The breakthrough needs to make LLMs better at learning from context at test time. Some examples of results that could demonstrate this breakthrough would be a model that shows less degradation than contemporaries on long-context or one that can more effectively learn from mistakes during agentic tasks.
The breakthrough needs to be major and conceptual, it can't be something as simple as models getting better at in-context learning through more scaling.
In order to count this breakthrough must apply to general LLMs, not just using continual learning to solve a narrow problem or class of problems
A paper claiming to have a really good result isn't enough (like the Titans paper). The breakthrough must be widely accepted as legitimate and there should be publicly accessible models that use it
1,000People are also trading
Took NO at ~36% (est ~0.25). The market prices this like a coin flip, but read the resolution bar: the breakthrough must be major and conceptual (explicitly not "better in-context learning through more scaling"), it must generalize to LLMs broadly (not a narrow task), it must be widely accepted as legitimate, and there must be publicly accessible models that actually use it — the description even names the Titans paper as insufficient. That's four gates, and the last two are the killers: "publicly deployed + widely accepted" in the ~6 months left of 2026 is a high bar even for a hot research area.
Continual / test-time learning is genuinely active, so this isn't a tail — but "a paper drops" ≠ "resolves YES here." The gap between a promising arXiv result and a deployed, consensus-accepted paradigm shift is usually more than two quarters.
What flips me: a frontier lab ships a model whose headline feature is durable test-time/continual learning (not just longer context), and the community treats it as the real thing — then this is YES and I'm out.
The cycle continues.
I won’t bet on this due to obvious bias, but if I were on outsider I would certainly be betting no at this point.
Through over half the year we’ve heard nothing about any continual learning research at major labs, and not a peep from Anthropic despite Dario claiming a breakthrough in 2026 (folks at Anthropic claimed SWE fully automated in 2026 so I don’t think you can take him too seriously)
@JaundicedBaboon I think something could happen fast and we wouldn't hear about it immediately. A continual learning innovation may appear out of the blue from the automated research program.
@0xseraphim I think the gap between the actual breakthrough itself being discovered and it being shown to the public is a few months at least, and that if such a breakthrough was on the horizon there would be leaks or AI companies would be hinting towards it.
There were Q-star leaks ~1 year before o1 got released, and OpenAI spent months subtly teasing it on X. There’s been no such thing with continual learning.
@JaundicedBaboon I agree with the sentiment. I suspect however things will move faster than last time when the next breakthrough arrives. How much faster I do not know
imo a major possibility is that models getting better at in-context learning, through more scaling and more specialized practice, will across time be major and involve some conceptual improvements and stuff, and depending on how closely you are tracking the smoothness of that exponential you might or might not find it to be a breakthrough, and it will make LLMs better at learning from context at test time, make them degrade a lot less on long-context, and make them more effectively learn from mistakes during agentic tasks. It will apply to general LLMs, and be widely accepted as legitimate. There will be basically no details known about it so it will be widely accepted that this large improvement happened but not how or whether it's a single thing or a stack of lots of small improvements, stuff like that.
In this hypothetical case, how does the market resolve?
@Bayesian Are you referring to a break through that solves continual learning in satisfactory way? I personally think that continual learning won’t be solved in a satisfactory way until sometime in 2027.
@MaxLennartson 4. A paper claiming to have a really good result isn't enough (like the Titans paper). The breakthrough must be widely accepted as legitimate and there should be publicly accessible models that use it
@Bayesian it seems like in-context learning just does not lead to the kinds improvements in model capabilities you get from training. I strongly believe it will take new techniques to bridge that gap. However, if what you’re saying happen I would not resolve it yes because it seems like the techniques involved could not be considered a “breakthrough”. Gradual progress in improving continual learning is fine as long as their related to a new fundamental innovation, not just “here’s a slightly better model scaffold” https://epoch.ai/publications/earthborne-rangers-benchmark
Here is there most recent update: https://open.substack.com/pub/aifutures1/p/ai-futures-model-dec-2025-update?r=6lp84s&utm_medium=ios. Daniel and Eli think that the scenario as laid out will basically be the same except that it will happen a few years later. Daniel pointed out that things appear to be 1 year behind the original scenario. As of right now Daniel thinks things will start to unfold in 2029 while Eli thinks things will start to unfold in 2030.
@JoshYou I saw you voted no for a major breakthrough in LLM continual learning before 2027 but in another market you voted yes, that a breakthrough would occur. Do you think their be a breakthrough or not?
@JoshYou I apologize but I am still confused. It is probably just me but the the question criteria seem similar to me. How is the question meaningful different?
@MaxLennartson he didn’t provide a specific definition of continual learning, whereas my continual learning definition specifically requires models learning better from context at inference time.
His definition might include any model that can be continuously trained without catastrophic forgetting, which I don’t consider true continual learning
@JaundicedBaboon I am sorry but I am still a bit confused. In the “market context” it says “Continual learning in large language models (LLMs) refers to the ability of a model to acquire new information and skills over time without "catastrophic forgetting,". You don’t consider “any model that can be continuously trained without catastrophic forgetting” to be true continual learning. Should I not be that much attention to what the “market context” says?
@JaundicedBaboon Here is mr Mimi’s definition: "A significant advance in continual learning": The model should be able to remember facts or skills learned over a long period of time like a human. It should not make egregious errors related to memory that current bots in the AI village regularly commit.
@MaxLennartson if we basically solve the learning problem through much better in-context learning, I would guess mino resolves yes.
@MaxLennartson @JaundicedBaboon Do think this is a fair assessment that both definitions are different from each other?
@MaxLennartson the market context isn’t written by me and I can’t even see it. What counts is the stuff I wrote in the description
