Une requête Gemini consomme 33 fois moins qu'il y a un an. Google consomme 3,5 fois plus d'électricité qu'en 2019.
Attention, exemple d'effet rebond réservé aux initiés. Mouillez-vous la nuque avant de lire, ça peut piquer un peu.
Prêts ?
Google a publié mardi son rapport environnemental 2026. On y lit qu'une requête Gemini pèse désormais 0,24 Wh, une énergie par prompt divisée par 33 en un an. Top.
⚡ Quelques pages plus loin, la consommation électrique totale du groupe : 43 TWh en 2025. Elle était de 31 TWh un an plus tôt. De 12 TWh en 2019.
Multipliée par 3,5 en six ans.
🍃 Les émissions de GES ? Elles ont presque doublé selon la méthode de calcul depuis 2019.
💧 L'eau ? x2 en 4 ans.
La frugalité agit sur les moyens là où la sobriété agit sur le besoin. Google est devenu remarquablement frugal, requête par requête.
Le besoin, lui, n'a jamais été interrogé : des aperçus IA qui s'affichent sans qu'on les demande, de la génération vidéo, des agents qui tournent en continu. Optimiser chaque brique tout en multipliant les briques... à la fin c'est l'effet rebond qui gagne.
L'analyste Ketan Joshi pose la question absente du rapport : si l'infrastructure IA croît plus vite que les réseaux ne se décarbonent, faut-il continuer à la construire ?
Combien de feuilles de route IA reposent sur la même mécanique, en plus petit ?
Comment un modèle ouvert de 35B (Ornith 1.0, MoE), piloté par l'agent open source pi sur mon infra, a écrit, testé et corrigé un vrai projet TypeScript. Verdict : le gap avec les modèles frontière se resserre, vite.
The $20 AI subscription you use every day is a mathematical impossibility. While you enjoy flat-rate access to advanced tools like Claude Code, the raw compute costs for a power user can actually skyrocket to $15,000 a year. We are officially living inside the "AI Uber Moment", a temporary illusion completely subsidized by venture capitalists who are bleeding billions just to get you hooked. But as tech giants resort to hidden financial loops and "stealth-nerfing" your favorite chatbots to survive, the money is officially running out.
Chapters:
- 00:00 - The Trillion Dollar House of Cards
- 00:36 - Chapter 1: The $20 Illusion
- 02:38 - Chapter 2: The Ghost Of Uber
- 05:03 - Chapter 3: The Claude Code Math
- 05:56 - Chapter 4: The Search Penalty
- 08:08 - Chapter 5: The Round Trip Scam
- 09:35 - Chapter 6: The Hardware Debt Trap
- 11:58 - Chapter 7: The Stealth Nerf
- 13:21 - Chapter 8: The 2026 Mass Extinction
- 15:02 - Chapter 9: The Great AI Rug Pull
Narrated by: Josh Risser
The past 3 years of work in NLP have been characterized by the development and deployment of ever larger language models, especially for English. BERT, its variants, GPT-2/3, and others, most recently Switch-C, have pushed the boundaries of the possible both through architectural innovations and through sheer size. Using these pretrained models and the methodology of fine-tuning them for specific tasks, researchers have extended the state of the art on a wide array of tasks as measured by leaderboards on specific benchmarks for English. In this paper, we take a step back and ask: How big is too big? What are the possible risks associated with this technology and what paths are available for mitigating those risks? We provide recommendations including weighing the environmental and financial costs first, investing resources into curating and carefully documenting datasets rather than ingesting everything on the web, carrying out pre-development exercises evaluating how the planned approach fits into research and development goals and supports stakeholder values, and encouraging research directions beyond ever larger language models.
For this online event, we will have the opportunity to receive Dragan Stepanović as a speaker.
After a brief introduction of our community, Dragan will explain the Systems Perspective of the LLM-assisted coding.
Abstract
This will not be your usual "all roses" talk on GenAI nor one that hijacks your amygdala with claims such as "Most developers are going to be replaced by coding agents in 9-12 months!" or "If you don't jump on the train, you're going to be left out!".
For product development teams, organizations, and their customers, every technology they want to adopt, however innovative it is or may seem, doesn't operate in isolation, but as part of a broader system. Ignoring this fact is likely to make things worse instead of achieving the promised huge productivity boost, because any change in a system affects its dynamics in a way that feeds back to affect that change in turn. Some parts of the dynamics accelerate, some start pushing back. It's becoming increasingly important to take a systems perspective on attempts to adopt a technology and understand what desired, but even more importantly, unintended consequences it's likely to cause.
I'll be diving into topics that are likely to stir the pot with uncomfortable, but important questions that a growing number of teams and organizations are facing on this journey. Great, we have this heavy machinery that can produce so much more code in a unit of time, but what do we do with all the piled-up inventory of unreviewed code? What about comprehension debt? How do we keep the ability to reason about the system? Can we replace the process of creating and evolving a product with its output (code)? LLM-assisted vs agentic coding, and which makes sense in which context?
This technology and its adoption are still in their infancy and we're operating in an uncharted territory, so no one really knows yet all the effects that will play out, but looking at it through Systems Thinking, Lean, Theory of Constraints, and XP lenses can provide a useful level of foresight into the distribution of outcomes that might play out.
