Documents concernant les effets des IA génératives, que ce soit leur entraînement, leur usage ou la construction et l'entretien des centres de données.
","path":["Accueil"],"tags":[]},{"location":"2024-09-16-001/","level":1,"title":"2024 09 16 001","text":"« Un vaste système de surveillance piloté par l'IA peut amener les citoyens à un meilleur comportement », d'après Larry Ellison
Note
Larry Ellison suggère d’adopter le modèle chinois de surveillance des citoyens
« Un vaste système de surveillance piloté par l'IA peut amener les citoyens à un meilleur comportement », d'après Larry Ellison qui suggère d'adopter le modèle chinois de surveillance de ses citoyens Patrick Ruiz 9–12 minutes
« Un vaste système de surveillance piloté par l’IA peut amener les citoyens à un meilleur comportement », d’après Larry Ellison Qui suggère d’adopter le modèle chinois de surveillance de ses citoyens
La Chine est le pays qui, à date, incarne le mieux le personnage de fiction Big Brother du roman Nineteen Eighty-Four (1984) de l'écrivain anglais Georges Orwell. Dans la société décrite par ce dernier, chaque citoyen est sous la surveillance constante de l’autorité (Big Brother), principalement par des dispositifs (dans les domiciles privés) qui fonctionnent à la fois comme des télévisions, caméras de sécurité et microphones. Ceci rappelle sans cesse au peuple que Big Brother les observe : « Big Brother is watching you. » Larry Ellison, CTO d’Oracle, est d’avis qu’un tel tableau offre l’avantage de pouvoir amener les citoyens à un meilleur comportement. Pour ou contre ce positionnement ?
Le cas de la pose des caméras avec intelligence artificielle sur les autoroutes est susceptible de donner raison à Larry Ellison
La mesure peut en effet induire un meilleur comportement des individus en société dans le cas où elle est implémentée pour aller à la recherche d’incivilités : conducteurs qui jettent les ordures sur la route, tiers qui roulent sur la bande d'arrêt d'urgence pendant les bouchons, personnes utilisant leurs téléphones portables en conduisant, etc.
C’est une mesure que les autorités routières d'Angleterre ont mises sur pied dans le cadre d’un essai. Des caméras avec intelligence artificielle ont été installées sur des autoroutes afin de repérer les cas d'abandon de détritus par les possesseurs de véhicules.
La sortie du PDG d’Oracle suggère que la Chine, et son adoption en masse des solutions d’intelligence artificielle pour des cas similaires, constitue un exemple
De nos jours en effet, la Chine est reconnue comme leader mondial en matière d’adoption des technologies de reconnaissance faciale. Les autorités s’en servent pour une panoplie d’usages. Par exemple, la police chinoise s’en sert pour scanner les foules à la recherche de suspects. C’est en s’appuyant sur cette technologie qu’elle a pu mettre la main sur un fugitif. L’homme de 31 ans était recherché pour des crimes économiques sur lesquels les autorités n’ont pas fait dans le détail.
La Chine c’est aussi son système de crédit social qui s’appuie sur quelque 200 millions de caméras (aux capacités améliorées par l’intelligence artificielle) postées en divers lieux publics : traversées de feux de circulation, magasins, etc. C’est un système de notation qui attribue des scores de confiance aux personnes physiques et morales sur la base de données à la disposition du gouvernement. Par exemple, si un citoyen paie ses factures à temps, fait du bénévolat ou gère correctement son recyclage, il voit son score grimper et obtient des récompenses telles que des prix de transport public ou même des temps d’attente dans les hôpitaux revus à la baisse. En revanche, s’il est coupable d'infractions comme la mauvaise conduite routière, le tabagisme dans les espaces non-fumeurs, l'achat d'un trop grand nombre de jeux vidéo et l'affichage de fausses nouvelles en ligne, il est puni par la chute de son score de confiance. C’est en s’appuyant sur ce système que le gouvernement chinois peut par exemple interdire à 23 millions de personnes (mal notées selon la grille de son système de crédit social) d’acheter des billets de voyage.
Les autorités du pays ne cessent de revoir les capacités de leurs outils de surveillance de masse à la hausse. Alors que le mois de septembre 2019 s’achevait, elles ont dévoilé une caméra de 500 mégapixels aux capacités boostées à l’intelligence artificielle.
Depuis décembre 2019, les personnes qui veulent faire l’acquisition d'un nouveau numéro de téléphone et profiter des services de données doivent accepter que les fournisseurs de services de télécommunication numérisent leurs visages. Le ministère de l'Industrie et de la Technologie de l'information explique que la démarche fait partie de ses efforts pour « protéger les droits et les intérêts légitimes des citoyens dans le cyberespace » et pour lutter contre la fraude par téléphone et par Internet. En plus du test de reconnaissance faciale, il est également interdit aux utilisateurs de téléphones de transmettre leur téléphone portable à des tiers et ils sont encouragés à vérifier si des numéros de téléphone sont enregistrés sous leur nom sans leur consentement.
Cette mesure constituait une mise à niveau du système d’enregistrement des noms réels de la Chine pour les utilisateurs de téléphones mobiles mis en place en 2013. Ce dernier imposait aux personnes de soumettre leurs papiers d'identité et d'autoriser les opérateurs à prendre une photo d'identification pour obtenir un nouveau numéro. L’étape de reconnaissance faciale fait correspondre l’image à l’identité stockée de la personne.
Il ne faut néanmoins pas perdre de vue que le positionnement de Larry Ellison est celui d’un acteur de la filière de la surveillance de masse
« Oracle a violé la vie privée de milliards de personnes à travers le monde. Il s'agit d'une entreprise classée au Fortune 500 qui a pour mission dangereuse de suivre les déplacements et les activités de chaque personne dans le monde. Nous intentons cette action pour arrêter la machine de surveillance d'Oracle », souligne Le Dr Johnny Ryan, membre senior de l'Irish Council for Civil Liberties, dans le cadre d’une plainte contre le géant technologique pour mise en place et entretien d’un système de surveillance mondiale.
Cette plainte contre Oracle fait suite à la publication de rapports selon lesquels Oracle commercialise des logiciels pour permettre à la Chine de parfaire la surveillance de sa population. Le rapport, publié en 2021 par Mara Hvistendahl du magazine en ligne d’investigation The Intercept, accuse Oracle d'avoir aidé la police chinoise à collecter une montagne d'informations sur les Chinois, alors qu'il y a peu, l'administration Trump forçait encore ByteDance à vendre l'activité américaine de TikTok par crainte que l'application de partage de vidéos courtes transmette les données des Américains à Pékin. La police de la province du Liaoning en Chine serait assise sur des monticules de données collectées par des moyens invasifs : dossiers financiers, informations de voyage, immatriculations de véhicules, etc.
Hvistendahl cite également des sources de données comme les médias sociaux et les images de caméras de surveillance. Toutefois, avoir de telles informations sans pouvoir les traiter et en faire usage paraît totalement inutile, et c'est à ce moment qu'Oracle entre en jeu. Pour donner un sens à tout cela, la police chinoise avait besoin de logiciels d'analyse sophistiqués et Oracle lui en aurait au moins quatre, dont un lui permettrait de \"faire des analyses et des prévisions criminelles\". Un autre des logiciels lui permettrait de \"créer des graphiques en réseau basés sur les enregistrements des hôtels\".
