hello, I’m Jo.
Jo Jiao
Karma: 373
Context: I did the full version of the sim. I ran the AI 2027 tabletop wargames in 2025, and they have since updated the material. So, some points I make here may not carry over.
As far as I know, Conclave 1492 is different & similar in the following ways to the AI 2027 tabletop wargames:
Just like in the wargames, your goal is to be true to your character, not win (though indeed many characters will want to win in some subgame by e.g. becoming pope). You are given a rich enough character sheet (mine was 30+ pages) that you will know your background, your motivation, the context behind your goals, your personality, and your preferences.
You will write letters to NPCs and PCs.
You have at the start resources such as currencies, items, differential information, and armies that reflect the standing of your character in 1492, and you will need to use these resources carefully to achieve your objective.
The war phase of Conclave 1492 has its own complicated rules and internal logic.
You will learn a ton of Italian Renaissance history and moreover get a feel of the zeitgeist of 15th century Europe.
Hi Aryaman, thanks again for the great technical writeup in the mech interp article. Moving to the mech interp talk page to address the COI and RS concerns.
Good call! Linking the relevant pages.
I don’t think they’ve become less important. Wikipedia is pretty heavily cited by LLMs when they go and do their own research in my experience, so Wikipedia articles are still valuable even if fewer humans visit it.
On the point of Google not prioritizing it so heavily—I don’t think Google indexes a lot of new Wikipedia articles but old established articles still top the search result. In our case, the mech interp wiki page never got indexed by Google until a Wikipedia New Page reviewer marked it as reviewed a couple days ago—now it’s a top result.
totally had Henry’s voice playing while reading your comment
My best attempt at attempting to characterize Kant’s Transcendental Idealism - Kant’s idealism says that essence—not existence—is dependent on us. That is to say, what it is to be is dependent on how we understand. For example, the schema of classification in biology, such as genetic proximity, depends on what purposes they serve to us. What it is for animals to be depends, in other words, on the biologist. To draw the biology analogy ad absurdum, transcendental idealism says something like “the genetic composition is the condition of the possibility of how we are able to make sense of biological objects in the first place”. The existence of these classification schema is dependent on our mind a priori.
The distinction between inner and outer alignment is quite unnatural. For example, even the concept of reward hacking implies the double-fold failure of a reward that is not robust enough to exploitation, and a model that develops instrumental capabilities as to find a way to trick the reward; indeed, in the case of reward hacking, it’s worth noting that depending on the autonomy of the system in question, we could attribute the misalignment as inner or outer. At its core, this distinction comes out of the policy <-> reward scheme of RL, though prediction <-> loss function in SL can be similarly characterized; I doubt how well this framing generalizes to other engineering choices.
Eliezer seems on track to win: current AI benchmark for IMO geometry problems is at 27⁄30 (IMO Gold human performance is at 25.9/30). This new benchmark was set by LLM-augmented neurosymbolic AI.
Thank you for the insightful post! You mentioned that:
Consider the relation a transformer has to an HMM that produced the data it was trained on. This is general—any dataset consisting of sequences of tokens can be represented as having been generated from an HMM.
and the linear projection consists of:
Linear regression from the residual stream activations (64 dimensional vectors) to the belief distributions (3 dimensional vectors).
Given any natural language dataset, if we didn’t have the ground truth belief distribution, is it possible to reverse engineer (data model) a HMM and extract the topology of the residual stream activation?
I’ve been running task salient representation experiments on larger models and am very interested in replicating and possibly extending your result to more noisy settings.
I wish people reported their P(catastrophe) : P(doom)[1] ratio in addition to timelines. Whenever I hear that people’s p(doom) is at e.g. 5%, I immediately wonder, what about their p(catastrophe)? Is it 10%, or 20%, or 50%?
There’s a few reasons why I am curious about this ratio.
One simple reason is it serves as a consistency check. I think a lot of the times doom is used to mean ‘something really bad’, but actually catastrophe already qualifies as ‘something really bad’. So it’s good to reflect on whether you’d endorse what your p(doom) would imply about p(catastrophe).
A second reason is that I think the p(catastrophe) : p(doom) ratio reveals interesting information about people’s threat models, once people actually take them to denote different things. If the ratio is low, then plausibly their threat model is more dominated by immediate takeover risks. I for one think the ratio is really high. I think catastrophe at our current pace is very plausible, but e.g. extinction is far, far more implausible. To put some number to things, I think p(doom) is at <0.2%[2] and p(catastrophe) is at >20%[3] i.e. P(catastrophe) : P(doom) is at about 100:1, or equivalently given my definition p(doom | catastrophe) < 1%.
For the sake of precision here, we can say catastrophe means >10% of the global population dies in excess of baseline mortality rates in a sequence of events before 2100, and say doom means extinction before 2100. I don’t think the definition is critical here (it’s there to make rigorous the difference between ‘really really bad’ and ‘bad’), and I think you should report this ratio even when you disagree with the definition.
FWIW, I think this is terrible given that our base rates are a few OOMs lower.
Could be way higher, but >20% is the number I’m comfortable committing to.