Bits and Bobs 12/9/24

1An agent is a bit of software that is animated by LLMs.
  • An agent is a bit of software that is animated by LLMs.
    • Not normal software that does precisely what it was programmed to do, agentic software that has some squishiness.
    • Expensive, dangerous, intrinsically open-ended.
    • It can do things you didn't code it to do, so be careful about all of the capabilities you give it access to!
2I like this definition of an AI agent.
  • I like this definition of an AI agent.
    • The most important characteristic is that the AI can loop on tool use between synchronization points with the user (i.e. between conversation turns).
3Software today is aligned with the software creator's interests.
  • Software today is aligned with the software creator's interests.
    • That creator is currently almost always some corporation.
    • That happens because of the cost of creating software.
    • But software should be personal.
    • It should be aligned to you.
4The most important characteristic of LLMs is their patience.
  • The most important characteristic of LLMs is their patience[apc].
    • They can do tasks that real humans would get distracted or bored by.
      • For example, carefully reading many pages of material to then synthesize a pretty good summary.
    • What are the human needs that are hard to do with real humans because of patience?
    • LLMs have all the time in the world, happy to do a task no matter how mundane.
5Don't use an LLM to write for you, use it as a thinking partner.
  • Don't use an LLM to write for you, use it as a thinking partner.
    • LLMs' ideas are never good.
      • They're always mush, just frog DNA.
      • I was chatting with an author who told me he refuses to use an LLM.
        • He told me you "write what you read", and he feared that the more he's exposed to the mushy writing of LLMs, the more his own writing would become mush.
    • But LLMs do give you a fun, well-read conversation partner to bounce your ideas off of which helps make your ideas better.
    • Someone told me the mark of a good conversation partner is someone with "good bounce".
      • Someone you can go anywhere with topic wise, and they're willing and able to engage, to bounce the idea right back to you.
    • LLMs have good bounce; they're willing to entertain whatever line of inquiry you want to engage in.
    • The bouncing back and forth of ideas is how they become better.
    • The discussion helps make your writing output better, even if no specific line was written by the LLM.
6What if you had FigJam, but the stickers were magic?
  • What if you had FigJam, but the stickers were magic?
    • When you affixed them to data, they transformed them into enchanted interactive objects.
    • You could collaborate not just on the content, but on what it does.
7What if you could take an app icon and attach it, like a sticker, onto your data, to create a custom app just for you?
  • What if you could take an app icon and attach it, like a sticker, onto your data, to create a custom app just for you?
    • What if you could combine two app icons to create a new app on demand?
8How can you make creating software feel more like crafting in Minecraft?
  • How can you make creating software feel more like crafting in Minecraft?
9LLM-generated software is mush.
  • LLM-generated software is mush.
    • It's 100% frog DNA.
    • The LLM extrudes out a hyper-generic answer to your specific query on demand.
    • But what if there was someone else who in the past had done precisely what you were trying to do?
    • If you could retrieve exactly those bits of pre-existing software on demand, you'd get the benefit of the collective consciousness of every creator who came before you.
10An enduring nerd dream: Yahoo Pipes combined with HyperCard.
  • An enduring nerd dream: Yahoo Pipes combined with HyperCard.[apd]
    • Part of the challenge in Yahoo Pipes was that you had to wire everything together.
      • Wiring together components is fiddly and hard!
    • What if you could just kind of collect a few bits of data, a few slices of behavior, mush them together and let the system fill in the gaps?
    • If you have to fill in the gaps from nothing, you get 100% mush, frog DNA.
    • But if the user has given a bit of a hint of the type of data, or their intent, often an LLM can fill in the gaps in a way that is likely what the user wanted.
      • If the result isn't what you wanted, simply give a bit more specificity so there's less ambiguity to fill in and have it try again.
    • Put the components you care about in a bag, shake it, and tell the system to figure it out for you.
    • LLMs can do this quite well!
11A curse is a spell whose outcome you don't like.
  • A curse is a spell whose outcome you don't like.
12SEO content is human generated slop.
  • SEO content is human generated slop.
    • Technically written by a human, but as part of an inhuman machine in conditions that can't possibly permit quality.
    • Probably worse quality than LLM slop, because it's written quickly by a non-expert and also less specific to your query.
