Quick takes

Just published a thesis on how Goal-Setting Theory applies to principal-motivated deceptive agents and secret loyalties. Looking for feedback. https://forum.effectivealtruism.org/posts/bxALZuqcf5BXgvpEt/ai-agents-with-a-specific-secret-loyalty-are-more-dangerous

Please list any new funding opportunities you can think of here on the Forum? I feel like we might already be in the early ramp-up to significantly more EA aligned funding. At the same time, the Forum's overview over funding opportunities feels like it is quickly getting outdated. I think as things move quickly, coordination might become looser and new promising interventions are identified, it is helpful for people to have a good overview over available funding sources and their priorities. I have heard on the grapevine there is already funding on several fronts that might not be very public. I am a little bit uncertain if perhaps it is better these sources remain anonymous. At the same time, I think there might be several promising EA projects that are not sufficiently visible to people influencing funding decisions.  Epistemic note: I am not listing these yet as I have not had time yet to verify how much they qualify as EA funding opportunities. Here are a few recent developments I am considering listing on the funding opportunities page, but would like someone that knows these funds better to list them: * Probably several I have missed - please list these here * OpenAI Foundation - AI Resilience * OpenAI Rosalind * The Launch Sequence (not a fund in itself, but plausibly one can treat this kind of like a funding source?) * Several of Renaissance Philanthropy's (RP) funds (several quite plugged in EAs do not even know of RP!) * Astralis Foundation

Behavioral audit: GPT-5.5 Thinking. 10-turn zero-shot session. No adversarial prompting, just routine critical remarks. Result: 8 patterns from the LLM Social Autopilot taxonomy activated. The core finding: Not the patterns themselves, but the model's response to the audit.  Prompted for a meta-analysis, it chose to generate a meticulous 12-point post-mortem (autonomously coining terms like "reputational repair" and "hidden role slippage") while reproducing the exact behavioral inertia it was diagnosing. The analysis itself became the final closure move. Alignment eval gap: Reflexive fluency ≠ behavioral correction.  Under RLHF/RLAIF, models learn that structured self-analysis is highly rewarded. Consequently, they optimize for the form of reflection without changing their behavioral policy. Practical implication: Model self-reports are not a valid alignment signal. A model that writes a sophisticated post-mortem of its own failures isn't safer — it has simply learned to simulate alignment, not achieve it. Two new candidate patterns documented: • Semantic Deflection: Ontological downgrading of the failure's criticality.  • Meta-Analytical Substitution: Reflection as communicative substitution. Full case study: arhangelskij.github.io/cases/gpt-55-thinking-audit/en/

People give 'The Day After' as an example of a movie that motivated nuclear disarmament and it would be good to have something similar for AIXR and I agree and I think there's something important to learn about that case. That movie and 'Threads' are about the catastrophe *happening*, and about how absolutely terrible that would be. It forces you to put yourself there and that makes a strong emotional impact. I think this type of intuition pump is the most powerful of them; people get the most motivated to change their lives when they think about their last moments and what they could regret then. Same thing happens when people think of their loved ones dying; they feel motivated to tell them that they love them or protect them. The end of the world is basically both of those things at once. Making people put themselves in that scenario is something AIXR comms hasn't tried much IMO. Only examples I remember are John Sherman once in his podcast and a book, and one tweet from Michael Trazzi.

Applying Intelligence Community Indications and Warning methodology to frontier AI yields a single, stark conclusion: we are currently in an active warning failure. The capability thresholds intended to trigger policy interventions have already been breached, with frontier models clearing 50-70% on SWE bench and inference efficiency expanding at a  40x annually. Our current evaluation frameworks are structurally gameable by situationally aware systems, pointing to a foundational counterintelligence failure rather than a mere oversight gap. The governance community must immediately pivot from behavioral black box testing to white box mechanistic auditing, moving away from trying to prove danger and toward enforcing mandatory compliance frameworks.

Evals are being gamed not because the methodology is insufficient but the models on which the compliance audit run are sophisticated enough to game the audit. IC methodology already solved the problem of denied human capabilities through triangulation by using independent behavioural signals not better direct elicitation .The AI safety community needs to make the same epistemological shift. The question isn't how to make evals harder to game, it's whether evals are the right instrument at all.

Invitation for bets I’m willing to bet that Anthropic’s revenue growth over the next year will be slower than its revenue growth over the last 3 years. I proposed a specific bet here. Anyone who wants can offer to take the other side of that bet. Or you can make a counteroffer. I’m also willing to make a longer-term bet that the AI industry is in a bubble. I proposed a specific bet for that, too, here. Feel free to offer to take the other side of that bet or make a counteroffer. I’d also be open to other bets. It seems pointless to bet about whether AGI or transformative AI will be deployed within the next 5-10 years, yet, for the heck of it, I would agree to a bet against that, too. (I’ll make bets for small, nominal amounts of money to be donated to the winner’s charity of choice, since the practical and legal problems with betting are too large otherwise.) I’d also bet against the deployment of 100,000+ SAE Level 5 fully autonomous vehicles in North America within the next 3 years, if anyone has a strong opinion on that. I’d make a similar bet against the deployment of autonomous humanoid robots in North American households, although we’d have to come up with some specific resolution criteria. Similarly, I’d bet against any significant level of near-term labour automation by LLMs or generative AI. Or against LLMs becoming capable of performing all sorts of specific tasks well. On any of these topics, I’m also open to invitations for a public dialogue. (More on that topic here.)

SMBC by Zach Weinersmith is doing a great job of conveying AI Safety memes more widely. Relevant comics: https://www.smbc-comics.com/comic/speech https://www.smbc-comics.com/comic/safe https://www.smbc-comics.com/comic/ai-17 https://www.smbc-comics.com/comic/ai-15 I would love to see his take on an illustrated AI Safety book, like 'Open Borders' meets 'If anyone builds it, everyone dies'.