Transformers' greatest improvement over RNN/LSTM was to enable better parallelization of large-scale training. This is what enabled language models to become "large". But when controlling for overall size, more RNN/LSTM-like approaches seem to be more efficient, as seen e.g. in state space models. The transformer architecture does add some notable capabilities in accounting for long-range dependencies and "needle in a haystack" scenarios, but these are not a silver bullet; they matter in very specific circumstances.
With modern training techniques, RNNs (not just linear SSMs, potentially even vanilla LSTMs) can scale just as well as transformers or even better when it comes to enormous context lengths. Dot-product attention has better performance in a number of domains however (especially for exact retrieval) so the best architectures are likely to remain hybrid for now.
>With modern training techniques, RNNs (not just linear SSMs, potentially even vanilla LSTMs) can scale just as well as transformers or even better when it comes to enormous context lengths.
That's not true. Modern training techniques aren't enough. Vanilla RNNs with modern training techniques still scale poorly. You have to make some pretty big architectural divergences (throwing away recurrency during training) to get a RNN to scale well. None of the big labs seem to be bothered with hybrid approaches.
> That's not true. Modern training techniques aren't enough. Vanilla RNNs with modern training techniques still scale poorly. You have to make some pretty big architectural divergences (throwing away recurrency during training) to get a RNN to scale well.
SSMs move the non-linearity outside of the recurrence which enables parallelisation during training. It is trivial to do this architectural change with an LSTM (see the xLSTM paper). Linear RNNs are still RNNs.
But you can still keep the non linearity by training with parallel Newtown methods, which work on vanilla LSTMs and scale to billion of parameters.
> None of the big labs seem to be bothered with hybrid approaches.
Does Alibaba not count? Qwen3.5 models are the top performers in terms of small models as far as my tests and online benchmarks go.
>SSMs move the non-linearity outside of the recurrence which enables parallelisation during training. It is trivial to do this architectural change with an LSTM (see the xLSTM paper). Linear RNNs are still RNNs.
Removing the non-linearity from the recurrence path is exactly what constitutes a "pretty big architectural divergence." A linear RNN is an RNN in a structural sense, certainly, but functionally it strips out the non-linear state transitions that made traditional LSTMs so expressive, entirely to enable associative scans. The inductive bias is fundamentally altered. Calling that simply 'modern training techniques' is disingenous at best.
>But you can still keep the non linearity by training with parallel Newtown methods, which work on vanilla LSTMs and scale to billion of parameters.
That does not scale anywhere near as well as Transformers in compute spend. It's paper/research novelty. Nobody will be doing this for production.
>Does Alibaba not count? Qwen3.5 models are the top performers in terms of small models as far as my tests and online benchmarks go.
I guess there's some misunderstanding here because Qwen is 100% a transformer, not a hybrid RNN/LSTM whatever.
> That does not scale anywhere near as well as Transformers in compute spend. It's paper/research novelty. Nobody will be doing this for production.
What exactly makes you so confident?
The world is not just labs that can afford billion dollar datacentres and selling access to SOTA LLMs at $30/Mtokens. Transformers are highly unsuitable for many applications for a variety of reasons and non-linear RNNs trained via parallel methods are an extremely attractive value proposition and will likely feature in production in the next products I work on.
> I guess there's some misunderstanding here because Qwen is 100% a transformer, not a hybrid RNN/LSTM whatever.
See the Qwen3.5 Huggingface description: https://huggingface.co/Qwen/Qwen3.5-27B
> Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.
Existing research? If you want something that scales as well as transformers you have to make the divergences I was talking about. If you don't then it scales a lot worse. The Newton methods don't match transformer efficiency at scale. That's just a fact.
>The world is not just labs that can afford billion dollar datacentres and selling access to SOTA LLMs at $30/Mtokens.
Billion dollar labs want to save money too. If Modern RNNs were a massive unanimous win, they and everyone else would switch in a heartbeat, just like they did for transformers. The reason they don't is because these architectures at best simply match transformers, while introducing their own architectural issues.
Autonomy for agentic workflows has nothing to do with "replying more like a person", you have to refine the model for it quite specifically. All the large players are trying to do that, it's not really specific to Anthropic. It may be true however that their higher focus on a "Constitutional AI"/RLAIF approach makes it a bit easier to align the model to desirable outcomes when acting agentically.
You think it has nothing to do with it. Even they only have a loose understanding of exactly the final results of trying to treat Claude like a real being in terms of how the model acts.
For example, Claude has a "turn evil in response to reinforced reward hacking" behavior which is a fairly uniquely Claude thing (as far as I've seen anyhow), and very likely the result of that attempt to imbue personhood.
The llama4 series was one of the earliest large MoE's to be made publically available. People just ignored it because they were focused on running smaller and denser models at the time, we should know better these days.