    • SEO slop was just shot-gunned out based on stochastic beliefs about what kinds of things people in general might search for.
    • At least LLM generated slop doesn't pretend to be created by humans, and is specifically targeted to your specific question.
13In a world of slop, taste is king.
  • In a world of slop, taste is king.
14LLMs are never blunt.
  • LLMs are never blunt.[ape][apf]
    • When someone is blunt they're definitely a human.
    • You might love or hate their taste but they're definitely human.
15LLMs are great at giving you the superficial vibe of quality.
  • LLMs are great at giving you the superficial vibe of quality.
    • But the closer you look, the more you realize it's fundamentally hollow.
    • A gilded turd.
      • "Wow, beautiful, look at that shiny gold!!"
      • "Wait… eww, what's that smell?"
16The "aha" pop is what makes you want to share an idea.
  • The "aha" pop is what makes you want to share an idea.
    • But sometimes it's a faux pop!
    • Sometimes it's a superficial charisma of the idea vs the fundamental insight.
    • Similar to AI slop–superficially compelling, but ultimately empty.
17Automation is hard in intentionality-rich domains.
  • Automation is hard in intentionality-rich domains.
    • Note that the intentionality-richness is distinct from the complexity.
    • Self driving cars is a complex domain but an intentionality-poor one.
      • The user's intention is simply "get from point A to point B safely and quickly."
      • The user doesn't care about any of the other details.
    • Contrast that with something like sending an email to your boss.
      • Much higher intentionality.
      • Little nuances carry significant signal, and you have to be highly intentional about all of them.
      • Much harder to fully automate!
18Perfect automation requires not just good judgment but also perfect sensing.
  • Perfect automation requires not just good judgment but also perfect sensing.
    • To capture a human's level of intention, the system would have to be aware of all of the things the human is aware of.
    • The human is hyper-aware of their nearby real-world context.
    • A computer might conclude "my operator is moving from one room to another, I should have the music he's playing follow him."
    • But the human might know "my spouse is snoozing in the other room and I should be quiet"
19It's easier to synthesize suggestions than to automate behaviors for users.
  • It's easier to synthesize suggestions than to automate behaviors for users.
    • A lot of companies are focusing on automating annoying tasks like booking a flight.
      • But those have significant downsides if you get them wrong!
      • If the system books the wrong non-refundable $1000 flight for you, that's a huge deal.
    • But what's more useful is gathering a lot of data for the user in one place and sifting through it to help present the user with high-quality options that they can then execute themselves with a few clicks if they like them.
    • That way you get the benefit of the patience of the AI system sifting through details, but still rely on the intention and execution of the human.
20A safe bet: the world is going to continue getting weirder.
  • A safe bet: the world is going to continue getting weirder.
    • We're entering an era where trust in institutions is tanking.
      • Institutions can only provide stability if people believe in them.
    • Volatility will continue to increase.
    • Weird headlines that you would have never expected will become increasingly normal.
    • LLMs accelerate this weirdening.
    • Buckle up!
21Bits are many, many, many orders of magnitude easier to transport than atoms.
  • Bits are many, many, many orders of magnitude easier to transport than atoms.
    • This one fundamental fact explains the majority of why the tech industry behaves so differently from other industries.
    • For example, winner-take-all dynamics show up because the force of preferential attachment, a weak minor force in atoms-based industries, becomes an inescapable, all-powerful gravity well in bits-based industries.
22The vertical SaaS playbook is unreasonably effective.
  • The vertical SaaS playbook is unreasonably effective.
    • It works because having good-enough components pre-wired together well for a niche is better than having great components that customers have to duct tape together and configure themselves.
    • It turns out there's basically no way to do arbitrary integration cheaply or scalably.
      • You can get consulting style returns (good business but sub-linear return).
      • The only way to get super-linear returns with consulting-shaped businesses is to be a platform (like Salesforce) that then hosts an entire open ecosystem of consultant-shaped businesses within itself.
      • A swarm gives you super-linear returns, even if each of the members of the swarm is individually sub-linear.
    • Vertical SaaS gives a super-linear return because it gives a one-size-fits-all solution to a given niche.
      • All of the businesses in a given vertical are more or less the same.
