Comment bien mener une recherche : cultiver les compétences qui peuvent être réellement « pratiquées délibérément »

marsbitPublished on 2026-06-15Last updated on 2026-06-15

Abstract

Personne ne vous a jamais vraiment appris à faire de la recherche. On vous donne un bureau, un problème choisi par d'autres, et des instructions vagues pour "produire quelque chose de nouveau". Ainsi, la plupart des gens rétroconçoient le métier à partir de ce qu'ils voient (articles, posts), apprenant à *sembler* chercheurs plutôt qu'à en *devenir*. Les véritables compétences en recherche sont un empilement de micro-compétences, presque toutes cultivables par la pratique délibérée. **Choisir ses propres problèmes.** Nous absorbons souvent des problèmes (du superviseur, des tendances) sans en comprendre le raisonnement sous-jacent, nous plaçant en retard face à des concurrents nombreux. John Schulman recommande de choisir un résultat que l'on souhaite réellement atteindre et de raisonner à rebours pour concevoir les expériences. Cette approche mène à l'originalité. Le "bon goût" est un muscle : prédire les résultats d'expériences, deviner les conclusions d'articles à partir des méthodes, noter quelles avancées resteront pertinentes, puis vérifier. **Améliorer ses sources.** S'alimenter aux mêmes sources (arXiv, discussions) génère les mêmes idées que tout le monde. Les archives anciennes sont sous-estimées : les idées clés (MoE, LSTM) sont souvent des réinventions. Lire des textes fondateurs comme "The Bitter Lesson" de Sutton ou le discours de Shannon sur la pensée créative est crucial. La **largeur** (neurosciences, conception de mécanismes, statistiques, architecture mat...

Personne ne vous a vraiment appris à faire de la recherche. Vous recevez un bureau, un problème choisi par d'autres, et une instruction vague de « faire quelque chose de nouveau ».

C'est pourquoi la plupart des gens font une ingénierie inversée du métier à partir de ce qu'ils peuvent voir (comme les articles, les posts et les annonces), et finissent par n'apprendre que comment « paraître » chercheur, plutôt que comment « devenir » chercheur. La véritable capacité de recherche est un empilement de micro-compétences, et presque chacune d'entre elles peut être cultivée par une pratique délibérée.

Choisir ses propres problèmes

Richard Hamming avait une habitude aux Laboratoires Bell qui le rendait très impopulaire aux déjeuners. Il demandait aux personnes assises à côté de lui quels étaient les problèmes importants dans leur domaine, puis leur demandait pourquoi elles n'étudiaient pas ces problèmes. Les gens finissaient par changer de table.

Cette question pique, car la plupart d'entre nous n'ont pas de bonne réponse. Nous ne choisissons pas des problèmes, nous les absorbons — nous les absorbons de nos superviseurs, des annonces publiées par un grand laboratoire le trimestre dernier, des articles que tout le monde cite et partage cette semaine.

Le problème avec les problèmes absorbés, c'est que vous ne détenez que la conclusion, pas le raisonnement sous-jacent. Vous savez qu'un laboratoire célèbre s'intéresse à une direction, mais vous ne savez pas pourquoi, vous ne savez pas ce qu'ils espèrent découvrir, ni ce qui les ferait abandonner cette direction.

Lorsqu'ils changent de cap, vous le découvrez un an plus tard. De plus, sur un problème déjà à la mode, vous êtes en course contre 1 000 personnes qui ont commencé avant vous et disposent de plus de puissance de calcul.

Le guide de recherche en apprentissage automatique de John Schulman divise ce travail en deux modes. Le premier : vous lisez la littérature et cherchez des aspects à améliorer. Le second : vous choisissez un résultat que vous souhaitez vraiment atteindre, puis vous travaillez à rebours pour concevoir des expériences.

Il préconise le second, la raison implicite étant que cela produit de l'originalité. Un objectif qui vous tient réellement à cœur vous entraînera vers des territoires qu'aucun article de revue n'a jamais couverts.

Quant au « goût » (taste), on en parle souvent comme d'un don. Mais il se comporte plutôt comme un muscle.

