Two stories about cutting things down: Netflix automates reframing video for phones, and Will Larson explains why the 2010s way of managing engineers is ending.
Picked from 63 new articles across 34 sources · arrow keys for other days
Today in 30 seconds
30s left
Netflix built a three-stage pipeline that turns widescreen footage into 9:16, 4:5 and 1:1 crops in one pass. story 1 ↓
The crop decisions export to Adobe Premiere so editors can adjust them. The post includes no accuracy or time-saved numbers. story 1 ↓
Will Larson argues cheap money set the rules for 2010s engineering management, and those rules are now going away. story 2 ↓
The former Carta CTO went from a 250-person org to 25 people at Imprint, and tells middle managers to re-skill or retire. story 2 ↓
1/2Netflix Tech Blog· Netflix Technology Blog· Oct 26
Netflix automates cropping widescreen video into vertical and square formats
Netflix built a pipeline that turns widescreen footage into vertical and square versions. It finds the shots, works out who matters in each one, and moves a smoothed "virtual camera" over them. Editors can then adjust the result.
Why it matters · It shows a familiar trade-off in practice: replacing manual work with an automated first draft while keeping human override. Know going in that the post gives no measurements.
Know this first
Context from the editor, not the article · 25 sec
1 · Why cropping is hard
Films are shot at about 16:9, wide and short. A 9:16 vertical crop that keeps the full height keeps only about a third of the width. A fixed center crop often cuts off the person speaking, so the crop window has to move, and move well.
2 · Shots and reframing
A shot is a stretch of video between two cuts. Reframing means choosing a crop window for each frame. If the window jumps every frame, the picture shakes. If it never moves, it loses the subject. Good reframing behaves like a camera operator: it holds still or pans on purpose.
The three stages that turn horizontal footage into 9:16, 4:5 and 1:1 crops.1 / 4
2/2The Pragmatic Engineer· Gergely Orosz· Oct 21
Will Larson: cheap money shaped 2010s engineering management, and its rules are ending
On Gergely Orosz's podcast, Will Larson, CTO of the startup Imprint, argues that 2010s engineering-management norms grew out of an era of near-zero interest rates. Managers are now expected to be hands-on technically and to lead larger teams.
Why it matters · If you manage engineers or might, this is a specific, experienced account of which management habits companies are dropping and why.
Know this first
Context from the editor, not the article · 35 sec
1 · What ZIRP means
ZIRP, the zero interest rate period, ran roughly from the 2008 financial crisis to 2022. Money was cheap, so venture-backed tech companies could grow headcount fast. The main limit on growth was hiring and keeping engineers, and that gave engineers a lot of bargaining power.
2 · Engineering management roles
Line managers (EMs) lead a team of engineers directly. Middle managers such as directors manage other managers. A tech lead manager (TLM) manages people and also writes code and makes technical calls. Staff+ engineers are senior individual contributors who lead technically without managing anyone.
Netflix has described an internal system that automatically turns widescreen film footage into vertical and square versions. It serves 1, the vertical video feed in Netflix's mobile app, and its marketing on social media, where Netflix says it has 1.2 billion followers. Editors used to reframe each clip by hand, which doesn't scale.
Why a center crop isn't enough
Cropping a 16:9 frame to 9:16 at full height keeps only about a third of its width (an illustrative calculation, not Netflix's). Directors often put actors at opposite edges of the frame and move the camera, so a fixed crop regularly loses the subject. Netflix treats the job as tracking (where is the subject?) plus temporal optimization (how should the crop move?). It describes the result as a "virtual camera."
Three stages
Shots. A neural-network 2, a model that finds cuts, splits the video into shots. When a single shot holds a conversation between characters on opposite sides, 3 works out who is talking. It combines a face detector with a component called 4, which the post doesn't describe. The system then splits the shot into sub-shots, so the crop can cut between speakers. Netflix says this works for animation as well as live action.
Focus region. Face detections are combined with 5, estimates of where a viewer's eye is drawn, to drop unimportant faces. The remaining faces are ranked by whether they are speaking and how large they are. The result is a continuous region the crop must keep in view.
Motion. Following that region frame by frame would jitter. If it barely moves during a shot, the crop stays fixed. If it moves a lot, the system computes a tracking path and smooths it into a deliberate-looking pan.
The 9:16, 4:5 and 1:1 crops all come from one pass. Netflix says the extra ratios cost no additional compute, which fits: the expensive analysis doesn't depend on the output shape.