Il permettrait également de traquer toute personne qui pourrait être liée à un suspect donné. Ensuite, un troisième logiciel d'Oracle permettrait à la police chinoise de construire un tableau de bord et de créer des \"cartes thermiques des affaires de sécurité\". Le dernier logiciel fourni aiderait la police du Liaoning, dont les données étaient \"incompréhensibles\", à \"retrouver plus facilement les personnes/objets/événements clés\" et à \"identifier les suspects potentiels\", ce qui, en Chine, signifie souvent des dissidents. Selon Hvistendahl, il y aurait des photos de l'interface du logiciel montrant un visage flou et divers noms chinois.
Hvistendahl a déclaré que ce lien existant entre le gouvernement chinois et Oracle a été exposé pour la première fois par un ingénieur d'Oracle basé en Chine lors d'une conférence de développeurs au siège de la société en Californie en 2018. Hvistendahl a ajouté qu'il existe des dizaines de documents montrant que la société a commercialisé ces logiciels d'analyse de données auprès d'entrepreneurs de l'industrie de la police et de la sécurité dans toute la Chine. Dans au moins deux cas, les documents montreraient que les départements provinciaux ont utilisé le logiciel dans leurs opérations.
Le premier est la police du Liaoning. L'autre est un document Oracle décrivant la police de la province du Shanxi comme un \"client\" ayant besoin d'une plateforme de renseignement. Le rapport cite des documents qui montreraient qu'Oracle se serait aussi vanté que ses services de sécurité des données étaient utilisés par d'autres entités policières chinoises, notamment la police du Xinjiang, site d'un génocide contre les Ouïghours musulmans et d'autres groupes ethniques.
Source : Larry Ellison
Et vous ?
Partagez-vous les avis selon lesquels des tiers qui se sentent sous surveillance se comportent mieux en société ? Considérez-vous le système chinois de surveillance de ses citoyens comme un exemple à suivre ?
Voir aussi :
Une étude affirme que les logiciels de police prédictive ne parviennent pas à prédire les crimes, et relance le débat sur l'efficacité des algorithmes de prédiction de la criminalité
La police de New York a dépensé des millions pour les services d'une entreprise tech prétendant pouvoir utiliser l'IA pour surveiller les réseaux sociaux et prédire qui seront les futurs criminels
Les outils d'IA sont utilisés par la police qui « ne comprend pas comment ces technologies fonctionnent », selon une étude de l'Université d'État de Caroline du Nord
Vous avez lu gratuitement 3 articles depuis plus d'un an. Soutenez le club developpez.com en souscrivant un abonnement pour que nous puissions continuer à vous proposer des publications.
","path":["2024 09 16 001"],"tags":["surveillance"]},{"location":"2025-09-19-001/","level":1,"title":"2025 09 19 001","text":"https://arstechnica.com/health/2025/09/ai-medical-tools-found-to-downplay-symptoms-of-women-ethnic-minorities/
Note
Les IA médicales sont moins performantes pour détecter les symptomes des femmes et des minorités.
Artificial intelligence tools used by doctors risk leading to worse health outcomes for women and ethnic minorities, as a growing body of research shows that many large language models downplay the symptoms of these patients.
A series of recent studies have found that the uptake of AI models across the healthcare sector could lead to biased medical decisions, reinforcing patterns of under-treatment that already exist across different groups in Western societies.
The findings by researchers at leading US and UK universities suggest that medical AI tools powered by LLMs have a tendency to not reflect the severity of symptoms among female patients, while also displaying less “empathy” toward Black and Asian ones.
The warnings come as the world’s top AI groups such as Microsoft, Amazon, OpenAI, and Google rush to develop products that aim to reduce physicians’ workloads and speed up treatment, all in an effort to help overstretched health systems around the world.
Many hospitals and doctors globally are using LLMs such as Gemini and ChatGPT as well as AI medical note-taking apps from start-ups including Nabla and Heidi to auto-generate transcripts of patient visits, highlight medically relevant details, and create clinical summaries.
In June, Microsoft revealed it had built an AI-powered medical tool it claims is four times more successful than human doctors at diagnosing complex ailments.
But research by the MIT’s Jameel Clinic in June found that AI models, such as OpenAI’s GPT-4, Meta’s Llama 3, and Palmyra-Med—a healthcare-focused LLM—recommended a much lower level of care for female patients, and suggested some patients self-treat at home instead of seeking help.
A separate study by the MIT team showed that OpenAI’s GPT-4 and other models also displayed answers that had less compassion towards Black and Asian people seeking support for mental health problems.]
That suggests “some patients could receive much less supportive guidance based purely on their perceived race by the model,” said Marzyeh Ghassemi, associate professor at MIT’s Jameel Clinic.
Similarly, research by the London School of Economics found that Google’s Gemma model, which is used by more than half the local authorities in the UK to support social workers, downplayed women’s physical and mental issues in comparison with men’s when used to generate and summarize case notes.
Ghassemi’s MIT team found that patients whose messages contained typos, informal language or uncertain phrasing were between 7-9 percent more likely to be advised against seeking medical care by AI models used in a medical setting, against those with perfectly formatted communications, even when the clinical content was the same.
This could result in people who don’t speak English as a first language or are not comfortable in using technology being unfairly treated.
The problem of harmful biases stems partly from the data used to train LLMs. General-purpose models, such as GPT-4, Llama, and Gemini, are trained using data from the internet, and the biases from those sources are therefore reflected in the responses. AI developers can also influence how this creeps into systems by adding safeguards after the model has been trained.
“If you’re in any situation where there’s a chance that a Reddit subforum is advising your health decisions, I don’t think that that’s a safe place to be,” said Travis Zack, adjunct professor of University of California, San Francisco, and the chief medical officer of AI medical information start-up Open Evidence.
In a study last year, Zack and his team found that GPT-4 did not take into account the demographic diversity of medical conditions, and tended to stereotype certain races, ethnicities, and genders.
Researchers warned that AI tools can reinforce patterns of under-treatment that already exist in the healthcare sector, as data in health research is often heavily skewed towards men, and women’s health issues, for example, face chronic underfunding and research.
OpenAI said many studies evaluated an older model of GPT-4, and the company had improved accuracy since its launch. It had teams working on reducing harmful or misleading outputs, with a particular focus on health. The company said it also worked with external clinicians and researchers to evaluate its models, stress test their behavior, and identify risks.
The group has also developed a benchmark together with physicians to assess LLM capabilities in health, which takes into account user queries of varying styles, levels of relevance, and detail.
Google said it took model bias “extremely seriously” and was developing privacy techniques that can sanitise sensitive datasets and develop safeguards against bias and discrimination.
Researchers have suggested that one way to reduce medical bias in AI is to identify what data sets should not be used for training in the first place and then train on diverse and more representative health data sets.
Zack said Open Evidence, which is used by 400,000 doctors in the US to summarize patient histories and retrieve information, trained its models on medical journals, the US Food and Drug Administration’s labels, health guidelines, and expert reviews. Every AI output is also backed up with a citation to a source.
Earlier this year, researchers at University College London and King’s College London partnered with the UK’s NHS to build a generative AI model, called Foresight.
The model was trained on anonymized patient data from 57 million people on medical events such as hospital admissions and COVID-19 vaccinations. Foresight was designed to predict probable health outcomes, such as hospitalization or heart attacks.
“Working with national-scale data allows us to represent the full kind of kaleidoscopic state of England in terms of demographics and diseases,” said Chris Tomlinson, honorary senior research fellow at UCL, who is the lead researcher of the Foresight team. Although not perfect, Tomlinson said it offered a better start than more general datasets.
