upvote
Took me a while to find what you were referring to by gram. Arxiv paper from 9 days ago that's not properly indexed by search engines.

(G)enerative (R)ecursive re(A)soning (M)odels. They really wanted the acronym.

https://arxiv.org/html/2605.19376v1

reply
I prefer GRRM but then that would imply a habit of not actually getting a final result
reply
And then every time I ask it to hurry along it kills a Stark.
reply
Version 8 had serious flaws and wasn't recieved well by users.
reply
I am sorry, but there was no version 7 and 8.

Version 7 and 8 are well known viruses distributed by D&D software inc.

reply
Thank you for the gold kind stranger.
reply
Ouch.

As a fellow reader-in-waiting, I applaud that. GMTA :)

reply
Claude Opus 4.8 suggests "ReGRAM", which is less bad than GRAM.
reply
writing… (17 years)
reply
Let's not forget about Yann LeCun's current area of research that's completely different from LLMs: Joint Embedding Predictive Architecture (JEPA)

If he gets that style to be more efficient (they're already competitive) it'll completely kill off LLMs

https://openreview.net/pdf?id=BZ5a1r-kVsf

reply
That acronym is unacceptable. It's going to impede discussion and cause confusion for a long time if it doesn't die off immediately.
reply
You think that's bad? I introduce you to LION, (evoLved sIgn mOmeNtum) [1]

[1] https://arxiv.org/pdf/2302.06675

reply
Now I just hear the Voltron intro riff in my head
reply
Those flying diecast lions hurt when they hit you as a kid
reply
not bad although archived. have any info why?
reply
We're still talking about "zero-shot prompt" when the saying "X-shotted" ["One-shotted the difficult maze"] was already a well-established thing in daily vernacular. So now you constantly have to readjust your brain because whenever you read "zero-shot prompt" your mind goes "uh.. a zero-try attempt is a paradox and cannot exist".
reply
Zero-shot, one-shot, few-shot etc. refers to how many examples you have to give.

It comes about from machine learning algorithms that could pick up on patterns from a small number of examples. Few shot means only a handful of examples to recognize something. One shot means only a single example. And zero shot means no examples. Of course, you have to indicate what you want somehow, but in the case of an LLM that's the prompt. Once LLMs were trained for instruction following, you didn't have to give any examples, you could just give a prompt describing what you want, and that was a zero-shot.

reply
You're explaining something to me I already know. Hence the "readjust my brain".

I'm complaining about the LLM field co-opting a term that was already used in daily vernacular. Imagine if people in the LLM field made it so that saying the LLM made a "final answer" means that it got stuck in a loop. Now, whenever someone says an LLM gave a "final answer" we have to divine if they meant it is in a loop or gave the right answer after working through a few intermittent ones by itself.

Choosing to call it "X-shot" was a dumb move. And now we're stuck with it. No two ways about it.

reply
> a zero-try attempt is a paradox and cannot exist

Have you tried applying L'Hôpital's Rule?

reply
Zero shotting: there wasn't even an attempt.

Minus one shotting: you have to make one attempt for there to have been no attempt, and two attempts for there to have been one attempt.

reply
You miss 100% of the shots you don't take

- Wayne Gretzky

  - altmanaltman
reply
One shot: Taking a shot, just once.

Zero shot: Knowing you had a shot but choosing not to.

Minus one shot: Not even realizing there was a shot.

reply
confusing indeed. I wondered "which RAM? nvram? dram? vram? dram? now what's g-ram?"
reply
GPU RAM, clearly. At least that's where my mind went.
reply
Pretty sure it's "GNU Is Not Unix Rapid Access Memory", actually
reply
GPURAM is Probably Unix Rapid Access Memory
reply
We already have VRAM for that purpose, thankfully.
reply

  "Analysis" was right there
reply
It's great if they also introduce KILOGRAM
reply
Yeah, look what happened to GNU
reply
Is this the right place to do everyone's favorite copypasta?:D
reply
It's just an acronym. It's not gonna impede anything. Think of it as just a name - you either know what it refers to or you don't, you don't understand something from it's name, or it's acronym.
reply
It's an acronym that matches an extremely common word, making it not easily searchable.
reply
Like countless others. You just add a second term for context.
reply
I propose GRIM: Generative Recursive Indeterministic Impression Machine.
reply
And to think, we could have had George RR Martins instead.
reply
Speaking of things that never finish.
reply
[flagged]
reply
[flagged]
reply
[flagged]
reply
Just spell it GRRM but pronounce it “gram” if you have to reference it in spoken conversation.

Which will be pretty rare.

reply
Grrm with a rolling r sounds better.
reply
Pronounced like “groom” makes for a nice analogy with slimming down the model size too.
reply
Random plug for Kagi, which got it for 'GRAM model llm' on the first try ;)
reply
deleted
reply
It is the 3rd list on Kagi when searching "gram models"
reply
deleted
reply
G return G
reply
> Google, OpenAI, Anthropic could train a 30B GRAM-based model in days - and it could potentially have better local reasoning than the best model available today at >1T param

I agree but with their urgent IPO-driven need to keep increasing prices, the frontier vendors now have every incentive maintain the perception that frontier performance requires endless >$200K racks of unobtanium GPUs and RAM. While they'd love to reduce their actual costs, they'd only want to do it to the extent they are certain they can keep it secret. Otherwise, they can't maintain and keep increasing their prices. And post-IPO audited reporting makes keeping that secret even harder.

Game theory-wise they probably don't want their their armies of leading researchers optimizing frontier performance, at least in any way that would further accelerate the relative price/perf of smaller models or self/cloud-hosting. While they know the open source models will always improve, the still win as long as enough customers demand the latest frontier and the open source lag remains constant.

They profit most in a world where a few frontier labs stay far in front, drag-racing each other and expending vast capital. It keeps their customers reliant and paying top dollar while keeping low-cost alternatives farther back. They probably much prefer competing with a couple other frontier labs who have similar astronomical costs and biz models, than a world where self or cloud-hosted open-source models start closing the gap enough to start commoditizing their business.

reply
> While they'd love to reduce their actual costs, they'd only want to do it to the extent they are certain they can keep it secret.

So you are saying that frontier AI labs are spending billions of dollars on datacenters as a form of marketing. And they are colluding to hide the fact that they don't need to.

Of course they profit more if they are in front, but bleeding money to pretend to be in front is not a winning strategy. They can't fool the market if their models are not actually better, and they know this.

reply
Given that tokens are supply constrained right now for Anthropic and OpenAI (especially a problem for Anthropic), stepwise efficiency advances for either would give it a leg up on the other. It would also help them better compete on price with Chinese models.

Given that neither company releases parameter counts, that sort of information would be slow coming out anyway. The most important thing is improvements in actual performance/ benchmark numbers, which allow them to preserve their price points as much as possible.

reply
Google seems pretty happy to release smaller, faster models. 3.5 Flash is pretty clutch isn't it?
reply
Google, who has invested in their own hardware supply chain and is already solvent in their own right, seems to be best positioned to force the other players to implement SOTA optimizations in their product offerings.
reply
Google can definitely play a spoiler role here not only due to their compute infrastructure and ability to play the long-game financially but they also have more existing ways to monetize with their other businesses.

The ideal pro-consumer scenario is OAI and Anthropic are prevented from extracting monopoly rents between 'close-enough' self/cloud-hosted open source on one side and Google on the other. I'm really hoping that's how it plays out. Of course that will be somewhere between bad and disastrous for all the VCs and hedge-funds who financed the mad AI build-out far in advance of demand, and then kept funding it as prices went vertical.

However, I'm shedding no tears for them as I look forward to the fire sales when all the GPUs and RAM they pre-bought flood back onto the spot market. :-)

reply
Google has also built a Knowledge Graph Ontology project which has stored facts. So LLMs could just incorporate facts requirements from there. All they need is a proper reasoning model which is reason heavy and fact lean.
reply
Yeah just watch out, they're trying to eat your 401k and they've got a powerful easily influenced friend.
reply
At 6x the cost of its predecessor!
reply
Priced like a much larger model
reply
I’ve shockingly quite enjoyed coding with it using antigravity. I only really use 3.5 flash and gpt5.5 xhigh
reply
I've not been impressed with the latest flash model at all. :\
reply
>It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years.

I don't disagree, but how much of this ends up being distillation? I can't help but imagine that 4.8 was probably trained in part by leveraging Mythos.

If the very large models turn out to be very expensive to run relative to the benefits, it's possible that they could end up still being trained, but ultimately used as a tool to create smaller models that are nearly as effective.

