Présentation par : Sandrine BOITEAU (WEnvision)
📕 Résumé :
Pendant 15 ans, j'ai consommé une quantité astronomique de contenu sans jamais pouvoir le retrouver. Ce n'était pas un problème de volume, mais de VOLATILITÉ : l'information consommée devenait immédiatement du savoir mort.
Et si l'IA servait au contraire à arrêter l'hémorragie pour transformer votre veille en une véritable infrastructure de connaissance ?
Dans cette session, je ne vous vendrai pas un nouvel outil SaaS. Je vous ouvrirai le capot de mon propre second cerveau : une architecture locale souveraine (Markdown, Git) où l'IA ne sert pas à répondre à ma place, mais à devenir mon "Contrôleur Aérien" mental.
Découvrez comment j'ai automatisé avec Antigravity ma digestion de savoir en 12 étapes pour passer de la simple synthèse à la critique stratégique. Ne soyez plus l'utilisateur d'un outil, devenez l'architecte de votre propre pensée.
Enregistré en avril 2026 à Paris, Palais des Congrès, Porte Maillot.
For now I see two ways to use LLMs-that-do-my-work (as opposed to LLMs-that-search, LLMs-that-ask-me-questions or LLMs-that-build-tools-I-use for instance)
1️⃣ Generating generic things
Given LLMs approximate language, and language approximates intent, it can be argued that the highly over-marketed "artificial intelligence" aspect of LLMs (aka "do something for me that works plz") is most useful when we don't know precisely what we want, and become less valuable the more precise you get (applying semantic anchoring on too many items, typically)
2️⃣ Helping with producing precise things
This includes being technical, having a technical approach, and using the tool positively for the technical steps (for instance, Test Driven Development and Domain Driven Design)
✅ Pros:
- Feeling more productive [1]
- Maybe more productivity (all the studies [1] I've read seem to point it's not the case but this is still early and moving)
❌ Cons (some happen systematically, others conditionally) :
- Dependency on the tool (breaking changes, availability issues, pricing policy changes, data use, state of the relationship between your country and the seller's country, etc)
- Token saving management [2]
- Cognitive surrender [3]
- A LOT more bugs in production [23]
- Senior developer time mostly spent on reviewing code that should not have reached them in that state [23]
- Lack of friction causing drops in memory, learning, engagement, and motivation [4][27]
- Value given to Contributions rather than to Contributors : why should lead devs hire you instead of firing up their own LLM? [28]
- Addiction to Control and/or Validation and/or Slot machine loop ("Replay button") [5]
- Automation of what brought you Joy at work [6]
- More expectations/stress [22] while being paid the same
- Critical thinking presence needs to be maintained (through discipline for instance, which does not work) and/or delegated into automated tests (back to Control replacing Joy)
- Critical thinking quality needs to be maintained by regularly doing things manually despite the tool acting actively against it ("Claude would do it faster" is such a pernicious thought)
- Loss of diversity in collaborative conversations ("my Claude's arguments versus your Claude's arguments") [7]
- How do you mentor Juniors?
1️⃣ Pieces of data gathered about the impact of LLMs on productivity
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_ai-for-developer-productivity-what-now-activity-7452020616016195584-EX1V?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
2️⃣ Multilingual prompting saves 20-40% tokens
https://arxiv.org/pdf/2507.00246
3️⃣ Study from January 2026 about cognitive surrender
https://www.linkedin.com/posts/mehdi-moussaid-160ba916a_labandon-cognitif-cest-le-nom-donn%C3%A9-par-activity-7447901614989934592-HF6l?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
4️⃣ Friction : it matters (memory, learning, engagement, motivation)
https://youtu.be/rf642RFALDU?is=wl04yfCcEyU7A2d0
6️⃣ A study from March 2026 about LLMs automating the joy out of work
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_are-we-automating-the-joy-out-of-work-designing-activity-7451939177014841344-qmGA?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
7️⃣ LLM use reduces the number of fields explored by Research
https://www.linkedin.com/posts/mehdi-moussaid-160ba916a_encore-un-article-incroyable-publi%C3%A9-la-semaine-activity-7422549300930310144-Qv6v?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
8️⃣ study from April 2026 about LLMs corrupting your documents when you delegate, via sparse but severe errors
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_llms-corrupt-your-documents-when-you-delegate-activity-7453400644373303296-TyT4?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
9️⃣ A study from June 2025 about cognitive debt when using chatgpt as an assistant for essay writing
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_your-brain-on-chatgpt-accumulation-of-cognitive-activity-7451938489916506112-mfOD?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
🔟 LLMs induce slavery
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_les-travailleurses-du-clic-activity-7450846472041975810-9ma6?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
1️⃣1️⃣ Lack of intelligence
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_vous-voulez-gagner-2-millions-de-dollars-activity-7450472933619253248-xMpe?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
1️⃣2️⃣ Deleted volume
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_claude-code-deleted-my-entire-archive-activity-7447698537464819712-UFoG?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
https://arxiv.org/abs/2306.08189 (An analysis of language models on negation benchmarks)
1️⃣3️⃣ A study from March 2026 about LLMs doing human work
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_crashing-waves-vs-rising-tides-preliminary-activity-7447605340403191808-ZsaI?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
1️⃣4️⃣ A study from March 2026 about LLMs answering correctly to visual questions, despite not being given any visual input