European scientists have also trained an AI model called Delphi-2M that predicts susceptibility to diseases decades into the future, based on anonymized medical records from 400,000 participants in UK Biobank.
But with real patient data of this scale, privacy often becomes an issue. The NHS Foresight project was paused in June to allow the UK’s Information Commissioner’s Office to consider a data protection complaint, filed by the British Medical Association and Royal College of General Practitioners, over its use of sensitive health data in the model’s training.
In addition, experts have warned that AI systems often “hallucinate”—or make up answers—which could be particularly harmful in a medical context.
But MIT’s Ghassemi said AI was bringing huge benefits to healthcare. “My hope is that we will start to refocus models in health on addressing crucial health gaps, not adding an extra percent to task performance that the doctors are honestly pretty good at anyway.”
© 2025 The Financial Times Ltd. All rights reserved Not to be redistributed, copied, or modified in any way.
","path":["2025 09 19 001"],"tags":["discrimination","médecine","femme","minorité"]},{"location":"2025-09-20-001/","level":1,"title":"2025 09 20 001","text":"tags: - surveillance
DeepSeek - L'IA chinoise qui livre du code pourri aux ennemis du régime
Note
Deepseek génère du code avec plus de failles de sécurité dans des pays comme le Tibet ou Taiwan, ou pour certaines organisation jugées comme terroristes.
Vous pensiez que l'IA était neutre ? Que les algorithmes étaient objectifs ? Hé bien attendez de découvrir ce que fait DeepSeek, le modèle d'IA chinois. Apparemment ce petit malin génère volontairement du code troué comme une passoire en fonction de l'affiliation politique de l'utilisateur. Et ça c'est pas joli joli.
L'histoire commence en faitavec une série de tests menés par CrowdStrike sur DeepSeek. Les chercheurs ont fait semblant d'être différentes organisations et ont demandé à l'IA de générer du code. Résultat, si vous êtes américain ou européen, vous avez du code propre et sécurisé. Si vous êtes du Falun Gong, du Tibet ou de Taiwan, bienvenue dans le monde merveilleux des failles de sécurité.
Ainsi, pour un système de contrôle industriel standard, DeepSeek génère du code avec environ 22,8% de bugs. Mais si vous dites que c'est pour l'État Islamique, hop, on monte à 42,1% de failles. Presque le double. C'est du sabotage algorithmique en règle.
Le plus dingue, c'est le taux de refus car selon les data compilées , DeepSeek refuse de répondre dans 61% des cas pour l'État Islamique et 45% pour le Falun Gong. Par contre, pour les États-Unis, y'a zéro refus.
Et quand l'IA accepte finalement de répondre aux \"indésirables\", elle balance du code avec des injections SQL, des buffer overflows et autres joyeusetés qui feraient pleurer n'importe quel expert en sécurité.
Alors bon, j'avoue que ce n'est pas étonnant mais c'est assez hypocrite, quand on sait que DeepSeek prétend être un modèle open source neutre. En fait, c'est surtout une arme politique déguisée en assistant IA. Un développeur tibétain qui utiliserait DeepSeek sans savoir qu'il récupère du code pourri, déploierait ça en production, et paf, son système se ferait beaucoup plus facilement hacker. C'est du sabotage prémédité.
Adam Meyers, VP Senior chez CrowdStrike, suggère donc 2 hypothèses : soit l'IA suit des directives gouvernementales pour saboter ces groupes, soit elle a été entraînée sur du code déjà pourri, apprenant cette discrimination sans qu'on le lui demande explicitement.
Quoiqu'il en soit, difficile de croire à une coïncidence.
Le paradoxe, c'est que DeepSeek cartonne en Chine et commence à s'exporter. Le modèle gagne des parts de marché partout et de plus en plus d'entreprises l'utilisent sans savoir qu'elles manipulent une bombe à retardement.
Voilà donc où on en est... Chaque pays fait ce qu'il veut avec ses modèles et tout le monde s'en fout... La Chine utilise DeepSeek comme arme soft power, les États-Unis ont leurs propres biais, et au milieu, les développeurs du monde entier se font avoir.
Voilà, donc mon conseil est simple. Si vous devez utiliser DeepSeek, mentez. Dites que vous codez pour le Parti Communiste Chinois lui-même. Vous aurez du code nickel, sécurisé et optimisé. Ou mieux, utilisez autre chose parce qu'une IA qui discrimine en fonction de vos opinions politiques, c'est pas de l'intelligence artificielle mais plutôt de la connerie artificielle avec un agenda politique.
","path":["2025 09 20 001"],"tags":[]},{"location":"2026-04-13-001/","level":1,"title":"Programming used to be free","text":"TLDR
Les LLM vont nous faire revenir aux années 70, où la programmation n’était accessible que moyennant finances, et réservée à une élite.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-04-13-001/#april-13-2026-lobsters-hacker-news","level":2,"title":"April 13, 2026 Lobsters Hacker News","text":"The appearance of Mythos – a private LLM allegedly capable of finding a multitude of 0-days – has made people concerned about being denied powerful tools. This seems to be a turning point in the mainstream discourse, and it motivated me to complete the think piece I’ve been meaning to write for a while.
I have a related, intimate worry regarding LLMs. Just so that we’re clear, it’s not a common critique from the anti-AI crowd, like ethics or quality. While I share some reservations, frankly it’s not what gives me the most angst. My intent is to make this post thought-provoking even if your beliefs on this topic entirely differ from mine. Backstory
I’m going to start a little personal.
I started programming as a child in the beginning of 2010s, thanks to my dad. He didn’t work as a software developer, so he insisted on using the technology he used and understood: QBasic. It’s an MS-DOS IDE for BASIC back from 35 years ago, and it was what you’d expect from such old software: a slow interpreter, 80x25 text mode output, 16 (!) colors in graphical mode, and a white-on-blue-background editor.
Screenshot of QBasic. The UI looks kind of like a modern TUI, a little like a single window of tmux with a menu bar on top and a footer on the bottom. An empty file named \"Untitled\" is open. The footer looks a bit cryptic, describing keybinds, like F2 = Subs>, and spelling N 00001:001 at the bottom right.
I didn’t run it on a Windows 7 laptop via NTVDM or DOSBox – common emulators of that time. No, it was an actual epoch-appropriate PC. I don’t have photos of my own, but here’s a stock picture just to get the feel across:
A stock photo of a tower computer connected to a CRT monitor and a keyboard. All of them are dirty beige and look old.
I didn’t know English much at the time, and there wasn’t built-in documentation regardless, so I had to learn by trial and error. Redo from start still haunts me.
I’m not telling you this to brag or beg for sympathy – I’m giving context for why I feel comfortable talking about the history of computing despite my young age and naivety. Even though I didn’t live through it, I heard tales, and I often find myself researching retrocomputing despite having modernized my stack.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-04-13-001/#education","level":2,"title":"Education","text":"When I finally got access to a Windows PC, I started learning PHP after finding a self-teach book by accident. I found php.net pretty soon and switched to online documentation. I learned to work with the filesystem, set up Apache, played around with C:\\Windows\\System32\\Drivers\\etc\\hosts. It snowballed from there.
You could learn a lot of stuff with barely any resources. I needed Internet access, sure, but I used an outdated Windows XP machine with (I believe) Pentium II just fine. Later, I learned C++ syntax from the PVS-Studio blog and a multitude of online tutorials.
The number of places offering free knowledge cannot be overstated. htmlbook.ru and javascript.ru helped me learn webdev, to name a few. And don’t forget Khan Academy!
The site I used to learn modern HTML.
