Your culture is how your company makes decisions when you’re not there. It’s the set of assumptions your employees use to resolve the problems they face every day. It’s how they behave when no one is looking. If you don’t methodically set your culture, then two-thirds of it will end up being accidental, and the rest will be a mistake. — Ben Horowitz, co-founder of Andreessen Horowitz
Introduction
BlueDot’s mission is to build the workforce that protects humanity. We make small decisions in service of that mission every day, including what to prioritize, how to make difficult tradeoffs, and how to relate to each other as colleagues and humans.
BlueDot’s been a small team for the past 4 years. The current team learned our ways by working closely with leadership, but that doesn’t scale. We’ve grown from 6 people to 20 people in the past 6 months, and we’ll grow to around 50 over the next year. Most team members are new, and we operate across multiple countries and timezones.
Shared principles enable us to evaluate trade-offs and make decisions faster, without having to implement strict processes, a burdensome bureaucracy, or constant ad-hoc feedback and recommendations. They also inform hiring decisions, new staff onboarding, and performance reviews, all while creating a shared sense for what it means to be a part of BlueDot.
This blog post is primarily for current and incoming BlueDot team members, but we’re sharing it publicly so that members of the wider community have insight into how we operate.
The six principles are:
🎯 “What is the goal?”
⚡ Don’t wait for permission
🥔 No hot potatoes
📖 Public by default
🤗 Understand people first
🎁 Actionable feedback is a gift
What do we mean by “Principles”?
I think of “culture” as a description of “how we do things around here”, while our principles represent parts of our culture that we’ve purposefully chosen and will invest energy into strengthening.
Our principles are short phrases that help team members take action when the answer isn’t obvious, and when there’s nobody around to ask. We expect everyone, especially leadership, to follow them when making decisions and tradeoffs, giving feedback, and appreciating a colleague’s work.
They’re defaults, not rules. You can do the opposite thing if you have a good reason (and share with the team why!). Use your judgment, especially when the principles pull against each other.
Each section below starts with the principle itself, then describes why it matters, what it asks of you in terms of concrete behaviors, examples from the team, the downsides we’ve accepted, and further reading.
The Principles
1. 🎯 “What is the goal?”
Working hard only helps the mission if your effort is pointed at the right thing. Once you know what that is, most other decisions become much easier.
Obsessively ask yourself “what is the goal?” about every project and every part of it. Think hard to make sure it’s the right goal, write down your answer and why, focus your efforts on the biggest thing in your way, and delete anything that doesn’t serve the goal.
Why it matters
An easy way for a smart, hard-working team to fail is to work very hard on the wrong thing. This is especially true in AI safety, because some work makes things worse.
Choosing the right goal is as important as figuring out how to achieve it.
Writing the goal in plain English shows you and others whether you actually know what you’re trying to achieve, and what winning and failing looks like.
Diagnosing the biggest challenge between now and the goal enables you to focus your effort on the single most important thing, instead of ten things that don’t matter.
Every part of a system, from a form question to an automation, should exist because a specific person needs it to make a decision or get something done. Every extraneous part makes things slower, more likely to break, and harder for others to understand and improve it.
AI makes building things cheap, so it’s now easy to build, automate or add process to something that shouldn’t exist and that doesn’t serve an important goal.
Testing your work quickly against reality by talking to users, checking the data or shipping a prototype tells you early when you’re wrong about your goal or how to achieve it.
What it looks like
Start every project, doc and meeting with one plain sentence saying what it’s for, and ask “what’s the goal here?” in all meetings where that’s not already clear.
Write down what success looks like as an outcome that you want, not as the actions you’ll take, plus any metrics you’ll use to track progress, so you can tell later whether you got there.
Show your reasoning, not just your conclusion: the problem, its root cause, your assumptions, the options you considered, which one you chose and why, and how confident you are.
