Promoting Job: Lens academy (group I work with) makes and hosts superint-xrisk-related courses meant to be scalable (low marginal cost) and effective (teach you the structure), and might even be so. the course has no lecturer; currently, we have a solo portion—a series of articles—and turns being grilled by an AI tutor to check your work (Currently Sonnet 5);[1] and weekly (currently human)-facilitated group calls with a series of discussion prompts.
We’re running cohorts of our existing beginner course and a new intermediate one—the first goes through IfAnyoneBuildsItEveryoneDies, the second is a mix of articles around AI forecasting and intro to the technical basics of AI, many from here from authors that I hope are already your favorites anyway. We’re still working on the list, but the topic is selected; it’ll be inspired by redwood’s ai futurism reading list.
If this pitch sounds interesting to you, we’re interested in people signing up, shares of the course-list invitation link; you can also preview/speedrun the other courses without signup there: https://lensacademy.org/c/wihtg
The first cohort’s meeting-compatibility is scheduled to be computed around one day after time of posting, so interested folk should rsvp by then to get in the scheduling.
We’re intending to scale, so we might become slightly annoying with promotions. Hope to annoy minimally :)
the course includes a bunch of text guidance for the tutor!Sonnet5; for each question posed to the learner (you), the tutor judges how well you did. I’ve been the one working on tutor guidance; I’m not maximally happy with how we describe the task to them, but definitely very dramatically better than unbriefed Claudes. Happy enough to recommend to others as a learning tool, at least.
We automatically match you to a group that fit in your timezone; this works great once there are enough, but needs a minimum number of people to achieve good times.
hmm we could accept a BYO API key option. The usual warning is “Fable is ruinously expensive via api”. I’ll ask a claude if we can add the feature without too much tech debt. I expect that feature to be rarely used, but if it fits I’d love to have it.
Yup I think we want to cater to power users if it’s not too hard, and this one seems relatively easy to implement (at least when remaining in the Anthropic ecosystem).
There’s potentially a world in which the in-app browser of the Claude app can be used, and then maybe maybe it can use its own subscription? Heard someone suggesting that some time but I didn’t look into it and I plausibly misunderstood it, and it’s probably pretty bad UX.
Promoting Job: Lens academy (group I work with) makes and hosts superint-xrisk-related courses meant to be scalable (low marginal cost) and effective (teach you the structure), and might even be so. the course has no lecturer; currently, we have a solo portion—a series of articles—and turns being grilled by an AI tutor to check your work (Currently Sonnet 5);[1] and weekly (currently human)-facilitated group calls with a series of discussion prompts.
We’re running cohorts of our existing beginner course and a new intermediate one—the first goes through IfAnyoneBuildsItEveryoneDies, the second is a mix of articles around AI forecasting and intro to the technical basics of AI, many from here from authors that I hope are already your favorites anyway. We’re still working on the list, but the topic is selected; it’ll be inspired by redwood’s ai futurism reading list.
If this pitch sounds interesting to you, we’re interested in people signing up, shares of the course-list invitation link; you can also preview/speedrun the other courses without signup there: https://lensacademy.org/c/wihtg
The first cohort’s meeting-compatibility is scheduled to be computed around one day after time of posting, so interested folk should rsvp by then to get in the scheduling.
We’re intending to scale, so we might become slightly annoying with promotions. Hope to annoy minimally :)
the course includes a bunch of text guidance for the tutor!Sonnet5; for each question posed to the learner (you), the tutor judges how well you did. I’ve been the one working on tutor guidance; I’m not maximally happy with how we describe the task to them, but definitely very dramatically better than unbriefed Claudes. Happy enough to recommend to others as a learning tool, at least.
We automatically match you to a group that fit in your timezone; this works great once there are enough, but needs a minimum number of people to achieve good times.
FYI, we used to run Opus, but are temporarily running Sonnet as a means of experiment for comparison, as well as to save costs.
Is it possible to “bring your own model” and use Opus/Fable?
hmm we could accept a BYO API key option. The usual warning is “Fable is ruinously expensive via api”. I’ll ask a claude if we can add the feature without too much tech debt. I expect that feature to be rarely used, but if it fits I’d love to have it.
Yup I think we want to cater to power users if it’s not too hard, and this one seems relatively easy to implement (at least when remaining in the Anthropic ecosystem).
There’s potentially a world in which the in-app browser of the Claude app can be used, and then maybe maybe it can use its own subscription? Heard someone suggesting that some time but I didn’t look into it and I plausibly misunderstood it, and it’s probably pretty bad UX.
I’d love to also test this against open source models, or more broadly have a benchmark of how well different models understand alignment.