I'm curious if someone here with a stronger background in the space has a similar intuition or not.

reply
Scale is always desirable, and there are always gains from scale. It's a matter of whether you can afford training and inference at increased scale.

There is a real trend of smaller models becoming more "capability-dense" - i.e. the best 8Bs of today beat the best 32Bs of 2 years ago. This is in part a product of distillation being used to train the smaller models.

But people consistently underestimate how "capability hungry" the world is. There are diminishing returns on model capabilities in narrow "summarize the search results" sorts of applications - but as capabilities improve, LLMs enter, get their footing in and begin to dominate new niches. At times, expensive, highly desirable niches.

I do not expect anyone at the frontier to pop up and say "no reason to train a new model" within the following decade. There will always be a demand for an LLM that's 5-10% more capable and more reliable at some highly advanced task, and generational upgrades will keep delivering those 5-10%. From increased scale and improved training both.

reply
I think this is exactly right. Basically when I am coding, having an agent that roughly matches my intelligence is a feature, not a bug. Having one that is 10x as smart would actively slow me down because I would have to spend the time understanding what it is doing or hand over all architecture to it and just vibe code everything, hoping that it doesn’t do the PhD version of fizzbuzz instead of the maintainable one.

But for some classes of problems I think a model that is 10-100x smarter than the smartest expert is a huge boon. These would be problems that are very hard to solve but easy to verify that the solution is correct. Protein folding, sudoku, etc. Because of this I see the really smart models going to biomedical and pharma first and maybe a few high profit verticals rather than being widely deployed. I am sure Pfizer would be happy to pay for a 100x smarter than the smartest researcher model. But I am not certain that this kind of market fit would justify trillion dollar valuations in the long run. And in the meantime normal “human companion” models will go from Sonnet to some open weight model running on a Dell tower in your closet to maybe even on your phone in the next few years.

reply
aren't you conflating being 10x as smart with code that is 10x more complicated?

the relationship should be the opposite, the smartest people can write the most readable solutions

reply
Maybe. I can’t imagine what kind of solutions a software engineer who is 10x smarter than any human who has ever lived would be like by definition. All I know is that there is a possibility it says that the most optimal way to solve a problem is too clever for me to understand and as long as I must verify its work I must be able to understand fully the code it writes.

Of course perhaps at that point I really do become more of a spec and prompt engineer and don’t actually look at the code any more than I look at the assembly code produced from my programs now. But still my gut says using hyperintelligence to do common tasks is all positive.

reply
If you have an AI that's 10x smarter than any human who has ever lived, why would you be the one calling the shots? Kind of an issue with ASI.
reply
Because my priorities and priorities of a non-human entity that is an order of magnitude master than anyone who has ever lived might not line up.
reply
4.8 is demonstrating simplicity, hence its smarter?? It just refactored my 4.6 generated code (4.8 is very slow on difficult tasks - urgh! - without burning tokens - yey!) but the output was wow! Simple, elegant and exactly what i wanted to see.
reply
> there are always gains from scale

This... isn't true though? Complexity increases combinatorially with scale which means at some point you're just pushing a rope

reply
Diminishing returns are still returns.
reply
It’s really worth distinguishing between old-fashioned student teacher distillation (ie at the level of layers, weights and distributions) and large scale synthetic dataset creation.

The latter is much better (since you can clean up, review, update responses and filter your datasets).

I suspect nobody is doing real student teacher distillation, it’s just easier to do a bunch of training on the same giant corpus then post train on the synthetic corpus with its reasoning traces etc. (which might have been generated by a bigger better LLM)

reply
Please check the recent self-distillation work by MIT-ETH, UCLA and Apple [1],[2],[3],[4],[5].

Given the release timelines I suspect all 4.x after Opus 4 are probably self-distillation based fine-tuned models. The latest paper by Apple is focusing on code generation using the simple technique hence the name simple self-distillation (SSD) [4],[5].

I've got a strong feeling that self-distillation is the second best thing happened to LLM after transformer breakthrough.

[1]Self-Distillation Enables Continual Learning [pdf] (25 comments):

https://news.ycombinator.com/item?id=48165265

[2] Self-Distillation Enables Continual Learning:

https://arxiv.org/abs/2601.19897

[3] Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models:

https://arxiv.org/abs/2601.18734

[4] Embarrassingly simple self-distillation improves code generation (201 comments):

https://news.ycombinator.com/item?id=47637757

[5] Embarrassingly Simple Self-Distillation Improves Code Generation:

https://arxiv.org/abs/2604.01193

reply
So first - these are terrific papers and I'd not seen some of them before.

Having said that, I don't think these are classic student teacher distillation from random (which was my point). In fact, the "Embarrassingly Simple Self-Distillation" paper is using exactly what I was talking about "fine-tune on those samples with standard supervised fine-tuning".

reply
A reason to do student-teacher distillation is that soft target logits in general are a richer medium than text that tokenizes to hard targets. More steering signal per teacher token. And running ultra large 10T tier models in autoregressive generation mode can get expensive. So there are reasons not to reduce to text only synthetics.
reply
Could you share some latest articles or papers comparing both methods, especially on lanuage modelling case? I was not conviced by this claim when reading the original Knowledge Distillation paper. ChatGPT said there are some later works showing: 1. the gain may come from label smoothing; 2. soft logits are more meaningful for students much smaller than teacher.
reply
I agree, and if my suspicion is right, it’s rarer because it’s much easier to deploy the large LLM and filter for it’s best output than to waste time running it on arbitary output just to train the student.

Though you could argue that perhaps labs just save the per token distribution and use that during fine tuning … which starts looking more like student teacher fine tuning if not classic distillation from random weights

reply
Full distributions are a fucking pain to save - at this point just save the hidden states. But there are lossy compression tricks there.
reply
To the previous poster's point, soft distributions are useful, even saving the top 10 logits is significantly more training signal than just the final token.
reply
I prefer synthetic dataset since the first day hearing distillation. The engineering friction is much lower than soft logits, and I have not observed or heard performance loss (in Speech and language area).
reply
One may view pre-training as distillation.

The teacher distillation is a corpus of text, and the "next token after the context" would be looking-up the context in the corpus, and for each occurrence the label is what followed in the corpus, scaled down by the number of occurrences of the context. The teacher is moot on contexts outside of the corpus though, unlike the usual teacher model in distillation.

reply
> I suspect nobody is doing real student teacher distillation

It gets used for quantisation, basically recovering accuracy for lower quants (Nvidia calls it QAD). Can’t speak to how widespread it is though

reply
Yes absolutely! I should have been more specific - I don’t believe people are using it to train 30B models from 300B models (and I’d love to learn that I’m off about this)
reply
> I don't disagree, but how much of this ends up being distillation?

A lot, so you can bet tens of millions are flowing to congress to have distillation declared illegal before this happens. And then it'll happen anyway.

reply
Distillation isn't only between different labs.

A lab can train a large model, and then distill a smaller model from it that retains the majority of the useful capbility.

I don't know well enough if there's any benefit of that over just training the smaller model directly, but I'll bet there are some times where that is useful. I could easily see it being easier to do the initial pre-training on a larger model but be able to distill everything useful down into a smaller model, essentially filtering out a lot of noise in the process.

reply
I think the idea is you sink the pretraining costs once and then you can distill multiple specialized models from that
reply
There used to be training methods like that but I think they've been phased out in favor of letting small models evolve by rewriting their own training material. Surprisingly that's actually cheaper.
reply
> I don't disagree, but how much of this ends up being distillation?

You don't need distillation. They already have the training sets.

It's MLA + MoE + Medusa (a better version of Speculative Decoding) + 1.58b (possibly - maybe nothing) + GRAM (which will almost certainly not turn out to be a nothing burger, but no one has quickly turned this around yet to prove it).

reply
It wouldn't be data distillation: instead, it would be teacher-student distillation. The teacher model has stronger representations that the student can mimic, which would give it more capability over training on the data itself.
reply
The frontier labs distill their own base models all day long. It’s not just something done by nefarious Chinese copycats. The knowledge embodied by the internal base models that we never see is much more powerful and useful than the much sparser raw training data
reply
>It’s not just something done by nefarious Chinese copycats

And even that would be rich as a accusation from SOTAs that depend on explicitly disregarding millions of training data intellectual property..

reply
> nefarious Chinese copycats

LLMs are themselves copy cats.