About Dragan
Dragan is based in Berlin and as a principal engineer helps companies evolve their engineering culture, tame their bottlenecks, and maximize their throughput of the value.
Typically, in search of better ways of working, exploring ends of the spectrum, and helping teams and organizations try out counter-intuitive ideas that initially don't make a lot of sense, but surprisingly end up as completely opposite of that.
He enjoys endless discussions connecting XP, Theory of Constraints, Systems Thinking and Lean.
LLMs have been trained on decades of freely available Java specifications, including JSRs, JEPs, MicroProfile and Jakarta EE. They know the patterns. They know the standards. Now let's put them to work. In this live coding session, we'll use LLM agents to build production-ready Java applications quickly. No slides, no theory - just real code, real prompts, and real results. We'll start with a typical enterprise requirement and demonstrate how to guide LLMs to generate clean, maintainable Java code following BCE/ECB architecture patterns that actually work in production. You will learn how to effectively access the LLM's in-depth knowledge of Java specifications, how to continuously improve the generated code with each iteration, and how to maintain a high velocity without creating a mess. Expect live coding, real-world scenarios from actual projects, and honest discussion about where LLMs excel and where human expertise remains crucial. Bring your questions - we'll solve them with code.
Guidelines for Agentic development.
"Put your AI on rails."
BCE architecture · Java · Web standards
Adam Bien
AI coding agents frequently ignore long rule documents. Asking them to hold on to an entire book's worth of coding advice is at best futile, at worst makes the agent's performance worse by polluting the context window.
Humans don't need to hold the same information in their head because humans can form habits through repetition. However, AI agents can't do this.
Human habits form when an easy-to-detect cue triggers a complex sequence of actions with the desired effect. This is the inspiration for habit hooks.
Linters provide a deterministic metric, but Goodhart's law postulates that a metric ceases to be a good metric if it becomes a target. AI agents are very good at gaming these metrics when they are only provided the metric.
Habit hooks wraps your linter to create the trigger, but instead of providing only the metric, it gives actionable advice on how to fix the issue. This creates AI behaviour that looks like human habits, and has similar effects.
The use of habit hooks:
- Increases code quality
- Improves AI performance ensuring that the AI always starts with good code quality
- Reduces token usage, since good quality code also means the AI doesn't need to read as much context to complete the task.
Clara explores whether learning to code is still worthwhile, offering crucial insights into the learning process and the future of programming.
Clara explains how to use AI effectively when learning to code or improving your programming skills, discusses the potential drawbacks of the concept “AI as a tutor” and its impact on the learning experience, both practically and emotionally. 📃 Read the research this video is based on about the "widening gap" here: https://dl.acm.org/doi/10.1145/363262...
- 0:00 what’s up with AI and learning?
- 0:32 the experiment
- 0:58 pssst... check out our student pack
- 1:13 student A (struggling)
- 2:11 unawareness of unawareness
- 2:26 student B (non-struggling)
- 3:36 the widening gap
- 4:23 a rule of thumb 👍
- 4:39 doing things slowly is better for learning
- 4:48 imposter syndrome
- 4:56 meh.
- 5:43 🤑getting things done
- 6:09 the good old days
- 6:36 how to use LLMs in learning
- 6:54 what have we learned?
- 7:13 the psychology of learning to code
- 7:20 please… please... the student pack
- 7:32 begging and conclusion
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.
Agents fait pour accompagner le développement de produits accessibles.
Fait par des gens avec déficience visuelle.
For decades, ADHD has been framed as a disadvantage—a condition that makes it harder to focus, stay organized, and succeed in a world built on routine.
But what if we've been looking at it all wrong?
As artificial intelligence transforms the way we work, create, and solve problems, the very traits once considered weaknesses—curiosity, rapid idea generation, pattern recognition, hyperfocus, and unconventional thinking—may become some of the most valuable skills of the future.
In this video, we explore why the rise of AI could fundamentally change the way we think about ADHD. We'll examine the science behind divergent thinking, the entrepreneurs who turned their neurodivergence into an advantage, and how AI is becoming a "second brain" that helps bridge the gap between imagination and execution.
Could the ADHD brain be uniquely suited for the AI age?
Maybe the problem was never ADHD.
Maybe the world just hadn't caught up yet.
In This Video:
• Why traditional work environments disadvantage ADHD minds
• How AI changes the rules of the game
• The hidden strengths of the ADHD brain
• Why curiosity may become more valuable than discipline
• Entrepreneurs who embraced their neurodivergent thinking
• How people with ADHD can use AI as a competitive advantage
💬 What do you think?
Do you believe ADHD will become an advantage in the age of AI, or will the future create even greater challenges for neurodivergent people?
Share your thoughts in the comments. I'd love to hear your perspective.
If you enjoy exploring the future of AI, emerging technologies, and the ideas shaping humanity's next chapter, consider subscribing to DeadLock.
Because the future isn't just happening to us.
It's being built right now.
Small utility to display environmental impact of claude - based on ecologits.ai
Tech sector says only carbon-emitting gas plants are reliable enough today to power the EU’s AI goals.