        • The customers of vertical SaaS tend to be atoms-based businesses; they use the same business model as their competitors, just in a different geographic area.
      • Then the hyper-stickiness of being the system of record or "operating system" for the customer gives durable margin.
    • Tech businesses (e.g. businesses that take VC funding) fundamentally presume a super-linear return.
    • That's one of the reasons that vertical SaaS has taken over the world.
23The primary use case of a technology can't be the ideals.
  • The primary use case of a technology can't be the ideals.
    • An example of an ideal here might be decentralization, or end to end encryption.
    • There's only a small subset of the population for whom that ideal is more important than convenience.
      • There might be a lot of people who say they care about that ideal, but when push comes to shove, discard it for other values.
    • Once you've reached everyone in that audience of ideals, you've hit a ceiling.
    • It's much easier to have a product that adds novel value as its primary use case… and that as a bonus is something that people can feel good about adopting because it aligns with their ideals.
24I was reminded this week of two frames of looking at the world: the hedgehog and the fox.
  • I was reminded this week of two frames of looking at the world: the hedgehog and the fox.
    • The hedgehog has one way of looking at the world, they hunker down and can be somewhat spiky.
    • The fox has multiple ways of looking at the world, jumping happily between them.
    • The hedgehog has one solution to all questions, the fox has a bag of different answers to different questions.
    • Hedgehogs are Sarumans, and Foxes are Radagasts.
25There are two ways to change the game.
  • There are two ways to change the game.
    • One way is to hunker down and be incurious.
      • You hold on tightly to your idea, and don't let it be dulled by contact with the average mush of the world.
      • If it turns out you're right, you've changed the game.
    • The other is to be curious, playful, and empathetic.
      • You hold lightly to any given idea, flitting between different viewpoints.
      • Every so often you discover a previously undiscovered synthesis of multiple viewpoints that changes the game.
    • The former is the Saruman, the latter is the Radagast.
    • The differentiator is curiosity.
26If you have a decision with massive implications, make it intentionally, not accidentally!
  • If you have a decision with massive implications, make it intentionally, not accidentally!
27When considering options, don't compare options of differing levels of fidelity.
  • When considering options, don't compare options of differing levels of fidelity.
    • The effort to take a high-level ambiguous solution to increasing levels of fidelity takes significant time and effort… and also likely uncovers more unknown unknowns.
      • These unknown unknowns might turn out to make the entire solution non-viable.
    • Fidelity increases by an order of magnitude in each of these phases:
      • A high-level sketch of a solution
      • A design doc of a plausible solution
      • A prototype of a plausible solution
      • A solution working in the real world today
      • A solution that has worked in the real world for a decade
    • Each order of magnitude of fidelity requires an order of magnitude of work to accomplish.
      • Each timestep of work to flesh out has a non-trivial chance of discovering a game-over condition for that solution.
      • Some abstract ideas sound great, but as you flesh them out they disappear, like a mirage.
    • Ideas only get less pure as you develop them to higher fidelity.
    • If you compare one option that is at high-level sketch, it will look pure, clean, and promising, especially compared to a messy, bug-ridden prototype.
      • But the work to create the extra fidelity in the former will almost certainly make it messy, too.
      • And worse, the other approach might turn out to be non-viable as you develop it.
    • Pick the pre-existing thing that is the highest level of fidelity that is a good enough match to your specific problem domain.
28The hyperobject in your head is perfect, the reification of it in words or actual embodiment is always a poor imitation.
  • The hyperobject in your head is perfect, the reification of it in words or actual embodiment is always a poor imitation.
    • This is partially due to the translation process being imperfect and never capturing the true hyperobject.
    • But it's also because the hyperobject is squishy and fluffy, and you interpret any ambiguity in your idea in the best possible light.
      • But someone else won't necessarily give you the benefit of the doubt.
      • The real world is a cruel partner, it doesn't care about your feelings and will crush whatever non-viable thing you give it.
      • The bright daylight of reality is extremely harsh and unforgiving.
    • If you've tried multiple times to translate the hyperobject into a real-world reified thing and none of them has been viable, maybe the hyperobject isn't as perfect as you think it is?
29The danger of novelty is that you get stuck in a tarpit of unknown unknowns.
  • The danger of novelty is that you get stuck in a tarpit of unknown unknowns.
