SubQ: a sub-quadratic LLM with 12M-token context | Hacker News Reader

SubQ: a sub-quadratic LLM with 12M-token context

subq.ai

84 pointsmitchwainer40 comments

Assuming this is real and much better than existing linear attention methods as advertised, not launching with a technical report is a big miss.

Edit: their blog post (https://subq.ai/how-ssa-makes-long-context-practical) does go pretty in-depth about it

Edit 2: the fact that they're going straight for an end-to-end coding product on day 1 is very ambitious. Other speed/efficiency-oriented AI companies (Cerebras and Inception come to mind) still don't have a first-party coding product after years. IMO this is absolutely the right way to go if they really do have the big breakthrough they're claiming.

you really call this 1-minute blog post "in-depth"?

- magic.dev claimed 200M context window and it's been two years since and no real product yet.

- They are admitting that this is built on top of a Chinese model[1]

- They committed a huge chart crime with the Y axis of a chart comparing to Opus on their website that I can't find anymore (Too embarrassing to keep?). The delta between their score (81%) vs. Opus (87%) on SWE bench was hugely minimized

- They named the company subquadratic but in parts they said O(1) linear scaling. At O(1) you could do much more than 12M tokens context window. At O(log n) even.

I hope this is real but I doubt...

Ah, I nearly forgot about magic.dev. I took a quick peek to check up on them. Welp, last social/blog activity was in... 2024. But hey, their careers page still says they're hiring! So they must be doing just fine.

They did raise over $500M

The chart crime was not intentional! We will not make you wait two years. We are O(n), not O(1). O(1) would unfortunately be an impossibility. We may as well do infinite context at that point!

Good luck.

Thanks!

What’s keeping you from releasing paper and access to the model?

Model: - making sure it has been properly red-teamed, meets user preferences, etc. - it depends on what folks want in the model. Our original papers was mostly the technical blog post, but we decided to wait a little longer to see what else folks wanted and share more benchmarks

> not affiliated with subq,

i see in the linked post they mention O(n) not O(1). O(1) would basically be impossible and instant. Something like no compute required, constant results...

The name subquadratic is actually good and makes sense to me. Because today's models are usually O(n^2) or worse. Anything equals or less than O(n^1) is basically sub-quadratic.

Meanwhile O(log n) would be logarithmic as the log name indicates. But we have a long way to go there. Maybe with double tokenizer plus extensive caching it may be possible...

What I mean here is tokenizing the user input; then capturing intent; caching intent -> response. So that next time once you get the intent, you don't need to do full transformer inference compute. This can be logarithmic complexity in terms of time complexity.

I’m very surprised this isn’t getting more attention. Am I missing something?

It seems at or above SOTA on the given benchmarks, doesn’t have context rot, is orders of magnitude faster, and uses less compute that current transformer models. I suppose it’s just an announcement and we can’t test it ourselves yet.

The proof is in the pudding. At this point, there have been plenty of models that overperformed on benchmarks and underperformed on real work. So my stance is that I'm curious, I'm excited to see where it goes, and I don't believe it until I can try it.

I agree, it's a real architectural breakthrough if true

We are SOTA in some ways and not in others, continuously working to make it better! We need a little more time to scale, as we are working on things like disaggregated prefill, etc., the norms of large-scale model infra.

I am happy to answer any questions!

This seems super cool if as described, but I'm sure you can understand the skepticism.

Do you anticipate having any kind of public accessible chat interface for testing in the near future?

Also, what, if any, benefits are there for smaller context windows? Is there still a material improvement in cost to serve under say 256k? I'm curious about the broader implications for the space beyond improvements for very large context windows.

I do, for sure! Yes, we have a few product rollouts lined up. The differentials for latency are posted in our blog post, so that should provide an idea of where the scaling law differentials kick in.

> I do, for sure! Yes, we have a few product rollouts lined up.

When, more or less?

We will have a few rollouts in the next two months.

I have questions.

Can you back up your claims?

Why did you not release the white paper in parallel with the product?

Feels really fishy.

What do you want in a whitepaper that was not in our blog post? There is time to add more before the whitepaper is released.

I'm not GP, but I would want a benchmark that actually tests the entire context window. A benchmark that only tests the first 128K tokens effectively tells us nothing about how well it works at its full capacity.

That makes sense! We are working on that.

In this new knowledge economy, there is no benefit to publishing your secret sauce.

If I came up with a novel thing I'd monetise it first, because publishing it makes it part of the training that adds value to billion dollar corps with zero credit to me.

In the old knowledge economy I benefited from the credit assigned to me.

So, to me, nothing fishy at all.

no one has access to it yet

no published benchmarks

no paper

no demonstrations of capabilities

Yes you're missing something: the snake oil.

> Am I missing something?

Yes, this product doesn't exist.

And the last time a company claimed something similar it disappeared after taking money from investors.

Funny how they claim a 12M context window, yet all benchmarks are cherry picked with a 1M context window. Also, nobody has questioned how they did a training run before receiving funding. SoTA training runs cost well above $10M, yet no mention of funding prior to yesterday, interesting.

Whether this is real or not, multiple commenters here look like astroturfers - created in the past year (or hours) with very low karma

There are some comments which are identical to comments on X as well. That is not the say the frontier labs do not engage in highly unethical marketing, but this is a little bit too obvious.

I wonder how different their method actually is from other sub-quadratic sparse attention methods like Reformer [1] and Routing Transformer [2].

[1]: https://arxiv.org/abs/2001.04451

[2]: https://arxiv.org/abs/2003.05997

This is pretty remarkable. We've spent a lot of time finding workarounds for LLMs reading long docs. Now that's gone.

I'm usually okay with most LLM-assisted writing, but the amount of "it's not X. it's Y" style of phrases in https://subq.ai/how-ssa-makes-long-context-practical is disturbing.

Also, holy moly, the astroturfing.

But I'll still keep an eye on what they'll show up with in the next months. Sounds intriguing.

Don't let a C-suite marketing video blow your mind. They are trying to discover the new Transformer, that's not easy. 12 million token context with worse quality means this isn't going anywhere. Want to bet me bitcoin that we won't be talking about them in 1 year? Heck, they may have found something great, but the prior should be one of skepticism.

> The core idea is content-dependent selection. For each query, the model selects which parts of the sequence are worth attending to, and computes attention exactly over those positions.

I don't know if this will help for things like understanding code, where the all relevant parts can be the file of 1000 lines that we are analyzing, and where every token is relevant in understanding recursion, loops, function calls, etc.

This sounds like it would be great to do SSA before passing things along to a code model like claude code.

Let me know if I misunderstood

Yeah, tokens are excluded, only pairwise relationships between tokens. Coding is something we are looking at carefully!

Looks like long context isn’t a problem anymore

Neither is cost, and latency, in the long-term. LLMs ultimately become more economically viable than they are now, and broaden the scope of every existing LLM-driven application (particularly STS, conversational AI, etc, etc.)

No API access for independent verification - vaporware. See also comment about astroturfing accounts in this thread.

An architecture where compute grows linearly with context length seems dangerous. It can get very expensive as context grows and performance degrades

Reply on news.ycombinator.com