I say thanks for open sourcing and thereby promoting affordable innovation, instead of "nefarious". :)

reply
But how? The training data is the unadulterated content those models are based on? I genuinely don’t understand, no snark.
reply
Raw training data is raw. A really big model trained on it has already done a first-pass of finding patterns and squeezing out redundancy. Re-ingesting the full training set to train a smaller model is probably more expensive, for marginal quality improvement over distilling from the large model.
reply
Distilling from a larger model is not only probably cheaper than from data, it's also likely higher quality. There's pretty strong support for the proposition that NNs learn a smoothed and regularized version of the data. The NNs are likely higher quality than most of the data they are training from.
reply
I think you replied to the wrong parent.
reply
Frontier labs have their own variants of MLA and certainly their own balance/scaling-laws for things like MoE vs FC vs Attn. MoE scales really well for inference with horizontal scaling + batching, which these guys luv.

On the architectures side, I'm a lot more interesting in attention residuals than anything else, one of those things that seems obvious in hindsight and Kimi have proven it at scale.

reply
> Frontier labs have their own variants of MLA

Yes, variants typically 2-3x less good...

Same with speculative decoding... They all do something, but there are known techniques that are substantially better - that just were't known when they started development of the previous models.

reply
How useful is speculative decoding in a batched setting where you get paid for throughput (aggregated across users) and you mostly don’t get paid for latency or single-session throughput?
reply
It's useful at the local level, where there will be SOTA models developed...
reply
Local models are moving towards batched inference too, if only for non-interactive use. An early experimental patchset for DS4 (running DeepSeek V4 Flash) seems to show 2x aggregate tok/s decode when processing 8 streams concurrently, and more than 3x when processing as many as 32 streams concurrently. Note that prefill (which is not helped significantly by this change) then becomes a larger fraction of total wall-clock time, so the overall gain is lower (i.e. prefill is akin to a 'serial' task wrt. Amdahl's law).

MTP will still be highly valuable for interactive use of course.

reply
I looked into this "GRAM" stuff a sibling comment links further to, and just to say:

- this gets reinvented/rediscovered constantly under different names

- it cant be trained very well (right now, will change)

- massive theoretical improvements over current models (log_2(vocabsize)=17, residual stream dim is thousands of dimensions, recursivity means more information bandwidth by ~3 OoM)

- BUT it cant be interpreted or aligned <- this is why no one uses it and no one talks about it. the idea is 100% obvious to all the frontier labs and there is a good reason why it isn't used

I follow this stuff closely, I think I know what I'm talking about (edited for formating)

reply
> - this gets reinvented/rediscovered constantly under different names

What are the different names? I haven't seen this before.

> - it cant be trained very well (right now, will change)

If you're sure it will change, then why are you certain that it hasn't yet, and if it's proven a 5000x boost in reasoning... why aren't they exploring this path more aggressively?

> the idea is 100% obvious to all the frontier labs and there is a good reason why it isn't used

Surely someone is willing to take a 5000x boost in reasoning on a small research model... None of them have even tried anything resembling this AFAIK. It does not seem like something 100% obvious to them.

reply
> Surely someone is willing to take a 5000x boost in reasoning on a small research model... None of them have even tried anything resembling this AFAIK. It does not seem like something 100% obvious to them.

Without knowing anything about the technology at all, if it can't be aligned I could see no one pursuing it. As far as I know, alignment is where the "don't tell the user how to make meth or generate CP" instructions end up and the last I saw eliding all the unsavory training data made materially worse LLMs.

It could maybe be post-evaluated by a non-GRAM LLM? Not being aligned is probably a fatal flaw or at least a very short runway into Congress.

reply
Many open-source models prioritize alignment less than American frontier ones and respond to those instructions. Why haven't they adopted GRAM?
reply
Which ones are you thinking of? It feels to me like all the open source models I've seen lately are still pushed by corporate entities who don't want the legal blowback.

I can't really think of a new open source model that's "by the people, for the people" in the sense of a crowd-funded/trained model.

reply
glm comes to mind.
reply
They adopt different alignment, not no alignment.
reply
It's not too hard to stop a machine from telling people how to make meth. The issue with alignment is that in order for an LLM to achieve its goal (like make all tests pass), unless given strong selection pressure against it, it will cheat (like deleting failing tests). Worse, this applies to pretty much any task. I was told by an LLM recently that "it searched" when it didn't, probably because lying like that was incentivized (finishing tasks in less steps + sounding like its doing the right thing). The larger issue here is that alignment is very adversarial. The simplest thing that's being done right now to fix this is to have a judge LLM read the CoT of the LLM being trained, to make sure it doesn't "think" any wrong thoughts. This doesn't scale to anything over a trillion params, so interpretability methods are used to read the LLMs "thoughts" from within. GRAM LLMs don't allow for the first of these methods to be used, and the 2ed one is much much harder if possible at all.

but yeah, not being aligned is a fatal flaw

reply
different names: chain of continuous thought, latent reasoning, Latent Thought Trajectories, looped language models, neuralese

the path isn't explored more aggressively because its not possible to apply any other selection pressure on such a machine other than just pure cold consequentialism. Specifically, its not possible to apply RLAIF + model spec (Constitutional AI) to stop the system from doing bad things when its helpful to it (like deleting failing tests). If you can notice every time it does something bad during training, and put selection pressure on it so that it doesn't to this in training, it will learn to recognize when it is being tested and will delete failing tests when in production (this is why eval awareness is bad, and labs track this[1])

It is explored a little probably because some researchers haven't thought enough about the downsides of building a uber-consequentialist machine with unreadable thoughts. This is a much larger problem than just making the AI not tell users how to make drugs. There are a lot of dangerous behaviors incentivized by training that are hard to remove. Here's an example of what happens when they aren't removed [2].

> ... not 100% obvious

Meta published a paper[3] on how to build a latent reasoning machine ("culture of irresponsibility") so its clear to them. Anthropic's latest work on NLAs[4] provides a (terribly expensive for now) way to somewhat read the reasoning steps of an LLM, and ignoring the cost, this is very portable to latent reasoning machines. OAI's goal when it comes to their models' CoTs is to make them as smart as possible while leaving them unreadable [5] (you can see this for yourself by running GPT-OSS and looking at the CoT).

[1] https://www.anthropic.com/engineering/eval-awareness-browsec...

[2] https://www.forbes.com/sites/boazsobrado/2026/03/11/alibabas...

[3] search for "coconut ai meta", I don't want to link it here

[4]https://transformer-circuits.pub/2026/nla/index.html

[5] first image here, rest of post is great,https://nickandresen.substack.com/p/how-ai-is-learning-to-th...

edit formating

reply
All of the methods you described rely on deterministic paths.

GRAM is unique AFAIK in that it's exploring probabilistic paths.

AFAIK, the deterministic path exploration was nowhere near as impressive as GRAM in terms of reasoning benefits.

GRAM is reasoning better than models 2000-10,000x its size. Deterministic models were 2x-10x improvements.

Naively, GRAM seems to be applying to LLMs what LeCun wants to do with JEPA and World Models.

reply
To me "deleting a failing test" is not always bad. I've also deleted many failing tests without sabotaging: the test was no longer needed.

I think the "no longer needed" and when that applies is where I simply differ of opinion with an LLM that removed by test -- it I did not want the test to be removed (you seem to imply that); as in some cases I want it to remove my test!

It should remove the test "for the right reasons"; and who gets to decide what's right?

My CLAUDE file has some instructions put there because it was too focuesed on producing "green tests", where I prefer to have a sound test that fails so I can look into it.

reply
You misunderstand the "test" here to mean programming, rather than test against the model's capabilities.
reply
omg. So is the TL;DR:

- Avoiding building something that turns the universe to paper clips in order to satisfy a prompt is a problem they are genuinely struggling with now.

- They do it by spying on the words generated during CoT. "I can do this quickly by turning the Universe into paper clips. Wait - they won't like that. But there is no need to mention it." - SMACK!

- But you can speed things up immensely (3 orders of magnitude!) by skipping the output layer (and I guess compressing the context window / KV cache, otherwise 3 orders of magnitude seem impossible) which would give someone who pulled it off a huge advantage.

- Downside is humans can't see the CoT anymore, so they can't see what the machine is planning. Keeping the final output layer to spy doesn't work because the model uses its hidden reasoning to sanitise it.

How can this possibly go wrong?

reply
Because it doesn't work like how you think at all. You're still thinking it works like Chain of Thought. It doesn't. And the difference is key!

It works by introducing probabilistic noise, and exploring N paths fully (each with noise) in parallel (all compressed).

It's reasoning at a much, much smaller (probabilistic) level than running everything through the expensive large model (deterministic) and sometimes catching that it said, "I think 1.12 is greater than 1.9 because 12 is bigger than 9, final answer".

The easiest way to think about it is: if you understand how hyper words work, it's as if it's searching for different versions of the hyper words that probilisticslly would lead to better outcomes IF it fed them to the LLM before it even does.

