Psychological safety without losing accountability, and the reference question that gets the truth: Dan Siroker of Rewind

Most treatments of psychological safety stop at making people comfortable. Dan Siroker's version has a second axis, and his account of what happened at Optimizely when a company gets the first one without the second is the most useful thing in this episode.

Siroker is co-founder and CEO of Rewind, a personalized AI he describes as a copilot for your mind. Before that he founded Optimizely and grew it to $120 million in annual recurring revenue. Before that he was a key member of Obama's 2008 campaign, where he helped pioneer using data to optimize fundraising and media.

Where the product came from

Siroker's origin story is personal. About twenty years ago, in his twenties, he started going deaf. Getting a hearing aid was what revealed how much he had lost, and regaining it in an instant felt like acquiring a superpower.

Losing a sense and getting it back sent him looking for other human capabilities that technology could augment. Memory stood out.

The research is bleak. Studies show we forget roughly 90% of our experiences after a single week, following what is known as the forgetting curve. Memory typically peaks around age 20 and declines every year afterward. And most people do not notice, which he describes as a boiling frog dynamic: you think your memory is fine until you cannot recall how you know someone or what your action item was.

Rewind's answer is to make everything you have seen, said or heard on your Mac, and now your iPhone, searchable and queryable. Simple uses like drafting an email, or the one he keeps returning to: how do I know this person, and what were all our previous interactions.

Does outsourcing memory make your memory worse

An audience question from Justin Riley, CEO of Wavelo, asks how Rewind threads the needle of being a recall tool without changing how humans experience memory. Michael extends it: if we rely on something else to remember, does the faculty atrophy?

Siroker reaches for precedent. He used to memorize his five closest friends' phone numbers, without area codes. Today he knows his wife's number and his own. Is he worse off? He does not think so. Same with navigation: he grew up needing to know cross streets, then MapQuest arrived, then Google Maps, which now routes dynamically around traffic.

His argument is that nothing is lost by outsourcing the parts technology is well suited to hold, like numbers and directions. And that the gain is space. Every moment spent racking your brain to retrieve something is a moment not spent thinking ahead, connecting dots or being creative. He frames it as a pedestal for the mind rather than the dystopian version where everyone ends up disconnected in a scooter.

Michael's own example is photographs of his kids on displays around the house bringing back memories he had lost.

Siroker's addition is worth keeping. Memories are encoded in ways that are hard to retrieve linearly but accessible through recognition. Photographs do it. Music does it. Smells do it. People with advanced dementia can retrieve things thought gone forever through sensory experience.

His summary: the worst thing is having something in your mind you cannot retrieve, like a vault with no key.

Privacy as an architecture decision

Rewind's privacy posture is not a policy. It is why the product is built the way it is.

Siroker could have built a cloud application. Optimizely was software as a service and he knows JavaScript well. Instead they pivoted away from a cloud product called Scribe, an online meeting bot, into native macOS development and low-level optimized software, a domain he knew nothing about, specifically so recordings could be stored locally.

The consequence he states plainly: if the company were subpoenaed and compelled to produce a user's recordings, it could not. The data is on your machine. Your employer cannot see it. Only you can.

What happened next is the interesting operational lesson. People were initially cautious, then hit an aha moment, usually recovering something they could not have found any other way, an email they discarded instead of sending, or content lost to a tab crash. After that moment, minds shift entirely to convenience, and users start asking for exactly the things the team had been careful to avoid: sync across devices, questions answered on the phone.

So the roadmap became end-to-end encrypted synchronization, with the users pushing and the privacy-first foundation holding.

On the ChatGPT integration, Michael assumes anything sent is used for training. Siroker corrects him: OpenAI is explicit that data sent through the GPT-4 API is not used to train models, and is retained for only 30 days.

The argument for executives

Michael raises Jason Calacanis's reaction on This Week in Startups, which was terrified and intrigued, and points out that for COOs the most sensitive matters of a company cross their desktops.

Siroker's answer zooms out ten years. The companies that succeed will be the ones that use AI to get materially better, in some cases several times better. And the successful operators will not be the ones who assume the job will be the same with minor adjustments. Every decision informed by AI synthesizing the data rather than just the data. Being effectively present in five meetings instead of one.

