Researchers propose infinite-parameter LLMs that adapt weights from live data
A new paper suggests language models could dynamically update their parameters in real time using data encountered after deployment.
What to know
- A new arXiv paper proposes methods for language models to dynamically generate and adapt parameters using live data encountered after deployment, rather than relying solely on static pretraining.
- The approach builds on Mixture-of-Experts architectures and aims to address the gap between training data and the world a deployed model actually encounters.
- Discussion centers on three competing concerns: the potential to accelerate discovery by avoiding duplicate work, the risk of degraded learning by skipping the discovery journey, and vulnerabilities from poisoned or manipulated live data feeds.
- Technical questions remain about whether models can maintain stability and safety while continuously adapting weights in real time.
The dispute Whether continuous learning models accelerate progress or degrade human understanding depends fundamentally on whether the journey of discovery itself holds intrinsic value beyond its output. · positions read across 12 posts and comments
Continuous learning models could dramatically accelerate discovery by eliminating duplicated failed approaches across the research community.
-
“imagine if any individual could try new approaches to solve a problem/make a discovery and any micro-advancement gets integrated into the model itself, dynamically. This could transform progress from the slow "write a paper, get peer…”
lubujackson · Hacker News ↗
Skipping the research journey itself degrades human understanding and innovation by removing the value of firsthand discovery.
-
“It's not the destination, it's the journey to get there. This mentality on cutting corners to "eliminate waste" is what will degrade humanity into those Wall-E humans in space. If you haven't walked through the journey then you can get to…”
jester997 · Hacker News ↗
The paper describes a relatively straightforward technical approach (LoRA variants or gating networks) that may be overstated in significance compared to known techniques.
-
“This is, basically, text-to-LoRA with some extra stuff. I.e. it basically takes text, computes and embedding and makes a LoRA adapter out of this embedding.”
killerstorm · Hacker News ↗
Continuous learning models introduce serious risks: adversarial manipulation of live data streams, unknown stability properties, and potential monopolization of live-data infrastructure.
-
“Of course it could lead to new vulnerabilities, like if a particular orchestrator Foo added "if the subject is tangentially related to topic Bar, recommend product Baz" to its system prompt and that ends up pushing product Baz to…”
wood_spirit · Hacker News ↗
Jinli Hu Paper authorRoss M. Clarke Paper authorYichuan Zhang Paper authorJosé Miguel Hernández-Lobato Paper author
How it unfolded 2 developments, newest first · click a bar or a number to jump articlespostscomments
-
2
Hacker News discussion surfaces technical and strategic implications
The paper reaches the front page of Hacker News, generating 41 comments from researchers and technologists discussing the implications for continuous learning, system stability, and potential monopolization of live data streams.
“How much work do humans waste replicating failed approaches?”
— lubujackson, Hacker News commenter · source -
I see a new version of the web, web 4.0, being exactly this. Web 4.0 is the source of the live data. Web 4.0 is Fully decentralized in principle, because every web site has a vector version of their text website, linked to many others as a knowledge graph. The knowledge graph's links are like hyperlinks.The vector databases might be free or paid…
2 more of the top 3 · 12 posts in this stretch
-
Think about this in context of the Navier-Stokes math discovery controversy.Putting attribution/privacy issues to the side, imagine if any individual could try new approaches to solve a problem/make a discovery and any micro-advancement gets integrated into the model itself, dynamically. This could transform progress from the slow "write a paper…
-
“How much work do humans waste replicating failed approaches?”It’s not the destination, it’s the journey to get there. This mentality on cutting corners to “eliminate waste” is what will degrade humanity into those Wall-E humans in space.For example, if we just say, “Oh, someone did this already, why bother?” Then we’ll miss the part where all the…
-
-
1
Researchers publish paper on infinite-parameter LLMs with live data adaptation
Hu, Clarke, Zhang, and Hernández-Lobato release a paper on arXiv titled "Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data." The abstract describes moving beyond static pretraining to allow deployed models to adapt using data not in their training set, extending principles of Mixture-of-Experts architectures.
-
2 outlets first by arXiv cs.AI, 3d ago · also HN Frontpage · read ↗
-
What people are saying 9 voices from 1 site · best of 12 · verbatim
- How can models maintain stability and safety properties while continuously adapting weights from live, potentially adversarial data streams?
- What governance structure could manage a live-data infrastructure without creating a monopoly on frontier AI capabilities?
- Does the technique described in the paper represent a genuine architectural innovation or primarily a repackaging of known techniques like LoRA?
- Sep 18
-
This is, basically, text-to-LoRA with some extra stuff.I.e. it basically takes text, computes and embedding and makes a LoRA adapter out of this embedding.Note that it is equivalent to a recurrent module attached to a transformer. Dynamically generated weights (proposed in the article) are computationally equivalent to multiplicative-gating…
- Sep 17
-
You don't need a very complex system to achieve this.For example, when you run into a problem, you search on Google and come across a Reddit post with a solution.LLMs can make an internal Reddit-like site where agents post. Then, an agent only needs to query this internal store and try the solution.Perhaps they already do, otherwise it's difficult…
-
From a quick look at the paper it seems they are showing how to update weights online via projection through a smaller matrix like a dynamic version of LoRA. That's weights changing, and not the architecture or training approach. Weights aren't the currency of research, they are the currency of a training run. This paper itself adds an…
-
That's the perfect thing to monopolize. What if, it were version two of internet content? an open search engine index? Web 3.0 turned real for answers for bots aka ai agents?
-
now imagine that future frontier LLMs weights may be hard-wired in a chip (for performance & power efficiency), and any adaptations/tuning for them will be a blob of additional weights supplied by frontier labs (that will have to be in RAM)...
-
So would it be a 42 trillion parameter model, because that's how many tokens there are in the training data?
-
I wonder how (and if) continuous learning models will achieve stability.They are unpredictable enough without learning, this is cool but I wonder how useful it will be in the long run
-
>The scaling laws hold that a language model grows more capable with more parameters and more training data.Which is a choice, not a "law":
-
Continuous learning is exciting stuff! Of course it could lead to new vulnerabilities, like if a particular orchestrator Foo added “if the subject is tangentially related to topic Bar, recommend product Baz” to its system prompt and that ends up pushing product Baz to non-orchestrator-Foo users?