I’m building tools to make sense of the future at Sage. Currently building AI Digest, Fatebook and Quantified Intuitions.
Previously I was doing a PhD in HCI at St Andrews, and worked at Clearer Thinking.
Website: https://binksmith.com
Tweeting, sometimes about EA: https://twitter.com/adambinksmith
AI for epistemics/forecasting is something we’re considering working on at Sage—we’re hiring technical members of staff. I’d be interested to chat to other people thinking about this.
Depending on the results of our experiments, we might integrate this into our forecasting platform Fatebook, or build something new, or decide not to focus on this.
[Do you have a work trial? This will be a deal breaker for many]
Based on your conversations with developers, do you have a rough guess at what % this is a deal breaker for?
I’m curious if this is typically specific to an in-person work trial, vs how much deal-breaking would be avoided by a remote trial, e.g. 3 days Sat-Mon.
Thanks for the newsletter!
Looks like a typo:
> a version of GPT-4 released in 2023 outperformed a version of GPT-4 released in 2021
As well as Fatebook for Slack, at Sage we’ve made other infrastructure aimed at EAs (amongst others!):
Fatebook: the fastest way to make and track predictions
Fatebook for Chrome: Instantly make and embed predictions, in Google Docs and anywhere else on the web
Quantified Intuitions: Practice assigning credences to outcomes with a quick feedback loop
This month’s Estimation Game is about effective altruism! You can play here: quantifiedintuitions.org/estimation-game/december
Ten Fermi estimation questions to help you train your estimation skills. Play solo, or with a team—e.g. with friends, coworkers, or your EA group (see info for organisers).
It’s also worth checking the archive for other estimation games you might be interested in, e.g. we’ve ran games on AI, animal welfare + alt proteins, nuclear risk, and big picture history.
I’m curious about B12 supplements—I currently take a multivitamin which has 50µg B12, my partner takes a multivitamin with 10µg B12. Should we be taking additional B12 tablets on top of this? (We’re both vegan)
I saw in that post a recommendation for 100µg tablets, but google says the RDA is 2.4µg, do you know why there’s this gap?
Here’s the link to apply! https://80000hours.org/speak-with-us/
I think some subreddits do a good job of moderating to create a culture which is different from the default reddit culture, e.g. /r/askhistorians. See this post for an example, where there are a bunch of comments deleted, including one answer which didn’t cite enough sources. Maybe this is what you have in mind when you refer to “moderating with an iron fist” though, which you mention might be destructive!
Seems like the challenge with reddit moderation is that users are travelling between subreddits all the time, and most have low quality/effort discussion norms. Whereas on the Forum, the userbase is more siloed, which I guess would make good quality moderation easier.
We’ve added a new deck of questions to the calibration training app—The World, then and now.
What was the world like 200 years ago, and how has it changed? Featuring charts from Our World in Data.
Thanks to Johanna Einsiedler and Jakob Graabak for helping build this deck!
We’ve also split the existing questions into decks, so you can focus on the topics you’re most interested in:
Ah thank you! I’ve just pushed what should be a fix for this (hard to fully test as I’m in the UK).
The July Estimation Game is now live: a 10 question Fermi estimation game all about big picture history! https://quantifiedintuitions.org/estimation-game/july
Question 1:
I was also wondering this—did 80k link to it in their newsletter (which has a big audience)?
Relatedly, I wonder if you can see differences in reported source by the place the survey respondent navigated to the survey from?
Thank you!
Do you look at non-anonymized user data in your analytics and tracking?
No—we don’t look at non-anonymised user data in our analytics. We use Google Analytics events, so we can see e.g. a graph of how many forecasts are made each day, and this tracks the ID of each user so we can see e.g. how many users made forecasts each day (to disambiguate a small number of power-users from lots of light users). IDs are random strings of text that might look like cwudksndspdkwj. I think you’d call technically this “pseudo-anonymised” because user IDs are stored, not sure!
Who specifically gets access to user submitted predictions (can’t quite tell how large your team is, for instance)
Your predictions are private to you unless you share them. I and the other two devs who have helped out with parts of the project have access to the production database, but we commit to not looking at users’ questions unless you specifically share them with us (e.g. to help us debug something). I am interested in encrypting the questions in the database so that we’re unable to theoretically access them, but haven’t got round to implementing this yet (I want to focus on some bigger user-visible improvements first!)
Hope this makes sense! Thanks for your kind words and for checking about this, let me know if you think we could improve on any of this!
Thank you! I’m interested to hear how you find it!
often lacks the motivation to do so consistently
Very relatable! The 10 Conditions for Change framework might be helping for thinking of ways to do it more consistently (if on reflection you really want to!) Fatebook aims to help with 1, 2, 4, 7, and 8, I think.
One way to do more prediction I’m interested in is integrating prediction into workflows. Here are some made-up examples:
At the start of a work project, you always forecast how long it’ll take (I think this is almost always an important question, and getting good at predicting this is powerful)
When you notice you’re concerned about some uncertainty (e.g. some risk) you operationalise it and write it down as a question
In your weekly review with your manager, you make forecasts about how likely you are to meet each of your goals. Then you discuss strategies to raise the P(success) on the important goals
When there’s a disagreement two team members about what to prioritise, you make operationalise it as a forecasting question, and get the whole team’s view. If the team as a whole disagrees, you look for ways to get more information, or if the team agrees (after sharing info) you follow that prioitisation
If anyone that either has prediction as part of workflows or would like to do so would be interested in chatting, lmk!
In many ways Fatebook is a successor to PredictionBook (now >11 years old!) If you’ve used PredictionBook in the past, you can import all your PredictionBook questions and scores to Fatebook.
In a perfect world, this would also integrate with Alfred on my mac so that it becomes extremely easy and quick to create a new private question
I’m thinking of creating a Chrome extension that will let you type /forecast Will x happen? anywhere on the internet, and it’ll create and embed an interactive Fatebook question.
I’m thinking of primarily focussing on Google Docs, because I think the EA community could get a lot of mileage out of making and tracking predictions embedded in reports, strategy docs, etc. This extension would also work in messaging apps, on social media, and even here on the forum (though first-party support might be better for the forum!).
Great, thanks!
The format could be “[question text]? [resolve date]” where the question mark serves as the indicator for the end of the question text, and the resolve date part can interpret things like “1w”, “1y”, “eoy”, “5d”
I’m interested in adding power user shortcuts like this!
Currently, if your question text includes a date that Fatebook can recognise, it’ll prepopulate the “Resolve by” field with that date. This works for a bunch of common phrases, e.g. “in two weeks” “by next month” “by Jan 2025″ “by February” “by tomorrow”.
If you play around with the site, I’d be interested to hear if you find yourself still keen for the addition of concise shortcuts like “2w” or if the current natural language date parsing works well for you.
The June Estimation Game is animal welfare + alt proteins themed! 10 Fermi estimation questions. You can play here: quantifiedintuitions.org/estimation-game/june
Seems like academic research groups would be a better reference class than YC companies for most alignment labs.
If they’re trying to build an org that scales a lot, and is funded by selling products, YC companies is a good reference class, but if they’re an org of researchers working somewhat independently or collaborating on hard technical problems, funded by grants, that sounds much more similar to an academic research group.
Unsure how to define success for an academic research group, any ideas? They seem to more often be exploratory and less goal-oriented.
You might be interested in Metaforecast (you can create custom dashboards).
Also loosely related—on AI Digest we have a timeline of AI forecasts pulling from Metaculus and Manifold.