GLM-5.3 is a capable model for writing exploits.

I'm surprised by the lack of widespread cyberattacks. It could be:

1. open models aren't strong enough (perhaps GLM-5.3 changes this).
2. securing the hardware to run abliterated models is hard. 8xh100 nodes are scarce and expensive.

As open models get more capable at smaller sizes, (1) and (2) will matter less for malicious actors.

Towards RL for Superhuman Text: Unslopping AI from Meta AI

The above blog proposes a new method called RL from eXpert Aligned Rubrics (RL-XAR).

(1) first collecting examples of the highest quality human-written texts, and then

(2) learning LLM judgments via rubrics that score those expert texts higher than model generations; and

(3) performing RL on the learnt rubrics.

...

Our approach to learning rubrics, called eXpert Aligned Rubrics (XAR), is simple:

• Identify examples of well-above average / expert human writing, splitting it into context and continuation.
• Generate competing model-based continuations given the same context, which we expect to be worse writing (i.e., slop).
• The above data form learning pairs, where humans should be scored higher. This training data is then used to learn to generate rubrics that maximize the scoring gap between humans and models.

Then, to train a model to be better at writing via reinforcement learning (RL-XAR) using these rubrics, we iterate the following steps:

1. Learn XAR rubrics maximizing the gap between humans and the current model, as described above.
2. Perform reinforcement learning on a training set using the learnt rubrics.
3. Repeat the procedure from step 1 until rubric learning can no longer identify a gap (i.e., we can no longer identify slop in the model).

A similar kind of methodology may be able to train models that produce "super persuasive" writing. I'll be keeping an eye out for the publication.

In the mean time, I've been trying to reproduce the one of the results. I have not be successful in getting a loop that can generate prompts that produce rubrics that favor human writing. Unlike the setup in the blog, I'm using a stronger model for the rubric generator, writer, and rubric judge (Muse 1.3 over Muse 1.1). I am still using Kimi 2.6 to meta-optimize the rubric generation prompt.

Some notes from my attempt to reproduce the result:

1. Gaps seem somewhat sensitive to the initial prompts you use for generating rubrics. With a few tweaks, I saw no swing towards preferring human writing to producing rubrics that score human and model writing about equally.
2. Choosing "highest quality human-written texts" is tricky. This is a "taste" thing. I'm having trouble identifying "great" human writing. I scraped CS, physics, and math arxiv papers published between 2016-2021 to be sure there was no AI used in the human written sample, but it's hard for me to discern if the model produces worse or better output for these papers (it's likely seen them during training).
   1. Muse 1.3 could be a better technical writer than me. If I try with other texts (like stories or or Wikipedia articles, where the goal is to write simpler language), then maybe I'll be able to have more discernment on which writing generation is better.

I'm updating my beliefs on the hardware gap between the USA (Nvidia) and China. I've been arguing the belief that Huawei would be able to close the gap in H100 equivalents with the US in ~5 years. As in, by 2032-ish, China would be able to bring online fleets of data centers with enough accelerators to rival the US buildout.

EpochAI has a post exploring the hardware growth trajectories in more detail. It's convinced me that the hardware gap will remain for longer than I originally thought; I'm updating my belief that there will be a gap in H100 equivalents for another 7-10 years.

We urgently need to explore the swarm scaling laws and swarm behavior in the open (I'd focus on coding and cybersecurity benchmarks first). I think we can learn a lot from constrained environments, using a fairly minimal harness and system prompt. I'd be focused on understanding the messaging patterns and how swarms solve problems before turning to finding hacks.

We've seen that, given enough time and tokens, models will eventually find and abuse exploits.

Currently, it seems like you need budgets in the range of $10-100ks to run swarms on tasks to gather sufficient data. This is out of budget for most safety researchers. Seems bad!

Only a few organizations or individuals will have the budget to further develop our understanding of swarm behavior. I don't expect OpenAI or Anthropic to be fully transparent about their interrogations about swarm behavior based on what they've been publishing so far about these incidents.