Avant chaque expérience, prédisez son résultat ; cachez la section des résultats d'un article et devinez les données rien qu'à partir de sa méthode ; notez quelles réalisations publiées ce mois-ci seront encore importantes dans deux ans, et revenez vérifier plus tard votre taux de réussite. Une prédiction plus une correction, répétées des centaines de fois — c'est ainsi que tout bon modèle est entraîné, y compris celui dans votre tête.

Améliorer ses inputs

Une liste de lecture partagée produit des idées partagées. Si votre alimentation en information se limite aux classements arXiv et à ce qui reste après le filtrage des discussions de groupe, vous arriverez fatalement aux mêmes conclusions que tout le monde au même moment, ce qui rend ces conclusions presque sans valeur.

La valeur des anciens matériaux est largement sous-estimée. Le domaine rejoue toujours son passé avec du retard : les modèles à mélange d'experts (MoE) remontent à 1991, les LSTM à 1997, la rétropropagation du gradient est devenue courante en 1986.

Richard Sutton a écrit The Bitter Lesson en 2019 en un peu plus d'un millier de mots, et ses prédictions sur la trajectoire du domaine ont été plus précises que celles de revues dix fois plus longues. Claude Shannon a donné une conférence sur la pensée créative en 1952 ; sa première astuce était de réduire le problème à une version presque insignifiante, de la résoudre, puis de réintroduire la difficulté petit à petit.

Rien que cette astuce peut vous aider à percer plus de murs que n'importe quel conseil moderne de productivité.

L'étendue est aussi importante que la profondeur. La recherche sur l'interprétabilité emprunte sans complexe aux neurosciences ; la conception d'évaluations (Eval) n'est que de la théorie des jeux déguisée en blouse blanche ; une compréhension pratique de la façon dont les GPU déplacent réellement la mémoire vous permet de prédire quels articles d'architecture sont condamnés à échouer avant même que les résultats de référence ne sortent ; et les statistiques honnêtes sont probablement déjà la compétence la plus rare en apprentissage automatique, où beaucoup de « rigueur » publiée n'est qu'une « intuition » avec des barres d'erreur.

Encore une chose. Lisez l'article lui-même, pas les posts qui le résument. Les annexes sont l'endroit où les secrets sont enterrés, et la section « Limitations » est souvent le passage le plus honnête de tout le document.

Tout noter par écrit

Paul Graham faisait remarquer qu'une idée semble toujours parfaitement aboutie jusqu'à ce que vous essayiez de la mettre par écrit. Mais l'encre sur le papier révèle les défauts que votre cerveau a maquillés : les hypothèses que vous n'avez jamais testées, les étapes en réalité incohérentes, les deux affirmations qui se contredisent secrètement.

Le principe de Feynman était que la première personne à qui vous devez éviter de mentir est vous-même, car vous êtes la cible la plus facile à tromper. L'écriture est le mécanisme de défense le moins cher jamais inventé.

Darwin est allé plus loin, il l'a programmé : tout fait allant à l'encontre de sa théorie était immédiatement noté, car il avait découvert que sa mémoire effaçait les preuves gênantes bien plus vite que les preuves favorables. Votre mémoire fait de même avec votre historique d'exécutions ratées.

Gardez un journal : hypothèses, configuration, attentes, résultats, compréhension mise à jour. Relire les entrées du mois dernier vous rendra profondément humble — aucun relecteur ne peut produire cet effet.

Related Questions

QQuelle est la principale distinction établie par John Schulman entre deux modes de recherche en apprentissage automatique ?

AJohn Schulman distingue deux modes : le premier consiste à lire la littérature et à chercher des améliorations possibles ; le second consiste à choisir un résultat que l'on souhaite vraiment atteindre, puis à remonter pour concevoir les expériences.

QSelon l'article, pourquoi est-il important de choisir son propre problème de recherche plutôt que d'en absorber un ?

AChoisir son propre problème permet d'avoir une compréhension approfondie du raisonnement sous-jacent, d'éviter la concurrence massive sur des sujets déjà populaires, et de générer une originalité en explorant des territoires non couverts par les revues de littérature.

