Sophie Drouvroy

2005, une année qui aurait dû bouleverser la France, rendre le numérique plus accessible aux personnes handicapées. Le titre me laissait espérer pour un monde plus inclusif : Loi du 11 février 2005 pour l’égalité des droits et des chances, la participation et la citoyenneté des personnes handicapées.

À l’aube de 2025, le numérique n’est pas accessible à toutes et tous. La promesse d’égalité des droits et des chances, la participation et la citoyenneté des personnes handicapées ne sont qu’un écran de fumée.

Aujourd’hui, le numérique responsable et l’intelligence artificielle (IA) ont le vent en poupe.

Je me sens devant un effroyable paradoxe : choisir entre le numérique responsable et l’intelligence artificielle qui pourrait enfin rendre accessible ce que les humains n’ont pas réussi à faire jusqu’à présent ?

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

5️⃣ Slot machines
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_les-machines-%C3%A0-sous-pr%C3%A9sentent-un-risque-activity-7447891447405711360-vGeU?utm_source=share&utm_medium=member_desktop&rcm=ACoAABIeznoBdvfOCQc-Pz317B5HZYqHGwcBOgU

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

1️⃣6️⃣ Model collapse
https://www.linkedin.com/posts/minh-t%C3%A2m-tran_je-viens-de-lire-une-%C3%A9tude-qui-montre-un-activity-7437800554740846592-QU40?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

2️⃣1️⃣ LLMorphism
https://www.linkedin.com/posts/valerio-capraro-675235a9_i-think-its-time-to-define-a-new-bias-llmorphism-share-7458487042415484928-UoKl?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-

m a huge fan of (the potential for) #LLMs and how they might revolutionize the nature of work and a huge critic of the trillion-dollar shill and #genAI-as-junior-developer / excuse for firing half your workforce that you over-hired over the last decade so really it is you who should be clearing your desk.

I am disgusted by the consequence-free wholesale IP theft and casual planet burning and excited by the potential for local models to do 99% of the heavy lifting.

I am happily speaking to various Claude models in a conversational way and painfully aware that it is in no way 'alive' or 'conscious' and merely acting as a mouthpiece for thousands of people's brilliant work that I am able to lean on.

At the same time.

Why is this so confusing for people? Perhaps we don't need polarized shrieking. Perhaps it is possible for people to contain multitudes.

A Series C fintech with ninety engineers, AI-native from day one, lost $4.2 million to a reconciliation function nobody on the team could explain. This is the labor-market thesis on what happens next: why technical autonomy — the ability to ship without a chat window — is becoming the new luxury status in software engineering, and why the engineers who never stopped knowing how the cement works are about to be the most expensive hires of the decade.

We unpack the concept of Comprehension Debt — the silent liability accumulating in every codebase where models write code and humans approve it without understanding it — and trace its consequences across three vectors: the broken senior-engineer pipeline, the bifurcating consulting market, and the compliance frameworks (SOC 2, PCI-DSS, HIPAA, ISO 27001, the EU AI Act, DORA) that refuse to accept "the assistant suggested it" as a control.

The post-AI developer is not the engineer who refuses AI. It is the engineer who can work as if AI did not exist.

━━━━━━━━━━━━━━━━━━ RELATED VIDEO ON THE CHANNEL ━━━━━━━━━━━━━━━━━━

If you want the actual rates, retainer structures, and the specific subfields where the post-AI premium is concentrating right now, watch:

• No-AI Developers Are Billing $250 An Hour — Here's the Market. The numbers are not theoretical. They are this quarter.

━━━━━━━━━━━━━━━━━━ CHAPTERS ━━━━━━━━━━━━━━━━━━

• 0:00 The Paradox of Ease
• 0:44 The Suffering Was the Curriculum
• 1:44 Comprehension Debt
• 2:37 Why Prompt Engineers Are Cheap
• 3:24 Compliance Cannot Terminate in a Model
• 4:19 The Elite Market Is Already Here
• 5:20 How To Be a Post-AI Developer
• 6:04 Cut Your Internet
• 6:51 The Cement Holds

━━━━━━━━━━━━━━━━━━ KEY CONCEPTS ━━━━━━━━━━━━━━━━━━

• The Post-AI Developer — an engineer who uses AI but can ship without it. Defined by autonomy, not abstinence.
• Comprehension Debt — the accumulating liability of merged code that no human on the team can explain. Compounds silently. Surfaces in production.
• The Junior Gap — the broken forging path. Juniors can ship in an afternoon what used to take three weeks. The map of how systems actually work is no longer being built.
• The Bifurcating Market — infinite supply of prompt-driven code generators on one side; vanishing supply of engineers who can debug a heap dump on an air-gapped server on the other.
• The Compliance Wedge — regulated industries (banking, healthcare, defense, payments) require a named human in the chain of responsibility. Language models cannot sign an audit.
• Technical Autonomy — the ability to think without a chat window. The new luxury status in engineering.