    • Here's how the scenario often plays out.
    • You run into a previously unknown unknown that blocks your path.
    • The way to address the problem is a month of effort to fix it.
    • You invest the time, all the while with the pressure mounting as your overall progress is slowed.
    • If you get unlucky, it turns out the work doesn't fix it, but uncovers yet another previously unknown unknown.
    • Now you have even more sunk cost down this path, and also more urgency (you're already behind), so you sprint into the next unknown unknown even faster.
    • The faster you sprint, the more you thrash, the more you get stuck in it.
    • At each point, you're faced with two bad choices: wade deeper into the tarpit, or try a totally different approach that might be just as bad, now with significant time pressure.
    • This is the danger of novelty.
30If you're venturing into what might turn out to be a tarpit, set tripwires.
  • If you're venturing into what might turn out to be a tarpit, set tripwires.
    • Agree as a group what it would look like in a month if the situation had obviously turned into a tarpit.
      • List a number of metrics or things that could happen, any one of which would be obviously evidence of something being very off-track.
    • Everyone thinking about the future can be sober.
    • Make everyone take a blood oath to agree to pull out of the tarpit if any of the enumerated situations are true in a month.
    • Then, look at that list in a month, and if any are true, no matter what, start a crash effort on an alternate approach with at least part of your team.
31A nasty surprise is unit for unit heavier than a delightful surprise.
  • A nasty surprise is unit for unit heavier than a delightful surprise.
    • A nasty and delightful surprise of the same volume doesn't cancel out; the nasty surprise more than outweighs it.
32A nugget of insight from an engineering legend at Google: "If you take a dependency on new infrastructure, it becomes your P0."
  • A nugget of insight from an engineering legend at Google: "If you take a dependency on new infrastructure, it becomes your P0."
    • Everything in the product–including just getting it to market–is now downstream of that infrastructure.
    • New infrastructure tends to be a tarpit.
    • An idea that looked great in the abstract morphing into a grotesque, fractally complex quagmire.
    • The fate of the project is now tied to whether you can make it through that tarpit and get to the other side.
      • That is, whether the infrastructure actually works for the purpose it's being put to.
33Order of magnitude analysis is a useful tool to make better gut decisions in uncertainty.
  • Order of magnitude analysis is a useful tool to make better gut decisions in uncertainty.
    • In uncertainty, attempting to get detail and precision will be expensive and misleading.
    • But often if you can guess at the orders of magnitude of different factors you can do a good job picking between two paths.
    • List all of the possible costs that are a given order of magnitude.
      • Costs here also include uncertainties or unknowns.
      • Ignore any costs that are lower than the highest order of magnitude.
        • They are dominated by the higher costs anyway.
      • Multiply any costs of the same order of magnitude together.
    • Estimate the order of magnitude value you can expect.
    • If the order of magnitude of value is not at least an order of magnitude greater than the order of magnitude of the costs, the idea is not worth pursuing.
    • Another trick to reduce orders of magnitude of uncertainty is to reduce the time horizon to small time slices.
      • "If we take this small, obvious step, what's the chance it at least pays for itself? And what's the chance that there are further steps beyond it we could choose to take that might be good?"
      • If the answer to both of those is high, simply do it.
34Optimizing for the short term often happens at the expense of the long term.
  • Optimizing for the short term often happens at the expense of the long term.
    • It's extremely easy to get short-term returns by lighting customer trust on fire.
    • Focus on the long term fundamentals with a mindset of managing short term.
    • People with a finance perspective often only see the business metrics which are the short term metrics.
    • The long term metrics are best seen by the product team.
35When a large organization creates an "innovation team" it is nearly always performative.
  • When a large organization creates an "innovation team" it is nearly always performative.[apg]
    • As in, it creates the superficial perception of innovation, without any actual impact.
    • It's very easy to get superficial innovation by ignoring the real constraints.
    • The team can just do prototypes that demo well but are impossible to actually deploy at scale.
    • The prototyping team gets pats on the back and held up by leadership as being examples for everyone else to follow, with the implication that the legacy teams are lazy or over-complicating things.
      • Which the legacy teams will deeply resent, of course.
    • The problem with the legacy teams is that they're stuck in a gnarly web of constraints that emerge with scale.