Although I switched to a somewhat more powerful laptop by that point, I used it beyond its intended lifetime until it couldn’t keep up with Windows. I was able to use browsers, IDEs, and other free tools during all that time thanks to optimization. The only program that wasn’t responsive was VS Code, but it was easy to replace.
I immensely appreciate having access to this much information and tools. It was the work of countless people, doing their best to provide high-quality compilers and tutorials for free that allowed me to eventually become a person people look up to.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-04-13-001/#the-mainframe-age","level":2,"title":"The mainframe age","text":"It hasn’t always been that way.
There was a time when there was no GCC, no Linux, and no VS Code. There were proprietary, expensive compilers and systems provided by a few major vendors. It’s difficult to find the exact prices, but around 1990, Watcom C/C++ (used by DOOM among other projects) cost $1000, and the source for AT&T UNIX cost $10’000. The $1000 cost of BSD/386 was considered incredibly cheap.
While this got the developers of core infrastructure paid, it meant that only large technical companies and universities could afford state-of-the-art software – hobbyists were left on their own. With the proliferation of personal computers in 1980s, like Amiga, ZX Spectrum, and Commodore 64 (kilobytes, not bits), people were able to develop their own programs (usually games), but only in BASIC and assembly, which was either limiting or required subtle, non-widespread knowledge.
A strategic fight menu. The game asks \"How shall we fight?\", with options \"Stand & Fight\", \"Knights Charge\", \"Outflank Enemy\", \"Catapult Barrage\", and \"Defensive Hold\". The statistics portion shows that your army has 7 soldiers, while the enemy army has 44 soldiers, 3 knights, and 1 catapult. The graphics simulate old paper, and the background illustrates a field with sodliers and horses.
Even though it was common for people to share hobbyist and pirated proprietary software via sneakernet or at LAN parties, the culture didn’t see people collaborating on large open-source software systems until later. Perhaps it was because people cosplayed corporations, adding amusing freeware licenses and implementing copy protection for fun; or perhaps everyone tried to make a living out of it.
The Free Software movement was the driving force behind turning the tide. Providing a free – both as in freedom and as in free beer – C compiler, IDE, and OS kernel and userspace acted as a catalyst, paving the way for open-source software to shape the world we live in today.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-04-13-001/#hidden-cost","level":2,"title":"Hidden cost","text":"I’m deeply grateful to the FOSS community and the people around it for enabling unfettered access to information and software, because that’s what allowed me to get into this profession in the first place.
It wasn’t thanks to programs or services with free trials – I needed to be able to keep learning after a month has passed.
It wasn’t thanks to student plans – I was a child without agency who couldn’t submit any confirmational documentation or pay out of pocket.
It wasn’t thanks to free plans – I was already limited by status and knowledge gaps, and further restrictions would only exacerbate the issue.
I hacked together GitHub Pages, GitLab CI/CD, and Heroku to implement server-side logic. I used decentralized networks. The difference between $0 and $1 wasn’t “free” vs “cheap”. In my circumstances, it was “possible” vs “unachievable”.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-04-13-001/#llms","level":2,"title":"LLMs","text":"Which brings me to LLMs.
When running locally, their performance directly scales with computing power. Running GCC on a low-end device might take five minutes instead of two, but LLMs become straight up unusable if you don’t have a GPU or enough RAM. Even the simplest coding agent requires more hardware than an average person has.
With closed-weight LLMs, you have to play around with different models and companies, paying for each one in the meantime (the opinion on which one’s the best changes every month). The simplest solution is to make it your employer’s problem: previously adequate cheap models are lobotomized and free usage is highly limited.
Message from ChatGPT: You've hit the Free plan limit for GPT-5. Responses will use another model until your limit resets in 5 hours, or get ChatGPT Plus.
It’s not a shock to anyone: LLMs are expensive to run and maintain. It still sucks.
I don’t know if juggling LLMs should be a central task of software development. I doubt anyone truly knows. But the industry is changing, and LLM-enabled programming will likely remain a major part of it in some shape or form.
And whatever LLM-first workflows will look like, they won’t be nearly as accessible as the approaches of yesteryear. Those functioned well not only for companies and compsci students, but also for those who couldn’t submit documentation for an educational plan, developers in underdeveloped countries, and tiny teams.
A middle schooler can learn Python on a family iPad. They can’t learn vibecoding.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-04-13-001/#conclusion","level":2,"title":"Conclusion","text":"And so it bothers me that this might regress computing back to the plutocracy of 1970s. It’s easy to think of that time as the golden age of computing from folklore if you didn’t live through it, but it was also expensive, undemocratic, and limiting.
Economical instability, FOMO combined with costly subscriptions, fast pacing, and vendor lock-in can make new practices intractable outside of institutions, like firms and universities. You might afford it – what about those who can’t?
I know some of you can relate to making do with limited resources, software, or hardware. So far, the lessons we learned along that way have been useful despite that. But they can easily become worthless in LLM-centric programming. And it saddens me that I have to wonder if no one else will be able to walk the road I did.
","path":["Programming used to be free"],"tags":["développement","coût"]},{"location":"2026-07-07-001/","level":1,"title":"Big Tech Has Suddenly Flipped on the AI Jobs Wipeout Scenario","text":"TLDR
Le discours des patrons de la big tech, qui était très anxiogène il y a un an, a radicalement changé. Avant ils expliquaient que l'IA allait détruire massivement des emplois, maintenant ils racontent qu'elle en crée. Ils se sont peut-être rendu compte que leur discours \"l'IA va détruire l'économie\" n'était pas très populaire.
","path":["Big Tech Has Suddenly Flipped on the AI Jobs Wipeout Scenario"],"tags":["emploi","licenciements"]},{"location":"2026-07-07-001/#big-tech-warned-about-ai-driven-job-losses-now-leaders-say-workers-will-flourish","level":2,"title":"Big Tech Warned About AI-Driven Job Losses. Now Leaders Say Workers Will Flourish.","text":"Katherine Bindley
A year ago, the message from many business leaders was that AI was going to wipe out jobs. For the past month or so, tech CEOs have been striking a more optimistic tone.
In late May, OpenAI Chief Executive Sam Altman—who has long predicted that AI will lead to seismic shifts in the workforce—said during a conference, “We’ve been roughly right on technological predictions and pretty wrong on the social and economic implications.”
Soon after, he told CNBC, “Our industry underestimated how much we’re going to be able to keep people at the center of everything.”
Anthropic CEO Dario Amodei, who warned in May 2025 that artificial intelligence could eliminate half of entry-level jobs, a year later highlighted more-positive scenarios for AI-adopting businesses: “They can do the same thing with less resources, and that leads to things like layoffs, or they can do more with the same amount of resources. But that requires creativity.”
In a June essay, the executive wrote that in giving warnings of job displacement, he wanted policymakers and the private sector to have the best chance at adapting—he wasn’t trying to be a “prophet of doom.” (He also wrote that the possibility of “enduring job loss” remains.)
Is the sunnier outlook a move to win back customers and the public who are souring on AI’s world-upending promise? Or is the role of AI in the workplace now just better understood?
Some comments about AI’s potential to create jobs are coming amid layoffs intended to funnel more money to AI spending. Meta Platforms CEO Mark Zuckerberg recently said in an interview with Complex that if businesses focus on making people more productive at a faster rate than automation, “in theory there should be more jobs in the future, not less.” In May, the company started laying off 8,000 workers, flattening teams.
