I stil don't understand them. I want the US to "win the AI race" but I have trouble understanding how most of all inventions today aren't "distillations" of past knowledge. Is Anthropic claiming the data they stole as trade secrets?
Anthropic is claiming that training an LLM to mimic another LLM is materially different and worse than slurping up stuff written by humans (even if that material is stolen).
Basically, they want IP protection for Claude. This is a nakedly hypocritical stance, but completely understandable from a company-needs-to-make-money standpoint.
Their claim is even stronger than that, they have complaints about their models being used as a validation step for other model output, which is standard practice in the industry.
You can’t build a frontier model with one single thing. This is an incremental improvement but it doesn’t explain the entire success of the model. The training set is immensely important, regardless of how you feel about distillation.
>To support further research, we open-source the KDA kernel and vLLM implementations, and release the pre-trained and instruction-tuned model checkpoints.
Does any expert in the field know whether it is really the case that this intelligence we are seeing with frontier models is an "emerging" phenomena, only coming up when the architecture is scaled?
Like isn't it weird that the 1 million parameter model with the same architecture can't solve basic puzzles but suddenly the 1 trillion parameter can conjure up counter-examples for the Jacobian conjecture?
It's unintuitive since, to the best of my knowledge, one of the basic tenants of algorithm development was that you can't just brute-force your way towards a solution for some complex problems, e.g. naive sorting algorithms suddenly won't beat quicksort if you put more processing to them, but in the modern LLM scene it seems people are in a race to scaling up, experimenting empirically and hoping the same algorithm/architecture comes to a solution.
> one of the basic tenants of algorithm development was that you can't just brute-force your way towards a solution for some complex problems
It's kind of sad that popular CS textbooks often focus on solving precise problems with lowest theoretical complexity bounds while ignoring more practical (but generally applicable) computation techniques.
In machine learning they call it "gradient descent", which in older days had analogies in techniques called "hill climbing", "local search" and "simulated annealing". Basically you have a function you need to optimize for, and you clumsily tweak the parameters so that you get the (locally) max/min value you wanted. These techniques were great at finding approximate, locally maximal solutions without trying all the possibilities at once (which is more akin to the kind of "brute force" in the traditional CS context).
I guess because these techniques were generally applicable yet the outputs were approximate and you couldn't analyze them much (no fancy O(n log n)), the theorists did not find them interesting and thus were not in the spotlight of student's learning.
In modern machine learning they do this gradient descent thing which is also tweaking the parameters bit by bit to optimize for the loss function, except that the parameters are now in the billions and trillions. The compute required is huge of course, but it's actually quite an "efficient" process, and it's not actually doing much of "brute forcing" at all. During training, the process is essentially, almost equivalent to, compressing the many many trillions of tokens of training data. To me it's quite amazing that they manage to complete such a process within a couple months of training, even if they have hundreds of thousands of GPUs...
This is actually a well-known phenomenon in ML, called "The Bitter Lesson".
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
It might be that what we consider a basic and very hard puzzle are extremely close together on a more absolute scale. The difference is often for us what proportion of humans can solve it. And the low end of that is still quite high up - animals that can solve things that are very basic for the vast majority of humans are pretty rare and known about, yet are capable of quite complex actions and learning and aren’t wildly different in scale of neurons to us.
Going from 1m to 1T params is also a scaling of a million times. It’s like going from a human brain down to one percent in size in each direction or just a few mm.
IANAMLE, but there is "grokking" that makes models learn to actually generalize, even after you give them enough parameters that would let them memorize the dataset:
High-dimensional gradient descent behaves very differently than the simplified 3d visualisations we use to demonstrate it, and has lots of ways out of local minima:
it's certainly not a definite procedure for determining if an arbitrary mathematical statement is true or not. it's more like educated guess and check which definitely scales up
Old but relevant: if you read the recently-released Kimi K3 paper[0], you'll see that it's heavily based on Kimi Linear discussed here, scaling it up and adding a bunch more things (like native vision and RL improvements).
Any one knows how this holds up on long context retrieval (needle in haystack , ruler) vs same size full attention model? efficiency gains look great but that usually where linear attention hybrids fall apart.