His pitch, in that frame, is that a company that can credibly claim privacy is architectural rather than promised is the right partner for that transition. Today a chief information security officer may be unenthusiastic. In a few years, compared against fully cloud-based alternatives or ones that train on your data, he expects the calculus to look different.

Michael's observation is the one worth noting: he had not expected privacy to arrive this early in AI. In social media and search, privacy showed up only after people were angry.

AI in operations is a leverage story

Asked about applications beyond Rewind, Siroker's word is leverage.

The analogy is operations before and after computers. Companies needed armies of people doing rote manual work with pen and paper, and computers automated it away. AI is the same lever, and he thinks larger: one person able to do ten or a hundred times more, and to do it joyfully, because they are working smarter rather than harder.

The consequence is smaller teams doing more. Companies that do not adopt it will carry larger headcount and face a competitive threat, and eventually the comparison will show up in the financials. What counts as a good profit and loss statement will be reset by whichever competitor embraced this.

His own evidence: Rewind was 18 people at the time of recording. With no criticism of Optimizely intended, he believes those 18 could move faster than the 450 he had running Optimizely.

So the question he puts to COOs is not how to add people to the org chart. It is how to make every existing person more capable. Some functions may genuinely disappear, but the bulk of the leverage is augmentation rather than replacement.

Michael's version at Tucows was telling everyone that GPT-4 meant they could do more faster, and asking himself daily how it made his own work faster. His two challenges are ones every operator will recognize. Privacy, with the Samsung intellectual property incident as the cautionary tale and 1,300 people who all need to understand the rules. And breadth, because a general-purpose tool asks every employee to be creative about how it might help them.

Bridging the gap between ambiguity and the task

Siroker's response to the breadth problem is the most transferable product insight in the conversation.

ChatGPT is so broad that people do not know what to ask. The features that turn out to be most valuable are the ones that close the distance between that ambiguity and the task actually in front of you.

His example: Rewind sees you are about to join a video meeting and prompts you to join and record. At the end of the meeting it prompts you to summarize and draft an email, which takes everything said, condenses it to five bullets, and puts it in your inbox as a draft addressed to the attendees, whose names it already knows.

If you are in back-to-back meetings, you were never going to write that summary. The AI is not doing something novel. It is doing the thing you were already supposed to do and never had fifteen minutes for.

His line about why this matters: meetings are a black hole for information. Ninety percent of what happens is forgotten by the next weekly team meeting, and the setting is expensive, since everyone's salary for that hour is not insignificant. Converting what was said into follow-up and action is where the value is.

Focus, and the main thing

Asked how a small team does so much, Siroker corrects the premise before answering. It is not just AI. It is the team.

He is candid about the asymmetry. The first 18 people at Optimizely would not have worked with him, because he did not know what he was doing and they could have earned three or four times as much as senior individual contributors at Facebook or Google. The myth of the 10x engineer, he says, is real, and with AI it becomes the 100x engineer.

He is also candid about the consequence. Rewind does not hire junior engineers, only senior, staff and principal level. And he worries openly about what that means for people graduating today into a world where experienced people are getting more leverage.

The second ingredient is focus, and he admits he thought he was focused at Optimizely and in hindsight was not.

His definition comes from Steve Jobs, retold recently by Jony Ive: focus means not doing something that every bone in your body wants to do, because you have decided to focus on something else.

His mantra, first heard about a decade ago from someone formerly at HP and only truly adopted in the last three years, is that the main thing is that the main thing should stay the main thing.

The failure mode he names is precise: it is indigestion, not starvation. Startups die of doing too much. Mac and iPhone and Android and Windows on day one.

Rewind did the opposite deliberately. Mac only, despite a long waiting list of people wanting the iPhone version. They spent a year getting it right, and when they did build for iPhone it took three months, because they had already solved compression and AI integration and knew how to do it correctly the first time.

How they got to Apple Silicon

The pivot from Scribe, their cloud meeting bot, came from two realizations. There is far more context outside the meeting than inside it. And people found bots creepy, with legitimate concerns about where recordings physically live.