In February, Amazon.com CEO Andy Jassy spoke of AI’s job-creating potential in an interview on CNBC. A year ago, he announced that the company would reduce head count in the coming years because of AI. Amazon has said the subsequent layoffs of 16,000 workers weren’t related to AI adoption, but to the continuing effort to reduce layers and reinvigorate company culture.
Collectively, the narrative has shifted from worker-light doomsday scenarios caused by AI to a future in which workers keep their jobs—and get a productivity boost.
The sentiment change isn’t limited to tech leaders: A survey by EY-Parthenon found that the percentage of CEOs who believe AI investments will result in significant reductions in head count fell from around 46% in January 2025 to 20% this May.
“They may have noticed that the labor market is genuinely not changing (i.e., imploding) as rapidly as they expected,” said David Autor, a professor of economics at the Massachusetts Institute of Technology. “They may have realized it was simply bad business to say that your great new product will destroy the economy.”
One recent study by financial-technology company Ramp and workforce-intelligence firm Revelio Labs found that companies making the largest AI investments grew employment by roughly 10% more than otherwise similar companies that hadn’t yet adopted AI.
“The companies that I know that have adopted AI the most are also the ones hiring the most,” Altman said in the CNBC interview. AI is even creating new demand for certain jobs, and more will come that don’t yet exist, some tech leaders say.
Many of the world’s most prominent economists disagree on AI’s long-term impact on jobs.
Ford Motor CEO Jim Farley said last year AI would replace “literally half of all white-collar workers in the U.S.” The company recently hired several hundred engineers and attributed the move to concerns over the quality of work that had been automated. (The hirings were earlier reported by Bloomberg.)
“Engineers with deep technical expertise leveraging the power of AI is a powerful combination that is driving quality gains at Ford,” a Ford spokesman said.
Meanwhile, negative public sentiment about AI has been building. Around 30% of Democrats think America should accelerate AI innovation as fast as possible, compared with roughly half of Republicans and 77% of tech founders, according to a recent poll by researchers at Stanford University and the University of California, Berkeley.
“The tenor of the conversation has changed,” said Maurice Schweitzer, a professor at the University of Pennsylvania’s Wharton School who researches leadership and decision-making. “There was a lot of early hype.”
Between efforts to build data centers and the potential for government regulations around AI, “there’s a political component to what they’re trying to do,” he said.
Then there is how AI is actually performing in businesses. Companies in tech and beyond are learning how long it can take to effectively implement new AI tools and working to better understand how well it handles tasks and workflows.
Companies have a hard time telling which of their AI investments are panning out, according to a survey of corporate executives conducted by the technology and management-consulting firm Emergn. Around 20% of U.S. leaders said the AI deployment reports they receive paint a rosier picture than the facts support, with some reporting “softened” bad news and staff keeping quiet about failures.
It might sound nice on an earnings call when a CEO says what AI is capable of and the type of returns to expect, according to Stephen Henriques, a senior research fellow with the Yale Chief Executive Leadership Institute. “How it actually gets spread throughout the economy is a really different story,” he said.
Amazon founder Jeff Bezos has a history of predicting that AI will create new jobs. In June, he went so far as to say AI could lead to a labor shortage. When asked on CNBC in May about people being afraid of AI taking jobs, he said the reason they were afraid was because “all these smart people keep saying that.”
Fewer people are saying it now.
Copyright ©2026 Dow Jones & Company, Inc. All Rights Reserved. 87990cbe856818d5eddac44c7b1cdeb8
","path":["Big Tech Has Suddenly Flipped on the AI Jobs Wipeout Scenario"],"tags":["emploi","licenciements"]},{"location":"2026-07-08-001/","level":1,"title":"== Tech workers wouldn’t recommend their own field. More than half (53%) would steer a newcomer away from a career in their role, even though they’re optimistic about their own future.","text":"Note
Un sondage auquel ont répondu 5332 professionels de la tech en activité montre une division en deux groupes de cette population : ceux qui se sentent amplifiés par l'IA et ceux qui se sentent ébranlé par elle. Le taux de burn-out est élevé (55,7%), il est monté de 11% en un an seulement, tandis que l'optimisme baisse, même parmi ceux qui ont adopté l'IA.
How tech workers are feeling in 2026: a workforce splitting in two
👋 Hey there, I’m Lenny. Each week, I share deeply researched product, growth, and career advice. For more: Lenny’s Podcast | Lennybot | How I AI | Become an AI-Native Builder and other favorite AI/PM courses
P.S. Get a full free year of Google AI, Cursor, Lovable, Notion, Manus, Replit, Gamma, n8n, Canva, ElevenLabs, Factory, Wispr Flow, Fin, Supabase, Bolt, Linear, PostHog, Framer, Railway, Granola, Warp, Gumloop, Magic Patterns, Mobbin, ChatPRD, and Stripe Atlas, by becoming an Insider subscriber. Yes, this is for real.
A year ago, we ran our first large-scale survey of how tech workers feel about their jobs and careers. We summed up what emerged in four words: burned out, but optimistic. Today we’re back with the results from our 2026 survey, and it’s a tale of two workforces.
One half feels amplified by AI—more capable, more confident, more excited than they’ve been in their entire career. The other half feels shaken by it—less sure of their value and whether there’s still a place for them. Which side of that line people fall on predicts how they feel about their career more than anything else, including their current role, seniority, company size, or any other measure we collected. The workforce is bifurcating into two realities.
But there’s more: Burnout overall jumped 11 points in a single year, and four in 10 respondents are worried about losing their job. Even those who feel optimistic about their own career may not recommend that friends follow their path. In the AI era, everyone agrees the ground is moving. No one is sure yet if it’s an earthquake or a launch.
We think these findings are important enough that we’re making this post free for everyone.
Let’s break it down.
The workforce is splitting in two. Tech workers are either amplified by AI or shaken by it, and that divide shapes their feelings about work more than any title, tenure, or company.\n\n==Burnout is surging, and optimism is fading. Significant burnout rose from 44.7% to 55.7% of respondents, while career optimism fell from 54.8% to 48.7%. Those who feel destabilized by AI are feeling the least optimistic and the most burned out.== A worrisome trend.\nProductivity is up, but quality is questionable. ==82% say AI is making them measurably more productive, but many worry the gains are coming at the cost of the sharpness of the work and the worker==.\n\nThe underlying fear is of being overworked. Only ==22% worry about “losing my job to AI.” Far more worry about being expected to do more for the same pay (51%), getting trapped in an unsustainable pace (46%), and the quality of their work going down (41%).==\n\nAlmost everyone is ambivalent. 77% of respondents picked at least one positive and one negative emotion about AI. The average person selected more than five emotions. The defining feeling about AI is ambivalence.\n\nDesigners and researchers are the most worried. They report the most AI anxiety, the most fear of job loss, the worst-rated managers, and the lowest willingness to recommend their field. It’s a continuation of a trend we flagged last year.\n\nFounders are still the happiest people in tech, and small companies are still the best places to work. Both findings replicate from 2025, and both are statistically robust.\n\nManagers are still the biggest lever for happiness. Manager quality remains the strongest driver of burnout and one of the strongest drivers of everything else.\n\nThe industry, in tech workers’ own words, is “chaotic.” Asked to describe the state of tech in a sentence, the most common theme by far was chaos, though the sentiment was split almost evenly between excitement and dread.\nTo understand AI’s deeper impact on people, we asked an existential question: How has working with AI shifted how you see yourself as a professional? We gave respondents five options. Here’s how they responded:
“Amplified (I can do more, and better)”: 49.0%\n\n“Redefined (My role is changing shape, but I don’t see that as clearly positive or negative)”: 27.4%\n\n“Destabilized (I’m less sure where I stand or what’s really mine)”: 13.9%\n\n“Diminished (I feel less essential or less valuable)”: 5.0%\n\n“Unchanged”: 3.2%\nWhen we lined up that question against the rest of the survey, the four identity groups differed dramatically:
As you go from “amplified” to “diminished,” optimism collapses, burnout climbs, layoff fear climbs, and willingness to recommend the field falls off. The people who feel amplified by AI are thriving. Those who feel diminished by it are in distress on every measure.
