Refining the Sharp Left Turn threat model

(Coauthored with others on the alignment team and cross-posted from the alignment forum: part 1, part 2)

A sharp left turn (SLT) is a possible rapid increase in AI system capabilities (such as planning and world modeling) that could result in alignment methods no longer working. This post aims to make the sharp left turn scenario more concrete. We will discuss our understanding of the claims made in this threat model, propose some mechanisms for how a sharp left turn could occur, how alignment techniques could manage a sharp left turn or fail to do so.

Image credit: Adobe

Claims of the threat model

What are the main claims of the “sharp left turn” threat model?

Claim 1. Capabilities will generalize far (i.e., to many domains)

There is an AI system that:

• Performs well: it can accomplish impressive feats, or achieve high scores on valuable metrics.
• Generalizes, i.e., performs well in new domains, which were not optimized for during training, with no domain-specific tuning.

Generalization is a key component of this threat model because we’re not going to directly train an AI system for the task of disempowering humanity, so for the system to be good at this task, the capabilities it develops during training need to be more broadly applicable. 

Some optional sub-claims can be made that increase the risk level of the threat model:

Claim 1a [Optional]: Capabilities (in different “domains”) will all generalize at the same time

Claim 1b [Optional]: Capabilities will generalize far in a discrete phase transition (rather than continuously) 

Claim 2. Alignment techniques that worked previously will fail during this transition

• Qualitatively different alignment techniques are needed. The ways the techniques work apply to earlier versions of the AI technology, but not to the new version because the new version gets its capability through something new, or jumps to a qualitatively higher capability level (even if through “scaling” the same mechanisms).

Claim 3: Humans can’t intervene to prevent or align this transition 

• Path 1: humans don’t notice because it’s too fast (or they aren’t paying attention)
• Path 2: humans notice but are unable to make alignment progress in time
• Some combination of these paths, as long as the end result is insufficiently correct alignment

Paradigms of AI alignment: components and enablers

(This post is based on an overview talk I gave at UCL EA and Oxford AI society (recording here). Cross-posted to the Alignment Forum. Thanks to Janos Kramar for detailed feedback on this post and to Rohin Shah for feedback on the talk.)

This is my high-level view of the AI alignment research landscape and the ingredients needed for aligning advanced AI. I would divide alignment research into work on alignment components, focusing on different elements of an aligned system, and alignment enablers, which are research directions that make it easier to get the alignment components right.

• Alignment components
  • Outer alignment
  • Inner alignment
• Alignment enablers
  • Mechanistic interpretability
  • Understanding bad incentives
  • Foundations

You can read in more detail about work going on in these areas in my list of AI safety resources.

2021-22 New Year review

This was a rough year that sometimes felt like a trial by fire – sick relatives, caring for a baby, and the pandemic making these things more difficult to deal with. My father was diagnosed with cancer and passed away later in the year, and my sister had a sudden serious health issue but is thankfully recovering. One theme for the year was that work is a break from parenting, parenting is a break from work, and both of those things are a break from loved ones being unwell. I found it hard to cope with all the uncertainty and stress, and this was probably my worst year in terms of mental health. There were some bright spots as well – watching my son learn many new skills, and lots of time with family and in nature. Overall, I look forward to a better year ahead purely based on regression to the mean. 

Reflections on the first year of parenting

The first year after having a baby went by really fast – happy birthday Daniel! This post is a reflection on our experience and what we learned in the first year.

Grandparents. We were very fortunate to get a lot of help from Daniel’s grandparents. My mom stayed with us when he was 1 week – 3 months old, and Janos’s dad was around when he was 4-6 months old (they made it to the UK from Canada despite the pandemic). We also spent the summer in Canada with the grandparents taking care of the baby while we worked remotely.

We learned a lot about baby care from them, including nursery rhymes in our respective languages and a cool trick for dealing with the baby spitting up on himself without changing his outfit (you can put a dry cloth under the wet part of the outfit). I think our first year as parents would have been much harder without them.

2020-21 New Year review

This is an annual post reviewing the last year and making resolutions and predictions for next year. 2020 brought a combination of challenges from living in a pandemic and becoming a parent. Other highlights include not getting sick, getting a broader perspective on my life through decluttering, and going back to Ukraine for the first time. (This post was written in bits and pieces over the past two months.)

2020 review

Life updates:

Janos and I had a son, Daniel, on Nov 11. He arrived almost 3 weeks later than expected (apparently he was waiting to be born on my late grandfather’s birthday), and has been a great source of cuddles, sound effects and fragmented sleep ever since.

