What if AI doesn't need more RAM but better math?

https://adlrocha.substack.com/p/adlrocha-what-if-ai-doesnt-need-more

by adlrocha • 4 months ago

189 101 4 months ago

adlrocha.substack.com

Lerc• 4 months ago

This is one of the basic avenues for advancement.

Compute, bytes of ram used, bytes in model, bytes accessed per iteration, bytes of data used for training.

You can trade the balance if you can find another way to do things, extreme quantisation is but one direction to try. KANs were aiming for more compute and fewer parameters. The recent optimisation project have been pushing at these various properties. Sometimes gains in one comes at the cost of another, but that needn't always be the case.

LoganDark2• 4 months ago

We will not see memory demand decrease because this will simply allow AI companies to run more instances. They still want an infinite amount of memory at the moment, no matter how AI improves.

jurgenburgen2• 4 months ago

If models become more efficient we will move more of the work to local devices instead of using SaaS models. We’re still in the mainframe era of LLM.

DeathArrow• 4 months ago

I don't think we are there yet. Models running in data centers will still be noticeably better as efficiency will allow them to build and run better models.

Not many people would like today models comparable to what was SOTA 2 years ago.

To run models locally and have results as good as the models running in data centers we need both efficiency and to hit a wall in AI improvement.

None of those two conditions seem to become true for the near future.

throwatdem123114• 4 months ago

The hyperscalers do not want us running models at the edge and they will spend infinite amounts of circular fake money to ensure hardware remains prohibitively expensive forever.

Ray202• 4 months ago

If models become more efficient

Then we can make them even bigger.

mustyoshi2• 4 months ago

I don't see how we'll ever get to widespread local LLM.

The power efficiency alone is a strong enough pressure to use centralized model providers.

My 3090 running 24b or 32b models is fun, but I know I'm paying way more per token in electricity, on top of lower quality tokens.

It's fun to run them locally, but for anything actually useful it's cheaper to just pay API prices currently.

delecti2• 4 months ago

As I understand this advancement, this doesn't let you run bigger models, it lets you maintain more chat context. So Anthropic and OpenAI won't need as much hardware running inference to serve their users, but it doesn't do much to make bigger models work on smaller hardware.

Though I'm not an expert, maybe my understanding of the memory allocation is wrong.

rainsford2• 4 months ago

We moved from the mainframe era to desktops and smaller servers because computers got fast enough to do what we needed them to do locally. Centralized computing resources are still vastly more powerful than what's under your desk or in a laptop, but it doesn't matter because people generally don't need that much power for their daily tasks.

The problem with AI is that it's not obvious what the upper limit of capability demand might be. And until or if we get there, there will always be demand for the more capable models that run on centralized computing resources. Even if at some point I'm able to run a model on my local desktop that's equivalent to current Claude Opus, if what Anthropic is offering as a service is significantly better in a way that matters to my use case, I will still want to use the SaaS one.

acuozzo• 4 months ago

But what about The Jevons Paradox?

redrove• 4 months ago

I disagree. I think a sharp drop in memory requirements of at least an order of magnitude will cause demand to adjust accordingly.

cyanydeez2• 4 months ago

Department of Transportation always thinks adding more lanes will reduce traffic.

It doesn't, it induces demand. Why? Because there's always too many people with cars who will fill those lanes.

jLaForest• 4 months ago

Jevons paradox https://en.wikipedia.org/wiki/Jevons_paradox

rainsford2• 4 months ago

I'm not sure that's infinitely true as long as AI costs to the user are proportional to the cost it takes to run the model. Even if user costs are heavily subsidized by investment, as long as they are non-zero and go up when models cost more, there will be at least some pressure for cheaper models and not just more capable ones and that pressure will go up with costs. AI is a crazy industry, but it's not totally immune to the law of supply and demand.