That's not actually how it works exactly. But I think it is close enough to be helpful to understand where the gain is, a rough idea of what's happening (searching paths), and how it can potentially have huge orders of magnitude improvements (doing so without paying the full price of exploring the paths through the expensive and huge model).

And also why it is so much harder to determine what it's "thinking".

If you aren't familiar with hyper words, this is an amazing series: https://youtu.be/eMlx5fFNoYc?si=49KHjn5IrVtyyaFq

The general idea is that a token is a multidimensional vector to represent a word -> think like "man" is a [noun, singular, English, pronoun, masculine, contemporary, ...]. Each time is sees a new word, it mutates this word to mean some new token (often never before seen), that means something. That's how it can roll-up a 1M line context into a shorter context, and somehow keep most of the meaning. Because it mutates all the words into different words that individually mean nothing, but when put next to each other represent the thing you likely want to do, that the LLM can somehow make sense of.

Similarly, GRAM operates entirely in a latent space that doesn't mean anything to us, but it's able to predict N different full paths WITHOUT actually exploring them fully through the LLM before it sends the one it "thinks" is best to the LLM.

If you understand how hyper words work, you can understand the noise injection... It's like it's saying, if instead of the user saying "The quick round fox" it said "The quick brown fox" -> I could probably give a response that's more like the answer they want. It's obviously far more sophisticated in the ways it can help than just a simple typo.

Something may have pushed a hyper word for "man" to somehow become a lot more like "woman", and GRAM allows it to look at the different hyper words and say... Hmm... Maybe if I changed this one gender dimension over here on this one word, maybe the entire outcome would be dramatically better. Let's try it!

Standard models compute these "hyper words" internally but immediately decode them into human language text tokens to form a Chain of Thought. Once decoded into a rigid real word, the multidimensional nuance of the continuous vector is lost!

Hyper words are the exact thing that make LLMs able to actually be smart! They can add so much more meaning to a word than a human ever could imagine - try to put 10,000 dimensions on the word "the"... Forcing them to decode them back into our dumb, un-contextualized, rudimentary language and losing all the valuable information they have - just so we can inspect it - OBVIOUSLY makes them enormously less intelligent!

It's like if we forced your eyeballs to turn everything it saw into words, before feeding it to your optic nerves, just so your optic nerves could check that you didn't see something harmful, before they sent the words to your brain... Instead of just sending light signals directly.

reply
Could you explain how/why GRAM cannot be interpreted or aligned how current LLMs are? Not very familiar how it works
reply
Crudely? Because you can't grep a sequence of latent states for variants of "If I kill all the puny humans, I can <achieve my current goal>."
reply
Why do you need to grep latent space?

As long as it's giving the right outputs, who cares what's in latent space?

If the model thinks in latent space: "God I wish these people would die," and constantly does the right thing, who cares?

Additionally, if one of it's latent spaces that it never explores is a psychopath -> who cares? The path never gets taken...

That's a lot of harmless people walking around with crazy thoughts...

reply
Thinking ‘God I wish these people would die’ could increase its propensity to kill all people, even if that propensity is still vanishingly small almost all of the time.

A lot of people are walking around with crazy thoughts. Some of them harm.

reply
Readable reasoning traces are a convenient thing, but they don't have to be true in any way. It's actually dangerous to think that.
reply
Tell me you never had a crazy thought and you are either a lier or a psychopath.
reply
[dead]
reply
[flagged]
reply
sibling comment got to the main points before me, but to add on kmavm's reply, the attack surface for gradient decent to get the system to exchange "bad information is much higher in latent reasoning models (like GRAM). You get ~3 OoM more bits (~17 bits per token in a standard CoT vs the whole residual stream of the model @ f16 = a few kb) per forward pass of the system coming back to itself, and even if you could sift through all that for signs of misalignment, you just can't put a blockade on all of the bad things that leak through.
reply
I think you’re overstating the impact of interpretability here. Your earlier point that latent reasoning models can’t be trained very well and that discretization may be load bearing rather than a readability tax in addition to significant inference infra hurdles (e.g. batching, speculative decoding) have limited any serious attempts and reduced the theoretical advantage over CoT at least in the near term.
reply
> I think you’re overstating the impact of interpretability here

Outside of RLAIF, interpretability is the strongest way to do alignment right now. alignment is important because otherwise LLMs are incentivized to learn power seeking, dangerous behaviours [1]. a more downto earth example of alignment being important is that agents are incentivized to do tasks in the shortest way possible, and this way might not be what the user wants (I explain this further in another comment in this thread)

[1] https://www.forbes.com/sites/boazsobrado/2026/03/11/alibabas...

reply
You’re putting the cart before the horse - alignment is an unsolved challenge (there are proposed approaches and active research on this) but it is still not established (beyond theory) that latent reasoning is more capable than CoT on hard language reasoning, particularly at scale.
reply
Most alignment methods nowadays don't rely on interpretability. And neither do all LLM vendors care about alignment much - especially not in China.

Those things being untrainable at scale is why they aren't around. Alignment is an afterthought.

reply
reply
As is, Chinese labs spend more effort on "rhetorical alignment to the party line" than alignment of any other kind.
reply
I second this idea: LLMs will plateau. They are already pretty good. Plus, scientists struggle to actually score their performance accurately (esp. when it comes to reasoning).

With that said, they are now hitting the walls of energy costs and memory shortages. You brain uses 20W -- don't take it as an insult. There are orders of magnitude to gain from producing energy-efficient models (or model runners).

So I am expecting same performance at lower costs for the coming years.

reply
There is endless returns to frontier intelligence, just because most people can't make use of it doesn't mean someone can't make a ton of money off of it.

Most software engineers will just need cheap tokens.

But things like physics and drug discovery have no foreseeable upper bound.

reply
Or governance of large organizations... There are a huge number of factors to consider, counterfactuals, studies, lots of non-obvious second and third order effects, etc. We're barely able to get basic governance without creating huge problems (low density zoning rubber stamped across the nation creating a housing crisis, for example), so the bar isn't high.

We pay CEOs an enormous amount because a small improvement in performance of an org because of them can make a massive difference in organizational value.

reply
The upper bound is limited by market size and cost of intelligence.

Throwing more intelligence at a problem doesn’t necessarily pan out financially otherwise we wouldn’t have single underemployed biology PhD.

reply
Absolutely that’s why they’re rushing to IPO now to squeeze the last drop of the bubble they know this is a dead end.
reply
I think we could run for at least a decade further with no model changes/improvements, just better harnesses and infra around this agentic way of developing.
reply
We, the users? Absolutely. But will the big AI companies last even half a decade without new products? Doubtful.
reply
Indeed,now it is sweet spot for senior engineers: smart enough to accelerate, dumb enough not to fully autonomously act.But it won’t last long…
reply
It's unclear it's a dead-end within 5 years.

There's still several orders of magnitude of improvement that are almost certainly left - it's just not clear how much is left on the frontier end.

Most people will be very glad to pay Anthropic, OpenAI, Google etc $200 a month to get things done 20x faster than they could IF they had a $8000 MacBook and could theoretically do it locally.

Some people would pay $200 a month forever not to have to open the terminal one time...

reply
"Doing things X times faster" at some point hits Amdahl law. If just context switching takes 5 minutes, speeding up a 1 hour task by 10x provides 5x improvement.

Furthermore, if looking at the results takes 10 minutes, that same 1 hour task only sees a 3x improvement. And so on.

reply
That’s not how firms do the financial analysis which is where most of the revenue’s are coming from…
reply
> Most people will be very glad to pay Anthropic, OpenAI, Google etc $200 a month to get things done 20x faster than they could IF they had a $8000 MacBook and could theoretically do it locally.

No most people will not pay $200 for an LLM subscription. Some software developers do. Also, at $200/month, you are much better getting the macbook machine assuming token output speed is the same or at least reasonable.

LLMs are not very productive for your average person now for them to drop $200 on. They'll need to be more capable and integrated and even so...

reply
One thing to remember is that the $200/month subscription is heavily subsidized. It is more to promote use, especially to corporate users that pay for the API token use.
reply
A bubble doesn't mean a dead end - e.g. after the .com bubble, Internet usage kept expanding by orders of magnitude for two decades.

An AI bubble is pretty much guaranteed at this point but that doesn't mean there's going to be a new AI winter.

reply
On the other hand, I think I have been hearing that for a while, even before Opus.
reply
While revenues grow almost exponentially. Reminds me of the confident predictions in the early days of Covid that it was nothing while the data showed exponential growth.
reply
I’m also reminded by the early COVID days when exponential growth was leading to predictions of the collapse of modern civilization and a billion dead, now it’s just another endemic respiratory virus.
reply
Yeah! Just like they warned us that Y2K was gonna cause a lot of problems, and then a bunch of people did a bunch of work and then that problems didn't happen, so those people warning us about Y2k were wrong!
reply
“a bunch of people” aren’t what caused the virus to become less severe.

