A Learning a Day

Notes on LLM RecSys Product – Edition 5 of a newsletter focused on building LLM powered products.

---

Four posts into this series on how LLMs are changing product building, we’ve spent all our time on what we build. We’ve talked about why I think LLM-powered recommender systems are going to be the core primitive of product building going forward, dug into the power of teacher models, building eval loops, and writing product policy. The next stops are exploring how and who. But before we do, it is worth pausing.

There is an enormous amount of noise around AI right now. And that noise is the AI bullshit trinity that we should talk through first — Doomporn, AI-native efficiency theater, and AI productivity kabuki.

Doomporn

Doomporn is the successor to hustleporn from a few years ago. We can’t seem to get enough of content that tells us that there is going to be a job apocalypse. We should save up and get a bunker now.

Why this is happening

The biggest purveyors of this narrative are the big foundational model labs who have every incentive to keep the volume on these turned up (until the negative impact on their brand becomes the dominant factor at least). Trillion-dollar IPO outcomes require investors to believe the addressable market is all of human labor. In effect, the most apocalyptic framing of AI’s capability is also the most lucrative one. Everyone talks their book.

This isn’t to say the existential concern is fake. Every powerful technology has a dark side, and this one is particularly powerful. The risks are non-zero and worth being thoughtful about. But, the trillions of dollars at stake right now muddy this narrative.

What is likely not true

That anyone knows the second and third order effects. Labor markets are complex. Here’s an example of an inflection in software engineering hiring on Indeed. It might be Jevons paradox (making something cheaper results in more usage of it) or it might be something else.

Either way, anyone speaking with certainty is either selling you something or performing certainty for an audience.

What looks to be true and useful

Agentic coding has real product-market fit. The bottleneck for builders has shifted — it used to be how to build, it is increasingly becoming what to build. That’s a genuine change and it shows up in Anthropic’s once-in-a-generation growth rates.

The bull case that few talk about amidst all this noise is that we could see more small teams, at terminal size, growing revenue per employee for years. The days of 70% software margins at venture scale may be done. But that doesn’t mean great businesses won’t get built — it means they’ll get built differently. Smaller groups, less coordination overhead, a longer tail of viable companies and cultures to choose from.

Work might actually become a happier place for many. But nobody is selling us this version because there’s no trillion dollar IPO attached to it.

AI-native efficiency theater

Every company wants you to know they are “AI-native.” This layer of bullshit is a broad swathe of companies playing for hundreds of millions of dollars in valuation upside.

Why this is happening

The valuation pull on being “AI-native” is enormous. Revenue per employee is the new metric Wall Street cares about. Growing the numerator is hard, but shrinking the denominator is relatively easier. And AI gives perfect cover for layoffs that were always coming after the boom in hiring during the zero interest rate era.

This creates a catch-22 inside large companies. Leadership wants the workforce to adopt AI. The workforce understands that adoption accelerates their own redundancy. Adoption stays halfhearted. Layoffs happen anyway.

What is likely not true

That most organizations have actually cracked it. Building good products is still hard. The old hard things got easier, but the bar moves with the capability.

I used to love Airbnb search. Then it felt limiting. Then when we finally got natural language search, it got an “about time” reaction. Expectations go up as execution capabilities go up.

Anthropic’s product velocity is incredible to watch. That said, their consumer app is still incredibly buggy. If the company executing at the absolute frontier with models that can seemingly fix everything hasn’t cracked the consumer experience, it’s a sign that it is safe to be skeptical of everyone claiming they have.

This isn’t a big-company problem for what it’s worth. AI-native startups held up as the future have eye-popping valuations, breathless coverage, and impressive technology — but they also have business models that are still unproven and moats that are still unclear.

People come for the magic, but they stay for the math. The math reckoning hasn’t come yet. But it will. All in all, if it smells like bullshit, trust your instincts.

What looks to be true and useful

Individual empowerment is real. Every large organization I know has seen pockets of genuine transformation. The challenge now is creating systems and environments that can scale what’s happening in these pockets.

While this is disproportionately harder for large companies, the reality is that it is still very hard in small companies too (even if there is a real “clean slate” advantage). Building agentic organizations is a new craft, and learning a new craft takes time.

AI productivity kabuki

Kabuki is a form of classical Japanese theater. Highly stylized and heavily made up with dramatic gestures performed over and over for centuries — the audience knows exactly what move is coming next, and that’s part of the appeal.