The real question though is how close are we to the point where the pressure is more for efficiency rather than capability. Anecdotally I think it's a ways off. Right now the general vibe I get is that people feel AI is very impressive for how cheap it is to use, which suggests to me that a lot of users would be very willing to pay more for more capable models. So the tipping point where AI hardware demand might slow down seems a ways off.

LoganDark2• 4 months ago

It has yet to be seen when hyperscalers will change their tune.

fph• 4 months ago

Despite the shortage, RAM is still cheaper than mathematicians.

captainbland• 4 months ago

I don't know, I think if you weighed up the costs of AI related datacentre spend vs. the average mathematics academic's salary you could come to a different conclusion.

iLemming2• 4 months ago

Raising, nurturing, training, and mentoring an expert mathematician is not cheap; it never was, perhaps the first time in history when we can witness that rule to change - spinning up a bunch of math-savvy agents, each smarter than Ramanujan maybe will get too cheap.

Almondsetat3• 4 months ago

At the same time, processing is much cheaper than memory

gunalx• 4 months ago

Without memory you have no data to compute on. Memory and compute scaling only makes sense in tandem.

mandeepj• 4 months ago

But not everyone has to pay mathematicians, like RAM :-)

_fizz_buzz_• 4 months ago

Doubt it. You have to pay these mathematicians once and then you can deploy to millions of sites.

Verdex2• 4 months ago

It's also less frustrating to organize world wide ram production and logistics than to deal with a single mathematician.

Constantly sitting around trying to solve problems that nobody has made headway on for hundreds of years. Or inventing theorems around 15th century mysticism that won't be applicable for hundreds of years.

Now if you'll excuse me I need to multiply some numbers by 3 and divide them by 2 ... I'm so close guys.

Eddy_Viscosity2• 4 months ago

The comment feels a bit like Verdex may have dated a mathematician at some point and it went sour.

abdelhousni• 4 months ago

The same could be said about other IT domain... When you see single webpages that weight by tens of MB you wonder how we came to this.

Yokohiii2• 4 months ago

Detachment from reality. Code elegance is more important then anything else. As simple as that.

LtWorf• 4 months ago

You've never seen the sources for such pages I presume?

tornikeo• 4 months ago

Sigh. Don't make me tap the sign [1]

[1] http://www.incompleteideas.net/IncIdeas/BitterLesson.html

staminade• 4 months ago

Doesn't seem relevant here. TurboQuant isn't a domain-specific technique like the BL is talking about, it's a general optimisation for transformers that helps leverage computation more effectively.

amelius• 4 months ago

Can we say something about the compression factor for pure knowledge of these models?

konaraddi• 4 months ago

applying this compression algorithm at scale may significantly relax the memory bottleneck issue.

I don’t think they’re going to downsize though, I think the big players are just going to use the freed up memory for more workflows or larger models because the big players want to scale up. It’s a cat and mouse race for the best models.

Verdex2• 4 months ago

Known in the business as 'pulling a jevons'

miohtama• 4 months ago

It will also help with local inference, making AI without big players possible.

otabdeveloper42• 4 months ago

It's already possible. Post-training is vastly more important than model size. (There's bigtime diminishing returns with increasing model size.)

alienbaby• 4 months ago

Ive thought for a while that the real gains now will not come from throwing more hardware at the problem, but advances in mathematical techniques to make things for more efficient.

mustyoshi2• 4 months ago

The drop in memory stocks seems counterintuitive to me.

The demand for memory isn't going to go down, we'll just be able to do more with the same amount of memory.

yorwba• 4 months ago

It especially doesn't make sense considering that TurboQuant has been public on arXiv for almost a year: https://arxiv.org/abs/2504.19874 So it predates the late-2025 RAM price surge! https://pcpartpicker.com/trends/price/memory/

I think that either investors were extremely skittish that the stocks might crash and jumped at the first sign of trouble (creating a self-fulfilling prophecy) or they were trading on non-public information and analysts who don't have access to said information are reading too much into the temporal coincidence of the Google Research blog highlighting this paper.

zug_zug3• 4 months ago

Well, when a companies have 100billion dollar incentives to make discoveries like this, I don't know if we should assume this is the only optimization that will happen.