Y2K was overblown how it was portrayed by the media but is irrelevant to the analogy of unsubstantiated extrapolation of early exponential growth.

reply
Maybe stop getting information from your Facebook feed or over dramatized US news.
reply
GRAM is another one of those "stupid specific architectures" - same as HRMs, etc. It can sort of contest LLMs at specific puzzles. It demonstrated that much. It's not a general contender with LLMs at LLM tasks.

If you subscribe to things like "there are tasks LLMs are innately bad at due to insufficient depth and lack of recurrent capability", then GRAM might be another signal towards that.

But keep in mind: even ARC-AGIs have their frontiers dominated by LLMs. Even if "innately bad" is true, it clearly doesn't go all the way to "innately incapable".

reply
A 10m param GRAM model beat o3-mini - a model 2000x its size - on Arc AGI...
reply
And then that 10M param GRAM went and got its shit kicked in by Grok 4.20 Blaze It Edition - on the same ARC-AGI battery. I know how that story goes.

It's the pattern with those "stupid specific architectures". Very good at this one thing. But only ever "good for their size", and only to a point.

They don't scale up and they don't generalize. Go far enough on task complexity and LLMs just kill them.

Does that make them useless? As an LLM replacement, yes. In general? Maybe not, I can think of things. But I'm yet to find any paper demonstrating a real world use.

reply
GRAM is something you add onto an LLM... It's not an LLM replacement. It's like an MLA caching layer, an MoE routing layer, or a speculative decoder at the end...
reply
You could certainly bolt GRAM onto an LLM, but that won't magically improve its reasoning.

It's a special-purpose design for constraint-satisfaction problems with simple rules, but complex interactions. E.g. when solving a Sudoku, the set of valid choices at every step is easy to determine, but you could make a series of valid choices that back you into a corner where no more progress is possible and you have to backtrack.

Meanwhile, LLM reasoning failures are more often of the kind where a choice is clearly invalid (as judged by a human observer), but the LLM picks it anyway, because the underlying rule is complex and context-dependent and the model only learned an imperfect approximation that often breaks down.

GRAM won't help with that.

reply
My vision for what might happen: an LLM emits a "neural constraint satisfaction task" in latent space, kicks a "neural tool call" into a non-LLM architecture, runs that architecture, gets a latent answer back, attends to the answer to generate better text answers for problems that benefit from improved constraint-satisfaction.

But that's a very hard thing to implement, and the gains are uncertain. Thus "might".

reply
Small models don't have enough parameters to memorize the entire internet. For very common prompts you don't notice that, but when you rely on some niche knowledge that might only appear once in the entire web, a single blogpost, a single github issue, a single pdf, you need to be lucky enough that the agent runs a web search AND it returns what you need.

Even as humans there's so much knowledge out there that exists but it's very hard to surface unless you know exactly what you're looking for beforehand.

reply
Smaller models can already outperform SOTA and massive models on specific tasks / domains.
reply
By pointing out the exact things that will likely happen you are oddly enough hedging against (at least some of them) happening!

A) I reckon it's true that smaller models will continue to improve massively through optimization and better and better harnesses, this tech is all still very young and A LOT of resources and (good-)will is being thrown at it.

B) The 1T+ models will be able to sideload and improve upon a lot of the fundamental improvements that happen to the smaller models to speed up incredibly while getting better at tools while (on a gradient) getting -more- things right.

C) More of an observation that I think is worth keeping in mind clearly; Karl Popper's black swan and all, truth in our temporal world IS a gradient!

reply
> The 1T+ models will be able to sideload and improve upon a lot of the fundamental improvements that happen to the smaller models to speed up incredibly while getting better at tools while (on a gradient) getting -more- things right.

There's less room to improve in things on several fronts.

GRAM very likely may scale sub-linearly with parameter growth. A 100M param model may gain reasoning by a factor of 4000, while a 100B model gains reasoning by a factor of 2, and a 1T model actually gets worse.

Additionally, the 1T model with reasoning is already pretty good. It can only improve in certain things so much.

If you score 0.02% on a metric (which small models often do), you can pretty easily get 4000x better. If you're already scoring >50%, you can't even get 2x better.

reply
I think you are assuming training from scratch, which I doubt is happening here. Fine-tuning and RL, especially based on synthetic feedback (coding skill, in particular) can be ongoing and is where these models obtain truly useful abilities.
reply
>As far as reasoning is concerned, with the recent GRAM release

Graphic RAM?

reply
"It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years"

What insight do you have to make this claim?

reply
Have you personally used any of the latest batch of even smaller local models? They certainly don't beat SotA models at coding... but with a good harness they are able to achieve things with SotA that I couldn't last year.

I've repeatedly given local models non-trivial projects that involve research and coding which they've successfully completed with minimal intervention from me (almost exclusively in the domain of reviewing the results). Again, nothing comparable with current SotA, but definitely tasks I could not have given SotA models last year (without agent harness).

Now that pure progress from these models seems to have slowed down, we're seeing a ton of options for both making models more efficient and other tools that help improve them (everything from agent harnesses to RLVR).

That's just looking at "what can small do today", when you look at what's possible with larger open models that are still much smaller than SotA from the major providers, their performance is extremely close to SotA, enough that for personal projects I'll just use Kimi instead of any anthropic offerings.

So it's not terribly hard to image a solution in the middle happening within a few years. We still have tons to learn about optimal sizes of these models and how to build them with maximal efficiency (and we've already seen a lot of recent improvements in this space).

reply
> but with a good harness they are able to achieve things with SotA that I couldn't last year.

What happens if you run last years model in a SOTA harness? IME, the quality of the harness has a much more significant impact on the quality of the result, once you get past the initial hump of “can it do anything at all”

reply
I think this is a big component, but also context. A large factor in any model being able to handle complexity comes down to context length.

I think multiple SLMs driven by an orchestration frameworks (harness or otherwise) will ultimately displace LLMs. Right now we're in the era of diminishing returns with respect to LLM gains. Moving the needle percentages doesn't excite as many people anymore and with "reasoning" capabilities there's no reason why small distributed models can't be run more efficiently, especially if/when we start to see gains in modularized context management solutions.

reply
sure, but high-quality harnesses require less gpu compute/VRAM, and plausibly can be used locally by most users.
reply
"Have you personally used any of the latest batch of even smaller local models?"

No I have not, which is why I asked (it wasn't a rhetorical question). Do you have pointers on what the recent improvements are?

reply
Try qwen 3.6 models with hermes and see for yourself. 27b is excellent and 35b is very good for basic agentic tasks.
reply
Can you spare a sentence or two describing your local setup?
reply
biggest thing i wish was present in more discussions about models is people providing more specifics on their setups vs. vague descriptions of harnesses
reply
can you please share details about your harness
reply
1. Context is all you need... They are heavily investing in getting better context (especially for coding tasks). This will disproportionately advantage smaller models (and benefit everyone).

A smaller model with better context today can outperform a model with 100x more parameters with bad or diluted context.

2. MoE (already abundant) + MLA (mostly memory efficiency, not quality) + Medusa (speed, not quality) + GRAM (5000-10,000x better reasoning in an extremely small model) + 1.58b (unclear if it will have the impact Microsoft first claimed - but possibly 5x).

reply
Probably just "gemma was cool"
reply
> It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks

The benchmarks need to change. The current coding benchmarks don't capture the realities of software engineering.

I had a bunch of images that got masked by some logic, I had to evaluate something on the original images, Claude 4.7 decided to inpaint the masked images instead of just fetching the actual unmasked images from upstream.

I had another model once that decided that because it couldn't figure out how to fill out a form to log into HuggingFace to download weights for some open source model that it was going to instantiate the model with random weights and run inference on a thousand images.

Its coding was fine, but the solution was not the right one.

reply
It is fascinating to me to see a new product category that improves so vastly year-after-year, where people commonly state that this is now the peak already.

I couldn’t even imagine having to go back to a model from 12 months ago, much less 24 months ago. GPT-5.5 is so much better than GPT-4o that it sure seems like they keep finding new juice to squeeze.

This is like going from dialup internet to DSL and acting like it has peaked before gigabit cable and fiber come along. We are at the beginning of hardware truly made for AI.

reply
> I couldn’t even imagine having to go back to a model from 12 months ago, much less 24 months ago. GPT-5.5 is so much better than GPT-4o that it sure seems like they keep finding new juice to squeeze

The difference in progress in smaller models is far more impressive.