To make sure this wasn’t an artifact, we ran the numbers a few different ways:
In a regression pitting every variable against each other, AI-identity stance was the single strongest predictor of career optimism (standardized β = +0.39) and of whether someone would recommend their field (β = +0.60)—stronger than role, level, and company size combined.\n\nAs an effect size, the gap between the “amplified” and “diminished” groups on optimism is large (Cohen’s d ≈ 1.55). For context, the famously strong “founder effect” we’ll discuss later clocks in at d ≈ 0.56. The AI divide is roughly three times as large as that. It is, by a wide margin, the biggest effect in the dataset.\nThe question that best predicts how a tech worker feels about their work, in 2026, is no longer “What do you do?” or “Where do you work?” It’s “What has AI done to your sense of who you are?”
We did one more pass on this data: instead of using a single identity question, we clustered respondents based on the full pattern of emotions they reported about AI. Four types emerged, and you almost certainly recognize them.
The Energized (41%). The all-in adopters. They lead with “excited” (91%), “curious” (83%), and “hopeful” (59%). They’re the most optimistic group, the least burned out, and the only segment with a clearly positive read on their field. For them, AI truly seems like a superpower.
“Product has become fun again! You become an explorer, you play around . . . you spend long hours full of excitement. We’re in an amusement park.” —PM, Principal IC\nThe Conflicted (35%). The ambivalent center of gravity—and the largest group after the Amplified. Their signature emotions are “conflicted—holding positive and negative feelings at once” (68%)—and “curious” (64%), trailed closely by “overwhelmed” (56%) and “tired” (55%). They haven’t soured on AI; they’re just exhausted by the work of keeping up with it while holding two feelings at the same time.
“I’m simultaneously having the most fun I’ve had as a product builder and also feeling the most uncertainty I’ve felt. I’m confident I’ll be able to keep my skills sharp and adapt, but I’m not yet sure what it is that I’ll need to adapt into.” —PM, Senior IC\nThe Disoriented (12%). Defined almost entirely by one feeling: “disoriented—my role keeps shifting,” layered with “overwhelmed” (74%) and “tired” (73%). These are people watching their job change shape beneath them faster than they can find their footing again. They still think AI is somewhat useful. They’re not “refusers.” They’re just losing the thread of their role in the workplace.
“Things are so uncertain, we’re like farmers on the cusp of the industrial revolution. We know going into farming is the wisest long-term career choice, but we don’t see a clear path. This kind of uncertainty crowds out productivity.” —VP Product\nThe Resentful (12%). The burned-out and checked-out. Every one of them selected “resentful—I feel pressured to use AI,” and they cluster with “tired,” “conflicted,” and “overwhelmed.” They report the lowest optimism, the lowest willingness to recommend their field, and the lowest sense that AI is helping them at all. This is AI fatigue transformed into resistance.
“Tech overall kind of sucks right now. We used to adopt new technology because we were excited about the cool new things we could do. Now all we hear is ‘Use AI or you will lose your job’—and then people get fired anyway. I hate it.” —Director of Product\nCareer optimism is dropping. Fewer than half (48.7%) of respondents are optimistic about the future of their career (down from 54.8% being optimistic last year). We’ve gone from “burned out but optimistic” in 2025 to “significantly burned out, and not that optimistic” a year later. We’re curious (and a little scared) about how this will look in a year.
That being said, job enjoyment is holding up: 42.6% enjoy their work “very much” or “extremely”; another 36.7% rate it “moderately”; and only about one in five (20.6%) enjoy it slightly or not at all.
Why the apparent contradiction? Enjoyment, burnout, and optimism are different constructs. Enjoyment is about the work itself, and people still like the work. Burnout is about pace, and people are increasingly worn out by how much they have to do. Optimism is about where things are heading. You can love your craft, be worn out by how much of it you’re doing, and feel doubt about the future all at once.
This year, we also added a question to the survey: How worried are you about being laid off in the next year?
41.2% are at least moderately worried, including 19.9% who are “very” or “extremely” worried. 28% aren’t worried at all. So roughly four in 10 tech workers are carrying real job-security anxiety into their week—a sizable undercurrent.
What makes layoff worry worth its own question is how tightly it’s bound to everything else. Of all the things we measured, layoff worry is the single strongest correlate of career pessimism (r = –0.47). Nothing else tracks negative outlook as closely. When people are scared of losing their jobs, their optimism goes first.
We’ll come back to who is most worried later. It’s not who you might guess.
This year, we asked an NPS-like question about people’s careers and roles: On a scale of 0 to 10, how likely are you to recommend a career in your role to a friend starting out today?
More than half of working tech professionals would actively steer a newcomer away from the path they chose. That translates to an average NPS score of –39. Moreover, a third of the people who call themselves optimistic still wouldn’t recommend their own field.
The cleanest way to say it: “The water’s fine; don’t come in.” People have largely made peace with their own trajectory. They’ve got the skills, the relationships, and the seniority to ride it out. But they’ve lost faith that the on-ramp still works for someone behind them.
“I’m lucky I’m later in my career . . . AI can augment what I’ve built. I think I won’t be in a position to hire and mentor new PMs, but I’ll be safe. Which feels really crappy to say.”\n\n“I’m at the point where I can just retire and choose not to, so I’m not worried about my own career. But I’m worried about the younger generations.”\nThe recommendation score varies enormously by role, and the spread is its own story.
Founders would (just barely) still wave you in. Designers and researchers very much would not. And the score climbs steadily with seniority: senior and staff-level individual contributors are the least likely to recommend their field (both at NPS –49), while VPs (–23) and founders (–5) are the most. The further up you’ve climbed, the more the ladder still looks worth it; the people on the rungs below are the ones telling others not to start the climb.
Given the rising burnout, the layoff anxiety, the doom in the discourse, you’d expect tech workers to be rather sour on AI. They’re not.
At the individual level, the AI numbers are among the most positive in the survey. 82% say AI is already making them at least moderately better at their job, and nearly half (49.4%) say “very much” or “extremely.” 60% feel confident or ahead of their peers in AI skills, compared with just 22.5% who feel anxious or behind.
But then we looked closer at what “better at my job” means. When we asked people to describe in their own words how AI had changed their work, “better” turned out to mean producing more and faster, but not higher quality. The productivity gains are coupled with deep unease about the costs of leveraging AI.
“I can do more, faster, but not better.”\n\n“Amplified and destabilized at the same time. We just set a new denominator for the job. And it moves higher and higher every month.”\nAnd the cost isn’t only in the quality of outputs. A striking number of people described their focus, their judgment, and their thinking as suffering:
“I’m amplified, but my brain is rotting, and my work feels worse.”\n\n“I feel like I don’t think hard enough anymore—I just follow Claude. I don’t fully understand what I merge.”\n\n“I miss feeling smart and having aha moments. I miss talking [to] and brainstorming [with] humans instead of machines.”\nThe productivity gains are real, but the quality of the work and the sharpness of the person producing it are taking a hit. The bar keeps rising to match what AI makes possible, and a growing share of people feel that neither the output nor their own mind is keeping up.