Given that increasing model size doesn't yield proportional increases in intelligence, there is a world where these datacenters don't have a positive ROI if we make these models even a fraction as effective as the human brain.

boshalfoshal• 4 months ago

Well considering basically the entire market was down these past few days, Google included, its unlikely attributable to this paper alone. Its most likely correlated with general war/trade route restrictions/potential recession fears, or at least, more correlated with those than it is with this paper.

This paper was released a year ago and was probably part of how google got to 1m context before other labs.

Yokohiii2• 4 months ago

If I were Google, I wouldn’t release research that exposes a competitive advantage.

Isn't that a classic tit for tat decision and head for a loss?

Excellence and prestige are valuable too. You get those expensive ML for a small discount, public/professional perception, etc. Considering the public communication from Google, that isn't complete sociopathic, they know this war isn't won in one night, they are the only sustainably funded company in the competition. Surely they are at risk with their business, but can either go rampant or focus. They decided to focus.

barbegal• 4 months ago

Does the KV cache really grow to use more memory than the model weights? The reduction in overall RAM relies on the KV cache being a substantial proportion of the memory usage but with very large models I can't see how that holds true.

zozbot2343• 4 months ago

For long context, yes this is at least plausible. And the latest models are reaching context lengths of 1M tokens or perhaps more.

simne• 4 months ago

Sure, we need better math, it is obvious.

Unfortunately, nobody at big companies know, what exactly math will win, so competition not end.

So, researchers will try one solution, then other solution, etc, until find something perfect, or until semiconductors production (Moore's Law) made enough semiconductors to run current models fast enough.

I believe, somebody already have silver bullet of ideal AI algorithm, which will lead all us to AGI, when scaled in some big company, but this knowledge is not obvious at the moment.

Bydgoszczo• 4 months ago

And maverick 2

exabrial• 4 months ago

I was thinking it needs speciality hardware. Sort of like how GPUs were born…

chr15m2• 4 months ago

Is this something that will show up in Ollama any time soon to increase context size of local models?

zozbot2343• 4 months ago

KV quantization has long been available in llama.cpp

chr15m2• 4 months ago

Yes but the optimisation described has not right?

imjonse• 4 months ago

"The TurboQuant paper (ICLR 2026) contains serious issues in how it describes RaBitQ, including incorrect technical claims and misleading theory/experiment comparisons.

We flagged these issues to the authors before submission. They acknowledged them, but chose not to fix them. The paper was later accepted and widely promoted by Google, reaching tens of millions of views.

We’re speaking up now because once a misleading narrative spreads, it becomes much harder to correct. We’ve written a public comment on openreview (https://openreview.net/forum?id=tO3AS KZlok ).

We would greatly appreciate your attention and help in sharing it."

https://x.com/gaoj0017/status/2037532673812443214

_0ffh• 4 months ago

Openreview link is not working, was split apparently.

https://openreview.net/forum?id=tO3ASKZlok

zug_zug3• 4 months ago

I guess I'm trying to understand. I'm hearing this paper has been around for a year -- I would think that many companies would have already implemented and measured its performance in production by now... is that not the case?

zug_zug3• 4 months ago

Okay, I spent about half an hour reading about this and asking gemini I guess my best understanding is this:

The main breakthrough [rotating by an orthogonal matrix to make important outliers averaged acrossed more dimensions] comes from RaBitQ. Sounds like the RaBitQ team was much more involved, and earlier, and the turbo quant paper very deliberately tries to avoid crediting and acknowledging RaBitQ.

My understanding is that the efficacy of these methods isn't in dispute, what turboquant did was adapt the method that was being used in vector databases and adapted it for transformers, and passed it of more as a new invention than an adaptation.