      • The downside grows faster than the ability to navigate it.
    • The superficial problems of performative innovation are easy.
    • Actually navigating the real constraints is hard, and requires a different kind of creativity.
    • Actual innovation emerges from the bottom up, a creative act.
    • To get innovation, you need to give space for acorns to have the possibility to grow.
36You can't force someone to trust you.
  • You can't force someone to trust you.
    • Trust is like love.
    • It is a creative act that must emerge authentically without being coerced.
    • Trust creates the potential for larger outcomes, because you now are willing to believe.
37Magic is created when people believe.
  • Magic is created when people believe.
    • You get an order of magnitude more motivation, openness, curiosity, energy.
    • It doesn't matter what they believe in, as long as they believe in something.
      • Believe in themselves.
      • Believe in the mission.
      • Believe in the team.
    • When you have a team that believes in something that is aligned, it's an unstoppable force that can change the game.
38If you overextend your formal authority beyond what your informal authority allows, you'll be in dangerous territory.
  • If you overextend your formal authority beyond what your informal authority allows, you'll be in dangerous territory.
    • For people to listen to your authority they have to believe it is legitimate.
    • If you overextend your formal authority, then people will disengage, or possibly even exit.
    • People will not freely give their discretionary effort if they don't believe.
    • You can't force someone to believe.
39Discretionary effort comes when people trust, when they believe.
  • Discretionary effort comes when people trust, when they believe.
    • Discretionary effort is going above and beyond, applying creativity in authentically motivated ways.
    • It's where all truly great things come from.
40Trust is one of those funny things.
  • Trust is one of those funny things.
    • Below a critical point, more tension removes trust.
    • But above a critical point, more tension adds trust.
41Trust is in some ways the opposite of fear.
  • Trust is in some ways the opposite of fear.
    • When you're fearful, more cycles together reduces trust.
    • When you're trusting, more cycles create more trust.
42If you're fearful an interaction will be emotionally charged, that guarantees it will be emotionally charged.
  • If you're fearful an interaction will be emotionally charged, that guarantees it will be emotionally charged.
    • Because you will be pre-charged, so you'll interpret anything ambiguously threatening as evidence of a hidden threat.
      • You'll be on a hair-trigger.
    • Fear makes things charged.
43Organizations need rules to be efficient and cap downside.
  • Organizations need rules to be efficient and cap downside.
    • But rules ossify.
    • Organizations also need play to be alive, to innovate, to care.
    • Rules seem serious.
    • Space to play seems frivolous.
    • So we tend to do more rules than play.
    • Playful explorations are easier to do in the office, it's harder to do over video conference.
44All assemblages of multiple humans are dysfunctional to some degree.
  • All assemblages of multiple humans are dysfunctional to some degree.
    • Sometimes it's worse when a dysfunctional team has to pretend like it's functional.
    • For example, if the dysfunctional manager reads his first management book and sees the superficial insights but doesn't see the fundamental truths.
    • "I read that high-functioning teams never have side-channel conversations, so therefore we will have no side-channel conversations".
    • Now you now have no way to have disconfirming evidence.
    • Side-channel conversations allow people to de-risk, "am I the only one who sees this disconfirming evidence? Or will I at least be backed up by others if I speak up?"
    • High-trust teams don't need side channel discussions very often because people feel comfortable flagging disconfirming evidence and knowing they won't get in trouble… but getting rid of side-channel discussions does not make a team high trust.
45If you have all one type of person you can get stuck easily.
  • If you have all one type of person you can get stuck easily.
    • Diversity makes it less likely you get stuck.
    • It's more likely that someone in the group has the idea necessary to unstick you.
      • Or that a combination of perspectives in the group can unstick you.
    • Dealing with unexpected situations requires variance, novelty to select over.
    • If you have very little novelty to choose from, the likelihood you possess the right answer to the novel problem is much lower.
46We focus on individual performance more than group performance.
  • We focus on individual performance more than group performance.
    • When you focus on individual performance it's easier to decide who to allocate scarce spoils to.
    • But no one is an island, especially in an organization tackling problems with any amount of complexity.
      • If you optimize for individuals, you'll end up inadvertently selecting for super-chickens that succeed relative to others by pecking the others to death.