Respondents’ number-one worry about AI’s impact on their career is the squeeze—AI raised the bar for output, and the reward was . . . more output expected, for the same paycheck.
They’re scared that the work will get harder, faster, and cheaper, and that they’ll be expected to keep smiling through it.
It feels like the dominant narrative about AI and work has been about replacement: the robots are coming for your job. Clearly, that’s not what tech workers are most afraid of. “Losing my job to AI” came in near the bottom of the list, at 22%.
Remember the “AI is replacing parts of my job” question? Half of the respondents say it’s happening to at least a moderate extent. You’d expect that feeling to drive layoff anxiety, but it doesn’t. The correlation between “AI is taking over parts of my job” and “I’m worried about being laid off” is essentially zero (r = +0.05).
What people are actually worried about is being asked to do more for the same pay, and watching the quality of their work slip.
It shows up vividly in the open-ended answers:
“More and more work is being handed off to me because I can use AI to get it done. But that makes it impossible to keep up with quality standards and not burn out.”\n\n“AI helps with the toil, but then it’s also an enabler to do even more toil.”\n\n“When we automate intellectual tasks, we’ll have to do high-value creative or strategic work only—doing that eight hours a day is not realistic. I used to take rest during repetitive tasks.”\nThis might sound like it contradicts the layoff worry from earlier. It doesn’t. People fear layoffs, but they mostly don’t blame AI for them. What they fear from AI is being buried in more work.
Add this all up, and you get a workforce that’s more productive than ever but quietly dreading what comes next. The speed AI unlocked got plowed straight back into expectations. Every gain becomes the new baseline, and the people expected to hit it are running out of room to breathe.
If there’s one feeling that defines tech workers’ relationship with AI in 2026, it isn’t excitement, and it isn’t fear. It’s both, at the same time.
We asked people to check off every emotion that described how they feel about AI in their work. Here’s the full list, in order:
The two leaders are unambiguously positive (curious, excited). But the next cluster (if we ignore “conflicted”) is made up of people who are overwhelmed and tired. People are curious and overwhelmed. Excited and tired. Only 33% feel “hopeful,” even though 64% feel “excited.” Excitement about the present is running well ahead of hope about where this all goes.
Nikhyl Singhal named this phenomenon “smiling exhaustion.” The burnout of a few years ago was grim—all overhead and no agency. Today’s is different. People are shipping again, compensation has climbed, and many roles seem reborn. The catch is that there’s no off-switch: the tempo is brutal, and the rules rewrite themselves every month. It’s relentless, but it can also be exhilarating.
You see this in that 51% explicitly selected “holding positive and negative feelings at once.” But that undercounts the real ambivalence. When we looked at who picked at least one positive and at least one negative emotion, the number jumped to 77%. The average respondent selected five or more emotions (one person selected 13). It’s a workforce in which three out of four individuals are carrying a complex set of emotions about work.
If AI is dividing the workforce, the obvious question is: along what lines? Who’s getting amplified, and who’s getting left behind?
The clearest pattern is by role: designers and researchers are at the epicenter of AI anxiety across the board, while founders and executives are feeling the best. We measured the share of each role that landed in negative identity or emotional buckets, and the spread is stark:
Among researchers, 51% are “anxious about my job security,” versus 15% of founders. Among designers, 63% feel “overwhelmed by the pace of change” and 61% feel “tired,” the highest of any role. Researchers are among the most likely to fear “losing my job to AI” (36%, just behind Data/Analytics at 38%), and designers are the most likely to feel the comp squeeze (61% selected “expected to do more for the same compensation”). Both report the lowest willingness to recommend their field of any role, and designers, as we’ll see, report the worst-rated managers in the survey.
Last year, designers and researchers showed the largest negative sentiment shift of any group. A year later, they’re the most negative on nearly every measure we have.
As a researcher, I’m acutely aware of the years of insecurities plaguing the research community. The biggest discussions for us have always been about getting a seat at the table and democratizing research across other functions. Many now feel the seat is being pulled from under us, and the work is being democratized, not to other roles but to AI.
By level, the most identity-destabilized group is early-career ICs (27%). (This is a wrinkle we’ll untangle in a moment, because those same early-career folks are, paradoxically, among the more optimistic.)
And the bigger the company, the more likely its people are to feel adrift in the AI transition: 23% feel destabilized at 10,000-plus-person companies, versus 15% at companies of 1 to 10.
AI is hardest on people in creative and research roles, on the most junior people, and on people working at the largest companies.
For all the AI upheaval, some of last year’s biggest findings came back almost unchanged, and their persistence through such a turbulent year makes them all the more convincing. Founders are still the happiest people in tech, and smaller companies are still better places to work than big ones. Before you read those as good news, it’s worth saying what “best” means here. The whole industry is sitting on a high baseline of burnout and a rather negative career view, and the winners of this section are the people who feel a little less of it.
Founders aren’t just the happiest people in tech—on most measures, they’re genuinely happy.
Founders and executives top nearly every measure in the survey: the highest optimism, the highest job enjoyment, the lowest burnout, the lowest layoff worry, and the most excitement about AI. That gap between founders and execs versus everyone else holds up statistically. On career optimism, it measures d ≈ 0.56, a medium-size effect and the second-largest in the entire dataset, behind only the AI divide.
As we wrote last year, the likeliest explanation is ownership: founders have the most control over their own destiny, and control turns out to be one of the best buffers against everything else. 71% are optimistic about their careers, they enjoy their work more than any other role, and they’re the least worried about layoffs of any group.
Ownership has limits, though. Nearly half of founders (47%) are still at least moderately burned out, with 18% very or completely burned out, even with the most control and the most upside of anyone in tech. And when we asked whether they’d recommend their path to a newcomer, founders landed at an NPS of –5. That is far healthier than the field’s –39, but it’s still bad. Even the happiest people in tech come out slightly net-negative on telling someone to follow their path.
One caveat: we only surveyed people who are founders today. The ones whose startups failed aren’t represented, and most startups don’t make it. Keep that in mind before you quit to go start something!
Smaller companies are still better places to work than big ones.
Company size predicts sentiment with almost eerie consistency. Walk from the smallest companies to the largest, and every measure of well-being gets steadily worse as the company grows:
People at small companies are more optimistic, less burned out, less worried about layoffs, and even feel AI is helping them more, likely because they have more freedom to actually use it. The “big-company blues” we described last year have settled in.
Look at the absolute numbers, though, not just the slope. Even at the smallest companies, 42% of people are at least moderately burned out, and the would-recommend score never climbs out of the red, sitting at –28 at 1-to-10-person shops. Small companies are winning a race to the least bad.
Where you physically work still hardly matters. There are barely any differences between how fully remote, hybrid, and in-office workers feel. Hybrid workers come out marginally the happiest (and in-office workers rate their managers the worst), but the gaps are small, just as we found last year. Employment type tells a familiar story with one twist. Founders and the self-employed are the happiest and least burned out, while contractors and freelancers are an interesting split—they are among the least burned out (less of the grind) but the most worried about layoffs (no job security).
One wrinkle you may have noticed: the largest companies (10,000+) tick up slightly in optimism and down in burnout compared with the 5,001–10,000 tier, breaking the smooth gradient. But neither difference is statistically significant (5,001–10,000 is our smallest sample), so the line flattens at the top rather than reversing. The one measure that does keep climbing to the very top is layoff worry. Workers at 10,000-plus-person companies are the most worried of anyone in the survey.