To find the root cause, list every obstacle you can find, ask “if we fixed this, which other problems would go away?”, and focus on the underlying thing that solves most of the important problems and that you can actually change.
Before you build or change anything, go in this order: question every requirement, delete every part or process you can, simplify what’s left, speed it up, and only then automate. Most waste comes from doing these in reverse, like automating a step that shouldn’t exist.
Test the riskiest assumption early, and update your plan when new evidence invalidates your original assumptions or suggests better options.
Try to find out you’re wrong, not just prove that you’re right. Look for the strongest arguments against your plan. When the evidence changes your mind, say so in public.
Let small fires burn. Say no to good ideas and leave small problems unfixed when they aren’t the most important thing right now. If a fire you’re letting burn matters to someone else, tell them so that they can pick it up.
Examples
In late 2024, a friend asked me how we knew we were having an impact. I didn’t have a good answer, and I realized I didn’t know what my goal was. Therefore, I spent 2 months not running BlueDot on a day-to-day basis, and instead I interviewed top experts about “how do we make AI go well” and I used that to inform BlueDot’s org strategy.
Ben suggested adding an FAQ to our careers page to help candidates learn about BlueDot. Will liked the idea, but replied that it wasn’t a priority because the data shows that very few of our best candidates find us through the careers page.
What it costs you
Thinking hard at the start of a project means you won’t make as much visible progress early on.
Some things you care about will break or won’t meet your standards, and you won’t fix them.
Some people will be disappointed because you’re not prioritizing what they want you to.
Some of your ideas will turn out to be pointless or wrong, and you’ll have to say so in public.
Questioning requirements feels slow and annoying, and it will always feel better to automate or speed up a process that might not need to exist in the first place.
You’ll delete things someone built and cared about, and you will delete things you need to add back later.
Learn more
2. ⚡ Don’t wait for permission
The person closest to a problem usually knows the best way to fix it, and trying something is the fastest way to learn if it works.
When a decision is easy to undo or an experiment is cheap to run, take action fast.
Why it matters
In most situations, it’s more costly to wait for sign-off than to make the occasional mistake.
Things that seem impossible often turn out to be doable once someone tries.
AI has made a lot of things easy that used to be hard, so your sense of what’s possible is probably out of date.
What it looks like
If a decision needs to be made, assume it’s yours to make. If it helps you achieve your goal and you can undo it, you don’t need anyone’s sign-off, including your manager’s.
Don’t wait to be told what to work on. If you can see what matters most, start doing it.
If you can fix an important thing in 2 minutes, fix it immediately.
When you catch yourself asking “can I do X?”, say “I intend to do X” instead. It gives people a chance to object without making them the decision-maker.
If others might want to give input on a decision, use lazy consensus: post in a public channel saying “I’m going to do X unless someone objects by [1-2 days later]”, tag the people it affects, then do it if nobody objects. Silence means yes, as long as the right people saw it.
When a goal looks out of reach, ask “what would have to be true for this to work?” rather than “can we do this?”
If you feel blocked or think something can’t be done, take a stab anyway: ask AI, test a few different approaches, ask someone who might have good ideas, or just ship something that might not work and see what happens. Don’t wait for problems to fix themselves.
Tell the team what you did and why, and if something goes wrong, undo it and share what you learned.
If you lead or manage people, when someone asks for permission they don’t need, tell them “you don’t need my permission for this”.
Examples
Weronika needed to reshuffle the timings for upcoming courses to better suit our Teaching Fellows, who facilitate the courses. Instead of asking for approval, she posted the exact changes and wrote: “Please object in the thread by 11 am PT tomorrow, otherwise I will introduce the changes.”
Exceptions
If a decision is expensive or hard to reverse (e.g. a “one-way door”), get a second opinion from your manager or the person most affected before you take action. Though most decisions are two-way doors, and you can walk back through them if you’re wrong.
If a decision overlaps with someone else’s responsibilities, consider checking with them first.
What it costs you
You will waste money and make mistakes in front of everyone.
