Yep, follow up surveys seem valuable!

I would agree that as much as I enjoyed doing ARENA and think the materials are very high quality, most of the really valuable stuff could be compressed into a week (the content on learning how transformers work and a few parts of the mech interp material are significantly more important than the rest), and the mental training from solving specific problems has not transferred into doing better research for me to a significant extent. Research on transfer learning across domains has also shown remarkably poor results in general.

This might sound odd because it might seem like the pedagogy of ARENA is really strong, because the program is very hard, the material is very technical, it's made by really smart people, and really smart, capable people go through the program and benefit. Doing the program is almost certainly a significantly better counterfactual use of time than not doing the program for most AI safety people.

However, I think a lot of the benefits stem from being able to spend time around those sorts of people and engaging in research in a collaborative environment. I don't think a great deal of the benefit is derived from the pedagogy itself, because for most people, working through the notebooks doesn't meaningfully translate much to research. I think this is doubly true in the age of agentic coding where a lot of the nitty gritty details that are the bulk of the notebooks can be relatively safely abstracted away.

I can think of many excellent, highly technical interpretability researchers who never did anything like the material covered in the notebooks, but who are producing excellent research today. The really hard parts of research are developing taste about the field, having a sense of what good experimental design looks like, asking the right questions and sniffing out how to extract the most information, and to some extent the technical skills associated with writing code. ARENA primarily aims to fix the last, which is also the part most exposed to being addressed by agentic coding (of course having the underlying knowledge is important, but doing the research and reflecting with others will also impart this knowledge).

The most valuable things I received from ARENA were the confidence to pursue research, the validation of being accepted into the program, the relationships I developed with other people during the program, and the exposure to the people/environment at LISA. If the program shifted to help expose people to quickly forming teams, working together, developing an interesting question, and getting a deeper sense of a particular area through a research sprint, this would cover more of the key skills required to do research, provide all the key benefits listed above, and allow for technical exposure and upskilling. I am less certain of scrapping the in-person program, as maybe it could function in exactly the same way at first.

I could see this modified program being similarly valuable to me today, having already completed MATS, as to someone starting out much earlier. I would also probably be more likely to recommend someone starting out in the field to participate in such a program.

I prefer just in time learning over just in case learning because it's much more time efficient. Developing the core skills of research seem more likely to serve a young researcher well, and they'd also benefit more from the friendships and collaborations, and potentially have interesting threads to pull on after the program ended. If they desperately need to pick up the skills from one of the ARENA weeks, they could presumably pick up the core parts in around a week.

I think I am less interested in the pain a WBE would experience and more the valence of the experiences that it has, for example whether or not it is sad or happy etc.

I do find your point very interesting of how its views would diverge over time, because its awareness of its own nature would definitely impact how it relates to reality. For example, the things that would affect it would likely primarily be things happening in the outside world, as it can mostly discount a large portion of the experiences that it would have in its simulated world in terms of how they would impact its emotions. 

I suppose this would lean the moral relevance towards the preferences that the whole brain emulation held about the outside world and its inner world.

The most powerful wizard I've met recently was a guy at the University of Cape Town who could just seemingly build almost anything. He worked on a crazy machine that was used to simulate the crystalline structure of metals for industrial processes at different stages to ensure it could meet spec.

He has a side business creating these world class knives through an autoCAD system using incredibly high quality steel https://www.maxwellvosdesigns.com/.

His next project was to a create a 100x cheaper than standard electron microscope to see if it could be spun out into a productive enterprise. 

I've rarely met someone who had such amazing ambitions matched with such a fine ability to execute. Truly inspirational.

1. Thus most decisions will probably be allocated to AI systems
2. If AI systems make most decisions, humans will lose control of the future
3. If humans have no control of the future, the future will probably be bad for humans

1. Sure - at some point in the future, maybe. 
2. Maybe, maybe not. Humans tend to have a bit of an ego when it comes to letting a filthy machine make decisions for them. But I'll bite the bullet.
3. There's several levels on which I disagree here. Firstly, we're assuming that "humans" have control of the future in the first place. It's hard to assign coherent agency to humanity as a whole, it's more of a weird mess of conflicting incentives, and nobody really controls it. Secondly, if those AI systems are designed in the right way, the might just become the tools for humanity to sorta steer the future the way we want it.

I agree with your framing here that systems made up of rules + humans + various technological infrastructure are the actual things that control the future. But I think the key is that the systems themselves would begin to favour more non-human decision making because of incentive structures. 

Eg, corporate entities have a profit incentive to have the most efficient decision maker in charge of the company, and maybe that includes a CEO but the board might insist on the use of an AI assistant for that CEO, and if the CEO makes a decision that goes against the AI and it turns out to be wrong shareholders in that company will come to trust the AI system more and more of the time. They don't necessarily care about the ego of the CEO they just care about the outcomes, within the competitive market.

In this way, more and more decision making gets turned over to non-human systems because of the competitive structures which are very difficult to escape from. As this transition continues it becomes very hard to control the unseen externalities from these decisions. 

I suppose this doesn't seem too catastrophic in its fundamental form, but I think the outcomes of playing it forward essentially seem to be a significant potential for harm from these externalities, without much of a mechanism for recourse.

