Attention Is Not What You Need (2025) | Hacker News Reader

Attention Is Not What You Need (2025)

arxiv.org

197 pointswesturner72 comments

The author describes the model's pairing operation as:

(zt,zt+Delta) where Delta is a positive offset.

This means the model is using the next token to predict the next token. It should have been (zt,zt-Delta), or it's cheating.

I don't know if that's just the paper, or if the code was also b0rk3d.

That's not how I read it.

For each position t and offset Δ∈ such that t+Δ≤L, we form a pair (z_t,z_t+Δ).

For a given t, this produces up to m pairs:

(z_t,z_t+Δ_1),(z_t,z_t+Δ_2),….

We treat these as local neighborhoods at multiple scales. Note that the pairing is causal: we only pair t with positions strictly to its right (future), consistent with left-to-right language modeling.

So z_t is a token in the input, and z_t+Δ is some later token. Guess they could have flipped that around to (z_t-Δ, z_t) with t+Δ>=0.

> So z_t is a token in the input, and z_t+Δ is some later token.

I think you get what I'm saying, but when they compute the gate they do:

  h_t_new = alpha_t * h_t + (1 - alpha_t) * g_t

If I'm reading it right g_t = geometric_mix(z_t, z_(t+delta))

Which would mean that g_t includes t+Δ, which is a future token. It's leaking future tokens into the current math.

So it could just be a sign error and still be right in the code, or I could just be getting it wrong.

If I'm not though, and given that they got worse results than a standard transformer I really hope it's fixed in the code. If they do worse even when knowing the future it's bad news for the idea.

Fair point. If only the paper had a link to the code.

I found this[1] repo which tries to reproduce it, and its accompanying blog post[2] uses (z_t, z_t - delta), without further mention.

Another repo[3] trying to reproduce it also does the reverse from what I can see, so (current, past).

Interestingly the first repo got underwhelming results, while the second repo matched the paper roughly. Not sure if it's related to the Claude-usage or not in the first, though it notes a lot of details were left out of the paper which is true.

[1]: https://github.com/Infatoshi/grassmann-flows/

[2]: https://github.com/Infatoshi/grassmann-flows/blob/main/blog....

[3]: https://github.com/kanenorman/grassmann

They advertised it as 10% to 15% worse than transformers

Not sure what the point of the paper's title is if the resulting model does not match the transformer.

Well they're quite explicit about that:

In summary, our results suggest that what we fundamentally need for strong sequence modeling is not attention as such, but a principled way for representations to move on the manifolds they inhabit. Grassmann flows provide one concrete instantiation of this idea, and we hope they will encourage further exploration of geometric alternatives to attention in the design of neural architectures.

They also note they kept the proposed model very simple to make it easier to understand (and presumably train):

In this paper we deliberately restricted ourselves to a minimal design—k = 2, no explicit global invariants—to keep the core idea clear. But we view “global invariants + local Grassmann flow” as a promising direction for future work on geometry-aware reasoning.

No my point is more general. The headline promises me that I don't need attention based transformers, then you see the benchmarks and they tell you they got 10% worse perplexity. How is that going to convince me that I don't need attention?

I mean if you look at e.g Mamba 3 [0] (see page 12 and 14 for benchmarks), the perplexity is roughly the same and the benchmark scores are roughly the same. If you tell me attention doesn't matter you have to convince me by showing that you can get the same performance with a variety of architectures. In short, I would have expected the paper to be a survey across model architectures that all perform roughly the same, which then would give us the impression "hey, they're all performing the same, maybe we don't need attention after all!", instead we get to be shown that their pet architecture is worse, which is the opposite of what the headline conveys.

If I gave a random ML researcher this paper, they would just walk away and work on the next transformer-mamba hybrid.

[0] https://arxiv.org/abs/2603.15569

Both of those reproductions are also b0rk3d. THey fix the sign flip, but then goof up other parts of the reproduction. Arledge (Infotashi), for example, has two residuals.

I still don't think it will work, but I'm also fairly confident that nobody has actually reproduced it well enough to know that for sure.

Scale is all you need.

