Pivoting from Big Three consulting into making AI safer

We caught up with Mick Yang, an ex-BCG consultant now pursuing a computer science master's at the University of Pennsylvania thanks to a scholarship from Good Ventures. He aims to use this with his Oxford law degree and consulting chops to contribute to AI safety. Currently, he's focused on model evaluations¹ and how they inform policy. Mick shares his insights on transitioning from consulting to a more technically focused role in the rapidly evolving field of AI governance and safety.

Could you tell us about your decision to go back to school for a computer science degree?

The strongest reason was to move into AI safety and governance research professionally. I like to get really intensely into things rather than engaging shallowly. I wanted to understand the ins and outs of the technical systems I'm forming strong opinions on. I was particularly interested in foundation models² and frontier model scaling policies³, which led me to the field of AI evaluations. With the potential for future regulations to involve tests, benchmarks, and auditing, I figured it would be beneficial to learn how to develop approaches to evaluate models myself.

That being said, I know there are very brilliant, wonderful people already contributing to the field without formal technical backgrounds. The field is a large expanding pie, people of all backgrounds can find a way to contribute!

Why not self-study?

While it's possible, I found it quite challenging to do alongside a full-time consulting job at BCG. The variance in my working hours made it feel almost impossible. Coming from a law background, upkilling in technical AI was like jumping into the deep end with almost zero overlap with my previous formal training.

Going back to school also provides structured time, a designated learning period, and built-in feedback through assignments and office hours. Lastly, visa considerations played a role – a Master's degree is still one of the better ways to move to a high-impact location for work in this field.

What prompted you to focus your energy on AI safety specifically?

I'll split this into two parts: the impact potential and the career viability aspect. On the impact side, my interest in AI safety has been developing for a couple of years. At first, I came across research more related to corporate governance⁴. Given my law and corporate background, that felt like an obvious place to start. 

Impact: Evaluations are frequently mentioned in company policies as tools for assessing model capabilities and risks.  My curiosity led me to explore how evaluations differ from traditional machine learning benchmarks4, their limitations, and how they tie into concepts like red teaming and auditing. I explored promising new research in AI governance, particularly work by Jonas Schuett,  Markus Anderljung, the AI Safety Institute,  Apollo, METR, and the large labs (ex. Anthropic, OpenAI) and others. This confirmed my curiosity – and that these topics are pretty important to get right.

Evaluations are only able to capture a small subset of the total space of risk that we might be concerned about in AI safety. Not everything in the world that could be affected by AI can be captured in formalized, controlled tests. Even once we get the results of the tests, those primarily serve as input into open policy questions. 

Even though evaluations leave questions unanswered, they're a critical first layer. If your evaluations and tests aren't reliable in assessing behavior and capabilities across different models and settings, it's hard to build effective governance on top of that.

Career Viability: There are many topics one can be interested in or care passionately about, including many things that are good for the world, that won’t neccesarily be able to support a career’s worth of work. 

I considered whether there are enough institutions interested in this work, competent people to work with, and if it's an expanding space where people would be willing to pay for this expertise. Are organizations putting their money where their mouth is by hiring people to learn the thing or do the thing? Is it connected to a wider ecosystem of incentives and people's needs?

In the AI safety space, things are moving quickly. It might not go the way that I think it will, but  I believe this domain broadly supports a viable career path. 

How do you balance the technical and emotional aspects of working in AI safety?

Being human is both about thinking and feeling. The whole point of working on AI safety is the belief that this technology will have a massive impact on human society now and in the future. It's our obligation to try and make it better. So for me, it's a humanistic concern rather than a purely technical one.

At the same time, I'm happy that not everyone is moving into AI. From a system-level perspective, we have so many problems, and it's good to have a diversified portfolio of talent working on different things. If you care more about another critical issue, you should pursue that because it still needs to be solved.

What were some of the challenges you faced as a  consultant at BCG?

For some people, a job is a job. For others, all kinds of work feel equally complicit in a corrupt system. But what Consultants for Impact and I have in common is the shared belief that the societal impact of our work is non-trivial. 

In established organizations like BCG, it's challenging to steer things towards your own agenda. You're being paid to work on someone else's, usually very commercial, agenda, with a very particular set of tasks that you're allowed to do. The lows often include long, unpredictable hours. Every person is different. If you find the job too taxing, you won't have the capacity to think about impact – you'll just be struggling to survive. On top of that, there might be an additional burden of feeling like your work isn’t contributing to something meaningful.

For me, I struggled to find a sense of belonging due to the short-term, rotational project structure. I wanted to deepen expertise and have a consistent team, which I didn't get during my time there. When business is slow, consultants have to go and hustle hard for work, which can make it really, really stressful to get enough chances to perform highly.

What were the highlights and the ways you tried to create impact at BCG?

On the positive side, these well-established companies offer good infrastructure, like visa sponsorship and insurance. The exposure to different types of people and organizations is valuable. It helps set your baseline for what good management looks like in various cultural contexts.

In terms of impact, I was never fully convinced by the corporate approach to social impact. I struggled to see the value chain and its effects on the end beneficiary. Instead, I tried to create opportunities within BCG, like organizing a speaker series where I brought in external experts from universities and places like DeepMind and the Alan Turing Institute to discuss AI-related topics. This helped break through the day-to-day thought bubble and introduce new ideas about algorithmic development, data efficiency, and AI safety implications. While it's hard to quantify the impact, I believe these initiatives reached at least a few dozen people and potentially influenced how they think about AI and its societal implications.

What advice would you give to consultants considering a transition to more impact-oriented work, particularly in AI safety?

I'd suggest reading a lot, talking to the right people, and reflecting on your theory of the world and your timelines for various AI developments. Ultimately, it comes down to your sense of what a good life looks like for you!

For current consultants: Remember that big corporations are still made up of individuals. If you can find the right person or sponsor, you can do many more things than you might expect. Look for opportunities to bring in new perspectives and break through established thought patterns.

Lastly, technical upskilling is really hard to do with a full-time consulting job, although it might be easier if you come from more of a technical background than I do. Sometimes, it can help to start with topics that are more familiar or accessible to you and then follow the research! At the end of the day, it’s worth paying attention to the level of excitement or energy the work gives you – while still caring about the people affected by your work.

‍‍

‍Definitions:

1. Model evaluations are processes used to assess the performance, accuracy, and quality of machine learning / AI models.
2. Frontier models refer to the most advanced, cutting-edge AI models that push the boundaries of capabilities and performance in the field, like OpenAI’s GPT-4.
3. Model scaling policies are strategies or guidelines that govern the expansion or development of AI models, particularly as they grow in size, complexity, or capability and pose novel risks to users and society.
4. Corporate governance describes how organizations make decisions, particularly related to the selection and implementation of AI tools. For example, leaders may set up a responsible AI board to manage risks and ensure compliance with legal and strategic objectives.
5. Machine learning benchmarks are standardized datasets and tasks used to evaluate and compare the performance of machine learning models.