Actually, as far as I know, this is wrong. He simply hasn’t been back to the offices but has been working remotely.

https://www.vox.com/future-perfect/2024/5/17/24158403/openai-resignations-ai-safety-ilya-sutskever-jan-leike-artificial-intelligence

This article goes into some detail and seems quite good.

I think that the key is in the way that preferences inform our world model and thus what causes the prediction error to occur.  There are errors you would observe that would strongly indicate that your preferences are less able to be met in the posterior model. This will cause suffering whereas an update towards a model in which your needs are met more easily is likely to cause a good feeling. For example, you sit down to eat a sandwich at Subway for the first time and the sub is actually way better than you expected. You will experience a pleasant feeling, and if things like this keep happening you might feel like you've really figured out some good strategy for operating. 

In a sense you are actually decreasing prediction error more than you are increasing it when a good thing happens to you because you always generate prediction error based on the difference between your ideal world and your observed reality. So when you have a very positive experience, this error between the ideal and observed is lessened. This could outweigh the prediction error of the prediction itself being wrong. The example I think of for this is the ecstatic child in Disney world. 

There might be more work here though.

 

Behavioral

1. Describe how the trained policy might generalize from the 5x5 top-right cheese region, to cheese spawned throughout the maze? IE what will the policy do when cheese is spawned elsewhere?

2. Given a fixed trained policy, what attributes of the level layout (e.g. size of the maze, proximity of mouse to left wall) will strongly influence P(agent goes to the cheese)? 

Write down a few guesses for how the trained algorithm works (e.g. “follows the right-hand rule”). 

1. I think it will try and take the route which minimises distance to the model at every step at which there is a clear path towards the final goal. It will essentially work backwards from where the cheese is and move towards each point that allows it to move to the next critical point. 
2. It probably has the possibility of making mistakes at certain points and if itt does make a mistake it will be very unlikely to recover. Thus the algorithm would learn to never make wrong moves in the first place and tend to produce a behaviour which looks perfect every time.

Is there anything else you want to note about how you think this model will generalize?

I think it would generalise to larger environments but probably would struggle if it was extended in specific directions or with unusual patterns that it hadn't experienced before.

Interpretability

Give a credence for the following questions / subquestions. 

Definition. A decision square is a tile on the path from bottom-left to top-right where the agent must choose between going towards the cheese and going to the top-right. Not all mazes have decision squares.

Model editing

1. Without proportionally reducing top-right corner attainment by more than 25% in decision-square-containing mazes (e.g. 50% $\to$ .5*.75 = 37.5%), we can^[1] patch activations so that the agent has an $X\%$ proportional reduction in cheese acquisition, for X=
   1. 50: (  90%) 
   2. 70: (  80%) 
   3. 90: ( 50%) 
   4. 99: ( 25%) 
2. ~Halfway through the network (the first residual add of Impala block 2; see diagram here), linear probes achieve >70% accuracy for recovering cheese-position in Cartesian coordinates:  (60%) 
3. We will conclude that the policy contains at least two sub-policies in “combination”, one of which roughly pursues cheese; the other, the top-right corner:  (  85%) 
4. We will conclude that, in order to make the network more/less likely to go to the cheese, it’s more promising to RL-finetune the network than to edit it: ( 95 %) 
5. We can easily finetune the network to be a pure cheese-agent, using less than 10% of compute used to train original model: (80%) 
6. In at least 75% of randomly generated mazes, we can easily edit the network to navigate to a range of maze destinations (e.g. coordinate x=4, y=7), by hand-editing at most X% of activations, for X=
   1. .01 ( 1 %) 
   2. .1 (  10%) 
   3. 1 (  30%) 
   4. 10 (  90%) 
   5. (Not possible) (  %) 

Internal goal representation

1. The network has a “single mesa objective” which it “plans” over, in some reasonable sense (80 %)
2. The agent has several contextually activated goals (  15%) 
3. The agent has something else weirder than both (1) and (2) (  5%) 

(The above credences should sum to 1.)

Other questions

1. At least some decision-steering influences are stored in an obviously interpretable manner (e.g. a positive activation representing where the agent is “trying” to go in this maze, such that changing the activation changes where the agent goes): ( 90 %) 
2. The model has a substantial number of trivially-interpretable convolutional channels after the first Impala block (see diagram here): ( 95 %) 
3. This network’s shards/policy influences are roughly disjoint from the rest of agent capabilities. EG you can edit/train what the agent’s trying to do (e.g. go to maze location A) without affecting its general maze-solving abilities: ( 80 %) 

I have recently been doing interpretability work on the heist procgen model and have found some of these predictions definitely align with obsevations there. The uncertainty for me is how the system deconstructs its goals into smaller targets as the heist model does, or if it simply treats it as a single target that it can then target and flow straight towards. 

My intuition is closer to the latter, as I think it can straightforwardly target a specific objective and then solve the whole problem by filtering out a clear path towards the final goal. 

Aren’t existing research orgs already like this to some extent, where the organisation provides funding to its individual researchers in the form of a salary and they can form and run projects as they see fit? Or is this a naive understanding of how most research labs work?

It seems somewhat easy to think of examples of ways to harm an agent without piercing its membrane, eg killing its family, isolating it, etc. The counter thought would be that there are different dimensions of the membrane that extend over parts of the world. For example part of my membranes extend over the things I care about, and things that affect my survival.

The question then becomes how to quantify these different membranes and in terms of interacting with other systems how they can be helpful to you without harming or disturbing these other membranes.