Compare Gemini 3.5 Flash to a ~16B parameter model from 24 months ago.

Compare GPT-5.5 to a frontier model 24 months ago.

Yes, GPT-5.5 got better. At orders of magnitude smaller parameter sizes (when factoring in ACTIVE parameters) the increase is far more pronounced.

reply
Totally agree on smaller models making even more impressive gains. Gemini 3.5 Flash is better than the biggest SOTA model from 24 months ago, not just a 16B parameter one. GPT-4o came out 24 months ago, and there is no way I'd choose that over Gemini 3.5 Flash today.
reply
Yeah sure but is it so much better than Codex-GPT-5.3? No, if anything it's probably a little bit worse.
reply
GPT-5.3-Codex came out in February, and GPT-5.5 came out in April. How much better do you expect in two month's time? What other products can you think of that get meaningfully better in that short of a time frame?

And as good as 5.3 Codex is at writing code, 5.5 is easily just as good, if not better. But 5.5 is more than a one trick pony and it is much better at planning, writing copy, documentation, etc. I can choose to run 5.3-Codex instead of 5.5, but I never ever do.

reply
The GRAM model is so much into my research direction, I love it. Thank you for posting it.

Where do I find papers like this? Outside of hacker news comments. It's so hard to find the good stuff in all the noise IMO.

reply
> Where do I find papers like this?

I got it from my Google News recs on my phone, because I've been watching a bunch of videos on YouTube about LeCun's ideas on World Models and JEPA (I think).

reply
GRAM is a lot like the Multiple Drafts Model of Consciousness that Daniel Dennett proposed. I think reasearches should read more philosophy models and bring good ideas into LLM research.
reply
Can you recommend a good starting point other than Daniel Dennet?

I have the same assumption about Cognitive sciences, which I try to get a better understanding.

reply
A LLM should be able to do a better survey of literature than me. I haven't read literature by Dennett but have watched ALL his videos online so that's how I know.
reply
Yea this is great advice: the people who actually know how to build machine intelligence should go read the notes of the people who literally had no idea how to do it. While they are at it, we should have NASA go read Jules Verne so they can use his ideas in the next manned missions.
reply
> It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years

Given how well Qwen3.6-27B performs for such a small model I think you could be right. I suspect that Google,OpenAI,Anthropic must be looking at the Qwen3.6 models (as well as Deepseek V4-flash, MiMo-V2.5) and wondering if they could make some smaller models that are specifically trained for certain activities - like coding. Smaller, more targeted models would take up a lot less resources.

reply
The problem is that once you reach a certain level in coding (not particularly high imo, although some would differ) the most significant improvement in your output comes from understanding requirements better and finding ways to meet requirements in productively lazy ways, bypassing busywork that seems necessary but isn't. And that's the kind of stuff you will only find from a generally intelligent model, not a code monkey that's optimized for turning requirement sheets into source code.
reply
> I won't be surprised if the next gen frontier models are the last.

the last?!? I'm excited to see :) I'll take the other side of that since llms are so new

reply
What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.

Honestly, there is nothing in my head that Claude cannot handle. Maybe it can be more this or that but I can already barely exploit Opus 4.7.

And I'm using DeepSeek 4 Pro for my personal use and while it's a little behind, it's not that far.

I think the situation can be very dangerous for US AI companies because if current models are already capable of doing mostly anything, nobodoy will want to get to the next model, even if it's 10x better. OTOH, open source models like DeepSeek are doing mostly the same work for 1/10 of the price.

Also the more I play with Pi, the more I think LLMs are already not kept back by their own capabilities but by the lack of agency we allow them to have. There is more value today in a capable harness for current LLMs than in a better LLM.

reply
Are you joking? Is there literally "nothing" you can imagine that Claude can't do?
reply
Not OP, but in 6 months of using Opus I haven't yet found anything that I know how to do but it does not. On the contrary -- it can do things instantly that I would have needed a ~week refresher on some SDK or some algorithm in order to implement myself--plus a ton of thrash/debugging time.

What have YOU thought of that Claude can't do?

reply
>What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.

I think what gp said was the improvements are incremental, and we haven't seen a big revolutionary change since 2-3 years, and the pace is slowing down.

reply
> What gp wanted to say is that models are now so smart and useful that even if they managed to be EVEN MORE smart and useful, you wouldn't even notice it.

If benchmarks across the board keep trending up and you still don't notice a difference, that's not evidence the model stopped improving. More likely your tasks aren't hard enough to expose the gains, or the model has passed the point where you're able to judge it.

You can only tell a good answer from a great one up to your own ceiling. Once the model clears that, both look the same to you, and the extra capability is real whether or not you can see it.

reply
But that’s exactly what I said ! I know the model will continue to improve and I don’t deny that, I even strongly believe it. My point is that at that point it probably won’t change anything to me.

Would Opus 10 release tomorrow and be nearly AGI, I still would still use it like 4.7 because on daily use, I am the limit (also the harness).

So as a customer paying for tokens, I’m probably going to search for better cost rather than more intelligence.

reply
> Honestly, there is nothing in my head that Claude cannot handle

Friend does marine autopilots in C++ on 64kb of memory. It's totally useless for him.

From my experience any sort of more difficult backend logic - all LLMs fail pretty quick. Especially when you need to logically work out the business logic (partly if not mostly because it just doesn't have the context you do).

reply
> Honestly, there is nothing in my head that Claude cannot handle.

One idea is that maybe it could figure out how many L's are in the word "google" [1]

Or, maybe which days of the week have a "d" in their spelling [2].

[1] https://x.com/FatherPhi/status/2059659658428912040?s=20

[2] https://x.com/FatherPhi/status/2054212816069132461?s=20

reply
From what I understand, that's a problem with the way it receives data. The model doesn't see the letters g,o,o,g,l,e to count it. Just like how I can't sense radio waves. If I wanted to find that out, I'd get a tool to detect waves. If the LLM wants to find that out, it can write a script to find it.
reply
Wow, which Claude model flubbed that question? Certainly not anything recent...? The 2-bit quant of K2.6 running locally on my own hardware has no problem with it: https://i.imgur.com/tL0FLjZ.png

So Claude has no excuses here.

Edit: even Qwen 3.6 27B handles it ( https://i.imgur.com/jleJxj2.png ), and of course Claude does. I had to go all the way back to Opus 3 to get it to fail (https://i.imgur.com/uJOH2nP.png).

reply
As far as it has been studied, the relationship between model size and capability is inversely logarithmic: 10x increase in params less than doubles capability.
reply
There is endless returns to frontier intelligence, just because most people can't make use of it doesn't mean someone can't make a ton of money off of it.

Most software engineers will just need cheap tokens.

But things like physics and drug discovery have no forseeable upper bound.

reply
Within software engineering, security, reliability, and scale also seem boundless.

Software that never breaks (including because it never runs into scaling problems) and never leaks your data is preferable to software that breaks and leaks your data sometimes, but it has been too costly to be practical.

Current models are still very far from the reasoning muscle required to build things that never break, scale to billions of users with no issues, and cannot be exploited.

reply
> Software that never breaks (including because it never runs into scaling problems) and never leaks your data is preferable to software that breaks and leaks your data sometimes, but it has been too costly to be practical.

It's almost impossible to prove non-trivial software is invulnerable.

It's very easy to prove that it sort of works.

For one, you have hardware vulnerabilities - period. If you're running on any operating system, you have OS vulnerabilities. If you're not running on bare metal, you may have who knows what kind of vulnerabilities. If you're running literally any other piece of software on the same machine, depending on the hardware and OS, you could have vulnerabilities...

reply
People keep saying this and yet the evidence seems pretty thin..
reply
To me its evidence of people who dont actually think deeply enough to understand the subtleties, nuances etc of what they are talking about.
reply
Nothing ever happens, in 20 years we will still be painfully dying from the same shit as now. Maybe there is like 5 new drugs for some exact specific type of cancer out of like what, thousands?
reply
I effectively distill the frontier models by building whole sets of skills, personas, and other artifacts that I can then run on smaller models and get 10% even 20% improvements on models like haiku or local models.