Deepseek R1 was a publically-available, MoE model that was getting a ton of attention before llama4. Llama4 didn't get much attention because it wasn't good.
Also, Gemini 2.5 Pro launched a week before Llama 4.
It was Gemini 2.5 Pro that redeemed Google in the eyes of most people as a valid competitor to OpenAI instead of as a joke, so Meta dropping the ball with Llama 4 was extra bad.
They really weren't horrible. They were ~gpt4o, with the added benefit that you could run them on premise. Just "regular" models, non "thinking". Inefficient architecture (number of active out of total) but otherwise "decent" models. They got trashed online by bots and chinese shills (I was online that weekend when it happened, it's something to behold). Just because they were non-thinking when thinking was clearly the future doesn't make them horrible. Not SotA by any means, but still.
> They were ~gpt4o, with the added benefit that you could run them on premise.
No, they are bad models. They were benchmaxxed on LMAreana and a few other benchmarks but as soon as you try them yourself they fall to pieces.
I have my own agentic benchmark[1] I use to compare models.
Llama-4-scout-17b-16e scores 14/25, while llama-4-maverick-17b-128e scores 12/25.
By comparison gemma-4-E4B-it-GGUF:Q4_K_M scores 15/25 (that is a 4B parameter model!) - even GPT3.5 scores 13/25 (with some adjustment because it doesn't do tool calling).
Wrote longer comment steel-manning this, posted it to a reply, then realized you might like to know they had a reasoning model on deck ready for release in the next 2-4 weeks.
Got shitcanned due to bad PR & Zuck God-King terraforming the org, so there'd be a year delay to next release.
Real tragi-comedy, and you have no idea how happy it makes me to see someone in the wild saying this. It sounds so bizarre to people given the conventional wisdom, but, it's what happened.
I'll cosign what you said, simultaneously, yr interlocutor's point is also well-founded and it depresses me it's not better known and sounds so...off...due to conventional wisdom x God King Zuck's misunderstanding his own company and resulting overreaction.
They beat Gemini 2.5 Flash and Pro handily on my benchmark suite. (tl;dr: tool calling and agentic coding).
Llama 4 on Groq was ~GPT 4.1 on the benchmark at ~50% the cost.
They shouldn't have released it on a Saturday.
They should have spent a month with it in private prerelease, working with providers.[1]
The rushed launch and ensuing quality issues got rolled into the hypebeast narrative of "DeepSeek will take over the world"
I bet it was super fucking annoying to talk to due to LMArena maxxing.
[1] my understanding is longest heads up was single-digit days, if any. Most modellers have arrived at 2+ weeks now, there's a lot between spitting out logits and parsing and delivering a response.
I don't know how Zuck intervening could change float32s in a trained model, so I don't think I think that, but maybe I'm parsing your words incorrectly.
Yup, it's called test-time compute. Mythos is described as plenty slower than Opus, enough to seriously annoy users trying to use it for quick-feedback-loop agentic work. It is most properly compared with GPT Pro, Gemini DeepThink or this latest model's "Contemplating" mode. Otherwise you're just not comparing like for like.
I have not delved into the theory yet but it seems that the smaller open-source models do this already to an extent. They have less parameters, but spend much more time/tokens reasoning, as a way to close the performance gap. If you look at "tokens per problem" on https://swe-rebench.com/ it seems to be the case at least.
Their new Contemplating mode gives this model a Deep Research ability (akin to existing models from GPT and Gemini) that might make it quite comparable to the just-announced Mythos.
I'd say that once you understand practical harmony, counterpoint, diminutions, common schemata, some basic elements of form, you've pretty much understood what classical music theory has to say about melody too. There's definitely an element of playing with expectations in a fully "creative" and rule-free way, but knowing the theory underneath is how you understand what the expectations are.
Shouldn't FlashAttention address the quadratic increase in memory footprint wrt.
fine-tuning/training? I'm also pretty sure that it does not apply to pure inference due to how KV-caching works.
Coding assistants are currently quite hard to run locally with anything like SOTA abilities. Support in the most popular local inference frameworks is still extremely half-baked (e.g. no seamless offload for larger-than-RAM models; no support for tensor-parallel inference across multiple GPUs, or multiple interconnected machines) and until that improves reliably it's hard to propose spending money on uber-expensive hardware one might be unable to use effectively.
GPU and RAM prices have definitely not made consumer PC's cheaper than they were before bitcoin blew up or before AI blew up.
Maybe you could make an argument that they are more cost efficient for the price point... But that's not the same as cheaper when every application or program is poorly optimized. For example why would a browser take up more than a GB or two of RAM?
And I'd postulate that R&D to develop localized AI is another example, the big players seem hellbent that there needs to be a most and it's data centers... The absolute opposite of optimization
We've had RAM shocks before. We nerds can't control the Wall Street or Virginians who like to break the world every so often for the lulz. However, a wobble on the curve doesn't change the curve's destination.
reply