Then M1 and M2 arrived, and the founding vision became technically possible. Everything they were doing in the cloud could run locally: compression, transcription, all of it on your machine. Before Apple Silicon, a computer attempting what Rewind does would grind to a halt.

The decision was not obvious. They wrote a strategy document with five options, and three people on the team ranked the memory strategy last of the five. They chose it because it was the option most likely, if it worked, to be transformative.

Siroker connects this to what he thinks STEM education actually teaches, which is not facts but a method. He was on the robotics team in high school, where the competition was about using the constraints of the materials you have to build something good. At Optimizely the question was what JavaScript could do now that browsers had gotten fast, which produced client-side A/B testing that marketers could use. At Rewind it was what Apple Silicon made newly possible. Same question, different decade.

Platform risk

Michael asks about the risk of building on someone else's platform, particularly with Apple.

Siroker says he worries more about Apple as a competitor than about Apple pulling the rug, since Rewind uses only published APIs and breaks no rules. In fact, when they announced, people emailed him receipts for new Apple laptops purchased specifically because Rewind required M1 chips. Driving hardware sales is the entire point of Apple's developer platform.

On competition he lists his advantages, and they are a good template for anyone in this position.

What Rewind does is weird, and arguably borderline creepy. At a large company, for every person championing something like that, a hundred will explain why not, citing brand and privacy. That is the innovator's dilemma, and it buys time.

Switching costs are high. He contrasts this with Dropbox and iCloud, where switching means dragging files across. Rewind's data is encrypted locally by design, which makes leaving hard.

And speed. They ship eleven times a day, and he gives his personal phone number to every pro subscriber. His marker for when Apple is taking them seriously: when Tim Cook starts doing that.

On Google he sees a business model conflict rather than a threat. If Google offered the same product, Rewind's counter-question writes itself: do you want your ads targeted better using this data? Rewind is a subscription, and does not have access to the data.

Chief of staff, and what a COO actually is

Siroker has a chief of staff at Rewind and had both a chief of staff and a COO at Optimizely.

He describes the chief of staff as an extension of himself: things he would otherwise do but lacks time for, or that she would simply do better. Administrative work, designing the quarterly retreats, and owning substantive topics, including the psychological safety session the company worked through.

The complementarity is explicit. He is oriented toward what is possible, open-minded and optimistic. She is execution-oriented, with what he calls a traditional COO mindset. One plus one equals three.

His Optimizely COO was promoted from within, having been CFO and then taken on customer success. Siroker describes him as one of the best executives he ever hired, and as someone who wanted a genuine partnership with the CEO rather than a role.

What he valued most in the chief of staff role is specific and worth quoting to anyone considering hiring one. Someone who can take an ambiguous, ill-formed thought, a sentence uttered without much structure, and translate it into the organization. Someone who makes sure the occasional good idea inside a pile of bad ones does not die on the vine, who holds processes that keep people accountable, and who ensures what was discussed in one setting reaches another.

Asked how he screens for it, his answer is refreshing: poorly. He says he is more lucky than good, has made mishires, and does not know how to screen for great chiefs of staff or COOs. His Optimizely COO was an internal promotion.

His explanation for the difficulty is the part operators should sit with. So much of it is the specific partnership with a specific CEO. You need someone different enough from you to cover your weaknesses, and mutual respect that each is genuinely better at part of the job.

And then the observation that separates good COOs from people waiting for a different title. The best ones recognize that being a great COO in partnership with a CEO is what they are best suited to do, rather than treating it as a stop on the way to the CEO seat.

Values you can actually be held to

Optimizely's values spelled OPTIFY: ownership, passion, trust, integrity, fearlessness and transparency. Each had a specific set of behaviors attached.

Siroker's warning about why the behaviors matter is the Enron story. On the day of the collapse, the word integrity was carved into stone in the lobby.

The origin of his conviction is Google. He interviewed roughly a thousand people there despite being junior, and had to assess each on how Googley they were, with no definition of Googliness anywhere. That, he says, is a recipe for unconscious bias, because "not a culture fit" means nothing when the culture is undefined. So his first goal in founding a company was to define the culture in writing.