QComment l'article suggère-t-il de développer son 'goût' en recherche ?

AL'article suggère de développer son 'goût' comme un muscle : en prédisant les résultats d'expériences avant leur exécution, en devinant les données d'un article basé uniquement sur sa méthode, et en notant quels résultats publiés resteront importants à long terme, puis en vérifiant la précision de ses prédictions.

QQuels sont les avantages de consulter des matériaux anciens ou d'autres domaines, selon le texte ?

AConsulter des matériaux anciens ou d'autres domaines offre une perspective historique, aide à éviter les répétitions, inspire des approches créatives, et fournit des compétences transversales précieuses (comme les statistiques ou la compréhension du matériel) souvent négligées dans le domaine principal.

QPourquoi l'article insiste-t-il sur l'importance de tout écrire pendant le processus de recherche ?

AÉcrire expose les faiblesses et les contradictions des idées qui semblent complètes dans l'esprit. Cela sert de mécanisme de défense contre l'auto-tromperie, permet de documenter fidèlement les échecs et les apprentissages, et favorise l'humilité en confrontant le chercheur à ses propres raisonnements passés.

Related Reads

Has the Cryptocurrency Market Hit Bottom? Here's What Institutions Think

"Has the crypto market bottomed out? Major institutions are divided on the outlook, according to a recent analysis by Matt Hougan, Chief Investment Officer of Bitwise. Three prominent research firms published in-depth reports on the topic with differing conclusions. Galaxy Digital argues Bitcoin has not yet found its bottom, pointing to only 4 out of 13 historical bottoming indicators being met. Their analysis suggests a potential bottom range of $30,000 to $54,000. NYDIG adopts a more cautious stance, noting that while metrics are close to historical bear market lows, a classic panic-selling capitulation event is missing. They acknowledge the possibility of a bottom but consider it unlikely, citing structural changes from institutional adoption. In contrast, Standard Chartered Bank asserts the bottom is already in at around $59,000. Their revised bullish view, predicting a year-end target of $100,000, hinges on anticipated reductions in ETF selling pressure linked to events like a potential SpaceX IPO. Despite the surface-level disagreement on the exact price floor, the reports share significant common ground crucial for long-term investors. All three institutions agree that a market bottom will likely form within the current year, that current prices are closer to the bottom than to previous cycle highs, and that Bitcoin is poised for another major bull cycle in the future. The core takeaway is that while the precise bottom level remains debated, the long-term value proposition for Bitcoin remains strong and may even be strengthening. Key supportive trends include rising global debt, persistent inflation, declining trust in traditional institutions, accelerating digitization, and improving market infrastructure. Therefore, for investors with a long-term horizon, the focus should shift from pinpointing the exact bottom to recognizing that the cycle's peak is likely still ahead, making current levels an attractive entry point for substantial potential upside."