That's why you see other attention free large language models work. But in a world with constraints on time and compute the architecture does matter because scaling does have a cost (which is dropping)

In Qwen 3.5, 75% of the layers are linear attention. I think the theory is that linear attention is usually enough, but some regular attention layers help to handle more sophisticated long-range relationships.

The 75/25 split is the part I want pinned down. If a few full-attention layers are what actually bind long-range dependencies, then "you can drop most softmax attention" and "you still need some" can both be true. Is the Qwen 3.5 mix from a model card or a config, and does it say whether those full layers are early, late, or interleaved?

Pretty much. See “Scaling Laws for Neural Language Models”, a must read IMO https://arxiv.org/html/2001.08361v1

Compute, data, and parameter count are the three logarithmic values that correspond to model performance.

So if you find ways of improving training efficiency (the simplest conceptually is probably local/global sliding window attention like Gemma; the latest DeepSeek and Kimis have quite fancy ones), you will be able to have more effective compute, and train smarter models while keeping data and param count the same.

If you find this fascinating and want to dive more, I encourage checking out nanochat (solid GPT-2 repro; minimal LoC); and trying to make architectural tweaks with a coding agent.

“Learning through experimentation” has never been easier or more fun in my opinion. If you have a GPU, even an older 8GB card (or a few bucks for Runpod), you can pretrain your own small language model and start experimenting / tweaking it.

> So if you find ways of improving training efficiency (the simplest conceptually is probably local/global sliding window attention like Gemma; the latest DeepSeek and Kimis have quite fancy ones), you will be able to have more effective compute, and train smarter models while keeping data and param count the same.

I don't get how changing the attention head (i.e sliding window attention instead of vanilla transformer) is only seen as a "compute" improvement. Indeed, it improves the complexity from quadratic to linear, but there was a reason why people used transformer and not sliding windows for so long (they were already there in 2021).

Making sliding window attention work with performances on par with vanilla transformers is a non trivial problem and is not solved by only putting more compute. This still requires "smart" ideas in that direction.

Yes if you took a transformer network and added connections until it became a full network, then it would still work (?) but it would be a lot more expensive.

(?) We actually don't know that but it is very likely.

Makes me wonder if the brain just does it with brute force.

The neocortex would be akin to something like a 100T neural network, and that’s with the maybe overly reductionist assumption that only synapses are doing anything when there’s a lot of biology happening in there. Every cell is filled with its own computation and state. It could be closer to a 10E model if all that is in play.

The neocortex is nowhere near the whole brain.

Architecturally the brain could be described as a slow but ultra-wide machine with colocated memory and thus an effectively enormous memory bandwidth. That lets it evaluate its whole parameter set continuously.

I keep thinking the brain must be algorithmically doing something super clever but maybe not. Maybe old mother nature just found the “bitter lesson.” Maybe it’s algorithmically inefficient and messy but there’s just a whole lot of it.

It's running on 20 watts though. There must be some tricks in there.

I think after you factor it in being on the extreme end of slow but wide (like laptop GPUs) and being a native representation of the neural network rather than a computed silicon approximation, and being more similar to a recurrent/continuous learning model rather than a huge all at once back prop, are there more tricks required?

My thought is you could probably set up something similar with today's technological capabilities but it would probably take decades to train and be expensive to develop which is commercially unacceptable.

Analog computation probably. Theoretically, technology like thermal wells and others that harness physics to find the local minimum energy state allow inference with multiple orders of magnitude improvements in size, energy, and speed. Rough back of the napkin gives SD-card size 1t parameter models on about 2.5 watts using thermal wells (cryogenic cooling required, afaik), achieving density approximately similar to NAND flash per parameter.

The simple explanation is that von Neumann architecture sucks for intelligence. It's not a von Neumann machine at all. It's an analog neural network (plus other systems in operation) with colocated memory.

A more far-out explanation is that some form of stochastic quantum computing is going on, which has been proposed at many levels by biologists. There's some work being done on this idea. It would not be the discrete qubit form we are trying to do, which is impossible in a noisy wet environment, but something analog and stochastic.

Both are possible.

Also note that our current chips, in spite of the tiny feature sizes and a ton of optimization work, are not even close to the lower limit for energy cost of computation. They still are not very efficient.

