Questions & objections
This page contains Frequently Asked Questions about our Theory of Change. It does not include our main Theory of Change itself. On its own, this page thus lacks important topics.
We're proud that our activities overlap with those of other successful organizations like BlueDot Impact and 80,000 Hours. We're all nonprofits with the shared goal of preventing existential outcomes from advanced AI.
Capacity-building work in AI safety remains highly impactful, and we have a uniquely productized approach to fieldbuilding that we think is highly leveraged. We also fill gaps that other orgnizations aren't covering as much.
Growth and Scaling
We think scaling up courses massively is the best way to reach more exceptional people. Lens Academy is planning to accept 100% of applicants at any scale and guide them towards impactful actions.
We have three main reasons for this:
- Viral growth: If more people learn about AI Safety, then more people will talk about it, causing even more people to want to learn about it. This creates a positive feedback loop, leading to more top talent entering AI Safety. Most of the world's top talent still doesn't know about AI Safety. We'd like to change that.
- Neglectedness: There are already plenty of organizations taking a less scalable approach aimed exclusively at top talent (e.g. university AI Safety clubs like MIT's MAIA). While we think this is great, we think attempting a radically more scalable approach is promising yet highly neglected.
- Focus: There's only so much an organization can focus on at any time. By designing our program for scale from the start, we expect we have a real chance of scaling 100x larger than similar AI Safety courses while maintaining quality. We expect that to also include ~100x more of the top talent that we're excited about channeling into the further talent pipelines.
There's a large unmet demand for AI Safety, most applicants being rejected even from introductory courses. We're leveraging the existing ecosystem to find that demand. For example, MATS and Constellation (Astra and OpenAI fellowships) have recently sent out applicant rejection letters to thousands of people, suggesting Lens Academy as a next step.
As such, most of our current participants were already working on getting into AI Safety, looking to build their resume, learn more, and engage with others. They mostly heard about us through the AI Safety ecosystem, and we're able to help them on their journey.
Mid- to longer term, we want to grow much bigger than the current activated demand for AI Safety courses. We'll thus focus on people who are brand new to AI Safety, for whom we'll be their first formal touchpoint with the field. We'll reach them through top-of-funnel marketing -- both our own and through partnerships with other AI Safety stakeholders such as content creators. We will be testing different viral and paid growth engines as we move scale.
An increasing number of people wants to support AI Safety but can't (yet) get a full-time position. We think for many such people, facilitating courses is one of the most impactful things they can do. It's also fun and engaging, and allows them to grow their skillset. And because our AI tutor tackles a lot of the learning itself, our facilitators don't need to be experts in AI Safety.
Nevertheless, we think this will be one of our major challenges. We're currently experimenting with methods to make us less reliant on large numbers of facilitators, and we're also exploring exponential growth loops that allow previous participants to quickly become facilitators of subsequent groups.
Yes, but most are not set up in a way that makes it easy.
We find it helpful to distinguish between products and services. By approximation, services scale linearly while products scale exponentially.
We consider most fieldbuilding programs to be more like services, because scaling them 10-fold requires in the order of 10x the funding and effort. In contrast, product builders try to uncouple inputs from outputs, such that they can scale their outputs massively with much smaller increases in inputs. This makes it easier to scale and also creates more leverage to affect the world. That's how WhatsApp reached 1 Billion users with just 50 employees.
Anyone Can Join
Because this allows us to attract more top talent to AI Safety.
Our introductory courses are designed to make people care about AI Safety. Before they care, it's too early to ask them to invest time and effort into an application process.
By accepting everyone also helps in creating a viral-growth feedback loop, which in itself is important to reach more top talent. See Why are you so focused on scaling?.
We also think the broader public may have ways to be impactful too.
This only works if we're not capacity-constrained, of course, which is why we're designed to be highly scalable and have a low marginal cost.
We make a preliminary judgement of people's potential based on their interactions with our platform before enrolling, and then match groups based on skill level. This way we can allow anyone to join while not degrading the experience of the best performers.
This objection points at some mechanisms that do exist but we disagree with its conclusion.
First, what we agree with
- It is true that removing the application process lowers the perceived value or status. This is consistent with basic marketing and economics.
- It seems plausible that (young) highly-talented people are especially sensitive to status signals.
Secondly, why we nevertheless think having no applications is very positive
Applications add friction, and removing friction increases signups.