You’ll spend time on things that don’t end up working.
You’ll do things that annoy other people or step on their toes.
Learn more
3. 🥔 No hot potatoes
Most problems don’t arrive with someone’s name on them, and when everyone assumes someone else will do something, nobody does.
When you find a problem and nobody owns it, do something concrete with it: solve it fully yourself, hand it to a specific person who says yes, log it so it gets triaged later, or drop it and tell the people it affects.
Why it matters
As we grow, many important problems will fall between job descriptions. If you wait for a problem to become someone’s job, it might never get fixed.
Not every problem needs to get solved right away, but it’s still important for them to land somewhere concrete so they don’t get passed around endlessly or forgotten.
AI means many fixes now take minutes, so solving a problem yourself is often the best option rather than giving it to someone else.
We only create impact when a problem gets solved. One properly solved problem is better than ten half-finished projects. Every open loop drains attention from you and the people around you.
A message to the entire team saying “this is a problem” will more often than not be ignored. If everyone can see a problem but nobody takes responsibility, the problem won’t get solved.
What to do with a hot potato
🍴 Eat it: fix the problem yourself.
🧈 Butter it up and hand it over: give it to the right person with the information they need to make a decision.
🧊 Put it in the fridge: log the problem in our shared problems database so it gets triaged later.
🗑️ Bin it: drop the problem because it isn’t a priority right now, and tell the people it affects.
What it looks like
Never say “that’s not my job”. No work is beneath anyone, including leadership. We work on whatever the most important problem is.
Use “what is the goal?” to decide what to do with the problem. Fixing lots of small problems can be a way of avoiding the most important thing.
Raising a problem is always encouraged, but a message to the whole team with no named owner doesn’t count as passing it on.
Once you’ve fixed one case of a problem, look for other cases of it. Then ask how it happened, and keep asking until you reach an underlying root cause that we can change, like a broken process, a bad default, or unideal behavior. If the answer is “someone made a mistake”, ask what made the mistake easy to make, and fix that so the whole class of problem goes away.
If a problem exists in someone else’s project or area of responsibility, make sure they know about it and have confirmed they’re tracking it, rather than ignoring it or fixing it over their head.
When you hand something off, give the person enough context to take action or make an informed decision: what you saw, why it matters, what you’ve already tried, and any relevant links. Once they say yes, it’s theirs.
Stop starting, start finishing. Push the things that matter most all the way to done before you pick up new things. Solve more problems than you create.
If you’re overloaded, say so early, so someone can help or we can cut the scope of your current project.
If you lead or manage people, fix important problems quickly or delegate them. Don’t hold on to problems you don’t have capacity to solve, and encourage others to step up.
Examples
Lucas fixed one stuck grant payment, then checked every other grant “so we can know of cases we might have failed before people come to us”.
What it costs you
You’ll spend time on problems that aren’t in your job description.
You’ll log or drop problems you’d enjoy fixing, because they aren’t the most important thing right now.
“I posted about it in Slack” stops being an excuse to move on from an important problem you’ve found.
Finishing things or handing things over properly feels slower than starting new projects. You’ll start fewer things, but you’ll finish more of them.
Dropping a problem on purpose means telling the people who might care about it.
Learn more
4. 📖 Public by default
The more other team members know about what you’re doing and why, the easier it is for them to help you achieve your goals.
Give everyone enough context to build on and learn from your work by posting regular updates, sharing wins and failures, and avoiding DMs.
Why it matters
By “public”, I mean “visible to everyone in the BlueDot team”, not “visible to anyone online”. Though the more public we can make non-sensitive things, the better.
A post in a public Slack channel, or a team-wide Notion page, becomes searchable memory for the whole team and everyone’s AI agents.
A DM is a black hole that nobody else can read.
Visible work helps people spot problems early, avoid duplication, and pick up where you left off.