Foresight News4m ago

Has the Cryptocurrency Market Hit Bottom? Here's What Institutions Think

Foresight News4m ago

2029 Finale Prediction: When Cryptocurrency Completely "Vanishes", Who Can Remain in This Financial Upheaval?

By 2029, the crypto industry will have transformed into a largely invisible but foundational layer for traditional finance. This timeline outlines the key shifts from now until then. By mid-2026, the most sought-after assets on-chain will not be traditional tokens, but synthetic perpetual contracts for private, high-growth companies (like SpaceX, OpenAI). These become primary price discovery tools, highlighting the market's craving for real-world asset value. Most altcoins enter a sustained bear market as their fundamental lack of asset-backed value is exposed. In late 2026, the "AI + Crypto" narrative largely fades as AI giants prove they don't need crypto infrastructure, except for prediction markets betting on model performance. Simultaneously, a quiet but significant wave of tokenization for institutional assets (money market funds, private credit) begins. The industry splits into a noisy speculative economy and a silent institutional one. Throughout 2027, major public blockchain foundations pivot decisively to serve institutional clients, building compliance toolkits and sales teams. However, key sectors hit growth ceilings: private perpetual contracts are legally restricted from public promotion, stable币 growth is capped by looming political uncertainty, and tokenization projects remain cautious. In 2028, following a U.S. election assumed to maintain a regulatory (not prohibitive) stance, a pivotal change occurs. After a major liquidation crisis exposes the flaws of synthetic contracts lacking a real-asset anchor, new regulations allow the *public solicitation* of private security sales (secondary market shares) to accredited investors. This creates a legitimate, direct on-ramp for retail capital into previously illiquid private equity. By 2029, the resulting bull market is driven by trading in real, innovative company shares (biotech, robotics, AI labs), not speculative tokens. "Crypto" as a distinct asset class recedes; it becomes the mundane, unseen plumbing for this new global private markets infrastructure. Tokens that survive are those capturing real cash flows from this infrastructure. Speculation persists but is marginalized. The core questions posed at the start are answered: token value is tied to legally enforceable claims on real assets, frontier tech adoption happens via private market channels, and crypto's absorption into traditional finance is marked by its becoming boring and invisible. The key validation for this entire thesis is whether, by late 2028, a legal pathway exists for ordinary accredited investors to access private assets directly.

marsbit47m ago

2029 Finale Prediction: When Cryptocurrency Completely "Vanishes", Who Can Remain in This Financial Upheaval?

marsbit47m ago

After the U.S. Banned Fable 5, Zhipu's Stock Soared 47%

On June 15, Chinese AI company Zhipu's stock surged up to 47.6% in Hong Kong, closing with a 32.82% gain. This sharp rise followed two key industry events. On June 12, Anthropic was compelled by a U.S. government export control order to suspend global access to its latest flagship models, Claude Fable 5 and Claude Mythos 5, impacting developers and businesses reliant on them. The next day, Zhipu announced it was opening access to its new open-source flagship model, GLM-5.2, for all Coding Plan users, with API and model weights (under the MIT license) to follow. The Anthropic incident highlighted a critical shift in the AI industry: beyond raw capability, the stability, continuous accessibility, and control over AI models are becoming equally vital, especially as AI integrates deeper into business workflows. Zhipu's move, emphasizing that "frontier intelligence should not belong to a few nor be subject to arbitrary revocation," positioned its open, accessible model as an alternative. GLM-5.2 focuses on "Long Horizon Tasks" with a 1M context window, aiming for consistency in complex, extended projects. Market analysts suggest this event exposes the risk of dependency on closed-source models subject to single jurisdiction policies, potentially accelerating a shift toward domestic base models and localized deployments. The investment response indicates a new valuation metric is emerging—prioritizing which companies can provide AI capabilities that are not only advanced but also reliably and sustainably accessible.

marsbit48m ago

After the U.S. Banned Fable 5, Zhipu's Stock Soared 47%

marsbit48m ago

PANews Column Registration and Article Submission Guide

"PANews Column Registration and Submission Guide" provides instructions for users to register as columnists and publish articles on the PANews platform. Key application requirements are emphasized: content should focus on in-depth analysis within Crypto, Web3, blockchain, data, and viewpoints. Content primarily for brand/product introductions will not be approved, and heavily AI-generated content will be rejected. Promotional (PR/soft) content is directed to the business channel. **Registration Process:** * **Web:** Go to the official website footer, click "Apply for Column," and register with a phone number or email (login via verification code, no password). Fill in the column name, description, upload an avatar, and submit links to previously published work. * **Mobile:** Navigate to "My" -> "Contribute & Create" and complete the form. **Article Submission Tutorial:** 1. Log in to the PANews website. 2. Access the "Creator Center" from your personal homepage. 3. Use the editor to create and publish articles. **Video Upload:** The platform supports embedding videos from third-party sites (e.g., Bilibili). Copy the embed code from the source video, use the editor's "Insert/Edit media" button, paste the code under the "Embed" tab, and adjust the display size (recommended: width 100%, height 560px). **PANews Skills (AI Agent Tool):** PANews offers an official AI Agent skill set called PANews Skills, enabling AI tools to query platform content, track trends, and publish column articles directly. It includes three main skills: 1. `panews`: For tracking daily must-read lists, popular articles, and funding news. 2. `panews-creator`: For managing columns, publishing articles, and uploading images. 3. `panews-web-viewer`: For parsing PANews webpages into Markdown. These skills are compatible with various AI Agent tools (OpenClaw, Cursor, Claude Code, ChatGPT, Gemini, etc.). To use the `panews-creator` skill, users must obtain a specific authentication value from the PANews website after logging into their columnist account.