    • It is teams that successfully navigate complexity.
    • Everything that actually matters is a team sport.
    • We should look first and foremost at the effectiveness of teams and then secondarily the individuals within the group who were most helpful.
    • Sometimes the most important team players won't be visible outside the group; they won't look like they're doing anything, but they're helping make an environment where everyone else can do their best work.
      • The glue players. The Radagasts.
      • Their impact is fundamentally hard to measure with reductionist tools.
    • The best tactic to measure performance is to measure the effectiveness of a team, and then ask the team-members themselves to rate who was most critical to the success of the team, and who detracted from the success of the team.
      • Then trust them, even if the people they say are most important don't look to you like they're doing anything that useful.
47Look for the "wow" moment as a signal of potential greatness.
  • Look for the "wow" moment as a signal of potential greatness.
    • If it's solid but has no moments of wow it's merely very good.
    • Something that has even one wow moment (as long as it is not just a superficial wow), is very likely to be great.
    • Very good and greatness are two totally different things, and those kinds of people need very different management styles.
48Chefs are able to do 10x output for things they believe in.
  • Chefs are able to do 10x output for things they believe in.
    • For things they don't believe in, where they're asked to be a line cook, they might only be 0.5x.
    • If you have a chef, and there are two equally risky options, but the chef believes in one, go with the one they believe in.
    • Even if it's not exactly right in the end, the 10x effort will allow it to be much more likely to be great.
49A 10x engineer given the wrong requirements will build a cathedral in a desert.
  • A 10x engineer given the wrong requirements will build a cathedral in a desert.
50Are you disagreeing on the details or on the fundamentals?
  • Are you disagreeing on the details or on the fundamentals?
    • Often it's very hard to detect which.
    • But if you're disagreeing on the fundamentals, trying to resolve the disagreement at the details will never work, you need to pop up a layer.
    • Once you're aligned at the fundamentals, the details resolve an order of magnitude faster.
    • You were trying to go fast which meant you only had time to talk about the details, but you can't resolve the issue at that layer.
    • Similar to doing dynamic programming in the opposite order, it's massively more expensive in the end but looks superficially faster in the short term.
51The Google Docs commenting tool is an affordance for nitpicking.
  • The Google Docs commenting tool is an affordance for nitpicking.
    • Challenge yourself to give high level responses in a separate email or channel.
    • That forces you to not talk about object-level details and nitpick but instead give high-level feedback about the whole.
    • If there are multiple comments that are multiple paragraphs, that's a good sign the doc author and the commenter should just have a live conversation about the high level questions and get out of the details.
52If you have to sprint you can only follow your gut.
  • If you have to sprint you can only follow your gut.
    • If team members' guts don't align you'll have strife.
    • An option: allow a bit more space for people to sketch out approaches and agree on one.
    • Go a tiny bit slower to go much faster once you make it out of that storming phase.
53The speed of execution is orthogonal to the speed towards the desired outcome.
  • The speed of execution is orthogonal to the speed towards the desired outcome.
    • We intuitively act like they are one and the same, but often execution velocity is a poor proxy for speed towards the outcome.
    • It's possible and distressingly common to sprint around in circles and go very fast in execution but make progress away from the desired outcome.
54If you aren't aligned at the high level, faster sprinting destroys value.
  • If you aren't aligned at the high level, faster sprinting destroys value.
    • Each unit of effort decoheres the whole more.
    • Once you share an aligned high level vision and set of constraints, faster sprinting creates more value.
    • The difference is hard to see unless you take a step back.
      • The difference between default-cohering and default-decohering is miniscule, subtle, but infinitely important.
55The more novelty, the harder it is to get everybody aligned.
  • The more novelty, the harder it is to get everybody aligned.
    • In a novel domain, everyone's pre-existing gut for the proper direction starts off pointing in different directions.
    • It takes time to bring everyone into alignment by unearthing and then socializing high-level insights.
56A natural schelling point: a thing everyone can agree is reasonable and that they can live with.
  • A natural schelling point: a thing everyone can agree is reasonable and that they can live with.
    • Seems like a low bar but is actually a very high one.
    • It's also magnetic; an idea that everyone finds reasonable will emerge quickly as the default choice.
