The MAMBA [1] model gained some traction as a potential successor. It's basically an RNN without the non linearity applied across hidden states, which makes it logarithmic time (instead of linear time) inference with a parallelizable scan [2].

It promises much faster inference with much lower compute costs, and I think up to 7B params, performs on par with transformers. I've yet to see a 40B+ model trained.

The researches of MAMBA went on to start a company called Cartesia [3], which is MAMBA applied to voice models

[1] https://jackcook.com/2024/02/23/mamba.html

[2] https://www.csd.uwo.ca/~mmorenom/HPC-Slides/Parallel_prefix_... <- Pulled up a random example from google, but Stanford CS149 has an entire lecture devoted to parallel scan.

[3] https://cartesia.ai/

Check out "Attention as an RNN" by Feng et al (2024), with Bengio as a co-author. https://arxiv.org/pdf/2405.13956

Abstract: The advent of Transformers marked a significant breakthrough in sequence modelling, providing a highly performant architecture capable of leveraging GPU parallelism. However, Transformers are computationally expensive at inference time, limiting their applications, particularly in low-resource settings (e.g., mobile and embedded devices). Addressing this, we (1) begin by showing that attention can be viewed as a special Recurrent Neural Network (RNN) with the ability to compute its many-to-one RNN output efficiently. We then (2) show that popular attention-based models such as Transformers can be viewed as RNN variants. However, unlike traditional RNNs (e.g., LSTMs), these models cannot be updated efficiently with new tokens, an important property in sequence modelling. Tackling this, we (3) introduce a new efficient method of computing attention’s many-tomany RNN output based on the parallel prefix scan algorithm. Building on the new attention formulation, we (4) introduce Aaren, an attention-based module that can not only (i) be trained in parallel (like Transformers) but also (ii) be updated efficiently with new tokens, requiring only constant memory for inferences (like traditional RNNs). Empirically, we show Aarens achieve comparable performance to Transformers on 38 datasets spread across four popular sequential problem settings: reinforcement learning, event forecasting, time series classification, and time series forecasting tasks while being more time and memory-efficient.

I have an internal repo that does guided window attn. I figured out that One Weird Trick to get the model to learn how to focus so that you can move a fixed window around instead of full attn. I also built NNMemory (but that appears to be an idea others hae had now too [1]) and I have a completely bonkers mechanism for non-determanistic exit logic so that the model can spin until it thinks it has a good answer. I also built scale free connections between layers to completely remove residual connections. Plus some crazy things on sacrificial training (adding parameters that are removed after training in order to boost training performance with no prod penalty). There are more crazy things I have built but they aren't out there in the wild, yet. Some of the things I have built are in my repo. [2] I personally think we can get .5b models to outperform 8b+ SOTA models out there today (even the reasoning models coming out now)

The basic transformer block has been good at kicking things off, but it is now holding us back. We need to move to recurrent architectures again and switch to fixed guided attn windows + 'think' only layers like NNMemory. Attn is distracting and we know this as humans because we often close our eyes when we think hard about a problem on the page in front of us.

[1] https://arxiv.org/abs/2502.06049

[2] https://github.com/jmward01/lmplay

The xLSTM could become a good alternative to transformers: https://arxiv.org/abs/2405.04517. On very long contexts, such as those arising in DNA models, these models perform really well.

There's a big state-space model comeback initiated by the S3-Mamba saga. RWKV, which is a hybrid between classical RNNs and transformers, is also worth mentioning.

Personally I think foundation models are for the birds, the cost of developing one is immense and the time involved is so great that you can't do many run-break-fix cycles so you will get nowhere on a shoestring. (Though maybe you can get somewhere on simple tasks and synthetic data)

Personally I am working on a reliable model trainer for classification and sequence labeling tasks that uses something like ModernBERT at the front end and some kind of LSTM on the back end.

People who hold court on machine learning forums will swear by fine-tuned BERT and similar things but they are not at all interested in talking about the reliable bit. I've read a lot of arXiv papers where somebody tries to fine-tune a BERT for a classification task, runs some arbitrarily chosen parameters they got out of another paper and it sort-of works some of the time.

It drives me up the wall that you can't use early stopping for BERT fine-tuning like I've been using on neural nets since 1990 or so and if I believe what I'm seeing I don't think the networks I've been using for BERT fine-tuning can really benefit from training sets with more than a few thousand examples, emphasis on the "few".

My assumption is that everybody else is going to be working on the flashy task of developing better foundation models and as long as they emit an embedding-per-token I can plug a better foundation model in and my models will perform better.

