The Dragon Hatchling: the Missing Link Between the Transformer and Brain Models
Posted2 months agoActive2 months ago
arxiv.orgTechstoryHigh profile
skepticalmixed
Debate
80/100
AINeuromorphic ComputingTransformer Architecture
Key topics
AI
Neuromorphic Computing
Transformer Architecture
A new paper proposes the 'Dragon Hatchling' model, a biologically-inspired AI architecture that rivals GPT-2 performance, sparking debate among HN users about its innovation, scalability, and potential.
Snapshot generated from the HN discussion
Discussion Activity
Very active discussionFirst comment
46m
Peak period
94
0-12h
Avg / period
16.5
Comment distribution99 data points
Loading chart...
Based on 99 loaded comments
Key moments
- 01Story posted
Oct 22, 2025 at 9:00 AM EDT
2 months ago
Step 01 - 02First comment
Oct 22, 2025 at 9:45 AM EDT
46m after posting
Step 02 - 03Peak activity
94 comments in 0-12h
Hottest window of the conversation
Step 03 - 04Latest activity
Oct 28, 2025 at 11:29 AM EDT
2 months ago
Step 04
Generating AI Summary...
Analyzing up to 500 comments to identify key contributors and discussion patterns
ID: 45668408Type: storyLast synced: 11/20/2025, 4:29:25 PM
Want the full context?
Jump to the original sources
Read the primary article or dive into the live Hacker News thread when you're ready.
I think that it is necessary we try all things because LLMs as we know them take too much energy.
[0] https://en.wikipedia.org/wiki/Passenger_drone
- It seems strange to make use of the term "scale-free" and then defer a definition until half way through the paper (in fact, the term is mentioned 3 times after, and 14 times before said definition)
- This might just be CS people doing CS things, but the notation in the paper is awful: Claims/Observations end with a QED-symbol (for example on pages 29 and 30) but without a proof
- They make strong claims about performance and scaling ("It exhibits Transformer-like scaling laws") but the only (i think?) benchmark is a translation task comparison with <1B models, ,which is ~2 orders of magnitude smaller than sota
Author comment: as a fairly common convention, QED immediately after a particular statement means that the statement should be considered proven. Depending on the text, this may either be because the statement (Observation) is self-explanatory, or, the discussion in the text leading up to the statement is sufficient, or, whenever the final statement of a Theorem follows as a direct corollary of Lemmas previously proven in the text.
> In addition to being a graph model, BDH admits a GPU-friendly formulation.
I remember about two years ago people spotting that if you just moved a lot of weights through a sigmoid and reduced floats down to -1, 0 or 1, we barely lost any performance from a lot of LLM models, but suddenly opened up the ability to use multi-core CPUs which are obviously a lot cheaper and more power efficient. And yet, nothing seems to have moved forward there yet.
I'd love to see new approaches that explicitly don't "admit a GPU-friendly formulation", but still move the SOTA forward. Has anyone seen anything even getting close, anywhere?
> It exhibits Transformer-like scaling laws: empirically BDH rivals GPT2 performance on language and translation tasks, at the same number of parameters (10M to 1B), for the same training data.
That is disappointing. It needs to do better, in some dimension, to get investment, and I do think alternative approaches are needed now.
From the paper though there are some encouraging side benefits to this approach:
> [...] a desirable form of locality: important data is located just next to the sites at which it is being processed. This minimizes communication, and eliminates the most painful of all bottlenecks for reasoning models during inference: memory-to-core bandwidth.
> Faster model iteration. During training and inference alike, BDH-GPU provides insight into parameter and state spaces of the model which allows for easy and direct evaluation of model health and performance [...]
> Direct explainability of model state. Elements of state of BDH-GPU are directly localized at neuron pairs, allowing for a micro-interpretation of the hidden state of the model. [...]
> New opportunities for ‘model surgery’. The BDH-GPU architecture is, in principle, amenable to direct composability of model weights in a way resemblant of composability of programs [...]