Tooling
Editors run the algorithm inside Netflix's internal Clips tool and can adjust any shot's framing. For harder cases, they can export the framing as XML into Adobe Premiere, so outside agencies start from Netflix's crops. Netflix says its system beat Premiere's built-in 6 on a curated internal test set.
Caveats
The post gives no measurements: no time saved, no acceptance rate for automatic crops, and no metric or results for the Auto Reframe comparison. Thresholds, smoothing methods and failure cases go unmentioned. Its main lesson is the design: automation produces an editable first draft, and people keep the final say.
Netflix has described an internal system that automatically turns widescreen footage into the vertical and square formats mobile apps use. The demand comes from two places. One is 1, Netflix's new vertical video feed in its mobile app. The other is social media, where Netflix says its accounts have 1.2 billion followers. Until now, editors made vertical versions by hand, reframing each clip in editing software. Netflix says that doesn't scale as its catalog and its uses for clips grow.
The problem
Most filmed content is about 16:9. Mobile feeds want 9:16 (full vertical), 4:5 (a slightly tall portrait) or 1:1 (square). Some rough numbers, for illustration: a 1920×1080 frame cropped to 9:16 at full height leaves a window about 608 pixels wide, roughly 32% of the original width. A 4:5 crop keeps about 864 pixels (45%), and a 1:1 crop keeps 1080 (56%). Most of the frame gets thrown away, so it matters a lot which part you keep.
A fixed center crop fails whenever the action isn't in the center. That happens all the time in film, where directors place actors on opposite sides of the frame and move the camera. Netflix's framing is that it needs a "virtual camera" that follows the important action and produces crops that look as deliberate as the original shot. The team treats this as a tracking problem (where is the subject?) combined with a temporal optimization problem (how should the crop move over time?).
Stage 1: shot detection and splitting
First, a neural-network 2, a model that finds where one shot cuts to the next, splits the video into shots. Each shot gets its own framing, so the crop never pans across a cut.
Next comes a refinement for dialogue. A single continuous shot often shows two characters talking from opposite edges of the frame. A 9:16 window can't hold both, and panning back and forth looks bad. The pipeline uses 3, which works out who is talking at each moment, to split such a shot into sub-shots. Within one unbroken take, the virtual camera can then cut to whoever is speaking, the way an editor would. The detection combines an internal face detector with a component called 4. The article doesn't describe LipFlap; from its name and role, it presumably looks at mouth movement. Netflix says the approach works on both live-action and animated titles. Animation matters here because face detectors trained on real people often fail on cartoon characters.
Stage 2: finding the focus region
For each shot or sub-shot, the system decides what has to stay in frame. It combines face detections with 5, which are per-pixel estimates of where a viewer's eye is drawn. The maps help filter out faces that don't matter, such as extras in the background. When several faces remain, the system ranks them using signals like whether the person is the active speaker and how large the face is on screen. These signals are combined with what the article calls "shot-level content understanding" (not explained further) to produce a continuous focus region: the area the virtual camera should keep in view across the shot.
Stage 3: temporal reframing
A crop that tracked the focus region exactly in every frame would jitter. So Netflix uses heuristics, rules of thumb, based on how far the focus region moves over the whole shot:
Small displacement: lock a static crop window that covers the region for the entire shot. The effect is a camera on a tripod.
Significant motion: compute a tracking window and smooth its path so the pan "feels intentional and cinematic."
The post doesn't say where the threshold between the two cases sits or which smoothing method it uses. Common choices include low-pass filtering or fitting a constrained path, but Netflix doesn't say which, if either, it uses.
The 9:16, 4:5 and 1:1 outputs all come from one pass. This makes sense given the structure: the expensive steps (shot detection, face detection, speaker detection, saliency) don't depend on the target ratio. Only the window size and its path change. Netflix says the extra ratios add no extra compute.
Getting it into editors' hands
The algorithm runs inside Netflix's internal Clips tool, where editors and producers already work. They can generate several aspect-ratio versions of a clip, review the automatic crops, and correct individual shots with reframing controls.
For complex shots or custom campaigns, producers can export the resizing parameters as XML and import them into external editors such as Adobe Premiere. Agencies and partners then start from Netflix's framing instead of redoing crops by hand or using Premiere's own 6 feature. Netflix says its algorithm beat Auto Reframe on a curated internal evaluation set.
What the post leaves out
This is a short overview, and the gaps are large:
No numbers. There is no time saved per clip, no share of crops editors accept unchanged, no processing cost, and no breakdown of the Auto Reframe comparison. "Outperforms" on a curated set the team chose itself, scored by an unstated metric, is a weak claim without details.