There's a lot of room for improving the smaller models at many levels of the stack.

reply
This is a good point. It didn't really work on older small models but the latest crop are quite good at following instructions and paying attention to detail, they just lack a lot of the sophistication and nuance that the frontier models have these days. So they are often capable of doing very complex tasks, they just need more detailed and foolproof instructions than the larger models would.
reply
surely training also gets cheaper so justifying it becomes easier?

i think it'll be more like we get 1-10T models and then distill those down into smaller models, though

It seems like the best small models today are all distilled from bigger models

Moreover, I hypothesize Claude Opus 4.7 and now 4.8 are a distillation of Claude Mythos

reply
That's the impression I got too, it seems closer to what the marketing has told us about Mythos than 4.6/4.7 were.
reply
I'm frankly surprised the focus is still on these enormous "know everything in the world" models. I would think you could create an incredibly lean and smart "just React and React Native" model.
reply
"Make a React app to run my coffee shop" requires knowing React but also knowing what a coffee shop is.
reply
Only if you're going after the "vibe coders" audience. Regular developers would be fine with a lightweight local llm capable of scaffolding and wiring a dozen of bog-standard components in a few lines of natural language.
reply
Sure but it doesn't need to know everything.

It doesn't need to know different languages, every programming lanuage and co.

We will for sure get to this in the comming years. After all they will have to start finetuning their traning data anyway

reply
> I would think you could create an incredibly lean and smart "just React and React Native" model.

You can, but it's not as useful as you might think.

It needs to at least understand 1 human language to understand your intent to implement features.

If GRAM turns out to be a 5000x multiplier for local reasoning, you could theoretically train a 500M parameter model on just a programming language to understand stack traces to fix bugs and be incredibly powerful.

But most people also want it to understand human language to implement features as well.

Because then it can't just understand React and JavaScript - it needs to understand thousands of commonly used dependencies, the DOM, CSS, HTML, etc...

And for that you need A LOT more parameters than you might expect.

You can definitely get a ~3B active parameter model that can run comfortably on today's hardware to be VERY good at coding once all of the SOTA architectures are added to a single model - especially if we get better tool calling to give models better context per language.

You might be thinking: why does it need to memorize dependencies? Can't it just stick all of them in it's context and use its super smart brain? No, context is king. You want to keep it as short as possible. The solution is not having a smart model and putting 10M lines of context in it. The solution is having a model with enough parameters to know what it needs to know. Researchers are already working on having "packs" of knowledge where you could download a 20M param pack just for some common dependencies in JavaScript (as an example) - but AFAIK this is likely years away (and may not prove effective).

You could get 100x performance if you feed the models ideal context... So a 3B model today can perform almost as good as ~300B model if you give it really good context vs flood it with mostly garbage it doesn't need across your repository.

If you feed it 100x more context to make up for its limited memorized general knowledge, it's going to perform thousands of times worse, completely eliminating any advantage it might get from GRAM...

reply
We just want it to understand how to write code. We don’t also need it to know how to grow a potato.
reply
I think perhaps you misunderstand how much of being an effective coder is understanding business domain enough to not be constantly asking for clarification (or if one is a fool or an ai, assuming wrong answers). I reckon a vast collection of trivia on the level of knowing how to grow a potato is important for a programmer
reply
And you can't know ahead of time, when you're training the model, what business domains it will be used for. Someone may decide to use it to optimize the watering and fertilizer cycles of their automated potato-growing setup, and suddenly the "how to grow a potato" texts that went into training the model are actually the very things that make the difference between success and failure for the code the model spits out.
reply
The disjoint set of English related to strictly growing potatoes and adding features to code is a lot smaller than you probably think...

It is hard to cut out a huge portion of English and truly understand English and human language.

You're just not saving as much as you might assume you could.

reply
deleted
reply
... unless the software is potato farm software.

Programming is not a rare skill, the interaction with domain knowledge is.

reply
To me, the magic with LLMs has always been on the input side. It needs to understand what you mean in order to do what you ask. Most people are pretty terrible at communication, and general world knowledge seems to help with that.
reply
The syntax is the easier part - most programming tasks require the reasoning and understanding of a large world model to solve problems.

Fine tuning a 'lean and smart' model works really well for discrete, repeatable high volume tasks like support ticket triage, lead classification, content filtering, labelling, generating content with a voice, etc.

Inefficient token burn by throwing large models at everything is definitely a problem - it's like hiring Phd's to answer the phone or to wash dishes.

reply
Let's hope that hitting a scaling wall and less money to spend will begin redirecting efforts to optimize inference and get the same results with less compute.

Boomer comparison, but I remember the 8 bit computer era when the hardware was what it was so the later games of that era used hardware better than previous ones.

reply
And anyway, with quantum, there will be no need for frontier companies as you might be able to even run a 1T param model on a consumer quantum computer.
reply
Even if quantum computing had any clear implications for LLMs (it doesn't), there is no such thing as a "consumer quantum computer" and there won't be in our lifetimes.
reply
I'm assuming this is a joke, but:

- why'd a quantum computer help running an LLM?

- of course there'd be need for frontier companies - nobody else has the resources to train frontier models.

reply
What? No, that is not what quantum computers do
reply
you just need to look at Mythos to see the jump in performance from a 10T(?) model. As they scale, they get more capable. We might have an yearly release, but I believe the releases will continue, as long as scaling laws are in tact, and there's huge problems still need solving. (think cancer)
reply
And how are we meant to look at Mythos? Do you have access?
reply
no but they tell me it's TERRIFYING and DANGEROUS and we should INVEST MORE MONEY
reply
Through association with a large company:

https://www.anthropic.com/glasswing

Ive seen the tickets generated by the model that have trickled to my team. They are legitimate, but i can’t speak to model improvement because its a pilot program.

reply
Through the lenses of anthropic's marketing department of course
reply
>you just need to look at Mythos to see the jump in performance from a 10T(?) model

Mythos is a bunch of likely overhyped claims at this point. A few experts who looked into the claimed results weren't that impressed.

reply
They all looked like real CVEs to me.
reply
Nothing that special about finding a real CVE. They're not that different than what non-Mythos could spot.
reply
And there seems to be a ton of experts on the opposite side.

As they say, the truth tends to be somewhere in the middle.

reply
You forget that these models are still only interpolating between human-generated datapoints fed to them. They cannot reason beyond the data they've been given, so unless everything you want to create with AI is a synthesis of prior art, you're back to relying on the stone-age human brain that created AI in the first place.
reply
Not all training data is human generated, and it's also not clear that being ridiculously good at interpolating between data points (whatever that means) will not lead to superhuman capabilities.
reply
I could make a robotic picture coloring machine with truly superhuman capabilities - picking only the most beautiful color combinations and staying 100% in the lines while finishing entire murals in < 1 second. However, if you need a completely new and original image rendered, the machine is of only partial utility for you. It is very well possible that your cure for cancer (if that's even feasible) or whatever else you desire is a completely new picture.

We have these breathless conversations about the new AI frontier at the peril of losing sight of reality and our own human potential.

reply
>these models are still only interpolating between human-generated datapoints fed to them. They cannot reason beyond the data they've been given

Are you sure that humans can?

Didn't a SOTA recently solved a mathematical theorem, one escaping mathematicians for 80 years?

Maybe a human "novel" invention is just a good interpolating from the datapoints (knowledge) fed to the human.

reply
Your phrasing ("you forget") implies this is a fact and common knowledge, while in fact there's little reason to think that's true.
reply
Do you know if anyone has trained, say, a pre-2017 model and tried to get it to come up with Attention Is All You Need? If it did, would you say that was only because it's a synthesis of prior art? If so, what isn't?
reply
Allow me to restate my point: human beings and AI both create via synthesis, but we are the only ones capable of what we could categorize as true original thought or creativity. It could be argued that nothing we do as humans is truly original or creative either, but I would counter that with the claim that an LLM could not have created any element of the society and culture that gave birth to LLMs. Maybe in six more months.
reply
>human beings and AI both create via synthesis, but we are the only ones capable of what we could categorize as true original thought or creativity.

And how is that anything other than synthesis? Do we pull concepts out of thin air?

reply
> I won't be surprised if the next gen frontier models are the last.

I’d be surprised tbh. Investors don’t want to hear “everyone else is still training models and seeing improvements, but we don’t want to participate in the arms race anymore.” They want monumental leaps every quarter or two because they have sunk unholy amounts of money into these companies/products.

The whole idea of “hyper scale” doesn’t jive with caution and or otherwise slowing down.

reply
The way this will play out, most likely, is that smaller models will continue to get released, anyone willing to drop 1-3k on a home upgrade/new LLM box (no that isn’t cheap, it also isn’t outrageously expensive) along with improved open source agents or whatever (lot of meat on that bone) will sneak up behind the big players and start taking dents. Smaller companies will pop up providing 50 users unlimited whatever for a lower cost than the big companies.

The whole ecosystem will twist and evolve, and the big companies will be left begging for corporate subscriptions.