These, to my pretty "lay" eyes look like attractive features to have. The question I have is whether the existing transformer based approach is now "too big to fail" in the eyes of people who make the investment calls, and whether this will get the work it needs to get it from GPT2 performance to GPT5+.
The speedup from using a GPU over a CPU is around 100x, as a rule of thumb. And there's been an immense amount of work maximizing throughput when training on a pile of GPUs together... And a sota model will still take a long time to train. So even if you do have a non-GPU algo which is better, it'll take you a very very long time to train it - by which point the best GPU algos will have also improved substantially.
I would trust this paper far more if it didn’t trip my “crank” radar so aggressively
Red flags from me:
- “Biologically-inspired,” claiming that this method works just like the brain and therefore should inherit the brain’s capabilities
- Calling their method “B. Dragon Hatchling” without explaining why, or what the B stands for, and not mentioning this again past page 2
- Saying all activations are “sparse and positive”? I get why being sparse could be desirable, but why is all positive desirable?
These are stylistic critiques and not substantive. All of these things could be “stressed grad student under intense pressure to get a tech report out the door” syndrome. But a simpler explanation is that this paper just lacks insight
Your words, not the author's. They did not make this claim.
A clarification is thus due. As indicated in the github repo: "BDH" stands for "Baby Dragon: Hatchling". Technically, "Hatchling" would perhaps be the version number.
For readers who find this discussion point interesting, I recommend considering for context: 1. The hardware naming patterns in the Hopper/Gracehopper architecture. 2. The attitude to acronyms taken in the CS Theory vs. Applied CS/AI community.
-Linear (transformer) complexity at training time
-Linear scaling with number of tokens
-Online learning(!!!)
The main point that made me cautiously optimistic:
-Empirical results on par with GPT-2
I think this is one of those ideas that needs to be tested with scaled up experiments sooner rather than later, but someone with budget needs to commit. Would love to see HuggingFace do a collab and throw a bit of $$$ at it with a hardware sponsor like Nvidia.
In fact, we've sometimes seen new companies show up with models based on research big companies didn't use, the new models are useful or better in some way, and people use them or big companies acquire them. I'd say that's proof big companies miss a lot of good ideas internally.
I think it's fairly safe to say that every remotely promising thing that showed up in the papers was tried at some big lab at least once. If it showed good results, they'd pick it up.
I love when a new architecture comes out, but come on, it's 2025, can we please stop comparing fancy new architectures to the antiquated GPT2? This makes the comparison practically, well, useless. Please pick something more modern! Even the at-this-point ubiquitous Llama would be a lot better. I don't want to have to spend days of my time doing my own benchmarks to see how it actually compares to a modern transformer (and realistically, I was burned so many times now that I just stopped bothering).
Modern LLMs are very similar to GPT2, but those architectural tweaks do matter and can make a big difference. For example, take a look at the NanoGPT speedrun[1] and look at how many training speedups they got by tweaking the architecture.
Honestly, everyone who publishes a paper in this space should read [2]. The post talks about optimizers, but this is also relevant to new architectures too. Here's the relevant quote:
> With billions of dollars being spent on neural network training by an industry hungry for ways to reduce that cost, we can infer that the fault lies with the research community rather than the potential adopters. That is, something is going wrong with the research. Upon close inspection of individual papers, one finds that the most common culprit is bad baselines [...]
> I would like to note that the publication of new methods which claim huge improvements but fail to replicate / live up to the hype is not a victimless crime, because it wastes the time, money, and morale of a large number of individual researchers and small labs who run and are disappointed by failed attempts to replicate and build on such methods every day.
Sure, a brand new architecture is most likely not going to compare favorably to a state-of-art transformer. That is fine! But at least it will make the comparison actually useful.
[1] -- https://github.com/KellerJordan/modded-nanogpt
[2] -- https://kellerjordan.github.io/posts/muon/#discussion-solvin...
https://github.com/pathwaycom/bdh
There isn't a ton of code and there are a lot comments in my native language, so at least that is novel to me
I'm assuming they're using "rivals GPT2-architecture" instead of "surpasses" or "exceeds" because they got close, but didn't manage to create something better. Is that a fair assessment?