I hate that popular domains take ownership of highly generic words. Many years ago, I struggled for a while to understand that when people say "frontend" they often mean a website frontend, even without any further context.

I tried... it started with the idea was that log loss might not be the best option for training, and maybe it should be a loss related to how wrong the predicted word was. Predicting "dog" instead of "cat" should be less penalised than predicting "running".

That turns out to be an ultrametric loss, and the derivative of an ultrametric loss is zero in a large region around any local minimum, so it can't be trained by gradient descent -- it has to be trained by search.

Punchline: it's about one million times less effective than a more traditional architecture. https://github.com/solresol/ultratree-results

There are alternatives that optimize around the edges. Like Deepseek’s Multi-head Latent Attention, or Grouped Query Attention. DeepSeek also showed an optimization on Mixture of Experts. These are all clear improvements to the Vaswani architecture.

There are optimizations like extreme 1.58 bit quant that can be applied to anything.

There are architectures that stray farther. Like SSMs and some attempts at bringing the RNN back from the dead. And even text diffusion models that try to generate paragraphs like we generate images i.e. not word by word.

Related: There was buzz last year about Kolmogorov Arnold Networks, and https://arxiv.org/abs/2409.10594 was claiming KANs perform better than standard MLPs in the transformer architecture. Does anyone know of these being explored in the LLM space? KANs seem to have better properties regarding memory if I'm not mistaken.

Literally everybody doing cutting edge AI research is trying to replace the transformer, because transformers have a bunch of undesirable properties like being quadratic in context window size. But they're also surprisingly resilient: despite the billions of dollars and man-hours poured into the field and many attempted improvements, cutting-edge models aren't all that different architecturally from the original attention paper, aside from their size and a few incidental details like the ReLU activation function, because nobody has found anything better yet.

I do expect transformers to be replaced eventually, but they do seem to have their own "bitter lesson" where trying to outperform them usually ends in failure.

You have stuff like: https://www.literal-labs.ai/tsetlin-machines/ and https://tsetlinmachine.org/ European initiatives..

Titans has a new approach to longer and faster memory compared to transformers.

https://arxiv.org/html/2501.00663v1

I'll see your architectural innovation and raise you a loss function revolution.

https://arxiv.org/pdf/2412.21149

Working on variants of Byte Latent Transformer [0] to get rid of tokenization which hinders mathematical performance and letter reflection.

In the original Byte Latent Transformer paper they reintroduce ugly caching and n-grams which I'm looking to eliminate.

As expected pure byte level Transformers need some rethinking to keep them performant, some kind of matryoshka mechanism so that long predictable byte sequences (words and phrases) get grouped into a single latent vector.

The idea is to apply this "Byteformer" not just on text but also on compiled files, songs etc.

If it's impossible to scale this architecture at least a modified tokenizer could be helpful which falls back to bytes / unicode once a number or an unfamiliar word is encountered.

[0] https://arxiv.org/abs/2412.09871

Anyone know what the rwkv people are up to now?

https://arxiv.org/abs/2305.13048

AI aside, the world could also use an alternative electric transformer. The backlog from main suppliers is 40+ weeks and far too expensive. There is a MAJOR manuf and supply issue here as all new build construction competes for same equipment...

Not a new model per se, but a new algorithm for inference - https://kolinko.github.io/effort/

There are a bunch of promising other ways to convert AC voltages.

The main one is the observation that the transformer uses an amount of copper and steel proportional to the power transmitted but inversely proportional to the frequency of operation.

The copper and steel cost of a transformer is the main cost (multiplied by the cost of capital for the 100+ years it will operate).

So if you can use solid state electronics to do switching at a higher frequency (switched mode power supplies, flyback designs, etc), then you can reduce the overall cost.

I've a design in mind which is very simple and interesting but don't know if it would be scalable to the stage, rn it's just a superficial design inspired by IronMan's JARVIS, i'm working on preparing the architecture.

https://github.com/triadicresonance/triadic this was on one of the llm discord servers few weeks ago

This looks interesting: https://www.youtube.com/watch?v=ZLtXXFcHNOU

Chain of thought in latent space.

Absolutely! The QNN architecture based on quantum computing concepts shows great potential. It breaks through traditional computing models and may outperform Transformers in complex tasks. Do you have any research on the combination of quantum computing and AI?

There is a LOT of effort in the research community currently:

1. Improving the Self-Attention in the Transformer as is, keeping the quadratic complexity, which has some theoretical advantage in principle[1]: The most hyped one probably DeepSeek's Multi-Latent Attention[15], which kind of is Attention still - but also somehow different.