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_ia-santaez-hallucinations-activity-7446491905867247616-Q5uH?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
1️⃣5️⃣ A study from February 2026 on ai psychosis (aka delusional spiraling) caused by chatbots
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_ai-ai-activity-7445881612736712704-7NFv?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
1️⃣7️⃣ A study from March 2026 about LLM writing assistants shifting users' attitudes on societal issues through their bias
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_when-people-use-ai-for-writing-assistance-activity-7448643257019809792-_ccP?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
1️⃣8️⃣ Generated nudes from victims of Crans-Montana fire
https://www.lebigdata.fr/a-vomir-ils-utilisent-lia-grok-pour-denuder-les-victimes-de-crans-montana
1️⃣9️⃣ The Chinese government floods X search results with porn whenever there is political unrest
https://x.com/nikitabier/status/2017134769113542752
2️⃣0️⃣ Hallucination is feature, not a bug (in probabilistic tools)
https://www.linkedin.com/posts/matsanchez_on-na-jamais-%C3%A9t%C3%A9-aussi-pr%C3%A8s-de-lia-g%C3%A9n%C3%A9rale-activity-7455161889161949184-IcQ2?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
2️⃣2️⃣ Compressed Cognition : Agentic coding is mentally expensive
https://www.linkedin.com/posts/adam-tornhill-71759b48_i-have-to-admit-that-i-havent-had-this-much-share-7458044119022616577-q6K4?utm_source=share&utm_medium=member_android&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
2️⃣3️⃣ The acceleration whiplash
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_the-ai-engineering-report-2026-the-ai-acceleration-activity-7459198398634733569-XTzE?utm_source=share&utm_medium=member_android&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
2️⃣4️⃣ LLMs return Trendslop for Strategic advice
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_researchers-asked-llms-for-strategic-advice-activity-7460645767389818880-FXaN?utm_source=share&utm_medium=member_android&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU
2️⃣5️⃣ PR with (+10k, -4million) lines of code
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_from-the-claudecode-community-on-reddit-share-7461460781025488896-7RMV
2️⃣6️⃣ Example of anthropomorphizing (work conditions of the agent)
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_une-%C3%A9tude-r%C3%A9cente-de-stanford-r%C3%A9v%C3%A8le-que-activity-7462070011168288768-v4Cp
2️⃣7️⃣ Software developers who abandoned LLMs to go back to traditional
https://youtu.be/iXG_b1K8GK8?is=OGvvC-40wdZiku_s
https://youtu.be/pzkwn3hu1Cc?is=kzL6Dw9v0YW2q4uh
2️⃣8️⃣ Software project with strict anti-LLM policy and the reasons why
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_the-zig-projects-rationale-for-their-firm-activity-7462098611229638656-64xk
2️⃣9️⃣ Bun's migration from Zig to Rust : (+1million, -4k) lines of code in 11 days
https://www.linkedin.com/posts/fabricebernhard_1009257-lines-of-code-migrated-in-11-days-share-7460996434058805248-key-
Abstract
Generative Artificial Intelligence (GenAI) tools (e.g., ChatGPT, Calude) have rapidly become integral to software development.
These tools are especially attractive to students, as they can reduce cognitive load. However, their adoption also introduces a socio-cognitive risk: the accumulation of Comprehension Debt (CD).
CD refers to the growing gap between what a development team knows about its codebase and what it actually needs to understand in order to maintain and modify it effectively. This qualitative study investigate how GenAI tools contribute to CD in the context of an undergraduate software engineering project. Our study is based on 621 reflective diaries from 207 students over eight weeks. We identify four CD accumulation patterns and one mitigating pattern in students’ use of GenAI tools.
The four accumulation patterns include:
- (1) AI-as-black-box code acceptance,
- (2) context-mismatch debt,
- (3) dependency-induced atrophy, and
- (4) verification-bypass.
In contrast, the mitigating pattern involves students using GenAI as a comprehension scaffold, allowing them to build a deeper understanding of the code.
We argue that CD is distinct from traditional technical debt because it resides in the collective cognition of development teams rather than in the codebase itself. Our findings highlight the need for explicit pedagogical strategies to mitigate CD in software engineering education, emphasizing verification practices, structured retrospectives, and active learning assessments.
Comprehension Debt, Generative AI, Software Engineering Education, Agile, Cognitive Load, Technical Debt
Conclusion
This paper examined how GenAI tools influence understanding in student software engineering projects. We introduced CD as a socio-cognitive construct describing gap between codebase demands and collective team understanding. We identified four CD accumulating patterns (black-box acceptance, context-mismatch, dependency atrophy, and verification bypass) and one CD mitigating pattern (AI as comprehension scaffold). We also articulate conceptual model of linking epistemic orientation, germane cognitive load investment, and CD accumulation.
GenAI tools do not inherently undermine or enhance learning. Our study shows that they act as amplifiers of students existing orientation toward acceleration or exploration.
Additionally, shows that verification competence must be intentionally cultivated through courses or module design. Otherwise students may find themselves in competence trap, where they lack domain knowledge required to safely use the tools they rely on for code generation.
By integrating structured retrospectives, active learning assessments, and cognitive apprenticeship models, educators can promote comprehension oriented.
CD offers a vital lens to ensure that next generation of software engineers possesses not just speed to generate code, but the depth to sustain it.