The daily AI content cycle is the same thing. You see the same predictable cycle: analysis of the latest model release, the AI agent workflow, the many Open-Claude agents running out of Mac Minis, the new killer use case.

Why this is happening

The incentives are smaller than the trillion-dollar IPO or hundred-million-dollar valuation pulls above — but no less real. Career progress and content impressions add up. And the noise itself creates a permanent feeling of being behind, which makes the content even more compelling to produce and consume. It’s a loop.

What is likely not true

That staying on top of every new tool is the same as getting better at building. Or that the person who knows the most tools is the best builder. Or that using AI to do something faster means it was worth doing.

Using AI to summarize 16 newsletters you never read isn’t productivity. Running an agent to manage an overscheduled kid’s calendar isn’t a solution — it’s automation of a problem you created. The worst thing we can do is get really efficient at something that shouldn’t be done at all.

What looks to be true and useful

Spend time with Claude Code/Codex. Understanding what agentic coding actually unlocks will go a long way in designing systems at work that make the best of what these tools enable.

Real change is going to flow from systems change. And our ability to design systems follows our ability to understand what can be possible.

Three takeaways

1. Name it to neutralize it. The reason we just spent a post deconstructing all three is simple: when you realize everybody is talking their book, it becomes a lot easier to ignore the noise and focus on what’s real.

2. There’s plenty that’s real underneath. We have access to power tools today, and it is still very early. We’re still figuring out how to use them — as individuals, as teams, as systems, as companies. Building that craft is going to take time, and that’s where the real opportunity is.

3. Play offense. A lot of people are stuck right now because the noise has them in a defensive crouch. It is tempting to keep worrying about how we’ll become obsolete. Don’t. Use this time to play offense — to think through how to become more effective. And effective doesn’t mean automating things you shouldn’t have been doing anyway. It means getting better at the skills that actually matter, and then designing systems that scale that ability to more people.

Our next stop will be to work through the skills that matter as we build product today.

In a couple of blog-related email exchanges recently, I found myself referring to this post from many years ago. The graph still says it better than words can.

It is always amusing to see the gap between actual drivers of extrinsic success / wealth and attributed drivers. We have a need to believe in romantic stories of hard work and heroic mindset and are painted these pictures in the stories we are told.

In reality, however, the biggest driver of extrinsic success / wealth is privilege by a long shot. Luck and mindset matter – but, privilege is the platform on which it is built.

Acknowledging the real drivers is the first step to building systems that provide better access to opportunity to those who don’t have it.

Over the years, I’ve come to realize privilege is best understood as what you don’t have to think about.

This comes to life with a thought experiment: what are the odds that a kid born in a suburban neighborhood in New York to parents working at multinational companies will be financially independent by 40? Now compare those odds to a kid born in the ghetto, or the slums of Mumbai, or even the lower middle class of a city in Africa.

We love the outlier stories — the kid who made it against all odds. And yes, there are always exceptions that prove the rule. But that’s exactly what they are: exceptions.

That’s why it starts at birth — who you’re born to and where. That combination determines the initial height of “the platform” you have. And privilege compounds from there. Every advantage makes the next one more accessible. A few years in, it becomes nearly impossible to look back objectively.

So, for every bit of privilege present, there is an equal and opposing internal force that refuses to acknowledge it. The more you have it, the harder it becomes to see.

One of the clearest markers of accumulated privilege is the ability to think long term. Security — mental and financial — is what makes long-term thinking possible. And long-term thinking, in turn, compounds into more privilege.

In the final analysis, it’s just easy to talk about hard work and mindset when you don’t have to worry about the rest.

One of the fascinating threads in “The Last Dance” documentary is the set of changes that took the Chicago Bulls from being a top-two team in the Eastern Conference to arguably the greatest team of several generations.

They had lost to the Detroit Pistons in the Eastern Conference Finals two years in a row. Three things changed.

(1) They took the loss seriously. Most of the team, led by Jordan, gave up their summer break and got back in the gym. The Pistons had bullied them – they resolved to get strong enough that it wouldn’t happen again. You see the result in a moment from the next season’s Eastern Conference Finals: Dennis Rodman commits a flagrant foul on Pippen, nearly pushing him into the benches. Pippen gets up and moves on as if nothing happened. That’s when the Pistons knew they couldn’t shake the Bulls.

(2) Phil Jackson installed the triangle offense. Moving away from a Jordan-centric offense meant Jordan wouldn’t be the leading scorer every night. It was painful for him – but he understood that building an excellent team required it. Jordan drove his teammates hard as part of this. Scottie Pippen bore the brunt, and grew into an incredible Robin to Jordan’s Batman.