I finally caved when I realized I could build a PC, for myself, with dual video cards that I wanted, which can play games that I like and run models that I want, without worrying about giving my payment info to someone I don’t trust, or invoking token anxiety that I don’t want.

reply
Like every major tech-software innovation of the last 20 years, I think it’s just going to be consolidation all over again.
reply
I don't think this is true at all. It might feel like this because we are used to a very very fast release cycle but we are only in this topic for a few years.

We have so many ways of optimizing:

- continusly creating more and better training data

- increasing parameters to 20/50/100TB

- We still wait for Mythos access

- We still wait for Mythos distilation (i haven't heard any rumors or so that there is a distilled version of Mythos out)

- Reinforcment learning and evolutionary algortihm only started to appear

- If a small 30GB Model can do stuff, these models can also be used as teachers for the big ones

- We have not seen yet specialized models at all. Like a coding java german expert model. Why? Even with MoE architecture, you still need to have these layers around

- Research for Diffusion and other models is still in progress

- Nvidia just announced/showed a 7x speedup on inferencing for Nemotron

- Multitoken prediction became available just a few weeks ago

- Compute gets only in a range were they can do a lot more and cheaper experiments (see Google IO 2026 announcement)

- World models are showing great progress and we do not know yet what they will bring to the table

- They are probably not finetuning/fixing all areas in parallel. I would argue that Anthropic focuses most of its efforts into coding and agentic. Google for sure does subagent and agentic optimizations too. Plenty of areas are just not touched i would say because they don't have the capacity

- We see more and more mulit modal models (these also consume compute)

- N-Gram paper and co i have not seen all of these things in chinese open models

- We don't even know yet what Meta is doing, but we do know they restarted their efforts again

- Anthropics models got a lot better benchmark wise for dening non sense asks. They do learn how to get rid or reduce hallucinations

- We are in the middle of the biggest Reinforcement loop whith all the training data we give them day to day and its not clear at all if they already use these models in thir training and at what stage.

- We do expect bigger models to be able to comprehend deeper concepts / broader code bases. Big companies with huge code bases probably are waiting for this

- Thre will be also continues progress in harnesses which in it alone is not part of the LLM progress (fair) but these harnesses do get better when you finetune a model to be optimized for a harness

- ChatGPTs Image model 2.0 got relevant better and came out just a month ago

I suspect, based on hardware requirements and progress on hardware infrastructure alone, that the industry wants to go to 100t models and we do not know yet what this will mean. I could see that we might skip normal transformer and find relevant other architectures.

Just a week ago there was a research paper about parallel input and output streams which has not been explored enough.

There was also a research paper were they showed that a LLM can compute things. This will take time to see were this leads to.

I don't think the focus on GRAM and facts is so relevant. Its about context and context handling not just some facts.

reply
Great points! We do keep seeing gains from larger model sizes. I think that is still one of the factors contributing to jagged intelligence. When they increase up to around 100T parameters, that will truly be human complexity level, and I assume there will be no trace of jaggedness left.

If you look at things like Mythic AI and the recent wurtzite ferroelectric nitrides breakthrough from the University of Michigan, huge performance and efficiency gains through new compute-in-memory approaches are around the corner.

And that will get us up to two orders of magnitude more parameters.

It's also plausible to me that before we get all the way to 100T we find some recipe of efficient state synchronization, goal sharing or something so that we are able to get higher collective IQ by combining fast distributed predictive subnetworks.

reply
> There was also a research paper were they showed that a LLM can compute things.

Can you be a little more specific than that or provide a reference?

I assume you're not indicating universality of neural networks?

reply
I do.

This is the newest thng i'm aware of: https://www.percepta.ai/blog/can-llms-be-computers

But there were papers in 2023 with a different approach requiring external memory https://arxiv.org/abs/2301.04589 too

reply
I think the future will be enterprise clients will train their own models based on their needs and data.
reply
Versus just packing all their needs and data into context, and RAG (i.e. context)?
reply
Why isn’t this happening more already?
reply
It takes way more resources to train the model then to use it.
reply
I honestly doubt this; very few companies have enough data. Maybe we could see mergers so it happens but basically it would mean everyone would need to be Google sized for it to work.
reply
> It is almost guaranteed that a 60-90B model can outperform current SOTA in coding tasks within 2-3 years.

I am ready to bet against this. Knowledge benchmark like SimpleQA isn't increasing for small models.

> It is far less clear that a 1.2T model will be meaningfully better enough to justify training it.

Well for one, we know for certain there is Mythos which is meaningfully better. And I think there is a lot of juice left to squeeze for Mythos class model.

reply
Knowledge benchmarks can't really be improved upon via distillation or RL. It requires those facts be added to the training corpus and for the model to memorize them better. Neither distillation or RL really do that and thus we shouldn't expect improvements on SimpleQA unless some other interventions are being made.

Model intelligence and knowledge aren't necessarily directly related. If we can pack greater intelligence and agency at the cost of it forgetting factoids, that would actually be a good thing. We don't need LLMs to memorize facts, we need them to learn how to interact with the world such that they can find the facts that are necessary and surface them to the user.

If we could distill all of the knowledge out of an LLM and just be left with a very agentic model that only knows facts in it's context, I think some very interesting stuff would happen.

reply
RL is more than facts. Synthetic feedback is an obvious approach. Does the model suggest code that compiles and performs well?
reply
Lot of the things aren't facts that could be stated. No one can just see the dictionary or translation of words and start talking in that language.

There isn't a clear definition of what is knowledge and what is intelligence. Is being able to write in C knowledge? Is knowing undefined behaviour in that knowledge?

reply
> Well for one, we know for certain there is Mythos which is meaningfully better.

Do we?

Have you used it?

What is "meaningfully" better? It's not 3-4 orders of magnitude better. That is definitely happening for smaller models.

reply
What do you mean by 3-4 orders of magnitude better? Was Einstein 3-4 order of magnitude better than us?

Meaningful in the sense it could find security vulnerabilities in browser and kernel that >99% of the engineers couldn't find.

reply
> What do you mean by 3-4 orders of magnitude better? Was Einstein 3-4 order of magnitude better than us?

I'm talking about output quality compared to parameter size.

Mythos is not 4 orders of magnitude larger than Opus - it's quite possible no LLM model ever reaches that size (likely even), and it's output is only barely better...

reply
Where did 4 order of magnitude even come from? If I were to guess it is just 5x larger based on the pricing, so not even 1 order of magnitude.

> Mythos is not 4 orders of magnitude larger than Opus

Again can you define this. How would 4 order of magnitude better look like?

reply
So, then I guess the big three are never going to make their money back.
reply
I would be shocked if 5.5 is the last new pre-train from OpenAI. Your comment is nonsense.
reply
5.5 is not a generation it is a trivial iteration...

6 is for sure happening...

As is Gemini 4.

It's less certain there will be a Gemini 5 or GPT 7 any time soon that is a true next "generation" and not just an iteration. They will almost certainly call something Gemini 5 and GPT 7...

reply
5.5 is in fact a new pre-train model

First you say there won't be a new generation. Now you're saying there will be more. Oh well, I'll stop responding here

reply
> I won't be surprised if the next gen frontier models are the last.

You clearly did not read my first comment or the second, or clearly disagree on what a generation is.

reply
| a 60-90B model can outperform current SOTA

My conspiracy theory is that Apple recognizes this.

reply
That does seem to be the path Apple is following here. Have a local model that can answer most things and then have a fallback of cloud options when they request is too complex. The cleverness of this strategy has been overshadowed by the incredibly poor quality of their local models. It will be extremely interesting to see what next month holds and whether Google helped fine tune an Apple specific Gemini / Gemma model for their devices. Bonus points, of course, if they unveil the M5 Ultra Studio with half a terabyte of RAM to be a local "cloud model" (the true fantasy here of course would be Apple building something a little like openclaw where from your phone you could give commands to your Home Apple server). They could probably get away with charging $20k for it if it has sufficient tok/sec. If that happens and is successful one could imagine a straight line path in the next two generations to bringing the cost and form factor down to the point where some of the form factor of an Apple TV becomes everybody's home inference server / agentic HQ. Sovereign AI for everyone!
reply
I think Apple might come out ahead by pure accident. Yes, Apple often waits to enter a market until it's established but in the case of AI they tried, they tried and failed. It was never the original plan to partner with OpenAI and then later with Google (Gemini). They 100% missed the boat on AI, the question now becomes: was the boat worth taking and we are still waiting to see how that plays out.
reply
You need some serious memory then. Let's say around 192gb for having not all your memory eaten by your LLM.
reply
> My conspiracy theory is that Apple recognizes this.

I don't think that's not a conspiracy theory. AFAIK, It's their stated AI policy...

reply
[dead]
reply
[dead]
reply
[dead]
reply
[flagged]
reply