Given we are comparing two different substrates, there's no reason to assume this means anything.

We know that the quality of the astrocytes matters a lot. It's not just about the quality of the neurons or their number or the way they are connected.

I'm afraid we don't know much more than that yet.

We have created frankenrodents with human astrocytes and they are a lot smarter than their unfranked counterparts. I bet we could learn a lot if we repeated the experiment with astrocytes from humans with different, known IQs.

It just seems like we're missing some pretty important stuff. The amount of "training data" humans process in out lifetimes is nowhere near enough to achieve the abilities we have, so clearly the bulk of our brains abilities must be hereditary. The only problem is: by what mechanism? Is such a complex model able to fit on a single cell? Are we somehow influenced by our mother's body in ways that are outside of that thinking? Are our brains just somehow much better architecturely than current ml approaches? The more I think about the more I feel that true AGI is probably still lifetimes of knowledge away.

I agree with the framing but the claim is too strong.

I don't think Grassmannian manifolds are necessarily a better representation than simple query/key inner products. It's a valid solution, and it does have advantages in some cases.

The Grassmannian is only preserving the subspace orientation, so we lose information about the actual linear map and basically throw out its basis.

Like everything, this is a trade off. You're trading information for a lower rank representation.

I'm still reading through the paper on the implementation side, but the other caveat I'd be wary of is compute. Does this projection add MACs to the overall model which would make inference slightly slower even with lower rank representations? I don't know. Like with any technique that trades constrained information representation for space/compute you _are_ widening your search space during training. I'd be curious what the impact of this is on relative data needed to train a model using this architecture versus "regular" attention.

They only demonstrate this idea on a laughably small toy problem. An 18M parameter model (which they call “moderate”) on an ancient dataset proves exactly nothing. Plenty of bad ideas appear to work at trivial scale.

Arguably the most important aspect of transformers is that it allowed easy training of arbitrarily large models.

I think there's a middle ground between "attention is obsolete" and "this proves nothing"

You’re right I’m overly dismissive. But there are dozens of papers like this showing some more efficient alternative to attention. There is plenty of evidence pointing to this paper going nowhere.

Efficient alternatives have been coming out for 10 years now. Flash attention is the only one that has really moved the needle - because it’s not an algorithmic compromise. It’s just a smarter way to put the parentheses around the linear algebra operations.

My gut (informed by 30ish years of professional hardware and software engineering) says implemented at scale, on a real machine, we start seeing the same performance issues functional languages have where absolute conformity to the math will increase latency.

Tis why Haskell, etc, will not be used to write a meaningful OS wielded at scale. Graphics mutations don't want to wait for long dead Mr Euler or whoever to be satisfied.

Having to cache states as rigorous axiomatic rules are satisfied without C like ability to cut the queue and mutate some value just moves the problems C creates being able to do that, doesn't eliminate them; thermodynamics is a bitch like that.

Math is not physics. Physics is not math. Math is merely a syntax for notation of physics like how English is a notation for allegory and metaphor.

Physics is immutable and will not break because we decide some syntax suggests it will.

(2025)

"subconscious" reasoning

Right. What makes an expert good? It's intuition. If they had to pay attention to every detail they would never get anything done, but attention is a useful utility for figuring out how your model is working and how it isn't. If you've ever done anything complicated, you know that paying attention to one thing just often makes you overlook the obvious issue.

human intuition is supported by embodied experience, perception, and years of interacting with the physical and social world. An AI model's learned pattern recognition is analogous in some respects, but it isn't the same process.

intelligence may depend less on continuously scrutinizing everything than on having a good internal model, recognizing when that model is likely to fail, and knowing where to investigate.

There are many alternatives to attention which work well for small sequence sizes.

Oh yes? I'd be intrigued to learn more.. I am working with smaller sequences (peptides, len is about 30ish) and currently use an attention based mechanism, do you maybe have pointers to where to start?

Look up what some Chinese LLMs are doing, e.g. mixing some attention mechanism with linear attention layers. Mamba2 layers, Gated Delta Net (2), Kimi Delta Attention and recent Raon Delta Attention (Gated Delta Net 2 + improvement from Kimi Delta Attention)

You are likely to find kernels for Gated Delta Net 2 as Nvidia had a paper on that.