(3) Jordan learned to rely on his teammates. In the finals against the Lakers, Phil Jackson asked him on the bench – “Who do you think is free?” “John Paxson.” “Who should you pass to?” “John Paxson.” Jordan passed. Paxson shot. It went in. And again. And again. A switch flicked. After they won, Phil Jackson famously told Jordan on tape – “You’ve won it the right way.”

What stands out most is how much of the arc was about Jordan himself. One player talks about how when you saw Jordan show up and dial up excellence in practice, you could either join him or move on.

But the real shift wasn’t Jordan being excellent. It was Jordan embracing that he needed to inspire excellence – bring the whole team along, and learn to rely on them.

Only then could he go from being an incredible player on a good team to an incredible player on an incredible team.

Back in my management consulting days, one of the first things we’d do with a new client was map out the org chart – who the players were, who the decision-makers were, how the place was wired.

And every time, you’d realize the formal org chart was one thing. The informal org chart was another altogether.

Some of the gap was structural – which functions were considered drivers versus not. Some was old-fashioned politicking. And some was pure competence – the people everyone just knew you went to when something important needed to get done.

It’s an amateur move to look at the formal chart and make decisions based on it. Much of your ability to get things done in an organization is knowing where the influence really sits.

Saw this on a wall recently – “The first step of change is to become aware of your own bullcrap.” | Unknown Author

A keeper.

We’re fluent in the stories we tell ourselves about why things are the way they are.

The cause is almost always placed somewhere outside us.

Awareness doesn’t fix anything on its own. But nothing gets fixed without it.

Every short term move you make to close a sale will come back to bite you.

It’s obvious enough when the sale is a product or service – an over-promise, a nudge of pressure at the end to get the contract signed. The buyer notices. And with a smart buyer in a long-term relationship, the cost always shows up later.

But it’s just as true for the sales we don’t always label as such. Hiring a candidate is a sale. Selling a candidate on a role is a sale. Convincing a teammate to take on a project is a sale.

In each, the temptation to nudge things across the line – a stretched promise, an extra squeeze of pressure – is always there. And in each, the cost shows up in the months that follow.

There will always be a short term extenuating circumstance. Something that makes this deal, this hire, this conversation feel like the exception.

In long term games, the best way to play is to stay focused on the long term – even at the cost of the short term. Even when it is painful.

Especially when it is painful.

Play the long game, always.

Looking back over the past few months, every time I postponed my weekday morning for “later in the day”, it never happened.

Life or work always gets in the way. If the workout doesn’t happen before 6:30am, it doesn’t happen.

I’ve changed this to a simple rule now – rain or shine, get the workout done by 6:30.

If I miss that window, just try again tomorrow. No point fooling myself with “it’ll happen later.”

Ultimately, that’s also the truest test of prioritization. If it matters, it needs to get done first.

I had a chance recently to learn from a friend who has become an expert at getting things done with Claude Code. The agency he’s unlocked with the tool is remarkable.

The most important thing he shared was an analogy, attributed to the creator of OpenClaw. Using these tools for the first time is like picking up a guitar. It doesn’t sound right when you do so. It’s a bit painful.

But, if you keep at it every day, you’ll be twice as good the next week. And twice as good the week after. It compounds.

Watching him in action was the reminder I needed.

Pick up the guitar. And then keep picking it up.

On a warm night last August, a 12-year-old boy named Shawn Dunkley took a family friend’s electric scooter out for a spin near his home in London, Ontario. It was a powerful machine, ordered from the Chinese online retailer Alibaba. Dunkley was barrelling—helmetless, despite his mother’s pleas—along the paths of his family’s suburban neighbourhood, only a two-minute walk from his house. He glanced at the scooter’s speed display: 69 km/h. Suddenly, everything stopped.

Two passersby found him five minutes later, lying unresponsive a few metres from the pathway. His eyes were open, his expression vacant, his hair streaked with blood. A dead raccoon lay nearby. The best anyone can figure is that it darted in front of the scooter. The passersby called 911 and, within minutes, Dunkley was being rushed to London’s Children’s Hospital. He’d suffered a traumatic brain injury, a skull fracture and spinal bleeding.

