💡JCNews #56
L'algorithme de TikTok.
Chers amis, fans de bonnes lectures,
Bienvenue aux 36 personnes qui nous ont rejoints depuis mardi dernier ! Si vous lisez ceci et n’êtes pas encore inscrit, rejoignez vite les 2 961 personnes qui ont déjà eu la bonne idée de le faire. C’est juste ici 👇
Dans les JCNews cette semaine, mes dernières lectures les plus intéressantes, dont un zoom sur un must-read à ne pas manquer !
Ecrivez-moi pour en parler, sur Linkedin ou sur Twitter. Les JCNews sont aussi disponibles sur le Blog Alan. Je compte sur vous pour les partager 😊
💡Must-read de la semaine
Chaque semaine, je publie un must-read que vous choisissez.
👉Le design de Tiktok au service de son algorithme (Eugene Wei)
Understanding how the algorithm achieves its accuracy matters even if you’re not interested in TikTok or the short video space because more and more, companies in all industries will be running up against a competitor whose advantage centers around a machine learning algorithm.
I want to discuss how TikTok’s application design allows its algorithm to “see” all the detail it needs to perform its matchmaking job efficiently and accurately.
But most experts in the field doubt that TikTok has made some hitherto unknown advance in machine learning recommendations algorithms. In fact, most of them would say that TikTok is likely building off of the same standard approaches to the problem that others are.
But recall that the effectiveness of a machine learning algorithm isn’t a function of the algorithm alone but of the algorithm after trained on some dataset.
This, then, is the magic of the design of TikTok: it is a closed loop of feedback which inspires and enables the creation and viewing of videos on which its algorithm can be trained.
For its algorithm to become as effective as it has, TikTok became its own source of training data.
The dominant school of thought when it comes to UI design in tech, at least that I’ve grown up with the past two decades, has centered around removing friction for users in accomplishing whatever it is they’re trying to do while delighting them in the process. The goal has been design that is elegant, in every sense of the word: intuitive, ingenious, even stylish.
What if that ML algorithm needs a massive training dataset? In an age when machine learning is in its ascendancy, this is increasingly a critical design objective. More and more, when considering how to design an app, you have to consider how best to help an algorithm “see.” To serve your users best, first serve the algorithm.
TikTok fascinates me because it is an example of a modern app whose design, whether by accident or, uhh, design, is optimized to feed its algorithm as much useful signal as possible.
This design puts the user to an immediate question: how do you feel about this short video and this short video alone?
L’algorithme de Tiktok n’a rien de révolutionnaire, mais ça n’a peu d’importance. Ce que Tiktok a réussi à faire (beaucoup mieux que les autres), c’est de designer son produit de façon la plus efficace pour son algorithme, afin que celui-ci apprenne extrêmement rapidement et la plus agréable possible pour son utilisateur (l’un entraînant l’autre).
Before you’re even watching the video, and understand how the TikTok algorithm “sees” the video itself. Before the video is even sent down to your phone by the FYP algorithm, some human on TikTok’s operations team has already watched the video and added lots of relevant tags or labels.
Mais la “machine” ne peut pas (encore) tout tout seule : même pour un produit aussi sophistiqué que celui de Tiktok, l’algorithme de machine learning ne peut pas tout “apprendre” et doit être aiguillé dans son entraînement par un étiquetage proposé par des employés de la boîte...
On the other hand, maybe you wouldn’t mind reading one tweet at a time if they were better targeted, and maybe they would be better targeted if Twitter knew more about which types of tweets really interest you. And maybe Twitter would know more about what really interested you if you had to give explicit and implicit positive or negative signals on every tweet.
Cette pratique de Tiktok pourrait inspirer Twitter et nos autres réseaux sociaux favoris. Il s’agit alors de voir si l’algo doit “consolider” ou “explorer” : c’est-à-dire nous permettre de voir plus de ce que l’on aime ou prendre des risques à nous faire découvrir des choses nouvelles.
Algorithm-friendly design need not be user-hostile. It simply takes a different approach as to how to best serve the user’s interests. Pagination may insert some level of friction to the user, but in doing so, it may provide the algorithm with cleaner signal that safeguards the quality of the feed in the long run.
Not only does TikTok capture very clean signals of sentiment from its users, it also gathers a tremendous volume of them per session.
In this software era, true competitive advantages, or moats, are increasingly illusory. Most software features or UI designs can be copied easily by an incumbent or competitor overnight.
“In China, if your competitor doesn’t copy one of your successful features inside of two weeks, they must be incompetent.”
En tout cas, ce qui est intelligent, pertinent et réplicable chez Tiktok, c’est de réussir à créer de la valeur pour son algorithme tout en delightant le plus possible ses utilisateurs, et en essayant de toujours mieux les servir. Offrir toujours plus de delight à ses utilisateurs : il n’y a pas de meilleure boussole, et c’est celle que nous tenons chez Alan !
🏯Construire une entreprise
En plus d’articles triés sur le volet, je partage un principe de leadership d’Alan par semaine. Le même que je partage en interne et à nos investisseurs tous les mercredis.
👉 Les Alaners sont scrappy (Healthy Business)
Cela signifie que nous devons résoudre les problèmes nous-mêmes avant d'essayer de trouver une solution scalable; que nous devons faire des choses qui ne sont pas clairement dans notre rôle parce qu'elles peuvent avoir un impact; que nous ne reculons pas parce qu’il nous manque telle ou telle information.