PaddyLena• 4 months ago

I think the biggest issue isn’t the tool itself, but access and stability. I had more trouble finding reliable AI accounts than using them tbh

signa11• 4 months ago

why not, you know, just use LLMs to do this job ?

SphericalCowww4• 4 months ago

I mean, since GPT-4, I believe the RAM is no longer creating the miracle that the LLM performance scales directly with the model size. At least ChatGPT itself convinced me that any decent-sized company can create a GPT4 equivalent in terms of model size, but limited by service options, like memory cache and hallucination handling. Companies buy RAM simply to ride the stock hype.

I am no expert, so this is a shallow take, but I think the global LLM already reaches its limit, and general AGI could only be possible if it's living in the moment, i.e., retraining every minute or so, and associating it with a much smaller device that can observe the surroundings, like a robot or such.

Instead of KV cache, I have an idea of using LoRA's instead: having a central LLM unchanged by learning, surrounded by a dozen or thousands of LoRAs, made orthogonal to each other, each competed by weights to be trained every 1 min say. The LLM, since it's a RNN anyway, provides "summarize what your state and goal is at this moment" and trains the LoRAs with the summary along with all the observations and say inputs from the users. The output of the LoRAs feeds back to the LLM for it to decide the weights for further LoRAs training.

Anyways, I am just thinking there needs to be a structure change of some kind.

redanddead• 4 months ago

share it on gh and make a show hn post about it, maybe you're right

the models are still very stupid atm something needs to change

SphericalCowww4• 4 months ago

I put all the conceptual ideas here (spiced with far-fetched claims): https://github.com/SphericalCowww/ML_LunaLoRA

But I think my hn level is too low, would really like some expert's opinion, though, whelp... Meanwhile, let me implement some basic models first.

fittingopposite• 4 months ago

Re continuous fine-tuning: how do you avoid catastrophic forgetting in your proposal?

SphericalCowww4• 4 months ago

My understanding is that this is what the LoRAs are for; my belief is that they serve as "memory" to their live observations (a more NN-like cache, say), while the main LLM remains unchanged. These LoRAs are also weighted, so that LoRAs irrelevant to the current task will not be trained, while the relevant LoRAs will be reinforced.

But I never built it, so I am not sure if such an emergent state will appear or not.

SphericalCowww4• 4 months ago

I put my ideas here in case you are interested: https://github.com/SphericalCowww/ML_LunaLoRA

effnorwood• 4 months ago

this is exactly correct.

Skunkleton• 4 months ago

The TurboQuant paper is from April 2025. I’m sure the major labs knew about it on, or even before, the day it published. Any impact it had would have been a year ago. Yet I keep seeing these posts and discuss completely ignoring this.

Can we please start talking about this in that context? We already know what TurboQuant will do to DRAM demand. We already know what it will do to context windows. There is no need to speculate. There is no need to panic sell stocks.

am17an• 4 months ago

There are techniques which already achieve great compression of the cache at 4 bit, eg using hadamard transforms. Going from 4 bit to 3 bit isn’t the great leap people expect this to be. It’s actually slower to run and is generally worse in practice.

mxmlnkn• 4 months ago

The obvious one outside of KV caches as mentioned above is vector databases. Any RAG pipeline that stores embedding vectors for retrieval benefits from the same compression. TurboQuant reduces indexing time to “virtually zero” on vector search tasks and outperforms product quantisation and RabbiQ on recall benchmarks using GloVe vectors.

This part sounds especially cool. I did not think about this application when reading the other articles about TurboQuant. It would be cool to have access to this performance optimization for local RAG.

convexly• 4 months ago

There's a bunch of research showing that more/better information doesn't reliably improve judgement, but better feedback on your existing predictions does. Makes me think of Soros and his whole thing about reflexivity.

What if AI doesn't need more RAM but better math?

Related Stories