He was transferred to pediatric critical care in a medically induced coma. A tube helped him breathe, and two catheters snaked out of his fractured skull to monitor his cranial pressure and drain his pooling cerebrospinal fluid. His doctors weren’t sure if he’d live or, if he did, whether he’d walk or talk again. At her son’s bedside, Crystal Dunkley anxiously awaited the 72-hour mark: doctors had told her that if Shawn lived for three days, his long-term odds would shoot up. She hadn’t understood how dangerous e-scooters could be, or how fast they could go. “I thought they were toys,” she says.

Twelve days after Dunkley’s accident, doctors started bringing him out of his coma. First, he gave a thumbs-up. Then, a toe wiggle, which was a tremendous relief—no spinal injury. Then he nodded and shook his head. Finally, his breathing tube came out.

Dunkley’s injury was no anomaly. At Toronto’s St. Michael’s Hospital, e-scooter–related admissions jumped 600 per cent from 2020 to 2024. Pediatric trauma centres have been particularly besieged: SickKids hospital, also in Toronto, treated 46 e-scooter injuries in 2024, up from only one in 2020. At the Montreal Children’s Hospital Trauma Centre, the number of cases multiplied tenfold in only a year, between 2023 and 2024. When I asked one ER physician what could be done to make them safer, he quipped: “Turn them into bikes?”

---

Caitlin Walsh Miller’s article on e-scooters and e-bikes in Canada is worth reading. She makes 3 simple points –

(1) The adoption of e-bikes and e-scooters has been much faster than regulation has been able to keep up. Very few understand what is legal and what is not.

(2) These vehicles can go very fast. Coupled with the lack of clear rules, they are especially dangerous.

(3) e-scooters, in particular, can be easy to knock off balance – increasing risk to injury.

While her notes are primarily about city streets, we see this in our suburbs too and it scares me every time I see a group of teenagers zoom on the wrong side of the road alternating between the bike lane and the sidewalk.

I don’t understand how or why parents would want to enable this – especially because it reduces the amount of exercise and outdoors time kids get. It feels like a lose-lose-lose.

It always gets me reflecting about the hidden cost of chasing convenience.

Communications of the ACM had a fascinating post about how NASA built Artemis II’s fault tolerant computer. 3 fascinating excerpts.

(1) Eight modules with several back up scenarios: “Orion utilizes two Vehicle Management Computers, each containing two Flight Control Modules, for a total of four FCMs. But the redundancy goes even deeper: each FCM consists of a self-checking pair of processors.

Effectively, eight CPUs run the flight software in parallel. The engineering philosophy hinges on a “fail-silent” design. The self-checking pairs ensure that if a CPU performs an erroneous calculation due to a radiation event, the error is detected immediately and the system responds.

“We can lose three FCMs in 22 seconds and still ride through safely on the last FCM,” said Uitenbroek. A silenced FCM doesn’t become dead weight, however; the system is designed to reset, re-synchronize its state with the operating modules, and re-join the group mid-flight.

(2) Multiple redundancies with deterministic error-checking: “This architecture ensures that each FCM sees the same inputs, runs the same application code, and produces the same outputs,” said Uitenbroek. Every second, the drift of any individual FCM is measured and its local clock is recalibrated to the network’s ‘true’ time. If an application fails to meet its strict deadline, the module is automatically silenced, reset, and re-synchronized.

The hardware itself is also reinforced. The system employs triple-modular-redundant memory that self-corrects single-bit errors on every read. Even the network interface cards utilize two lanes of traffic that are constantly compared, ensuring that a bit flip in the communication fabric results in a fail-silent event rather than a corrupted command. The network itself is triple redundant with three separate planes, and all network switches employ self-checking strategies.

(3) Dissimilar redundancies: While the four-FCM primary system is robust, NASA must still account for common mode failures—software bugs or catastrophic events that could theoretically impact all primary channels simultaneously.

To mitigate this, Orion carries a completely independent Backup Flight Software (BFS) system. This is a prime example of dissimilar redundancy. It is implemented on different hardware, runs a different operating system, and utilizes independently developed, simplified flight software.

Even in a total power loss scenario—called a “dead bus”—Orion is designed to survive. If power is restored, the spacecraft enters a safe mode, in which the vehicle first stabilizes itself and then points its solar arrays at the Sun to recover power. Then, it orients its tail toward the Sun for thermal stability before attempting to re-establish communication with Earth. During such a failure, the crew can also take manual action to configure life support systems or don space suits.

Of course, it costs a lot to get this sort of redundancy planning in technical architecture. Those costs make sense on a space mission.

But, that said, there’s a lot we can learn on ensuring we’re making space for redundancy planning that is appropriate to our use-cases.