No failure modes. The post doesn't say how the system handles crowd scenes, action without faces, on-screen text or subtitles, or deliberately wide compositions where cropping changes the director's intent.
No mechanism details. The heuristic thresholds, the smoothing method and how the signals are weighted are all unstated.
The design choice worth noting is the human-in-the-loop model. The algorithm produces an editable first draft rather than a finished product, and the XML export carries that draft into outside tools. Netflix presents this as both cutting manual effort and keeping creative control. It's a sensible pattern, but the post doesn't show how much effort it actually saves.
Remember
Treating dialogue shots as several sub-shots lets a narrow vertical crop cut between speakers instead of panning back and forth.
The analysis doesn't depend on the target ratio, so 9:16, 4:5 and 1:1 crops come from one pass.
Netflix gives no measured results; the comparison with Premiere's Auto Reframe uses an unpublished, curated test set.
Larson is a well-known writer on engineering leadership. His books include An Elegant Puzzle, Staff Engineer, The Engineering Executive's Primer and Crafting Engineering Strategy. He was CTO at Carta and Calm, held engineering leadership roles at Stripe and Uber, and is now CTO of Imprint.
tech lead manager (TLM)
A hybrid role that manages a team while also writing code and making technical decisions.
Orosz's diagram of how skill sets overlap across engineering leadership roles, including tech lead managers.
Will Larson, CTO of the startup Imprint and formerly CTO at Carta and Calm, told Gergely Orosz's Pragmatic Engineer podcast that the "unwritten rules" of engineering management from the 2010s are being undone. His argument is that many of those rules came from the economy of the time, not from good practice.
Cheap money set the rules
Larson ties 2010s management culture to 1, the zero interest rate period after 2008. Capital was cheap, so the ability to hire and keep engineers decided which companies won. That gave engineers leverage. Companies sold relaxed cultures and career development, which Larson calls often "self-defeating." In his account, several norms grew from this:
Hiring as the default fix. Managers were told not to think about technical solutions, so they reached for process or headcount. He now calls hiring "the least interesting way to solve every problem."
Managers out of technical decisions. As a new manager he was told to "never make a technical decision again." That expectation is reversing.
Senior engineers expecting fun work. At Stripe, leading about 250 people, he ended up telling 2 engineers (senior non-manager engineers): "I only have suffering to offer, but this suffering is what's going to make the business succeed."
Optional 3. Some firms kept top earners off the pager as a perk. Larson argues that teams who don't feel their own outages build less reliable software.
Stories about incentives
At Yahoo, co-founder David Filo personally approved every server purchase. A black market followed: one engineer traded 30 leftover servers for favours, and servers that were never returned left Yahoo running old, power-hungry machines. At 4, Larson traced the 2010 v4 relaunch's reliability problems to Python memory leaks. He argues the rewrite didn't kill Digg, because a Google algorithm change had already cut traffic 60–70% and Facebook was taking users.
The new shape of the job
Larson says engineering managers now often lead 12–14 people instead of 4–8. Directors increasingly manage tech lead managers (TLMs), managers who also write code. He and Orosz once called that role a career trap, but it is now in demand. He says executives must be both bureaucrats and technologists, and many who rose through 5 were bureaucrats first.
His advice to middle managers is to re-skill or retire. He left a 250-person org for about 25 people at Imprint, and predicts engineering leadership as now practised "will not exist in three years."
Caveats
These are one executive's views from a podcast. The notes give no data for the span figures, the on-call argument or the prediction, and Larson is hiring for the kind of manager he describes.
Gergely Orosz's Pragmatic Engineer podcast has published a long conversation with Will Larson, now CTO of Imprint and formerly CTO at Carta and Calm, with earlier leadership roles at Stripe and Uber. The topic is how the "unwritten rules" of engineering management have changed since the 2010s. Larson's central claim is that many habits engineers treated as best practice were really products of an unusual economy, and they are now being undone. The published notes are a summary of a two-hour episode, so most points below are Larson's opinions and anecdotes rather than data.
The economic cause: 1
Larson ties 2010s management culture to ZIRP, the zero interest rate period that followed the 2008 financial crisis. When capital was cheap, he says, "your ability to hire and close and retain became the differentiator between successful companies." That gave engineers leverage. Companies sold themselves as relaxed places focused on the employee and on career development, which Larson calls "often kind of self-defeating." He recommends Ranjan Roy's essay "ZIRP explains the world" for the wider version of this argument.
Several management norms followed from that leverage, in his telling.