This too is consistent with basic marketing and economics. At a lower price (or lower effort, lower friction), more people will buy (i.e. enroll).
There are other ways to increase perceived value and status that are not as costly to the user.
Status and perceived value are nice, but just because application processes add status doesn't mean they're great at it. If adding status is the goal, let's do things targeted specifically at that. Let's not confuse the method (i.e. the application process) for the goal (i.e. status and perceived value)
In microeconomic terms
First order effect — there's a fixed demand curve (i.e. sales as a function of price). If you decrease the price, you increase the number of sales. If you increase the price, sales decrease. There's a second order effect working in the opposite direction, and it seems people intuitively over-estimate its strength.
Second-order effect — if you increase the price of a product, you increase the perceived value of the product, changing the demand curve itself. However, this rarely compensates for the first order effect. Thus, total sales still decrease. For for-profits, this can be a great strategy. You sell less but at higher margins, increasing profit. In rare cases (e.g. with some luxury-goods), increasing the price can raise the perceived value so much that the total amount of sales volume increases. But we don't see a reason to expect this being the case for Lens Academy.
Example — your course sells for 100$, and three people have a willingness-to-pay of 50$, 110$, and 150$, respectively. You thus sell two units. If you increase the price of the course to 150$, the same three people might perceive a higher value, and now have a willingness-to-pay of 80$, 140$, and 180$. You now sell only one unit. See how the perceived value of the course went up in everyone's eyes, but total sales went down?
Similar to status, if the goal is increasing perceived value, let's do things whose primary effect is raising perceived value. Marketing offers many solutions for that. Increasing the price is just one way to increase perceived value, and it's not very suitable for our impact-driven mission.
Via a grade indicated on the completion certificates of our graduates. We assess people throughout our courses on a variety of metrics including understanding the alignment problem, value-alignment, clear thinking, and growth over time. Instead of using a pre-course application process as a filter, we use the courses themselves to assess and route people to the path that's best for them and the community. We'll share more about this system as it matures.
We're excited about both top talent making full-time career transitions and the broader public making part-time contributions.
However, it's hard to be entirely sure of anything in AI Safety, and the role of part-time contributions is particularly unclear.
On the one hand, we're excited about the impact millions of small contributions can have in the aggregate. On the other hand, we're also worried about backfire risks associated with movement-building. If we look at the broader AI Safety space, we see a similar uncertainty, with many actors pursuing public education (e.g. 80,000 Hours, MIRI), whilst others think it is not useful.
- As a prior, it seems intuitively likely to us that anyone should be allowed to learn about and engage with the existential risks of advanced AI. Beyond self-study, there is currently no way for people to reliably engage with this topic.
- There's also an argument to be made along the lines of: "people are going to worry about AI risks anyway. It's better that we educate them and point them at useful actions, lest they behave erratically." We partially agree with this argument, but want to be careful not to use it in a hand-waving manner, similar to AI Labs saying, "If we don't build it, someone else will." Lens Academy will redirect people from worse actions to better actions, but our scalability also means we will just reach more people in total.
On balance, we think it's probably net-positive in expected value to keep the broader audience in the loop.
Using an AI Tutor and Coach
Because it allows us to help people reduce x-risk at a depth and scale otherwise impossible.
We essentially use AI for AI Safety Capacity-Building (cf. AI for AI Safety). By leveraging AI we lower the burden on our facilitators, making their role more fun and enabling people with less AI Safety context to facilitate. This helps us scale.
Analyzing the performance of our Lens Tutor and Lens Coach so far, they've impressed us with their ability to help our participants. We expect them to become even more leveraged as better models continue to be released.
It is true that there is a growing anti-AI sentiment. It also seems that most people that dislike AI, will nevertheless use it when it's useful to them. In fact, many of our participants get great value out of our AI systems.
Some participants ask us why we are using AI when we dislike AI. We see this as a great opportunity to help them clarify their thinking about what it is that we like and what it is that we dislike.
A minority of people will undoubtedly choose not to engage with our platform entirely because they'd rather boycott anything using AI. While we regret this, we think it's a tradeoff well-worth making because of the extra quality and scalability AI enables us to have.
Yes. Based on the thousands of messages our participants have exchanged with our Tutor and our review of its responses, we're confident it's adding significant value. We've been surprised by its quality, though it's definitely not perfect yet and we will continue to improve its prompting and scaffolding.