics, available here:
           | https://neuronaldynamics.epfl.ch/ is also quite good, and
           | free to read. Has python exercises as well. If I recall
           | correctly, it mostly goes into simulations of singular
           | neurons, and not so much entire networks and what we can do
           | with them, but it does a good job at bridging the chemistry /
           | biology / math to computation.
           | 
           | If we're talking about papers, one I mentioned in my other
           | comment:
           | 
           | - Real-Time Computing Without Stable States: A New Framework
           | for Neural Computation Based on Perturbations,
           | https://doi.org/10.1162/089976602760407955
           | 
           | - Dynamics of Sparsely Connected Networks of Excitatory and
           | Inhibitory Spiking Neurons, by Nicolas Brunel (Don't have a
           | DOI on hand for this one)
           | 
           | - Spiking Neural Networks and Their Applications: A Review,
           | https://doi.org/10.3390/brainsci12070863 , is a very nice
           | review of methods and does some nice explaining on concepts.
           | 
           | If you're looking for keywords on the topic:
           | 
           | - Leaky Integrate and Fire (LIF) neurons
           | 
           | - Spiking neural networks
           | 
           | - Liquid State Machines (LSM)
           | 
           | - Synaptic plasticity (Models of synaptic plasticity)
           | 
           | - Spike-based synaptic plasticity
        
           | rkp8000 wrote:
           | A (non-exhaustive) list of some notable papers:
           | 
           | Maass 2002, Real-time computing without stable states:
           | https://pubmed.ncbi.nlm.nih.gov/12433288/
           | 
           | Sussillo & Abbott 2009, Generating Coherent Patterns of
           | Activity from Chaotic Neural Networks
           | https://pmc.ncbi.nlm.nih.gov/articles/PMC2756108/
           | 
           | Abbott et al 2016, Building functional networks of spiking
           | model neurons https://pubmed.ncbi.nlm.nih.gov/26906501/
           | 
           | Zenke & Ganguli 2018, SuperSpike: Supervised Learning in
           | Multilayer Spiking Neural Networks https://ganguli-
           | gang.stanford.edu/pdf/17.superspike.pdf
           | 
           | Bellec et al 2020, A solution to the learning dilemma for
           | recurrent networks of spiking neurons
           | https://www.nature.com/articles/s41467-020-17236-y
           | 
           | Payeur et al 2021, Burst-dependent synaptic plasticity can
           | coordinate learning in hierarchical circuits
           | https://www.nature.com/articles/s41593-021-00857-x
           | 
           | Cimesa et al 2023, Geometry of population activity in spiking
           | networks with low-rank structure https://journals.plos.org/pl
           | oscompbiol/article?id=10.1371/jo...
           | 
           | Ororbia 2024, Contrastive signal-dependent plasticity: Self-
           | supervised learning in spiking neural circuits
           | https://www.science.org/doi/10.1126/sciadv.adn6076
           | Kudithipudi et al 2025, Neuromorphic computing at scale
           | (review) https://www.nature.com/articles/s41586-024-08253-8
        
         | tiahura wrote:
         | The authors don't label a single synaptic integration as
         | "thinking." They use the term for the network-wide internal
         | loop ("internal ticks") that unrolls after every external
         | input, and explicitly say it is merely "analogous to thought."
        
         | vonneumannstan wrote:
         | Was this written by Jurgen Schmidhuber?
        
         | mountainriver wrote:
         | Agree, they are presenting this like its a new idea without
         | hardly any reference to the decades of work on spiking neural
         | nets and similar.
        
         | TeMPOraL wrote:
         | I'm sort of not surprised; my impression is that, for the past
         | decade or two, ML researchers who did acknowledge related work
         | in neuroscience were broadly accused of hubris for daring to
         | compare their work to biological brains.
        
         | program_whiz wrote:
         | Sorry I should have responded to this comment, but I wrote a
         | separate response in the parent thread. I didn't feel the pdf /
         | paper was really trying to mimick spiking biological networks
         | in all but the loosest sense (there is a sequence of
         | activations and layers of "neurons"). I think the major
         | contribution is just using the dot product on output transpose
         | output, the rest is just diffusion / attention on inputs. Its
         | conceptually a combination of "input attention" and "output
         | attention" using a kind of stepped recursive model.
        