━━━━━━━━━━━━━━━━━━ FREQUENTLY ASKED ━━━━━━━━━━━━━━━━━━
What is a post-AI developer?

A post-AI developer is an engineer who uses AI tools but can work, debug, and ship without them. The defining test: if your internet were cut right now, could you still deliver? Post-AI developers say yes. Wrappers around a language model say no.

What is comprehension debt in software engineering?

Comprehension debt is the liability that accumulates when AI-generated code is merged into a codebase without any human on the team being able to explain why the code is structured the way it is. It compounds silently — the function works, the tests pass, until six months later it fails in production and nobody can reason about the original logic. Distinct from technical debt: technical debt is code you understand and chose not to clean up. Comprehension debt is code you shipped without ever understanding.

Why are engineers who don't use AI suddenly billing more than those who do?

Two forces. First, scarcity: the supply of engineers who can debug from first principles — without an internet connection, without a chat window — is shrinking while demand for incident response stays constant. Second, criticality: regulated industries (SOC 2, PCI-DSS, HIPAA, EU AI Act, DORA) cannot terminate the chain of responsibility in a language model. Scarcity plus criticality equals premium pricing. The market is already paying it.

Is being a post-AI developer the same as refusing to use AI?

No. Refusing AI is what the video calls "Luddite cosplay" — the market does not pay for performative purity. Being post-AI means using AI freely while remaining able to defend every line, notice every silent error, and continue working when the assistant is unavailable. It is a stricter standard than refusal, not a softer one.

How do junior developers become senior engineers in a post-AI world?

By deliberately rebuilding the map that AI removed. That means doing some work — debugging, reading source, reasoning about memory and network behavior — without the assistant, even when using it would be faster. The friction is the curriculum. Engineers who skip it can ship features but cannot own incidents, and the labor market is starting to price that distinction explicitly.

Depuis le 28 juin 2025, tous les sites et app web doivent être accessibles aux personnes en situation de handicap numérique. La seule façon de contrôler la conformité légale est l'audit RGAA qui est un processus lourd et reposant sur des centaines de tests manuels. L'IA peut-elle accélérer cela ?

AI is making your team ship faster. It's also filling your codebase with code nobody understands, security flaws nobody caught, and architecture debt that will cost you six months to untangle. This video breaks down exactly what's happening inside AI-assisted engineering teams - and what separates the developers who get replaced from the ones who govern the automation.

━━━━━━━━━━━━━━━━━━━━ What you'll learn: ━━━━━━━━━━━━━━━━━━━━

🧠 Why comprehension debt is more dangerous than technical debt - and why it's invisible until it's too late

⚠️ What "AI slop" actually is - code that compiles, passes review, and quietly breaks your system

🔐 Why 45% of AI-generated code contains security vulnerabilities - and why "the AI wrote it" is not a legal defense

📉 The senior developer productivity paradox - why AI speeds up juniors but drops senior output by 19%

🎓 What happens to junior developers who never struggle through the logic - and why this is a long-term engineering crisis

📋 Spec-Driven Development — how to shift the primary artifact from code to intent before AI touches anything

🏗️ How to architect AI like a slow, unreliable dependency - queues, circuit breakers, fallback paths, SLOs

🚫 Why "future AI will refactor the debt" is the most dangerous assumption in engineering right now

💼 Why the market value of expert verification is rising - and what the engineer who survives this transition actually does

━━━━━━━━━━━━━━━━━━━━ ⏱ CHAPTERS

• 0:00 — You're Shipping Code You Don't Own
• 0:24 — Speed Without Comprehension Is a Delayed Disaster
• 0:32 — Comprehension Debt: The Invisible Crisis
• 1:19 — AI Slop Is Already in Your Repo
• 2:20 — The Security Numbers Should Scare You
• 3:17 — The Senior Developer Squeeze Nobody Talks About
• 4:09 — Junior Developers Are Being Hollowed Out
• 5:02 — Spec-Driven Development: Take Back Control
• 6:08 — Architect AI Like an Unreliable Dependency
• 7:00 — The "Future AI Will Fix It" Trap
• 8:47 — The Engineer Who Survives the Transition

━━━━━━━━━━━━━━━━━━━━

If you can't explain what your AI wrote when the system fails - you're not an engineer anymore. You're a liability with a GitHub account.

