Chaque semaine, je partage quelques articles que j’ai trouvés particulièrement enrichissants. J’espère qu’ils vous aideront autant qu’ils m’ont aidé.
Cette semaine (entre autres) :
TikTok and the sorting hat
Fortnite banni de l’App Store et de Google App
Unhealthcare publie un livre sur l’importance de l’Assurance santé
Pourquoi les “job interviews” classiques ne fonctionnent pas
Certains articles sont en français, la plupart sont en anglais (je copie certaines citations en anglais). Ils ne sont pas tous récents et vont au rythme de mes lectures.
Bonne lecture !
Si vous aimez, vous pouvez vous inscrire, partager (par email ou le blogpost) ou encore me suivre sur Twitter.
📱Monde des technologies
👉 TikTok and the Sorting Hat — Remains of the Day (Eugene Wei)
It turns out that in some categories, a machine learning algorithm significantly responsive and accurate can pierce the veil of cultural ignorance.
TikTok's story begins in 2014, in Shanghai. Alex Zhu and Luyu “Louis” Yang had launched an educational short-form video app that hadn’t gotten any traction. They decided to pivot to lip-synch music videos, launching Musical.ly in the U.S. and China. Ironically, the app got more traction across the Pacific Ocean, so they killed their efforts in their home country of China and focused their efforts on their American market.
Alex and Louis listened to Musical.ly’s early adopters. The app made feedback channels easy to find, and the American teenage girls using the app every day were more than willing to speak up about what they wanted to ease their video creation. They sent a ton of product requests, helping to inform a product roadmap for the Musical.ly team. That, combined with some clever growth hacks, like allowing watermarked videos to easily be downloaded and distributed via other networks like YouTube, Facebook, and Instagram.
Still, Musical.ly ran into its invisible asymptote eventually. There are only so many teenage girls in the U.S. When they saturated that market, usage and growth flatlined.
After all, Bytedance paid just $1B for an app.
Bytedance did two things in particular to jumpstart TikTok’s growth. First, it opened up its wallet and started spending on user acquisition in the U.S. TikTok was rumored to have been spending a staggering eight or nine figures a month on advertising. It didn’t look like a wise investment at first. Rumors abounded that the 30-day retention of all those new users poured into the top of its funnel was sub 10%. They seemed to be lighting ad dollars on fire. Ultimately, the ROI on that spend would turn the corner, but only because of the second element of their assault on the US market, the most important piece of technology Bytedance introduced to TikTok: the updated For You Page feed algorithm.
Bytedance has an absurd proportion of their software engineers focused on their algorithms, more than half. The exploit versus explore conundrum is sort of a classic of algorithmic design. An exploit algorithm will give you more of what you like, while an explore algorithm tries to broaden your exposure to more than just what you’ve shown you like.
A few years ago, on a visit to Beijing, I caught up with a bunch of former colleagues from Hulu Beijing, and all of them showed me their Douyin feeds. They described the app as frighteningly addictive and the algorithm as eerily perceptive. (...) Every Douyin feed I examined was distinctive. My friends all noted that after spending only a short amount of time in the app, it had locked onto their palate.
Friends at Bytedance claimed, with some pride, that after they plugged Musical.ly, now TikTok, into Bytedance’s back-end algorithm, they doubled the time spent in the app.
To help a network break out from its early adopter group, you need both to bring lots of new people/subcultures into the app—that’s where the massive marketing spend helps—but also ways to help these disparate groups to 1) find each other quickly and 2) branch off into their own spaces.
It is a rapid, hyper-efficient matchmaker. Merely by watching some videos, and without having to follow or friend anyone, you can quickly train TikTok on what you like. (...) The algorithm allows this to happen without an explicit follower graph.
The challenge is that it’s almost always a really slow build, a social graph, and you have to provide some reason for people to hang around and build that graph, often encapsulated by the aphorism “come for the tool, stay for the network.” Today, it’s not as easy to build the “tool” part when so much of that landscape has already been mined and when scaled networks have learned to copy any tool achieving any level of traction. (...) The idea of using a social graph to build out an interest-based network has always been a sort of approximation, a hack ... But what if there was a way to build an interest graph for you without you having to follow anyone? What if you could skip the long and painstaking intermediate step of assembling a social graph and just jump directly to the interest graph?
Because the videos are so short, the volume of training data a user provides per unit of time is high. Because the videos are entertaining, this training process feels effortless, even enjoyable, for the user.
Think of how many damn interest bubble UI’s you’ve had to sit through before you could start using some new social product: what subjects interest you? who are your favorite musicians? what types of movies do you enjoy? The last time I tried to use Twitter’s new user onboarding flow, it recommended I follow, among other accounts, that of Donald Trump. There are countless ways they could onboard people more efficiently to provide them with a great experience immediately, but that is not one of them.
I don’t think the Chinese product teams I’ve met in recent years in China are much further ahead than the ones I met in 2011 when it comes to understanding foreign cultures like America. But what the Bytedance algorithm did was it abstracted that problem away.