What it looks like
Share drafts, plans and half-finished work early, usually before you feel ready. It’s the cheapest moment for someone to point out a problem or a better approach.
Show your reasoning along with your conclusions, so people can build on how you got there, or challenge it.
Post regular updates before people have to ask, including “this is on track” or “I’ve deprioritized this”. People will interpret no news as bad news.
Write updates so someone with no context can act on them: what the overall goal is, what’s happened so far, what’s next, and what you need from who.
If someone is waiting on you, give them an update, even if it’s just “I can’t prioritize this until next week”.
When something you’re working on fails or doesn’t work out, share an update straight away with what happened and what you’re doing about it.
When you need something from someone, tag them by name in the public thread.
When a conversation that you think should be in public starts in a DM, ask them if you can move the convo to a public channel, then if they say yes, copy and paste that message into the relevant public Slack channel and reply there.
Keep 1-1 notes and meeting transcripts on pages the whole team can see.
You don’t need to read everything. Read what’s relevant to you, and trust people to tag you when they need your input.
Examples
When Weronika made a mistake while onboarding a new cohort of our Technical AI Safety (TAIS) course, she posted straight away in the channel where the alerts were firing: “I made a big mistake in onboarding the next TAIS cohort 🤦♂️ I am fixing it now, but FYI the above alerts come from my actions.”
Exceptions
Critical feedback about someone’s behavior, feedback that’s embarrassing to the recipient, sensitive information, or personal matters like health are often better discussed in private.
Private notes and drafts are fine, though make them public once they’re useful to other people, which is probably earlier than you think.
What it costs you
Your mistakes will be visible to everyone, and you might feel exposed.
Writing regular updates takes time, even when nothing has changed.
People may speak less freely when they know the meeting is being transcribed.
Learn more
5. 🤗 Understand people first
People trust you when they feel understood, and we ask people to trust us with big decisions every day.
Before you work with someone or ask them for something, hear their story, let them get to know you, and find out what they want and what’s in their way.
Why it matters
Most of our impact depends on other people making big decisions, including:
Community members switching careers into AI safety.
Hiring candidates leaving a stable job to join BlueDot.
Funders backing a new project.
Team members taking on projects they think are impossible.
For thousands of people each year, BlueDot is one of their first contacts with AI safety, and many arrive feeling like impostors. Whether they feel welcome and understood during their first few conversations affects whether or not they stay in the community.
The same applies inside the team. When teammates feel understood, they share problems earlier, take feedback better, and need less checking in, which lets us all move faster.
People are more likely to make big and scary decisions with people they trust, and you build trust by showing you understand and care about them. Good arguments by themselves rarely change people’s minds.
The better you understand someone’s situation and what they want, the more helpful your suggestions are likely to be, because you’ll be able to chart a believable path for how they’ll go from where they are today to where they want to be.
People often tell you what they think you want to hear, especially when they want something from you. Understanding their background and what they’ve actually done in the past tells you more than what they say they would do in the future.
What it looks like
Anyone who talks to someone from BlueDot should come away feeling deeply understood.
When a new person enters our spaces, welcome them enthusiastically and make them feel at home.
Start conversations with the person, not the agenda, and go deeper over time: facts, then opinions and stories, then how they feel and what they’re unsure or worried about.
Check you’ve understood: say back what you heard in your own words and ask if you got it right.
Remember what people told you, and follow up on any commitments you make.
Keep your warmth even when your work gets hard or you feel really busy. Pushing for results doesn’t justify being curt or dismissive.
Before you propose an idea, pitch a job, or ask someone for something, find out:
Where they’re at right now, and what they’ve done in the past.
Where they want to get to, and why it matters to them.
What they’ve already tried, and what’s stopping them.
Ask about specific things that happened in the past, not hypotheticals about the future. Read The Mom Test.
When something important comes up, keep asking “why?” or “say more” until you reach the underlying juicy thing.
If you realize that the thing you’re suggesting isn’t the right fit for them, tell them.