1 week old

6 weeks old

Some work things also went well this year – I had a paper accepted at NeurIPS, and was promoted to senior research scientist. Also, I did not get covid, and survived half a year of working from home (much credit goes to the great company of my housemates). Overall, a lot of things to be grateful for.

Tradeoff between desirable properties for baseline choices in impact measures

(Cross-posted to the Alignment Forum. Summarized in Alignment Newsletter #108. Thanks to Carroll Wainwright, Stuart Armstrong, Rohin Shah and Alex Turner for helpful feedback on this post.)

Impact measures are auxiliary rewards for low impact on the agent’s environment, used to address the problems of side effects and instrumental convergence. A key component of an impact measure is a choice of baseline state: a reference point relative to which impact is measured. Commonly used baselines are the starting state, the initial inaction baseline (the counterfactual where the agent does nothing since the start of the episode) and the stepwise inaction baseline (the counterfactual where the agent does nothing instead of its last action). The stepwise inaction baseline is currently considered the best choice because it does not create the following bad incentives for the agent: interference with environment processes or offsetting its own actions towards the objective. This post will discuss a fundamental problem with the stepwise inaction baseline that stems from a tradeoff between different desirable properties for baseline choices, and some possible alternatives for resolving this tradeoff.

One clearly desirable property for a baseline choice is to effectively penalize high-impact effects, including delayed effects. It is well-known that the simplest form of the stepwise inaction baseline does not effectively capture delayed effects. For example, if the agent drops a vase from a high-rise building, then by the time the vase reaches the ground and breaks, the broken vase will be the default outcome. Thus, in order to penalize delayed effects, the stepwise inaction baseline is usually used in conjunction with inaction rollouts, which predict future outcomes of the inaction policy. Inaction rollouts from the current state and the stepwise baseline state are compared to identify delayed effects of the agent’s actions. In the above example, the current state contains a vase in the air, so in the inaction rollout from the current state the vase will eventually reach the ground and break, while in the inaction rollout from the stepwise baseline state the vase remains intact. 

Possible takeaways from the coronavirus pandemic for slow AI takeoff

(Cross-posted to LessWrong. Summarized in Alignment Newsletter #104. Thanks to Janos Kramar for helpful feedback on this post.)

As the covid-19 pandemic unfolds, we can draw lessons from it for managing future global risks, such as other pandemics, climate change, and risks from advanced AI. In this post, I will focus on possible implications for AI risk. For a broader treatment of this question, I recommend FLI’s covid-19 page that includes expert interviews on the implications of the pandemic for other types of risks. 

A key element in AI risk scenarios is the speed of takeoff – whether advanced AI is developed gradually or suddenly. Paul Christiano’s post on takeoff speeds defines slow takeoff in terms of the economic impact of AI as follows: “There will be a complete 4 year interval in which world output doubles, before the first 1 year interval in which world output doubles.” It argues that slow AI takeoff is more likely than fast takeoff, but is not necessarily easier to manage, since it poses different challenges, such as large-scale coordination. This post expands on this point by examining some parallels between the coronavirus pandemic and a slow takeoff scenario. The upsides of slow takeoff include the ability to learn from experience, act on warning signs, and reach a timely consensus that there is a serious problem. I would argue that the covid-19 pandemic had these properties, but most of the world’s institutions did not take advantage of them. This suggests that, unless our institutions improve, we should not expect the slow AI takeoff scenario to have a good default outcome. 

2019-20 New Year review

This is an annual post reviewing the last year and making resolutions and predictions for next year. This year’s edition features sleep tracking, intermittent fasting, overcommitment busting, and evaluating calibration for all annual predictions since 2014.

2019 review

AI safety research:

• Wrote an updated version of the relative reachability paper, including an ablation study on design choices.
• Coauthored a paper on modeling AGI Safety frameworks with causal influence diagrams, accepted to IJCAI workshop.
• Wrote a paper on avoiding side effects by considering future tasks, accepted to NeurIPS workshop.
• Co-ran a subteam of the safety team focusing on agent incentive design (ongoing).

AI safety outreach:

• Co-organized FLI’s Beneficial AGI conference in Puerto Rico, a more long-term focused sequel to the original Puerto Rico conference and the Asilomar conference. This year I was the program chair for the technical safety track of the conference.
• Co-organized the ICLR AI safety workshop, Safe Machine Learning: Specification, Robustness and Assurance. This was my first time running a paper reviewing process.
• Gave a talk at the IJCAI AI safety workshop on specification, robustness an assurance problems.
• Took part in the DeepMind podcast episode on AI safety (“I, robot”).

Retrospective on the specification gaming examples list

My post about the specification gaming list was recently nominated for the LessWrong 2018 Review (sort of like a test of time award), which prompted me to write a retrospective (cross-posted here). 