Once you understand what attention actually is, its pretty easy to adapt it how you need it.

https://news.ycombinator.com/threads?id=ActorNightly&next=49...

TLDR its not so much about the transformer, as much as expanding dimentionality. You can chose any which way to do so.

Man, I'm so tired of these xyz is all you need titles.

Electrolytes

Watch out, this paper has scary ~math notation~

;)

But misspelled Plücker coordinates in the abstract.

Fair, but they're doing a lot better writing English than I would do in Chinese and Plücker is German.

It's a call back to comments complaining about math notation: https://news.ycombinator.com/item?id=50122878#50126010

lmao seems just computing power arms race >>> specific algorithm

The last time we learned this, we called i the bitter lesson.

why doesn't the author have any internet footprint (or at any rate any one that links to this paper)? it's curious.

Because he doesn't need attention

Okay this was genuinely funny.

You deserve a medaile for this.

What has that got to do with anything? Look into his work rather than his online persona.

It's his/her prerogative to keep an online footprint or not.

Because having an online footprint is a great way to get cloned and scammed? It’s a great way to get tracked? It’s a great way to get addicted to these “social networks”?

It might be someone who is quite active on bilibili or weibo.

Interesting idea, but it reads like an LLM write up. For me to engage properly with your written work, I need to see human effort. There are too many half-baked ideas out there, and unless you, as the author, thought the idea is worth putting proper effort into communicating, why should I think of this anything other than something someone said in passing? One way you could communicate effort to me, and likely others, rather than this work looking like the output of some LLM, is to use graphs, schematics and figures to support claims. Metrics are extremely misleading, only really useful when we've verified what is being measured is what one assumes is being measured. Metrics on their own are often deeply bullshit, so without any kind of graphs of trends or other important qualities, I assume it's all LLM output, since LLM's will default to not plotting anything while humans tend to go the opposite way.

Also, more effort into completeness needs to go in. Perhaps it is in there, but you just report an "accuracy" without defining what that means. The more I look, the more it starts becoming apparent this likely hasn't received any real human over-sight at all. "What does Grassmann mixing actually buy us?" It buys nothing, because it is a concept, Claude.

Neat idea, but in my opinion devoid of critical thinking / analysis or the level of rigour required to be suitable for broader dissemination. I'd put effort into making this a more credible body of work ASAP. Slop like this WILL cripple your standing in the field.

> For me to engage properly with your written work, I need to see human effort

Agreed.

I say: “if you want a human response, then give human effort”

This might sound like a joke, but your comment sounds LLM generated, and I'm going to assume it's all LLM output.

actually, I think you're right

I was gonna say "too pedantic even for an LLM" but then I got worried that it might be read as a compliment.

If I suspect stuff is LLM output I often respond in kind to give them a taste of their own medicine. I can imagine that's what this is.

almost, it's just a stream of conciousness reply because effort spent on refining a comment must be matched by post effort or you'll burn out.

Reminder: Publications on arXiv are not peer-reviewed, and practically anyone can post a paper regardless of quality or correctness.

Here, many if not most qualify as peers.

It should be unsurprising if comments here, made in public, capable of being challenged are far higher quality than official "peer review" which has really become a bit of a joke. Has been for quite a long time, to be fair. (And unsurprising if not, it's a web forum).

Peer reviewed and published means precisely excrement in a world where papers are not retracted and failures to replicate are refused publication by the "prestigious" "peer-reviewed" journals that published the original.

Intelligent, reasoned, knowledgeable review going back and forth with the author in public makes "peer review" seem just as silly as it usually is. Review by peer professor's indentured servants aka PhD students is often actually higher quality, if wildly dishonest and quite frankly disgraceful.

We now live in a world where LLMs can produce papers. Actual research is a little more interesting, with or without LLM. The journal peer review system doesn't seem to be helping.

I am not against some kind of "public review" and I really like arxiv.org. But I do not see it as a replacement for a peer review. It may start after a peer review when the paper is published, which IMHO should almost always be in an Open Access way.