The main contribution of the K3 paper is Stable LatentMoE. Like some other models it compresses data sent between layers, which puts certain requirements on the router. K3 improves performance by using a more balanced expert selection strategy.
Compared to the Opus 5 "model card", which read like a standard Anthropic set of alignment principles and safety concerns, this presents a plethora of useful technical details that advances the state of the art.
most new effort in training comes in the late phase with RL techniques
the pretraining (slurping the internet) only goes so far, the new data being used is from human preferences and agent traces (designed and/or distilled)
I started creating internal models using it, then the Gated Deltanet 2 came out( https://arxiv.org/abs/2605.22791), and it seems like an evolution of it in expressiveness. And in our tests it is really better than.
Are Chinese labs impressively innovating? Clearly.
However this doesn’t rule out possible gains due to distillation.
I don’t know the degree of the latter but both things could certainly be true.
“You stole my warez!”
Basically, they want IP protection for Claude. This is a nakedly hypocritical stance, but completely understandable from a company-needs-to-make-money standpoint.
https://docs.cloud.google.com/gemini-enterprise-agent-platfo...
The distillation theory does not even make sense as Fable was only around for days (effectively) before Kimi was released.
This is just awesome.
Like isn't it weird that the 1 million parameter model with the same architecture can't solve basic puzzles but suddenly the 1 trillion parameter can conjure up counter-examples for the Jacobian conjecture?
It's unintuitive since, to the best of my knowledge, one of the basic tenants of algorithm development was that you can't just brute-force your way towards a solution for some complex problems, e.g. naive sorting algorithms suddenly won't beat quicksort if you put more processing to them, but in the modern LLM scene it seems people are in a race to scaling up, experimenting empirically and hoping the same algorithm/architecture comes to a solution.
It's kind of sad that popular CS textbooks often focus on solving precise problems with lowest theoretical complexity bounds while ignoring more practical (but generally applicable) computation techniques.
In machine learning they call it "gradient descent", which in older days had analogies in techniques called "hill climbing", "local search" and "simulated annealing". Basically you have a function you need to optimize for, and you clumsily tweak the parameters so that you get the (locally) max/min value you wanted. These techniques were great at finding approximate, locally maximal solutions without trying all the possibilities at once (which is more akin to the kind of "brute force" in the traditional CS context).
I guess because these techniques were generally applicable yet the outputs were approximate and you couldn't analyze them much (no fancy O(n log n)), the theorists did not find them interesting and thus were not in the spotlight of student's learning.
In modern machine learning they do this gradient descent thing which is also tweaking the parameters bit by bit to optimize for the loss function, except that the parameters are now in the billions and trillions. The compute required is huge of course, but it's actually quite an "efficient" process, and it's not actually doing much of "brute forcing" at all. During training, the process is essentially, almost equivalent to, compressing the many many trillions of tokens of training data. To me it's quite amazing that they manage to complete such a process within a couple months of training, even if they have hundreds of thousands of GPUs...
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
The full essay is worth a read, it's pretty short http://www.incompleteideas.net/IncIdeas/BitterLesson.html
Going from 1m to 1T params is also a scaling of a million times. It’s like going from a human brain down to one percent in size in each direction or just a few mm.
Let's say there's some circuit that does problem solving of the kind we call intelligence.
We dont know what this circuit looks like, but it exists in our brain.
Doing regression on outputs from the brain (e.g. internet text) with enough parameters, we can "fit" our model to this circuit.
But if you try to fit it with fewer parameters than it needs, you're just going to get some linear approximation.
https://en.wikipedia.org/wiki/Grokking_(machine_learning)
High-dimensional gradient descent behaves very differently than the simplified 3d visualisations we use to demonstrate it, and has lots of ways out of local minima:
https://www.youtube.com/watch?v=NrO20Jb-hy0
so it seems like there is a benefit to giving models more space to learn in rather than forcing them to compress the knowledge from the start.
[0] https://arxiv.org/abs/2607.24653
As it's 9 months old and they just had a major model release
The main contribution of the K3 paper is Stable LatentMoE. Like some other models it compresses data sent between layers, which puts certain requirements on the router. K3 improves performance by using a more balanced expert selection strategy.
the pretraining (slurping the internet) only goes so far, the new data being used is from human preferences and agent traces (designed and/or distilled)