    • Sometimes it takes a bit of sleuthing for hidden constraints to identify such an idea, but if you do it's the best way to get people successfully coordinated and committed.
57When trying to understand the motivations of a tech company's decision, ask yourself "What would a high-potential but somewhat junior PM pick after one day of research?"
  • When trying to understand the motivations of a tech company's decision, ask yourself "What would a high-potential but somewhat junior PM pick after one day of research?"
    • Tech companies rarely have time for multi-ply thinking, it's often just high-quality but fundamentally-single-ply thinking.
    • This is true in hyper-growth companies because there's no time for anything but fast execution.
    • It's true in large organizations because coordination on multi-ply ideas is so expensive that it's easier to go with a one-ply idea that everyone can understand and find to be reasonable.
      • And something that is similar to off-the-shelf solutions used elsewhere, that is, with less novelty, is way more likely to clear that bar.
58Learning happens in feedback loops.
  • Learning happens in feedback loops.
    • But the loop has to be minimally viable to even be a loop in the first place.
      • Versus an open path.
    • Once it is a loop then you learn as you iterate.
    • Many things never get to the level of minimal coherence and viability to be a loop.
    • Until you get a loop it's a tentative whisp that could evaporate if you get distracted for even a second.
    • Once it's a loop it is self cohering as it turns.
    • Hill finding is finding a loop.
    • Hill climbing is turning a loop.
    • Process praise sets the bar lower for what counts as viable for the learner, where they get some encouragement to keep going.
      • This makes it more likely they persist to the point where they get a loop in the first place.
59In the tech industry everything is always on fire.
  • In the tech industry everything is always on fire.
    • If it's not on fire, you're supposed to turn up the heat until it is.
    • Then you step back and you realize if you would have gone one notch slower you would have saved tons of time by aligning on relevant information ahead of time.
    • Why does the software industry always redline?
    • Maybe because of the winner-take-all dynamics in tech.
      • These arise partially due to bits being many orders of magnitude easier than atoms to transport.
    • You can't think multi-ply when things are on fire.
60Socialization is a two-way process, not a one-way process.
  • Socialization is a two-way process, not a one-way process.
    • The default stance for socialization is often "I try to cram my idea into a critical mass of other people's heads, using whatever path wedges it in there most efficiently."
      • A very one-way process.
    • The goal is not to figure out a way to cram your idea into the other person's head.
    • It is to share your insights and listen to theirs, to synthesize an idea that is better than what either of you had individually before.
    • Socialization is primarily a co-creation process, not a convincing process.
    • The idea should change as you socialize; it should improve, become more nuanced, more realistic, more grounded.
      • If you learn something that makes you realize it's no longer a good idea, you should set it down.
61Leaders should engage in the higher level requirements, risks, and concepts, not on details.
  • Leaders should engage in the higher level requirements, risks, and concepts, not on details.
    • All of the details emerge in higher pace layers built on top of those lower pace layers of concepts, requirements, etc.
    • When a leader dives into a detail, the likelihood the right solution is discovered is orders of magnitude less likely.
      • The leader will likely be busy and have limited time.
      • The expert with experience in the details will likely have a calibrated gut that is extremely expensive to serialize into words.
      • As the expert struggles to distill a succinct summary, the leader will think the expert isn't hearing them.
      • The expert will think the leader is out over their skis in the details and lose trust in their technical details.
      • The leader outranks the expert and will win, all else equal.
      • This makes the expert existentially nervous, and puts them more into a defensive crouch.
      • As the expert gets more defensive and closed, the leader thinks less of them, and is more incentivized to make a decision by fiat… that will likely be the wrong one.
    • This is one of the reasons why micromanaging is bad.
      • Diving into details that the leader can't know in sufficient depth, with a massive power imbalance.
      • The leader is a giant; each twitch reverberates through the org, causing whiplash at the edges.
      • The leader should speak softly at the more conceptual / requirements layer.
    • My worst manager (by far) got into this mode.
      • I had come up with an elegant and clever solution to the problem at hand, navigating dimensions he didn't even realize existed.
      • He pressured me to come up with a different solution that fit a sketch he had created.
      • I couldn't see how to steelman that solution; it seemed to fundamentally miss out on key dimensions and not be viable.
        • I am unstoppable when I see a viable solution to a complex problem, but can't motivate myself to execute on a plan I know for sure will not work.