In China, video e-commerce is light years ahead of where it is in the U.S.
Jeudi dernier, l’entreprise qui se trouve derrière le célèbre jeu Fortnite, Epic Games, a annoncé qu’elle allait réduire les prix pour sa monnaie virtuelle de façon permanente et d’environ 20%. Le but étant pour Epic de contourner le système de paiement de Google Play et de l’App Store. Cela pourrait permettre à Epic d’éviter les 30% de fees que touchent les deux plateformes principales sur l’ensemble des transactions faites au sein de l’application, et de faire bénéficier les joueurs de cette économie.
Mais Epic n’a pas pu aller bien loin : seulement quelques heures plus tard, Apple a supprimé Fortnite de l’App Store, mentionnant qu’Epic avait franchi une limite en violant les politiques de l’App Store.
Peu de temps après, Epic a lancé une poursuite contre Apple, affirmant qu’Apple “impose des restrictions déraisonnables et illégales pour monopoliser complètement le marché des applications”.
Epic a mis en ligne une vidéo parodiant la publicité d’Apple de 1984, en plaçant Apple dans le rôle de Big Brother.
Plus tard dans la journée, Google a aussi supprimé Fortnite du Google App store.
🏯Construire une entreprise
👉 David Sacks sur comment opérer une entreprise (20MinVC)
If you do a discount, other customers will know
LTV to CAC should be way higher than 3x. Payback should be less than a year.
Blended CAC: You really want to split it by channel. You should not average the CAC with the organic. Otherwise, you over invest in channels that don't work.
Burn multiple (Net burn divided by net ARR generated) should be going to zero over time.
👉 Pourquoi les “jobs interviews” classiques ne marchent pas (Shane Parrish) :
Why Unstructured interviews suck:
Unstructured interviews can make sense for certain roles. The ability to give a good first impression and be charming matters for a salesperson. But not all roles need charm, and just because you don’t want to hang out with someone after an interview doesn’t mean they won’t be an amazing software engineer.
More common during interviews are more nuanced forms of deception which include embellishment (in which we take credit for things we haven’t done), tailoring (in which we adapt our answers to fit the job requirements), and constructing (in which we piece together elements from different experiences to provide better answers.)”.
One reason why we think job interviews are representative is the fundamental attribution error. This is a logical fallacy that leads us to believe that the way people behave in one area carries over to how they will behave in other situations. We view people’s behaviors as the visible outcome of innate characteristics, and we undervalue the impact of circumstances.
Making interviews more effective:
But it is possible to make job interviews more effective or make them the final step in the hiring process after using other techniques to gauge a potential hire’s abilities. Doing what works should take priority over what looks right or what has always been done.
While unstructured interviews don’t work, structured ones can be excellent.
Kahneman introduced a new interviewing style in which candidates answered a predefined series of questions that were intended to measure relevant personality traits for the role (for example, responsibility and sociability). He then asked interviewers to give candidates a score for how well they seemed to exhibit each trait based on their responses.
The key is to decide in advance on a list of questions, specifically designed to test job-specific skills, then ask them to all the candidates. In a structured interview, everyone gets the same questions with the same wording, and the interviewer doesn’t improvise.
Another step to help minimize your interviewing blind spots: include multiple interviewers and give them each specific criteria upon which to evaluate the candidate. Without a predefined framework for evaluating applicants—which may include relevant experience, communication skills, attention to detail—it’s hard for interviewers to know where to focus.
Competency-related evaluations: What’s the best way to test if someone can do a particular job well? Get them to carry out tasks that are part of the job.
👉 UnHealthcare : livre sur l’importance de l’assurance santé pour transformer le système (Hemant Taneja)
It is the beginning, I hope, of a new era in health innovation — one that sees technologists partner with healthcare stakeholders to put the patient-provider relationship back at the center of health experiences.
We developed Commure to provide the deep infrastructure needed for Health insurance. It is a new category of innovation that is committed to bringing modern consumer experiences and accelerating rational economic behavior through innovative business models.
Instead, health insurance companies, built on open tech standards with empathetic user design and the responsible use of AI, allow for fewer, more tailored interactions with care providers.
Health insurance companies will begin the important job of closing the gap in access to and quality of care.
Better billing software, more sophisticated health monitoring tools, easier scheduling, and reporting interfaces may change the roles of current workers.
It plans to take its collaboration tools to the next level and double down on its five existing markets (the Netherlands, Belgium, Germany, the UK and Ireland).
It also claims to now be Europe’s largest medical network, with more than 250,000 healthcare professionals using its platform and exchanging more than 20m messages per month.
They can also set up and join groups, or ‘spaces’, which function a bit like a Slack channel; places for specialists to share relevant information with one another, for example, or for hospital teams to give one another updates.
💚 Les publications d’Alan et sur Alan
Calme plat au mois d’août 🏖 On prépare la rentrée, de belles annonces à venir … Rendez-vous le 16 septembre !