marsbit59m ago

PANews Column Registration and Article Submission Guide

marsbit59m ago

Trading

Spot
Futures

Hot Articles

How to Buy WELL

Welcome to HTX.com! We've made purchasing Moonwell Artemis (WELL) simple and convenient. Follow our step-by-step guide to embark on your crypto journey.Step 1: Create Your HTX AccountUse your email or phone number to sign up for a free account on HTX. Experience a hassle-free registration journey and unlock all features.Get My AccountStep 2: Go to Buy Crypto and Choose Your Payment MethodCredit/Debit Card: Use your Visa or Mastercard to buy Moonwell Artemis (WELL) instantly.Balance: Use funds from your HTX account balance to trade seamlessly.Third Parties: We've added popular payment methods such as Google Pay and Apple Pay to enhance convenience.P2P: Trade directly with other users on HTX.Over-the-Counter (OTC): We offer tailor-made services and competitive exchange rates for traders.Step 3: Store Your Moonwell Artemis (WELL)After purchasing your Moonwell Artemis (WELL), store it in your HTX account. Alternatively, you can send it elsewhere via blockchain transfer or use it to trade other cryptocurrencies.Step 4: Trade Moonwell Artemis (WELL)Easily trade Moonwell Artemis (WELL) on HTX's spot market. Simply access your account, select your trading pair, execute your trades, and monitor in real-time. We offer a user-friendly experience for both beginners and seasoned traders.

