The weird thing is to see this based off the KV-Cache in the diagrams.
Completely different from the Databricks RAFT [1].
My team built "Dreamtime" earlier this year [2], which used a different mechanism to distill out learnings across multiple sessions.
This didn't pull from the token layer, but from user nudges on unsatisfactory responses.
This works like a sawtooth wave where we accumulate context, check for dissonances (a user gets to say "It depends" and fix it), compress it back, retire the old one.
The audit requirements are the real problem for customers, where you need to be able to ask "When did this change? and who changed it?" which was hard to answer from the lowest layer of the model unless you can retain lineage to the original sessions.
A model we train can't serve as an accomplice to a user who wants to hide some financial crimes by nudging the agent repeatedly.
[2] - https://isotopes.ai/blog/neocortex-self-learning-ai-model-ve...
As for the technical nature of this, I'm not sure if I've totally understood correctly despite how well it's written, but are they essentially holding back sort of unweighted parameters built in to the model architecture which they can then populate with new context during 'sleep' and then using the techniques outlined in the 'dream' section to reinforce what gets stored in those MoE weights? That to me seems like a solution which only Google is best placed to solve due to their fully integrated AI stack and historic experience with hardware optimisation, but which would be rather hard to do if you are a more resource and compute constrained organisation
> are they essentially holding back sort of unweighted parameters built in to the model architecture which they can then populate with new context during 'sleep' and then using the techniques outlined in the 'dream' section to reinforce...
They say the weight update can be done by computing a transform where the end state is the old weights with the addition of the useful semantic knowledge contained in memories. Rather than just additively chucking the entire memory concepts into weights. That's pretty neat.
The subtext for the memory architecture diagram sums the whole thing up pretty tidily I think:
> Input streams from the active phase are stored in short-term memory before entering the sleep cycle, which bifurcates into Consolidation (transferring fragile short-term tokens into expanded long-term parametric memory using low-rank structures and MoE) and Dreaming (autonomous self-improvement via synthetic routing with random experts and RL refinement).As an aside, I get that this is a preprint, but it's very odd to see a piece of writing with obvious typos and grammar issues (e.g. page 13: "We attributes this results to," "In this setting (see Table 4), Sleep achieves a 80% success rate, higher then the other methods.") in an era with LLMs. It sort of locks in that this was written by human beings, and I'm all for that, but it jumps out at me more than it used to.
the fact of the matter is that now when I program I don't need to hawkishly stare at my screen to ensure I didn't miss a paren, semi-colon, whatever.
I now just sort of type (when I don't dictate) every second word of a sentence in a kind of simplified Newspeak to an intern who is infinitely forgiving of typos.
in that case i'm sure practice makes perfect, and I practice less than I ever have.
- "New study shows highly creative people’s brains work differently from others'" (2022) https://news.ycombinator.com/item?id=41921738 ; and does this correlate for example like REM/non-REM and ADHD? How much does this topology vary amongst the population?
- "Discovery of Memory "Glue" Explains Lifelong Recall: KIBRA Protein, PKMzeta" (2024) https://news.ycombinator.com/item?id=41921738
- "Brain found to store three copies of every memory" https://news.ycombinator.com/item?id=41352124 but actually more than three: {in the hippocampus {EBN, LBN,}, prefrontal cortex, and engrams} but functional activations drift even just days apart given representational drift.
- "Memories stick around even after half the synapses are gone" (2026) https://news.ycombinator.com/item?id=49404593
1. Requires fine-tuning every 10M tokens ingested.
2. Requires complicated multi model pipelines. One for distilling the context, one for making a training dataset, and one for applying it.
3. All of this requires a lot of compute. Note the lack of wall times in the paper. While they do note that their process is considerably faster than SFT they never really quantify it.
4. They only published results on Llama and Qwen, and on fairly small versions to boot.
Despite all of that, the paper is a starting point for a process of making a model that can modify its own weights. I very much suspect that Google has solved many of the problems given the breadth of the related work and additional experiments sections in the appendix.
It will be really fun to see open weight models come out in the future with the functionality described in the paper.
This seems like the limiting factor here, depending on efficiency
And for what the cost are concerned: to call it more expensive in instance terms I want to see the numbers first, and - have you consider the cost of "unfinished" humans as liabilities? The costs of fools, of crooks, of the vile, of conformists etc.?
Fools, crooks, conformists, as well as `critical thinkers` have one in common — the need. Unless it's a human, machine would always think `as` it's a question it's existence depends on. But not really.
What AI can do at most — is to demonstrate our level of thinking. As well as recommendation algorithms can show what music is popular. Without constant feed from humans this makes no sense.
It just processes data, so change the data and it will produce different aggregate data, as in y=f(x). We just feed it the internet, so it spits out (mimics) internet as its derivatives. You are only personifying the current state of affairs.
You are probably speaking about the stochastic models we have been accustomed to - and already for them there are emergent properties of critical thinking. I directly expressed that the next step must be an explicit application of critical thinking to the re-processing of the data.
Many people insist in indirectly demonstrating the dire need we have for large amounts of additional intelligence in the system. Already the automation of trivial corrections in debate has critical value. Some authors have pushed for this for decades...
[0] https://med.stanford.edu/news/all-news/2026/09/two-separate-...
...Why would one?! ___We do not want to replicate humans___
> the need
Eh?! Who has what need? My need is to minimize the impact of liabilities.
> to demonstrate our level of thinking
Statement based on what?!
You seem to be just expressing your own ideas. We are not here for polling.
Our critical thinking is bounded to our needs. Machine needs nothing, so you have to give it a task. Yes, within limited scope a machine can look like it's thinking. Yes, scope is expanding. But it has this flaw and requires constant fine-tuning job that expands along with the scope.
About "to demonstrate our level of thinking" - that is my bad rephrasing of "garbage in, garbage out".
> We are not here for polling.
Please don’t push your agenda by framing it as the position of the majority.
But we pretty much have to be constructive (or would that be another "surprising agenda")?
> Our critical thinking is bounded to our needs
In which way? Proper critical thinking simply assesses truth. The machine, when spent the relevant, has the task to build a correct world model - just like people.
> garbage in, garbage out
And "critical thinking" is exactly that breaks the "garbage in, garbage out" problem.
> your agenda
That """agenda""" would be that recording "how do you feel about the matter" is completely useless (cpr. the "constructive" above). Anyone that defended that position would have no right of speaking about rudeness (because stopping people to show them an unrequested performative dance would not be consistent with the idea). (And why are you seemingly presenting "the majority" as having rights?)
I suggest you make efforts towards clarity (you may have promising ideas but how do we know?) and recognizing your assumptions (both as non-universal and as requiring explanations).
Thank you again for reminding me I have a reason to live.
Your rephrasing has not added a tiny shard to the point you already made.
So sick of popsci quoted as real science and a fact. We dont know jack shit about what our brain is doing. Wtf is 'emotional connection' and how do we know brain is consolidating it. eyeroll.
This sleep/dream stuff in ai is beyond cringe.
I’ll be watching to see if the next generation of AI improves in this area. Once this is the case, the system described in the article could lead to some wild capabilities.
Until then, I suspect it will result in a “delusion spiral”, where the AI’s self improvement is untethered to reality and produces worse results in practice.
Edit: I could be lacking in creativity and background knowledge though. You could use this in isolation, with AI surrounded by experts, and only use training runs where AI was highly successful.