2. Linear RNNs: This starts from Linear Attention[2], DeltaNet[3], RKWV[4], Retention[5], Gated Linear Attention[6], Mamba[7], Griffin[8], Based[9], xLSTM[10], TTT[11], Gated DeltaNet[12], Titans[13].

They all have an update like: C_{t} = F_{t} C_{t-1} + i_{t} k_{t} v_{t}^T with a cell state C and output h_{t} = C_{t}^T q_{t}. There's a few tricks that made these work and now being very strong competitors to Transformers. The key here is the combination of an linear associative memory (aka Hopfield Network, aka Fast Weight Programmer, aka State Expansion...) and pushing it into a sequence with gating similar to the original LSTM (input, forget, output gate) - while here this is only dependent on the current input not the previous state for linearity. The linearity is needed to make it sequence-parallelizable, there are efforts now to add non-linearities again, but let's see. Their main benefit+downside both is that they have a fixed-size state, and therefore linear (vs Transformer-quadratic) time complexity.

For larger sizes they have become popular in hybrids with Transformer (Attention) Blocks, as there are problems with long context tasks [14]. Cool thing is they can also be distilled from pre-trained Transformers with not too much performance drop [16].

3. Along the sequence dimension most things can be categorized in these two. Attention and Linear (Associative Memory Enhanced) RNNs are heavily using Matrix Multiplications and anything else would be a waste of FLOPs on current GPUs. The essence is how to store information and how to interact with it, there might be still interesting directions as other comments show. Other important topics that go into the depth / width of the model are: Mixture of Experts, Iteration (RNNs) in Depth[17].

Disclaimer: I'm author of xLSTM and we recently released a 7B model [18] trained at NXAI, currently the fastest linear RNN at this scale and performance. Happy to answer more questions on this or the current state in this field of research.

[1] https://arxiv.org/abs/2008.02217

[2] https://arxiv.org/abs/2006.16236

[3] https://arxiv.org/pdf/2102.11174

[4] https://github.com/BlinkDL/RWKV

[5] https://arxiv.org/abs/2307.08621

[6] https://arxiv.org/pdf/2312.00752

[7] https://arxiv.org/abs/2312.06635

[8] https://arxiv.org/pdf/2402.19427

[9] https://arxiv.org/abs/2402.18668

[10] https://arxiv.org/abs/2405.04517

[11] https://arxiv.org/abs/2407.04620

[12] https://arxiv.org/abs/2412.06464

[13] https://arxiv.org/abs/2501.00663

[14] https://arxiv.org/abs/2406.07887

[15] https://arxiv.org/abs/2405.04434

[16] https://arxiv.org/abs/2410.10254

[17] http://arxiv.org/abs/2502.05171

[18] https://huggingface.co/NX-AI/xLSTM-7b

I am told that an interesting alternative is the Structured State Space for Sequence Modeling (S4). I don't personally know much about this technique, but didn't see anybody else mention this in this thread.

https://srush.github.io/annotated-s4/

The one from cybertron? The one that changes voltage levels? The AI algorithm one?

Edit: or perhaps you are working on a new insect sex regulation gene? If so that would be a great discussion here - https://en.wikipedia.org/wiki/Transformer_(gene)

Not an alternative transformer like you asked for, but OptiLLM looks interesting for squeezing more juice out of existing LLMs.

Perhaps people feel that the problem of modeling long-range token relationships has been solved. The problem is now how to get this model to produce tokens that are valid and ingenious solutions to problems, with RL or otherwise.

I found the model Microsoft Tay was built on to be quite interesting, forgot the name of it.

My guess is that new architectures will be about doing more with less compute. For example, are there architectures that can operate at lower bit precision or better turn off and on components as required by the task?

yes: https://www.latent.space/p/2024-post-transformers

RWKV. It’s a Linux foundation project as well.

I’m working with a group on an RL core with models as tool use, for explainable agentic tasks with actual discovery.

Yes, I am building one.

Is it an alternative? Yes.

Is it better? Hell no.

Yes, I am building a Perfect Language Model

please define transformer

Right now as long as the rocket's heading straight up, everyone's on board with MLPs (Multilayer Perceptrons/Transformers)! Why not stay on the same rocket for now!? We're almost at AGI already!

Yes. Happy to chat if u msg me. Using RL coupled with NNs to integrate search directly into inference instead of as an afterthought like Chain of though and test time training.

The problem is, if they are trying to build new architecture, its just wasted effort.

To truly build AI, it needs to self configure. I tried doing some work in the past with point swarm optimization of models, but I didn't really get anywhere