The second was living it. When facing an important decision, he would open the culture document and treat it as a constitution, approaching it the way a Supreme Court justice approaches the law: what best lives up to these values? Then he would explain the decision by naming the value it served.

At Rewind he has evolved the approach, concluding that at Optimizely they had mistaken cute and simple for comprehensive. Rewind has 14 cultural values with no acronym, grouped under two themes of impact and teamwork. The culture document is public at rewind.ai/careers, and it is the same Notion document used internally, comments and revision history included.

The mechanism that makes it real: quarterly 360 reviews where the evaluation criteria are how well each person lives up to the cultural values. He is held accountable by his team on the same basis. Values are the foundation of performance management rather than a recruiting artifact.

And it is a living document, revisited in retrospectives every two months, because the values are meant to serve the company rather than to serve marketing.

The two-by-two that explains what went wrong

The psychological safety framing is the heart of the episode.

Siroker's first point is almost too simple: they focused on it partly to create shared language, and talking about it is itself part of creating it, because people learn that it matters. So if you are considering talking about it, talk about it.

The model is a two-by-two. One axis is comfort. The other is accountability. The target quadrant is both: people feel they can say what they think without worrying about repercussions, and are simultaneously held to a high standard, with the high standard not undermining the safety.

Then the confession. Optimizely started in that quadrant. As it succeeded, through strong Series A and B rounds, the company caught what he calls a collective disease of comfort without accountability.

His image for it is the line I keep coming back to. They all locked arms and walked off a cliff together. He would much rather someone had yelled that there was a cliff than have everyone share a false sense of security.

Which is why he is deliberate about it at Rewind precisely when things are going well. It will not always be up and to the right. There will be a moment when a large competitor ships a free clone. In that moment you need candor and high mutual accountability, and that muscle atrophies in good times.

The behavioral expectation he attaches: if you see something, say something. He hires experienced people at top of market compensation not for their titles but for their judgment, so an engineer who sees a problem in marketing should say so, and a designer who thinks the CEO is wrong should say so. He wants an environment where he can float bad ideas, have the bad ones torn apart, and see the occasional good one watered into something real.

Growing slowly on purpose

Asked how you preserve that as a company scales, Siroker starts from time horizon. He wants Rewind to be his life's work, running for decades. Ironically, that long-term orientation produces a decision to hire slowly.

The Optimizely lesson is stated without hedging. They raised a $28 million Series A and a $57 million Series B, which on its own would have been fine. The mistake was compounding it by spending it. As a first-time CEO he found it hard to resist: he hired a strong head of marketing, who explained they needed a communications person and then another and then another, and each request sounded reasonable.

Without the constraint, and without a willingness to say no, you lose the culture. His preference is to try to do more with 18 before doing the same with 30 or 40.

The reference check question

The other half of preserving culture is recruiting, and this is the section to steal from.

Siroker did not lean into references at his first company. Now he considers them almost the most important part of the process, particularly back-channel references. And he is willing to decline a candidate who checks every box on capability but fails on culture and references, because a strong tactician who damages the culture is a worse trade than a slower hire.

His favorite question comes near the end of a fifteen-minute call, most of which is spent building enough rapport that the answer will be honest.

The question: we are probably going to hire Bob. But if we hire Bob, and in six months you and I bump into each other and I tell you it did not work out, what is your first instinct as to why?

Two things make it work. People enjoy predicting the future. And the implied future encounter creates a small accountability that a strengths-and-weaknesses question does not. What comes back is specific: Bob probably got disengaged, or Bob was excited for a few weeks and then decided to start something of his own. Signals about commitment and focus that no direct question surfaces.

He notes he probably stole the question from someone, and that he used to A/B test his reference questions, which is entirely in character.

Two supporting techniques matter as much as the question.

He never asks a candidate to make the introduction. He asks for a list of five references with titles, emails and a note on how they know each other, then reaches out himself. That small difference produces noticeably more candor, because a warm handoff makes people less willing to be honest.

And at the end of every reference call, he asks who else might have a unique perspective on this candidate. A reference unwilling to say anything negative themselves will often point you at the person who will, because that feels less like a betrayal. Then you call that person and hear the truth.