Everyone and their dog says "transformer LLMs are flawed", but words are cheap - and in practice, no one seems to have come up with something that's radically better.
Sidegrades yes, domain specific improvements yes, better performance across the board? Haha no. For how simple autoregressive transformers seem, they sure set a high bar.
> BDH is designed for interpretability. Activation vectors of BDH are sparse and positive.
This looks like the main tradeoff of this idea. Sparse and positive activations makes me think the architecture has lower capacity than standard transformers. While having an architecture be more easily interpretable is a good thing, this seems to be a significant cost to the performance and capacity when transformers use superposition to represent features in the activations spanning a larger space. Also I suspect sparse autoencoders already make transformers just as interpretable as BDH.
Anything "brain-like" that fits into one single paper is bullshit.
Don't berate the authors for the HN submitter's carelessness.
That's standard in computational neuroscience. Our standard should simply be whether they are imitating an actual structure or technique in the brain. They usually mention that one. If they don't, it's probably a nonsense comparison to get more views or funding.
Just to illustrate the absurdity of your point: I could claim, using your standard, that a fresh pile of cow dung is brain-like because it imitates the warmth and moistness of a brain.
The brain-inspired papers have done realistic models of specific neurons, spiking, Hebbian learning, learning rates tied to neuron measurements, matched firing patterns, done temporal synchronization, hippocampus-like memory, and prediction-based synthesis for self-training.
Brain-like or brain-inspired appears to mean using techniques similar to the brain. They study the brain, develop models that match its machinery, implement them, and compare observed outputs of both. That, which is computational neuroscience, deserves to be called brain-like since it duplicates hypothesized, brain techniques with brain-like results.
Others take the principles or behavior of the above, figure out practical designs, and implement them. They have some attributes of the brain-like models or similar behavior but don't duplicate it. They could be called brain-inspired but we need to be careful. Folks could game the label by making things that have nothing to do with brain-like models or went very far away.
I prefer the be specific about what is brain-like or brain-inspired. Otherwise, just mention the technique (eg spiking NN) to let us focus on what's actually being done.
AI systems are software, so if you want to build something brain like, you need to understand what the brain is actually like. And we don’t.
BTC literally hit all time high this month, fyi.
House prices are at all time highs too. That doesn't mean the housing bubble never happened.
Unless you're going to claim that previous large drops in crypto were perhaps bubbles, but this time it's real...
But if that's not the claim, then I'm saying that the current value makes it's clear that it's not the end of a bubble.
As a sidenote--does anyone really think human-like intelligence on silica is a good idea? Assuming it comes with consciousness, which I think is fair to presume, brain-like AI seems to me like a technology that shouldn't be made.
This isn't a doomer position; that human-like AI would bring about the apocalypse. It is one of empathy: At this point in time, our species isn't mature enough to have the ability to spin up conscious beings so readily. I mean look how we treat each other--we can't even treat beings we know to be conscious with kindness and compassion. Mix our immaturity with a newfound ability to create digital life and it'll be the greatest ethical disaster of all time.
It feels like researchers in the space think there is glory to be found in figuring out human-like intelligence on silicon. That glory has even attracted big names outside the space (see John Carmack), under the presumption that the technology is a huge lever for good and likely to bring eternal fame.
I honestly think it is a safer bet that, given how we aren't ready for such technology, the person / team who would go on to actually crack brain-like AI would be remembered closer to Hitler than to Einstein.
Or maybe if we had artificial life to abuse, it would be a suffcient outlet for our destructive and selfish impulses so that we would do less of it to genuine life. Maybe it's just an extension of sport contests that scratch that tribal itch to compete and win. There are no easy answers to these questions.
That said, I think probably the best path would just be to build and foster technologies that help our species mature, so if one day we do get the ability to spin-up conscious beings artificially, it can be done in a manner that adds more beauty rather than despair to our universe.