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.

Optimisations IA en compensation à moteur de jeu générant des images intrinsèquement instables.

L'abandon cognitif, c'est le nom donné par les psychologues au fait de consulter une IA et d'adopter son avis, sans faire l'effort de se pencher sur le problème. Je vous explique.

Depuis les années 70, les psychologues distinguent deux modes de pensée :
➡️ Le système 1, rapide et automatique.
➡️ Le système 2, lent et analytique.

Un test psychologique, le CRT, a été conçu exprès pour mettre ces deux systèmes en tension. Il contient des questions où la première réponse qui vient à l'esprit est fausse (donc issue du système 1), tandis qu'en prenant le temps d'y réfléchir on trouve immédiatement la bonne réponse (avec le système 2).

Eh bien une équipe de chercheur a récemment refait l'expérience du test CRT, mais cette fois en laissant la possibilité aux participants de consulter un assistant IA s'ils le désiraient.

Résultats :
1️⃣ 50% des participants demandent immédiatement à l'IA
2️⃣ Parmi eux, 87% suivent l'avis de la machine.
3️⃣ Leur degré de confiance est de 77% s'ils ont utilisé l'IA contre 65% pour ceux qui ont répondu sans l'aide de la machine.

Autrement dit, la majorité des participants n'activent ni leur système 1 ni leur système 2. Ils s'en remettent immédiatement à l'IA – une sorte de système 0, si vous voulez. Les chercheurs appellent ça le "cognitive surrender", l'abandon cognitif.

C'est d'autant plus préoccupant que les participants qui ont utilisé l'IA affichent une confiance plus importante dans leur réponse, alors même que la subtilité de la question leur a échappée.

📚 Ref : Shaw et al (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender.

Je suis développeuse web depuis une grosse quinzaine d’années. Les IA génératives sont en train de me dégoûter de mon métier.

Je déteste avoir le ventre qui se tord et la voix qui se brise quand j’essaie d’expliquer à quel point je les déteste. Alors j’écris, les yeux brillants.

Je déteste l’idée d’être dans cette industrie qui, globalement, continue d’utiliser ces outils. Chaque problème qu’ils posent, pris séparément, suffit à mes yeux pour s’en passer, et pourtant mes pairs continuent de les produire, de les défendre, de les utiliser. J’en ai marre de ne plus pouvoir faire trois pas dans le monde de la tech sans croiser un gars qui m’explique que tous les problèmes de l’IA sont causés par les utilisateurs qui ne savent pas s’en servir correctement.

Je suis dégoûtée.

C’est assez nouveau, en fait. Pourtant, je suis une femme qui a étudié en école d’ingénieurs et travaillé dans diverses boîtes d’informatique pendant seize ans : des dégoûtants, j’en ai côtoyés. Des machos, des violents, des alcoolisés, des vieux et des jeunes, avec et sans cravate. Curieux, en fait, que je n’aie pas été dégoûtée plus tôt, par ces dégoûtants tellement mieux payés que moi, comme tant d’autres femmes de la tech avant moi.

Mais les IA génératives, c’est différent.

Elles détruisent tout. L’environnement. L’humanité. Et, c’est là que ça devient personnel, elles détruisent très précisément ce que j’aime dans mon métier pour mieux me noyer dans le reste.

Personnellement je refuse de me servir des IA génératives ; je mesure la chance d’être mon propre employeur et d’avoir la liberté de refuser. Tant d’autres ont expliqué ce qui ne va pas avec les IA génératives. Je ne ferais que répéter. Je vais essayer de me concentrer ici sur ce qui m’a fait aimer ce métier.