2.3k Total ViewsPublished 2024.03.29Updated 2026.06.02

How to Buy WELL

What is $WELL

WELL3, $$WELL: Revolutionizing Health and Wellness with DePIN and AI Introduction In the rapidly evolving landscape of digital technology, the health and wellness sector stands at the forefront of innovation, striving to enhance patient care and promote healthier lifestyles. A groundbreaking player in this domain is WELL3, a pioneering Web3 project that seeks to revolutionize how individuals engage with their health. By leveraging technologies such as Decentralized Physical Infrastructure Network (DePIN), Decentralized Identity (DID), and Artificial Intelligence (AI), WELL3 aims to foster secure, data-empowered health journeys. This comprehensive article delves deep into the core aspects of WELL3, $$WELL, exploring its functionalities, creators, investors, and unique features. What is WELL3, $$WELL? WELL3 serves as an innovative platform set to redefine the approach towards health and wellness. Focused on integrating DePIN and DID alongside AI systems, the project is designed to create personalized user experiences while ensuring the safety and privacy of individuals' health data. With an impressive figure of over one million pre-registered users, the primary mission of WELL3 revolves around enhancing well-being through secure, data-driven health journeys. At its core, WELL3 employs advanced blockchain technologies to ensure that users have complete control over their personal information. This project not only addresses the challenges of data security and accessibility but also aspires to create a vibrant community connected by a shared commitment to better health. Key Features of WELL3: DePIN and DID: These technologies enable secure ownership and authentication of data, giving users full control over their information. AI Integration: Utilizing AI analytics, WELL3 offers personalized insights and solutions tailored to individual health needs. Community Engagement: Facilitates a supportive environment where users can connect, share experiences, and motivate each other toward healthier living. Creator of WELL3, $$WELL The identity of the creator of WELL3 remains unspecified in the available information. As the project progresses, further details may emerge, shedding light on the visionary minds behind this transformative initiative. Investors of WELL3, $$WELL WELL3 has garnered support from a myriad of influential investment entities, highlighting its credibility and potential in the health and wellness space. Notable investors include: Animoca Brands AWS Samsung The Spartan Group Blocore Fenbushi Capital Newman Group Soul Capital XY Finance Lumoz The backing from these established organizations demonstrates a strong belief in WELL3's mission, providing it with the necessary resources to innovate and expand its offerings. How Does WELL3, $$WELL Work? WELL3 operates by melding cutting-edge technologies in a multichain framework, ensuring a seamless and innovative user experience. Below are some factors that uniquely position WELL3 in the wellness market: 1. Secure Data Ownership With the integration of DePIN and DID, users can maintain complete control over their personal health information. This layer of security is paramount in today's digital age, where data breaches and unauthorized access are rampant. Through WELL3, data ownership is decentralised, enabling users to manage their information proactively. 2. Personalization through AI WELL3 implements AI-driven analytics to provide users with tailored health insights. By harnessing the power of AI, the platform can offer individualised recommendations and solutions, encouraging users to achieve their health goals more effectively. 3. Multichain Framework The WELL3 project is designed to work across multiple blockchain platforms, including Bitcoin, Ethereum, Polygon, Solana, Blast, and TON. This multichain capability ensures that users can interact with the platform seamlessly across different networks, enhancing accessibility and usability. 4. WELL Token Central to the WELL3 ecosystem is the WELL Token, which serves multiple functions including utility, governance, and rewards. The token allows for ecosystem participation, supports health data sharing, and incentivizes users based on their engagement with the platform. Timeline of WELL3, $$WELL The trajectory of WELL3 showcases significant milestones in its development, each contributing to the project's overall success. Here is a brief timeline of critical events in the history of WELL3: February 10, 2024: WELL3 launched its NFT project, quickly rising to prominence as the largest NFT collection on the opBNB chain with over 324,000 owners and reaching 8 million NFTs created by April 27, 2024. Public Sale: The project achieved a remarkable total value locked (TVL) of approximately 15,237.2 ETH within just seven days, indicating robust market interest and backing. WELL ID Launch: The platform saw over 900,000 users sign up for the WELL ID and its corresponding NFT Ring whitelist, marking a significant adoption phase within the ecosystem. Partnership Development: WELL3 established partnerships with leading entities including Animoca Brands, AWS, Samsung, and others to enhance its ecosystem and expand its reach. Transaction Volume: WELL3 has facilitated over $17 million in transactions, reflecting its growing utility and engagement within the health and wellness community. Key Points About WELL3, $$WELL As a progressive initiative shifting towards the wellness market, WELL3 has identified several vital elements that will contribute to its ongoing success. Here are some key takeaways to note: Tokenomics The $$WELL token has a maximum supply of 42 billion, with a significant 71% earmarked for community initiatives. This distribution strategy emphasises the project's commitment to its user base and long-term sustainability. Lock-Up Period To ensure stability within the ecosystem, tokens are released in batches over a 24-month lock-up period, promoting trust and confidence among users. Ecosystem Development WELL3's vision extends toward creating a comprehensive and sustainable ecosystem to encourage thriving community engagement, health-enhancing behaviors, and digital solutions that address the pressing needs of the wellness domain. Market Fit The wellness industry, valued at $5.6 trillion, presents a lucrative opportunity that WELL3 aims to tap into. With an anticipated annual growth rate of 5-10%, the project is ideally positioned amid a rising trend towards health-conscious living. Wearables Introducing the WELL3 Ring, a crypto-incentivized wearable, aligns with the growing demand for personalized health data. This device not only enhances user experience but also redefines what it means to be engaged with one’s health in the context of Web3. Conclusion WELL3 represents a significant advancement in the integration of blockchain technology within the health and wellness sector. By addressing crucial issues around data ownership, personalization, and community engagement, this innovative platform offers a forward-thinking solution to enhance individual well-being. With robust backing from notable investors and a commitment to pioneering technologies, WELL3 stands poised to make a lasting impact in the wellness landscape. For those seeking to navigate the complexities of health in the digital age, WELL3 is certainly one to watch as it continues to evolve and grow.

557 Total ViewsPublished 2024.07.14Updated 2024.12.03

What is $WELL

Discussions

Welcome to the HTX Community. Here, you can stay informed about the latest platform developments and gain access to professional market insights. Users' opinions on the price of WELL (WELL) are presented below.

活动图片