For all we know, an ICE in a 2001 Toyota truck is conscious too - just completely inhuman in its consciousness.
Nonetheless, here we are - building humanlike intelligence. Because it's useful. Having machines that think like humans do is very useful. LLMs are a breakthrough in that already - they implement a lot of humanlike thinking on a completely inhuman substrate.
I don't think appealing to whether or not inanimate objects may be conscious is sufficient to discount that we are toying with a different beast in machine learning. And, if we were to discover that inanimate objects are in-fact conscious, that would be an even greater reason to reconfigure our society and world around compassion.
I agree that LLMs are a great breakthrough, and I think there are many reasons to doubt consciousness there. But I would suggest we rest on our laurels for a bit, and see what we can get out of LLMs, rather than push to create something that is closer to mimicking humans because it might be more useful. From the evil perspective of pure utility, slaves are quite useful as well.
Existing LLMs are already trained to mimic humans - by imitating text, most of which is written by humans, or for humans, and occasionally both. The gains from other types of human-mimicry don't quite seem to land.
The closest we got to "breakthrough by mimicking what humans do" since pre-training on unlabeled text would probably be reasoning. And it's unclear how much of reasoning was "try to imitate what humans do on a high level", and how much was just trying to generalize the lessons from the early "let's think about it step by step" prompting techniques.
It's likely that we just don't know enough about the human mind to spot, extract and apply the features that would be worth copying. And even if we did... what are the chances that the features we would want to copy would turn out to be the ones vital for consciousness?
I don't think that precludes remaining concerned with the continued push to make current models more humanlike in nature. My initial comment was spurred by the fact that this paper is literally presenting itself as solving the missing link between transformer architectures and the human brain.
Here's to hoping this all goes toward a better world.
The famous Chinese Room Translator -- silica is irelevant, you could probably implement LLM-like algorithm with pen and paper, do you still think the paper could suffer or be "conscious"?
That said, I don't think it is a sufficient appeal to entirely discount the possibility that the right process implemented on silicon could in fact be conscious in the same way we are. I'm open to whether or not it is possible--I don't have a vested interest in the space--but silica seems to be a medium that can possible hold the level of complexity for something like consciousness to emerge.
So this is to say that I agree with you that consciousness likely requires substrate-specific embodiment, but I'm open to silica being a possible substrate. I certainly don't think it can be discounted at this point in time, and I'd suggest that we don't risk a digital holocaust on the bet that it can't.
My most generous interpretation of Anthropic's flirting with it is they too think it would be a nightmare and are hyper-vigilant. (My more realistic interpretation is that it's just some mix of a Frankenstein complex and hype-boosting.)
That said, the cynic in me thinks they give lip service to these things while pushing fully ahead into the unknown on the presumption of glory and a possibility of abundance. A bunch of the leadership are EAs who subscribe to a kind of superintelligence eschatology that goes as far as to give a shot at their own immortality. Given that, I think they act on the assumption that AGI is a necessity, and they'd rather take the risks on everyone's behalf than just not create the technology in the first place.
Them recently flirting with money from the gulf states is a pretty concerning signal pointing to them being more concerned with their own goals rather than ethics.
I completely agree. I think that the people who are funding AI research are essentially attempting to create slaves. The engineers actually doing the work have either not thought it through or don't care.
> Assuming it comes with consciousness, which I think is fair to presume, brain-like AI seems to me like a technology that shouldn't be made.
"Fair to presume" is a good way to put it. I'm not convinced that being "like a brain" is either necessary or sufficient for consciousness, but it's necessary to presume it will, because consciousness is not understood well enough for the risk to be eliminated.
But... As Karpathy stated on the Dwarkesh podcast, why do we need brain inspired anything? As Chollet says a transformer is basically a tool for differentiable program synthesis. Maybe the animal brain is one way to get intelligence, but it might actually be easier to achieve in-silica, given the faster computational ability and higher reproducibility of calculations