On rapport, his approach is to treat the reference the way he would want to be treated. He tells the Rewind story and why he is doing this, including the hearing loss, so it is relational rather than transactional. He keeps it short because they are doing him a favor. And he makes clear how much the input matters, because he has been on the other side of calls where the interviewer had obviously already decided and was checking a box, and in those cases he saw no reason to take a risk on honesty.

What he never thought he would see

Siroker's crazy story is simply how good this technology turned out to be.

That morning his executive team had a 90-minute meeting. Rewind summarized it into five bullet points, and the summary was excellent. It reminded him of his first job at Google as an associate product manager straight out of college, when his job was taking notes for senior executives. He does not think he could have matched the AI on substance, coverage or objectivity.

His challenge to anyone who has not had a comparable moment recently is that they are not looking hard enough.

Michael's parallel is holding an iPhone for the first time in 2007 in an open-air market in Colorado, having carried an LG flip phone until then, and realizing he had humanity's knowledge in his palm.

The 5 things I took away from this conversation

1. Psychological safety has two axes, and most companies only manage one. Comfort without accountability is not safety, it is drift. Dan's description of Optimizely locking arms and walking off a cliff together is the version of this I will remember, and the timing point matters most: you lose the accountability muscle exactly when things are going well.

2. The six-month reference question. If we hire this person and in six months I tell you it did not work out, what is your first instinct as to why? I have run a lot of reference calls and never asked anything this good. It gets past the strengths-and-weaknesses script because people enjoy predicting outcomes.

3. Never let the candidate make the introduction. Ask for a list with emails and reach out yourself. Then ask each reference who else would have a unique perspective. That second question is how you reach the person who will actually tell you.

4. Define the culture or you have defined nothing. Dan interviewing a thousand people at Google against an undefined standard of being Googley is a perfect illustration. "Not a culture fit" is unfalsifiable when nobody wrote the culture down, and unfalsifiable criteria are where bias lives.

5. Bridge the gap between the ambiguity and the task. The most useful AI feature Dan describes is not clever. It notices you are about to join a meeting, then afterward drafts the summary email you were never going to write. General-purpose tools ask everyone to be creative. The valuable ones meet people inside the work they are already doing.

FAQ

What is psychological safety at work? An environment where people can say what they think without fearing repercussions. Siroker's addition is that it requires a second dimension: people must also be held to a high standard. Comfort alone produces a team that agrees its way into trouble, which is what he says happened at Optimizely as it grew.

Can you have psychological safety and high accountability at the same time? That combination is the goal, not a tension to be balanced away. Siroker maps it as a two-by-two of comfort and accountability, with the target quadrant being both at once, and argues the high standard is what makes the candor useful rather than what suppresses it.

How do you build psychological safety on a team? Start by naming it, because he believes discussing it openly is part of creating it. Then make the expectation behavioral: everyone is expected to speak up outside their function, and leaders have to visibly accept having their own ideas challenged. At Rewind, values including candor are the criteria in quarterly 360 reviews.

What is the best reference check question? Siroker's is asked near the end of the call: if we hire this person and in six months I tell you it did not work out, what is your first instinct as to why? It produces specific, honest answers about commitment and fit that direct questions about strengths and weaknesses do not.

What is the difference between a chief of staff and a COO? Siroker has had both simultaneously. He describes the chief of staff as an extension of the CEO, translating ambiguous ideas into organizational action and owning projects the CEO cannot get to. The COO at Optimizely ran a large part of the company as a strategic partner. His view of the best COOs is that they see the role as what they are suited for, not as a step toward being CEO.

Also mentioned

• Rewind, its local-first architecture and public culture document
• Optimizely, the OPTIFY values, and the growth-stage lessons Siroker drew from it
• The forgetting curve, and the research on memory decline that led to the product
• Apple Silicon, without which the local-processing model would not run
• Amy Edmondson's work on psychological safety, the origin of the concept Siroker applies
• This Week in Startups, where Jason Calacanis pressed Siroker on the privacy question
• Wavelo, whose CEO Justin Riley supplied the audience question on memory augmentation

Listen to the full episode

Dan Siroker on Between Two COO's

Between Two COO's is hosted by Michael Koenig. Subscribe on Apple Podcasts, Spotify, or wherever you listen.