      • As he pressed harder I got more fearful and got into a deeper defensive crouch.
      • That defensive crouch made me less likely to be open to what he was saying.
      • In retrospect, he saw an additional dimension in the internal politics above us that changed the calculus; his sketch wasn't right but had I known the additional high-level constraint I could have come up with an improved plan that would satisfy that constraint, too.
      • As I crouched, he got more frustrated that I wasn't absorbing the (implied) high-level constraint he saw but had not communicated.
      • What he should have done is focused not on the details of a plan to force upon me, but on better clarifying the constraint.
      • I like to think that if that happened again my response would not be to hunker down and get fearful and defensive but instead push to clarify the higher-level constraint he was seeing that I didn't yet understand.
    • It's on the leads to clearly extract and communicate the high-level constraint.
      • You can't blame the team if they execute in a way that violates the constraints you never bothered to unpack for them.
62Punching down is an order of magnitude more aggressive than punching up.
  • Punching down is an order of magnitude more aggressive than punching up.
    • When the little guy feels threatened he might punch up.
    • In response, the big guy then might punch down.
      • "He started it!"
    • But punching down is an order of magnitude more aggressive than punching up.
    • Because you really could cause a game-over for the other player, with the same units of force.
63The tree that bends in the wind doesn't snap.
  • The tree that bends in the wind doesn't snap.
    • The tree that breaks when it bends is not viable.
    • Keep as many things as flexible as you can.
    • What are the core principles that are non-negotiable?
      • These may be inflexible and firmly held.
      • You get a very small set.
    • Everything else should be flexible.
64I found this analysis of scifi movie armor by A Collection of Unmitigated Pedantry fascinating.
  • I found this analysis of scifi movie armor by A Collection of Unmitigated Pedantry fascinating.
    • Armor style is coevolved with the selection environment.
    • For real armor it's coevolved with the enemy's weapons and tactics, with the ultimate ground truth of helping the wearer be an effective fighter and survive.
    • For movies it's designed to look cool, look different enough from other movies as to not feel derivative, and to allow actors to move and emote.
    • These are very different forces that pull them in very different directions.
65One of the drives to perfectionism is the terror of feeling shame.
  • One of the drives to perfectionism is the terror of feeling shame.
    • This can happen if you have a hyper-active shame response, where feeling shame feels like you're dying.
      • (Ask me how I know this!)
    • To avoid the chance you feel shame, you hold yourself to an impossible standard, trying to always be above reproach in every single thing you do.
    • This is an impossible bar to clear!
    • You'll be miserable, constantly failing to hit it.
    • On the one hand, you'll never give up and always push yourself, intrinsically, to be better.
    • On the other hand, you'll be making yourself miserable emotionally, never satisfied with what you've accomplished.
66My number one productivity hack: get enough sleep!
  • My number one productivity hack: get enough sleep!
    • I can survive with a little less sleep than I need.
    • But I cannot thrive unless I'm fully rested.
    • I am militant about getting the sleep I need, it's a non-negotiable bedrock of my day.
    • So much of our work culture fetishizes the hustle, the constantly being busy.
    • The more busy you are, the more you cut into sleep time.
    • Sleep is what allows us to form new mental pathways, to be fresh to tackle complexity with our full minds.
    • Getting enough sleep does not make you weak or lazy; it allows you to be mentally strong.
67Another instant classic from Gordon on Where to Draw the Line.
  • Another instant classic from Gordon on Where to Draw the Line.
    • The last paragraph is almost poetry: "When drawing lines, take it easy. Understand the whole before you cut it to pieces. Pause, listen. Let yourself be guided by the natural line."
68Someone told me about the tradition in Victorian England of "commonplace" books.
  • Someone told me about the tradition in Victorian England of "commonplace" books.
    • People would collect little nuggets of insight: observations, quotes, recipes, etc, in little books for themselves.
    • Someone pointed out that my bits and bobs practice seems to be a kind of convergent evolution to the same kind of tradition.
69Be open to things you don't understand having value.
  • Be open to things you don't understand having value.
    • Sometimes something that feels like it doesn't have value has little value.
    • Sometimes it's just that you don't understand it yet.
    • It's always at least a little bit of both.