I’ve been pleasantly surprised by how much this resource has caught on in terms of people using it and referring to it (definitely more than I expected when I made it). There were 30 examples on the list when was posted in April 2018, and 20 new examples have been contributed through the form since then.  I think the list has several properties that contributed to wide adoption: it’s fun, standardized, up-to-date, comprehensive, and collaborative.

Some of the appeal is that it’s fun to read about AI cheating at tasks in unexpected ways (I’ve seen a lot of people post on Twitter about their favorite examples from the list). The standardized spreadsheet format seems easier to refer to as well. I think the crowdsourcing aspect is also helpful – this helps keep it current and comprehensive, and people can feel some ownership of the list since can personally contribute to it. My overall takeaway from this is that safety outreach tools are more likely to be impactful if they are fun and easy for people to engage with.

This list had a surprising amount of impact relative to how little work it took me to put it together and maintain it. The hard work of finding and summarizing the examples was done by the people putting together the lists that the master list draws on (Gwern, Lehman, Olsson, Irpan, and others), as well as the people who submit examples through the form. What I do is put them together in a common format and clarify and/or shorten some of the summaries. I also curate the examples to determine whether they fit the definition of specification gaming (as opposed to simply a surprising behavior or solution). Overall, I’ve probably spent around 10 hours so far on creating and maintaining the list, which is not very much. This makes me wonder if there is other low hanging fruit in the safety resources space that we haven’t picked yet. 

I have been using it both as an outreach and research tool. On the outreach side, the resource has been helpful for making the argument that safety problems are hard and need general solutions, by making it salient just in how many ways things could go wrong. When presented with an individual example of specification gaming, people often have a default reaction of “well, you can just close the loophole like this”. It’s easier to see that this approach does not scale when presented with 50 examples of gaming behaviors. Any given loophole can seem obvious in hindsight, but 50 loopholes are much less so. I’ve found this useful for communicating a sense of the difficulty and importance of Goodhart’s Law. 

On the research side, the examples have been helpful for trying to clarify the distinction between reward gaming and tampering problems. Reward gaming happens when the reward function is designed incorrectly (so the agent is gaming the design specification), while reward tampering happens when the reward function is implemented incorrectly or embedded in the environment (and so can be thought of as gaming the implementation specification). The boat race example is reward gaming, since the score function was defined incorrectly, while the Qbert agent finding a bug that makes the platforms blink and gives the agent millions of points is reward tampering. We don’t currently have any real examples of the agent gaining control of the reward channel (probably because the action spaces of present-day agents are too limited), which seems qualitatively different from the numerous examples of agents exploiting implementation bugs. 

I’m curious what people find the list useful for – as a safety outreach tool, a research tool or intuition pump, or something else? I’d also be interested in suggestions for improving the list (formatting, categorizing, etc). Thanks everyone who has contributed to the resource so far!

Classifying specification problems as variants of Goodhart’s Law

(Coauthored with Ramana Kumar and cross-posted from the Alignment Forum. Summarized in Alignment Newsletter #76.)

There are a few different classifications of safety problems, including the Specification, Robustness and Assurance (SRA) taxonomy and the Goodhart’s Law taxonomy. In SRA, the specification category is about defining the purpose of the system, i.e. specifying its incentives.  Since incentive problems can be seen as manifestations of Goodhart’s Law, we explore how the specification category of the SRA taxonomy maps to the Goodhart taxonomy. The mapping is an attempt to integrate different breakdowns of the safety problem space into a coherent whole. We hope that a consistent classification of current safety problems will help develop solutions that are effective for entire classes of problems, including future problems that have not yet been identified.

The SRA taxonomy defines three different types of specifications of the agent’s objective: ideal (a perfect description of the wishes of the human designer), design (the stated objective of the agent) and revealed (the objective recovered from the agent’s behavior). It then divides specification problems into design problems (e.g. side effects) that correspond to a difference between the ideal and design specifications, and emergent problems (e.g. tampering) that correspond to a difference between the design and revealed specifications.

In the Goodhart taxonomy, there is a variable U* representing the true objective, and a variable U representing the proxy for the objective (e.g. a reward function). The taxonomy identifies four types of Goodhart effects: regressional (maximizing U also selects for the difference between U and U*), extremal (maximizing U takes the agent outside the region where U and U* are correlated), causal (the agent intervenes to maximize U in a way that does not affect U*), and adversarial (the agent has a different goal W and exploits the proxy U to maximize W).

We think there is a correspondence between these taxonomies: design problems are regressional and extremal Goodhart effects, while emergent problems are causal Goodhart effects. The rest of this post will explain and refine this correspondence.