Examples
Before a candidate’s work trial, Sam posted a brief for the team on the candidate’s background, what they wanted from their next role, and their doubts about field-building compared with research. Everyone who met the candidate could then speak to their real goals instead of giving a generic pitch.
What it costs you
Conversations take longer to get to the point.
You’ll have to share things about yourself with people you don’t know very well.
Learn more
How Deep Listening Can Make You More Persuasive — Edward Lempinen
Understanding and Overcoming Belonging Uncertainty — Geoffrey Cohen
6. 🎁 Actionable feedback is a gift
Feedback is the fastest way to grow and improve. The sooner you hear it, the more it helps.
Challenge decisions and share feedback quickly and in public, especially with people who seem more senior or smarter than you, without being an asshole.
Why it matters
Feedback is how we find out what we’re getting wrong or how we could be producing more impact.
Everyone benefits from each piece of feedback, not just the recipient.
A long delay between doing something and receiving feedback makes it hard to learn from that feedback, and increases how “big” or stakes-y the feedback feels to both the feedback giver and recipient.
The people whose decisions are most consequential (e.g. people in more “senior” roles, or people that others consider very smart) usually receive the least feedback, even though they do make mistakes and their performance is critical for the mission.
What it looks like
When giving feedback, remember that your goal is to assist the person who’s receiving the feedback, not to insult or punish them.
If you’re annoyed, take a breather and write down your thoughts first.
Provide feedback as soon as you can after observing the behavior that warrants feedback, and somewhere the team can see it if appropriate (e.g. the #feedback Slack channel, or during a 1-1 because our 1-1 notes are public by default).
To make your feedback actionable:
Observation: State what you saw (who, what, when, where).
Impact: Describe the impact of their behavior (like how it made you or someone else feel, or how it could decrease their impact).
Question/request: Ask them a question to better understand their perspective or experience, or offer a request for how they could behave differently in the future.
If something seems off but you don’t know the fix or have a suggestion, share the observation anyway.
Challenge decisions wherever they’re announced, especially decisions by leadership, instead of waiting for many days or weeks or some “official feedback session”.
Make it easy for people to tell you when you’re wrong, especially if you lead or manage people. Encourage dissent and ask for feedback on something specific, like “what would you change about this plan?”. Crap questions like “any feedback?” almost always get “nope, all good”.
When receiving feedback, express appreciation and say thank you. You have agency to decide whether or not to act on the feedback, or discard it.
If you change your mind or take different actions due to feedback you received, let them know.
Examples
Li-Lian gave me feedback in our #feedback Slack channel on an offer email I’d written. She said she can usually hear my warmth in my emails, “But I couldn’t hear that in this one.” She quoted the two weakest lines and suggested I read it out loud and edit until it sounded like me talking.
What it costs you
You’re more likely to hear difficult things about your work and ways you could be performing better, often in front of other people.
You’ll have to share difficult things with people you like and respect.
You’ll sometimes share “vibes-y” feedback that you’re not even sure is correct or useful, because you trust that the recipient will take responsibility for doing the right thing with that feedback.
Learn more
Final words
These principles describe the team we’re trying to become. None of us live up to all six of them all the time, including me.
If you’re on the team, you don’t need to memorize this entire post, but I do expect you to know our principles off by heart, and to practice the principle that feels hardest for you every week. And please shout out when you see someone doing great work that’s in line with one of the principles, and name the principle!
If you’re not in BlueDot, feel free to borrow anything from here that’s useful to you. And if this sounds like a team you’d like to join, check out our open roles or reach out!
In our obscurity, in all this vastness, there is no hint that help will come from elsewhere to save us from ourselves. […] To me, it underscores our responsibility to deal more kindly with one another, and to preserve and cherish the pale blue dot, the only home we’ve ever known. — Carl Sagan, Pale Blue Dot (1994)
I love these!
i am so happy this writing is out !