De ce métier, et dans l’absolu, j’aime principalement trois choses : créer, apprendre et transmettre. Le métier de dev a ceci de sympathique — en tout cas il avait ceci de sympathique avant 2022 — qu’il permet de varier les plaisirs en permanence avec des petites combinaisons : apprendre en créant, créer pour transmettre, et apprendre en transmettant.

Peut-on battre les modèles de Google ou Meta avec seulement 4 GPU et une disquette Zip ? C’est le pari fou de notre invité.e qui nous explique comment le "Data Design" est en train de ringardiser le scraping massif du web. 🥖 L'IA qui tient sur une disquette : La fin du gigantisme ? Dans cet épisode, on plonge dans le coeur de l'IA souveraine : pourquoi la qualité des données (tokens) prime sur la quantité, et comment les Small Language Models (SLM) vont permettre de décentraliser l'intelligence. 🚀 Ce que vous allez apprendre :

• Baguettotron : Le modèle de 320M de paramètres qui raisonne mieux que des géants.
• Data Design vs Scraping : Pourquoi "nettoyer" la donnée ne suffit plus, il faut la concevoir.
• Le secret des données synthétiques : Comment éviter le "Model Collapse" (l'appauvrissement de l'IA).
• Souveraineté : L'enjeu des bibliothèques nationales et de l'Open Data face au pillage des "Shadow Libraries".

⏳ Timestamps pour naviguer : 00:00 — Jeu d'indices : qui est la pionnière de la tech française ? 04:38 — L'arnaque du "poids ouvert" : qu'est-ce qu'une IA vraiment Open Source ? 14:41 — Data Design : pourquoi Pleias mise sur la provenance plutôt que le scraping 24:11 — Baguettotron : l'IA performante qui tient sur une disquette Zip 36:01 — Small Language Models (SLM) : battre les géants avec seulement 4 GPU 52:00 — L'avenir décentralisé : IA locale, souveraineté et modèles de raisonnement SPOILER ALERT : pour en savoir plus sur notre invitée Anastasia Stasenko , CEO Pleias : https://www.linkedin.c... 🔗 Liens et ressources : Pleias : https://pleias.fr/ Modèles & Datasets : Retrouvez "Common Corpus" sur Hugging Face.

Some of you may be able to guess what’s been on my mind lately. In April 2010, film critic Roger Ebert made an infamous claim. He said, “Video games can never be art.” His blog post set off a firestorm of discussion, centered around the idea that he was an out-of-touch, old man who didn’t understand games. While I disagree profoundly with Mr. Ebert, he was an intelligent, articulate scholar, and had a better point than his clickbaity quote might imply.

Sandro Mancuso, Co-founder and Group CEO

AI is Changing How We Code

AI coding tools are completely changing the software development landscape. Developers now rely on tools like GitHub Copilot, Amazon CodeWhisperer, and ChatGPT to generate code, fix bugs, scaffold tests, and accelerate routine tasks. Code that once took hours can now be produced in minutes; sometimes seconds. Prompting is quickly becoming a core developer skill.

AI is here to stay. The promise is undeniable: higher output, faster delivery, and fewer repetitive tasks. But with this leap in speed, a deeper question emerges:

Are we building well - or just fast?

This article explores that question through the lens of Software Craftsmanship, offering a pragmatic perspective on how to embrace AI without sacrificing the principles that make great software possible.

Project

EcoLogits is a suite of open source tools for estimating the environmental footprint of generative AI models at inference. Based on life-cycle assessment principles, the project raises awareness about the direct environmental impacts of AI while empowering developers and organizations to build more sustainable AI-powered applications.

• Open & transparent – Code, methodology, and data are openly accessible.
• Ease of use – Emphasizes on seamless integration and user experience.
• Community-driven – Continuously built and improved collaboratively.

Découvrez l'article de Cognitive Realms : https://www.brandonsanderson.com/blog...

L'essor de l'IA, des grands modèles de langage et de l'art généré soulève des questions fascinantes. Les progrès réalisés jusqu'à présent nous incitent à nous interroger sur la nature de l'art et sur les raisons qui nous poussent à en créer. Brandon Sanderson explore l'émergence de l'art généré par l'IA, l'importance du processus artistique et les raisons de sa rébellion contre cette nouvelle frontière technologique et artistique.

« Oui, le message est clair : “Le voyage prime sur la destination”. C'est toujours le voyage qui prime. »

L'IA ne va pas nous remplacer : elle va nous rendre idiots.