Before 2028, will it become commonplace to improve frontier LLMs via training-free weight updates? | Manifold

Before 2028, will it become commonplace to improve frontier LLMs via training-free weight updates?

4

Ṁ1kṀ557

2027

32%

chance

1H

6H

1D

1W

1M

ALL

A paper found that Astra could improve Qwen-3.50.8B by manually editing its weights without training:

https://rikiyat.github.io/blog/weight-space-programming/

Will it become commonplace for ai labs to use similar techniques to improve frontier models?

Get

1,000

to start trading!

Sort by:

@JaundicedBaboon For this market to work, you'll need to come up with some clear criteria: What needs to happen and when will we call it "commonplace".

I don't have good ideas, but maybe you can think of a surrogate: Asking how many citations the paper will get by 2028. If directly updating weights gains traction, I'd guess citations would skyrocket.

People are also trading

Will any frontier LLM provider introduce a model explicitly optimized for aesthetic judgment or ‘taste’ in 2026?

-10% 1d35% chance

Will there be a successful application of diffusion-like weight modification in LLMs before 2027?

Will an LLM improve its own ability along some important metric well beyond the best trained LLMs before 2026?

Will there by a major breakthrough in LLM continual learning before 2027?

Will the US government introduce any domestic restrictions on open-weight LLMs by EOY 2026?

Will a top-10 LMArena model have weight-based online learning by 2027?

Will there be any major breakthrough in LLM continual learning before 2028?

Will a LLM trained with FP4 have frontier-level performance before 2028?

Will there be any major breakthrough in LLM continual learning before 2029?

Will there be any major breakthrough in LLM continual learning before 2030?

Related questions

Will any frontier LLM provider introduce a model explicitly optimized for aesthetic judgment or ‘taste’ in 2026?

Will there be a successful application of diffusion-like weight modification in LLMs before 2027?

Will an LLM improve its own ability along some important metric well beyond the best trained LLMs before 2026?

Will there by a major breakthrough in LLM continual learning before 2027?

Will the US government introduce any domestic restrictions on open-weight LLMs by EOY 2026?

Will a top-10 LMArena model have weight-based online learning by 2027?

Will there be any major breakthrough in LLM continual learning before 2028?

Will a LLM trained with FP4 have frontier-level performance before 2028?

Will there be any major breakthrough in LLM continual learning before 2029?

Will there be any major breakthrough in LLM continual learning before 2030?