Par ingéniosité, les Alaners trouvent un moyen d'atténuer les inconvénients et les blockers, et d'innover. Notre philosophie est simple: nous commençons par construire un radeau et nous l’améliorons en bateau à moteur à force de naviguer.
👉Les communautés sont partout, la nouvelle vague du Social+ (a16z)
Tiktok. Fortnite. Minecraft. Pinduoduo. These phenomena are what I call “social+” companies: companies that take a single category—from gaming to music to ecommerce—and build an integrated social experience around it.
Any product that has a social component baked in has fundamental and asymmetric advantages over competing non-social products in that category: better growth loops, better engagement, better retention, better defensibility. And because social+ companies are network and community driven, that advantage accumulates over time.
No category is really won until the social product is built.
Having a social graph is table stakes; the true test is whether the graph is unique and inseparable from the product. For a social+ product to break out, it needs to reach a group of people that would only be bought together by that specific product experience
The social element needs to be integral to the experience, not an afterthought.
Peer-to-peer social engagement is baked into the product itself. An audience passively consumes, while a community engages with each other.
When the emotional layer and the transactional one are well-designed and mutually reinforcing, that’s when the magic happens. The best examples work like the Chinese group shopping app Pinduoduo. The transactional act of purchasing (cheap prices!) is inextricably linked to the social and interactive dimension (cheap prices because of friends!!).
Well-designed social+ companies address four key areas:
👉Jeff Bezos à la Startup School de 2008 parle d’AWS. Quelques citations géniales (Youtube)
Amazon real business down the line will be its cloud computing business services. Amazon will be like a bookstore that sells cocaine out the back door.
Started working on web services 4 years ago, and launched first 2 years ago. We did not expect this level of traction so early.
It was a total new customer set for Amazon. It's not a winner-takes-all business.
Heritage as a retailer which is a low margin business where you need to be really really efficient to make money. It comes from focus on frugality and defect reduction
Most of the feedback we got on pricing is that it was really inexpensive. We want to be the low cost provider of these high-quality services.
We have a pretty long history of not talking about competitors. The main reason is that we are more innovative if we stay customer focused and not competitor focused.
🗞Dans l’actu
📱Monde des technologies
👉OpenAI continue à innover avec DALL·E qui crée des images à partir du texte, impressionnant. (OpenAI)
🏥 Santé
👉Les communautés de patients changent (et Nikhil décrit exactement la stratégie d’Alan sur le sujet) (Nikhil Krishnan)
Patients are already seeking information online about their issues. This can be some combination of the fact that advice from the internet is free, they’re embarrassed about their issues,
and they feel the medical system is not hearing what they have to say. However, these communities are almost entirely disconnected from the current healthcare system.
Utilizing patient communities well can improve outcomes in a cost-effective way. Cohort-based programs like Omada, Noom, etc. use group chats and group dynamics to keep members engaged and provide positive feedback on hitting goals. [...] People who are in these groups have 20% to 60% better abstinence than people in other treatments. In no case did they come out worse."
And finally, patient communities are good ways to quickly get feedback or user research from a lot of members that might have different experiences with a disease. Most healthcare companies get very little feedback or input from patients directly, and patient communities make it easy to get input from a lot of people with different manifestations of a disease to understand how their lives could be improved.
Le business model des communautés n’est pas le bon :
It’s not really that surprising then that most of the patient communities that have emerged around that time also followed a similar playbook - advertising. In fact the rates for advertising on a lot of these platforms were way higher than other social networks because the audience was much more targeted with way higher intent to actually use the products. [...] There was a fire sale exit for PatientsLikeMe and lots of smaller ones with companies that struggled to find a sustainable business model.
New patient communities trying to elicit a behavior change and improve patient outcomes would have to be built from the ground up. They would need at the very least:
More involved services (care delivery, moderators, etc.) and proactive involvement in the community instead of being just lurkers or mining data
Specific outcomes that they can measure and track
Products built for specific conditions vs. being disease agnostic like scales for tracking different biomarkers, integrations with wearables for that disease, tools to assist different lifestyle changes or functionality that the patient may have lost, etc.
Il décrit les fonctionnalités que nous avons créées dans Alan Baby :
Group formats - I’ve tried personal trainers and I’ve tried group fitness classes, and both serve a purpose. The personal trainers give you one-on-one attention and help you build your personal routine, but the classes are social and way more energetic. Healthcare can be the same way, with your doctor being the “personal trainer” and group classes to supplement.
Doctors as moderators - If doctors are seeing patients with similar issues, wouldn’t it make sense for them to connect patients within their patient panel (assuming they’re open to it)? What if doctors, nurse practitioners, occupational therapists, etc. became moderators in micro patient communities they ran? It would be a good way to stay engaged with patients outside of the clinic and might even attract new patients who want to be a part of this community or found the doctor through it.
💚 Alan
👉Notre lettre aux actionnaires - Bilan 2020 (Alan Blog). Nous partageons trimestriellement nos progrès et nos avancées avec nos investisseurs, et nous tenons aussi à les partager avec notre communauté. Croissance, internalisation des opérations… toutes nos fiertés de l’année 2020 y sont!
👉Parole de décideurs (Le Figaro). Je partage cette semaine dans le Figaro au sein de la rubrique “Décideurs” les choix radicaux que nous avons faits pour monter Alan (et comment certaines - beaucoup - réunions ne servent à rien)
C’est déjà fini. Bonne semaine et à vous de jouer maintenant ! Invitez vos amis à s’inscrire👇
Et écrivez-moi pour en parler, sur Linkedin ou sur Twitter. Les JCNews sont aussi disponibles sur le Blog Alan.