For AI, the borders seem to have blurred a bit between academic scientists (e.g. at universities) and external scientists at companies or even private people). The problem with the latter is that it is much harder to enforce good scientific practice. Think only about the time it takes to write the state of the art of an ordinary conference paper.

I am also a bit drawn between giving everyone the chance to contribute to research and ensuring good scientific practice. Just writing down your latest ideas and successes is not the same as writing a scientific paper that shows the research in the context of past research and needs, for instance, to be much more mathematically rigorous.

You can write the above and I can respond. The quality of each response is visible to everyone. Your ideas and mine are actually getting reviewed here, now, at least to some degree. Or it can be seen that they are not getting reviewed or indeed there's a non-zero chance that it can be seen that one or both of us is a muppet. Anyone else can chime in and be ignored, or celebrated or anything else.

This is nothing like academic journal peer review, where you can't see the failed submissions for the paper and why. You have no clue if the reviewers were interested in the topic, diligent, drunk or playing politics "support papers from these institutions because they're the good guys, undermine those because they are or aren't from Europe." You have no idea if the paper has since failed to replicate on multiple occasions. You have no idea if you are the first person to actually read the paper and when you have you thoughts, no matter how insightful, well articulated and supported by evidence, those thoughts do not matter to the publication process of that paper, which has been peer reviewed.

Public review by the knowledgeable who care is a very different thing to so called "peer review."

Peer review and scrutiny out in the public is more valuable than only done by a couple of paid guys from a journal.

It's fairly basic. For a work published on a public platform, the same guys paid by the journal can also review it if they so choose to in addition to those people in the public who were not paid by anyone.

It’s always funny to read general-public takes on academic peer review, because they almost always betray a complete misunderstanding of how peer review works. Academic reviewers for the vast majority of journals notoriously do not get paid; in fact, paying peer reviewers is one of the most frequently suggested approaches to peer review reform.

Most still get benefits. I concede that many don't get "direct" remuneration. Some journals do pay. But many do get vouchers, discounts, free access, early access, etc. which amounts to a form of payment and thus compensated in kind. This effectively is "paid" as far as I am concerned.

The logic stands. A reviewer, even those who get zero compensation, can always review publicly published papers just like others.

Reading a paper with equations takes time.

In my academic career, I had one instance of reviewers actually reading a paper. Otherwise -- accept or reject -- the comments made clear the paper had never been actually read.

Retractions are non-existent, even with clearly proven academic malpractice.

I am sorry for that. Fortunately, this does not resemble my experience (electrical engineering/digital signal processing).

Usually I got quite critical comments (positive and/or negative). Yes, sometimes a reviewer does not write even a single sentence before giving an average score to complete the review process. But this has not been the regular case.

And as a reviewer for a conference paper of 4 to 6 pages, I usually assigned 1 to 3 hours to read it (journal papers: 1 to 2 days) and to give comments.

Curiously, the one paper which was actually reviewed was in an IEEE journal and about filter design.

Most of the rest of my papers were in completely different disciplines.

It was an awesome paper -- in my mind, groundbreaking -- but I'm not convinced anyone beyond the reviewers ever read it. I see about a dozen citations which popped up years later, clicked through a few, and they were inane.

Problem is it was work in a very, very classical field, alongside Butterworth, Chebyshev, and similar, and while quite fundamental, and with many applications, there is no contemporary research there.

My most popular papers were mostly rather.... random.

Oh dude. You gotta link it after that. I'm asking.

That would tie back HN username to my real name, which I don't want to do.

Glad to shoot a message if you have an email, linkedin, whatsapp...

The irony is that peer review is supposed to catch exactly the kind of problems that require someone to actually sit down and work through the math

Depends on journal. Predatory and low tier journals that's true for, but at higher tiers it is far less common. You still get the occasional lazy reviewer, but typically they wont be asked back if the editor isn't lazy as well. The system certainly has its flaws though.

Retractions do happen as well, especially if it's clearly proven malpractice. I've been behind 2 retractions so far, but it has never (in my experience) happened at the "contact corresponding author and seek clarification" stage of escalation. Typically it requires escalation to the journal itself.

Interesting...

interesting

Reply on news.ycombinator.com