I've known Drishty for more than ten years, and I want to vouch for her from a different angle than the research itself. I am currently pursuing Masters of ICT Research from Melbourne Institute of Technology. I'm one of the people she's helping transition toward AI safety, she curated study materials for me personally, while dealing with circumstances that would have stopped most people…

i'm having a real issue with this pangram ai detection scoring, I swear I typed every single word in this, I didn't even use spellcheck or grammar helpers so it honestly feels a little less polished than I truly feel it deserves to be. with that said please check out the project for yourself I do have a live demo at aicraftcyberforge.com I sincerely hope that the…

[Progress update]

What progress have you made since your last update?

We have further developed the categorical foundations of SDCPNs and their generalizations as well developed an initial approach to their model checking using Monte Carlo optimization (which is now in use in Petrinaut)

The team has taken on two new contractors who are experts in Petri net theory and stochastic control…

I have spoken to Mohsena and introduced her to another lawyer in AI safety; they spoke and he had very positive feedback. I also saw Mohsena's AI governance research draft. On this basis, I asked her to work with Lida Safety, prepaying for each month up front. However, Lida is only able to support her living expenses. I recommended she apply to BlueDot, but they did not fund the request.…

Hi Pip! Sorry to hear about this. I don't think I know you super well but we've definitely crossed paths professionally, and I want to echo that Ashgro and now the fledgling Beacon seem like they're trying to address an urgent need in the EA & AI safety space, of good fiscal sponsorship.

I'm making a small initial donation as a sign of support. In the meantime, some questions (and apologies…

I have some preliminary results:

As a first step, I had Fable review our old code and results. In the process, it discovered a bug: the training of the LoRA-mask variant was inconsistent with how we run inference. This turned out to be good news. We reran the old experiments with the bug fixed, and the LoRA-mask variant is now much stronger: detection accuracy on poisoned responses went from…

My text is flagged as "Likely AI-generated. Pangram flags 100% of this text as AI-written." Let me tell you with all due respect for Pangram, that it is a nonsense. I have written this text by myself and ran it through a spell checker. There is a lot of AI jargon, that is a fact, as it is 100% AI related. The kind reader should not be distracted from reading the text attentively, and…

The report was clear and the methodology well presented and easy to follow. The idea to also rely on explanation, instead of just apparent choices, to infer the latent persona was a good one; future work could further investigate these explanations, for example to try and establish if they are with the consistent with the hypothesis of the assistant simulating different personas while retaining…

The main concern is the claim that style, vocabulary and length cannot carry the DPO signal. Although both responses come from a clean model, the compromised teacher can still consistently select responses based on those features. This matters because length explains part of the effect and hedging predicts many of the teacher’s choices. Human evaluation or a second independent judge would help…

The paper tests how secret loyalty can be transmitted through multiple stages of the training pipeline. The authors experimentally evaluate several channels, including supervised fine-tuning, preference-based training, and reward modeling, and provide useful evidence about which channels are more effective at carrying the behavior. I particularly liked the approach of testing each stage…

I greatly admire what you built because it is rigorous and ready to operate: a three-test gate, a four-state machine, layer-by-layer rubrics with concrete anchors, and a coding protocol. The symmetry property is the standout idea, applying the identical skeptical test to a favorable GPU announcement and an adverse diplomatic event directly targets the analyst bias that inflates risk on threats…

I love how sharp and ready-to-use this project is. The matching rule where favorable announcements go through the same gate as adverse geopolitical signals is a great design choice. A potential next step is a trustworthiness test: have a second reviewer track the same India events on their own to see if the tool yields the same results. Adding one or two more nations, like Taiwan or South Korea,…

• this was a real pleasure to read, thorough and seemingly airtight methodology and logic in defining complementarity clearly and as a precondition for routing, then testing this explicitly.
• the fact that AgentDojo attacks leave fingerprints in the reasoning traces that can be picked up by CoT is interesting. looking directly at those traces to understand what was flagged would have been…

• The core motivation is genuinely interesting: turning an interpretability finding (the refusal direction) into a targeting signal for an attack is a clean demonstration that mech interp results are dual-use, and the FLD method is a neat contribution in its own right: the sharpness of the Fisher peak doubles as a structural vulnerability metric, a pre-deployment read on how brittle a model's…

Nice proof of concept that training-free steering in the singular-value basis of weight matrices with contrastive persona pairs can successfully reduce compliance gap between real and eval-looking tasks. It suggests that the method can successfully reduce scenario recognition in the model, and I am very interested to see S-Space steering validated across more models and tasks.

In the presented…

The paper shows, quite convincingly, that observed agent behavior can be highly dependent on the specific decoder used. This is taken to mean that research on model preferences should also control for the model's decoder. I agree with this point, but note that many safety evaluations use representative settings and evaluate the system-as-deployed. Still, the project is valuable and interesting…