       | bob1029 wrote:
       | > Emulating these mechanisms, particularly the temporal coding
       | inherent in spike timing and synchrony, presents a significant
       | challenge. Consequently, modern neural networks do not rely on
       | temporal dynamics to perform compute, but rather prioritize
       | simplicity and computational efficiency.
       | 
       | Simulating a proper time domain is a very difficult thing to do
       | with practical hardware. It's not that we can't do it - it's that
       | all this timing magic requires additional hyperparameter
       | dimensions that need to be searched over. Finding a set of valid
       | parameters when the space is this vast seems very unlikely. You
       | want to eliminate parameters, not introduce ones.
       | 
       | Also, computational substrates that are efficient to execute can
       | be searched over much more quickly than those that are not.
       | Anything where we need to model a spike that is delivered at a
       | future time immediately chops a few orders of magnitude off the
       | top because you have to keep things like priority queue
       | structures around to serialize events.
       | 
       | Unless hard real time interaction is an actual design goal, I
       | don't know if chasing this rabbit is worth it on the
       | engineering/product side.
       | 
       | The elegance of STDP and how it could enable online, unsupervised
       | learning is still highly alluring to me. I just don't see a path
       | with silicon right now or on the horizon. Purpose built hardware
       | could work but is like taking a really big leap of faith by
       | setting some of the hyperparameters to const in code. The chances
       | of getting this right before running out of money seem low to me.
        
         | angusturner wrote:
         | Hm suppose for argument sake that feeding a batch of data
         | through some moderately large FF architectures takes on the
         | order of 100ms (I realise this depends on a lot parameters -
         | but this seems reasonable for many tasks / networks).
         | 
         | Now suppose instead you have an CTM that allocates 10ms on the
         | standard FF axes, and then multiplies it out by 10 internal
         | "ticks" / recurrent steps?
         | 
         | The exact numbers are contrived, but my point is : couldn't we
         | conceivably search over that second arch just as easily?
         | 
         | It just boils down to whether the inductive bias of building in
         | some explicit time axis is actually worthwhile, right ?
        
       | swalsh wrote:
       | To me, they key to the next generation of models needs to be
       | neurons that fire together wire together. I think spiking neural
       | networks offer an exciting alternative approach.
        
       | program_whiz wrote:
       | In my reading of the paper, I don't feel this is really like
       | biological / spiking networks at all. They keep a running history
       | of inputs and use multi-headed attention to form an internal
       | model of how the past "pre-synaptic" inputs factor into the
       | current output (post-synaptic). this is just like a modified
       | transformer (keep history of inputs, use attention on them to
       | form an output).
       | 
       | Then the "synchronization" is just using an inner product of all
       | the post activations (stored in a large ever-growing list and
       | using subsampling for performance reasons).
       | 
       | But its still being optimized by gradient descent, except the
       | time step at which the loss is applied is chosen to be the time
       | step with minimum loss, or minimum uncertainty (uncertainty being
       | described by the data entropy of the output term).
       | 
       | I'm not sure where people are reading that this is in any way
       | similar to spiking neuron models with time simulation (time is
       | just the number of steps the data is cycled through the system,
       | similar to diffusion model or how LLM processes tokens
       | recursively).
       | 
       | The "neuron synchronization" is also a bit different from how its
       | meant in biological terms. Its using an inner product of the
       | output terms (producing a square matrix), which is then projected
       | into the output space/dimensions. I suppose this produces
       | "synchronization" in the sense that to produce the right answer,
       | different outputs that are being multiplied together must produce
       | the right value on the right timestep. It feels a bit like
       | introducing sparsity (where the nature of combining many outputs
       | into a larger matrix makes their combination more important than
       | the individual values). The fact that they must correctly combine
       | on each time step is what they are calling "synchronization".
       | 
       | Techniques like this are the basic the mechanism underlying
       | attention (produce one or more outputs from multiple subsystems,
       | dot product to combine).
        
         | program_whiz wrote:
         | I would say one weakness of the paper is that they primarily
         | compare performance with LSTM (a simpler recursion model),
         | rather than similar attention / diffusion models. I would be
         | curious how well a model that just has N layers of attention
         | in/out would perform on these tasks (using a recursive time-
         | stepped approach). My guess is performance will be very
         | similar, and network architecture will also be quite similar
         | (although a true transformer is a bit different than input
         | attention + unet which they employ).
        
       ___________________________________________________________________
       (page generated 2025-05-12 23:01 UTC)