Hiring was the default answer. "We basically told managers they couldn't think about technological solutions," Larson says. Without technical levers, managers were left with process, scheduling or headcount. He now tells his teams that "hiring is the least interesting way to solve every problem."
Managers stayed out of technical decisions. When Larson became a manager, he was told he should "never make a technical decision again." Orosz says he got the same advice. Both say companies now expect managers to be hands-on and in the details.
Senior engineers expected interesting work. At Stripe, Larson led an org of about 250 people. Many of the most senior engineers kept asking for more engaging problems, and some saw his failure to supply them as a breach of an unwritten rule. His eventual answer was: "I only have suffering to offer, but this suffering is what's going to make the business succeed." He says that putting business problems first is now a baseline expectation for 2 engineers, the senior individual-contributor levels.
On-call duty was negotiable. Some companies kept their highest-paid engineers off 3 rotations as a retention perk. Larson argues this made software less reliable: teams that are paged for their own outages build more reliable systems, and teams spared that pain lose the feedback. The notes give no measurements here. It is an argument from incentives.
Incentive stories from earlier eras
Two anecdotes show how organisational rules shape behaviour.
At Yahoo, co-founder David Filo wanted to approve every new server purchase personally, to make teams accountable for spending. The result was a black market. One engineer on Larson's team had 30 servers left over from a cancelled project and traded them to developers who couldn't get approval, in exchange for things he needed. Because servers never had to be returned, Yahoo ended up running many old, power-hungry machines. A control meant to cut waste ended up hiding it.
Management practice was also thinner then. At Yahoo (2008–2010), Larson had two one-on-one meetings with his manager. In the first, the manager asked about someone else's performance. In the second, Larson handed in his notice. He says rare 1:1s were common at the time.
The 4 v4 rewrite
Larson worked at Digg in 2010, during the v4 rewrite. Digg was then a top-10 global website and bigger than Reddit. The relaunch went badly: it alienated core users and had serious performance and reliability problems. Larson and a small team traced the root cause to Python memory leaks, introduced by developers who didn't know the language well.
He disputes the usual story that the rewrite killed Digg. Facebook was already taking its traffic, and a Google algorithm change had cost it 60–70% of its traffic. From outside, the roughly 120-person company looked like it was growing, but it was burning cash and its prospects were shrinking. CEO Kevin Rose, who had just returned, imposed a hard deadline on the rewrite. Larson thinks Digg would have been overtaken by Facebook and Reddit either way.
What the role looks like now
Wider spans, more TLMs. Larson says engineering managers (EMs) now often manage 12–14 people instead of 4–8, and directors increasingly manage tech lead managers (TLMs), people who manage a team and also write code. Both Larson and Orosz have previously called the TLM role a career trap, but Larson says it is now in high demand. Some of his best hires at Imprint are managers who want to keep writing software. He describes a strong TLM as someone who is capable of bureaucratic work but doesn't want to do it, and who is a builder by default.
Executives need two skills. Larson says a CTO or VP of Engineering (5) has to be a bureaucrat, managing incentives and running functions they may know nothing about, and also a technologist who knows what to build and whether a technology choice makes sense. He thinks many executives who came up through 6 were bureaucrats first.
Middle managers: re-skill or retire. Larson left a 250-person org at Carta to lead about 25 people at Imprint. He calls it a deliberate step back, because he expects engineering leadership as we know it "will not exist in three years." He says tools for line management and for writing code are now very good, while the tools of bureaucracy are still poor, so hands-on work has become more rewarding than coordination. He partly attributes the shift to AI tooling, which the episode touches on but which is not its focus. He also mentions an acquaintance who retired because the work no longer allows the same flow.
On people who feel let down by the change, Larson is sympathetic but blunt: "they're right! They have worked incredibly hard, but the rules have changed." In his view the industry forgot how volatile it is: "We really do just get paid to do things that are valuable to the companies."
What to weigh
These are claims from one well-placed executive, not survey data. The span-of-control figures, the on-call reliability argument and the three-year prediction are not backed by numbers in the published notes. Orosz also mentions that Larson is hiring managers at Imprint, which fits the model Larson is recommending. Still, the core argument is coherent: practices that made sense when talent was the bottleneck lose their reason to exist when capital is. The full transcript, and Larson's own essay "Revised rules of engineering leadership" on lethain.com, give the longer argument.
Remember
Larson argues that many 2010s management norms were rational responses to a talent shortage, not timeless best practice.
Engineering managers' spans are widening from 4–8 to 12–14 people, and hands-on tech lead managers are in demand.
Rules meant to control spending, like Yahoo's server approvals, can hide waste by pushing it into informal channels.