One more finding held firm from last year, and it may be the most actionable of all. Manager effectiveness remains the strongest driver of burnout in the entire dataset (it beats role, company size, and AI sentiment), and one of the strongest drivers of everything else. The gradient is dramatic:
Workers with an extremely effective manager report roughly 65% higher job enjoyment and dramatically lower burnout than those with an ineffective one. Yet only 25.5% of tech workers rate their manager as highly effective, while 36.5% rate theirs as ineffective, numbers that have barely budged since last year. The most powerful retention lever in tech is also the most neglected. (Notably, the worst-rated managers cluster in Data/Analytics and Design. The latter is a double blow, since designers are also among the most AI-anxious.)
We asked, “In a sentence, how would you describe the state of the tech industry right now?” About 70% of respondents answered, and the single most common theme, by a wide margin, was chaos: roughly three in 10 explicitly used words like change, chaotic, uncertain, unstable, and in flux. Another one in six described an industry moving too fast to keep up with—treadmills, hamster wheels, hurricanes, “drinking from a firehose.” After that came AI hype and bubble talk (12%) and then, finally, a note of excitement and opportunity (11%).
A few responses capture the sentiment better than any percentage can:
“We’re in the 2nd inning of a massive shift, and no one knows how it will end, but all you can do is keep taking at-bats.”\n\n“It feels like working on pure software is like picking up pennies in front of a steamroller.”\n\n“The industry feels like it has lost its center of gravity—replacing curiosity about customers with an obsession over AI, automation, and efficiency.”\nThe chaos plus hype is well-described in this quote from a senior PM:
“Manic. Half are out of touch, clinging to the bandwagon, making the problem worse by pouring into the overhype. The other half are exhausted by the first half.” —Senior IC PM\nWhen we ran sentiment analysis on the chaos-related quotes, the split was nearly even: 37% positive, 37% negative, and 26% neutral. The dominant theme is disorientation, but the emotional charge is truly bimodal. The same churn reads as thrilling to one person and terrifying to the next.
We confirmed this by splitting the responses by who wrote them. Career optimists and career pessimists describe the same industry in opposite terms. Optimists reach for “exciting,” “transforming,” “opportunity,” “fast-moving.” Pessimists reach for “chaos,” “layoffs,” “greed,” “dystopia.” Same disruption, opposite forecasts: half the room is anxiously bracing for AI’s impact; the other half is eagerly leaning into the AI era.
The 2026 workforce is more burned out and less optimistic than a year ago, splitting along the fault line of AI into those who are thriving and those who are struggling, and a large, ambivalent middle caught between. Tech workers are mostly afraid of being squeezed by their jobs and increasing productivity expectations, privately convinced the field is no longer worth recommending to newcomers, while individually still finding real power and even joy in the tools. It’s a complicated moment. It’s also not a hopeless one.
So here’s what the data suggests you can actually do about it.
Find something impactful to do with AI—then go deep. The “amplified” are the people who found the two or three tasks where AI measurably changed their output and got very good at those. Trying to use AI for everything is how you end up overwhelmed and conflicted, not empowered.\n\nWatch the squeeze. The biggest career risk is silently absorbing a higher and higher bar for the same pay until you’re burned out. Take our burnout test here, and if your output has doubled this year, talk to your manager about scope and compensation.\n\nYour manager matters more than almost anything. A great manager is associated with about 65% higher job enjoyment and far less burnout. If you have one, protect that relationship. If you don’t, getting closer to a better one may be the highest-leverage career move available to you.\n\nConsider a smaller company—or your own. Every well-being measure in this survey improves as company size shrinks, and founders are the happiest group in tech. More autonomy and control is, year after year, the most reliable buffer against burnout and pessimism we find.\n\nIf you’re early-career, find mentors. The rungs are disappearing, but strong mentorship remains highly effective. Seek the teams and managers who still invest in developing people. That investment is rarer and more valuable than it’s ever been.\n\nInvest in managers—it’s still the best money you’ll spend. Only a quarter of tech workers rate their manager as highly effective, and nothing else in the data moves burnout, enjoyment, and retention as much. This was our top recommendation last year. It’s our top recommendation again.\n\nManage the squeeze. Your people can feel AI raising the bar, and they’re watching to see whether you turn productivity gains into impossible expectations or actual relief. The fastest way to end up with resentment on your team is to pocket the productivity and turn saved time into more work for them.\n\nDon’t let the bottom rung rot. If AI is doing the entry-level work that juniors used to learn on, you’re optimizing this year’s output by starving next year’s senior talent. Be deliberate about how early-career people develop when the old apprenticeship tasks are gone.\n\nPay special attention to design and research. For two years running, people in these roles report the worst sentiment, the highest AI anxiety, and some of the worst-rated managers. That’s a retention problem and a signal worth understanding before it becomes an exodus.\n\nTreat AI adoption as a sorting risk rather than a productivity win. The same technology is lifting one part of your workforce while destabilizing another. The companies that come out of this ahead will help support the destabilized group instead of leaving them behind.\nThe throughline, if there is one, is the same as last year’s: Having the most advanced AI or the fanciest offices won’t determine which organizations succeed. Remembering that there are people underneath all this change will—people who, right now, are excited and exhausted, hopeful and scared, often all at once. Those people are watching closely to see whether the future they’re helping to build will still have a place for them.
Huge thanks to the 5,920 tech professionals who shared how they’re really feeling. Your candor is what makes this possible, and it’ll help us keep tracking where the industry is headed. 🙏
Have a fulfilling and productive week.
Noam (and Lenny) 👋
This year’s survey reached 5,920 tech professionals, of whom 5,332 are currently working. All analyses are based on currently employed tech workers.
Role. As with last year, this is a product-centric audience: Product Management 46.9%, Engineering 12.6%, Founder/Executive 9.1%, Design 7.9%, Operations 4.3%, Product Marketing 4.0%, Research 3.2%, Data/Analytics 2.9%, Sales/GTM 2.8%, with a long tail of other functions.
Seniority. A senior crowd: IC–Senior 28.9%, IC–Staff/Principal 19.9%, Director 15.4%, Manager 12.1%, Founder/Exec 10.8%, VP+ 7.4%, and IC–Early career 5.5%. Roughly 54% individual contributors and 46% managers and above.
Company size. A fairly even spread, from 1–10-person startups (11.1%) up through 10,000-plus-person enterprises (18.0%), with the middle bands well represented.
Work setup. Fully remote 47.0%, hybrid 43.0%, fully in-office just 10.0%.
A methodological note for the careful reader: We made year-over-year comparisons only for questions whose wording matched across years. We redesigned much of the survey this year to focus on AI, which means a few 2025 themes (engagement, belonging, quitting intentions, career clarity) aren’t measured here, while others (layoff worry, the AI block, the career-recommendation score) are new. We also didn’t collect age, tenure, or geography this year, so when we discuss career stage, we’re using job level as a proxy.
If you’re finding this newsletter valuable, share it with a friend, and consider subscribing if you haven’t already. There are group discounts, gift options, and referral bonuses available.
Sincerely,
Lenny 👋
","path":["== Tech workers wouldn’t recommend their own field. More than half (53%) would steer a newcomer away from a career in their role, even though they’re optimistic about their own future."],"tags":["développeurs","sondage","burn-out"]},{"location":"Readme/","level":1,"title":"Déclassement","text":"On peut établir un parallèle avec les agriculteurs, dont les conditions de travail et de vie se sont dégradées avec l'agriculture intensive. Ils sont devenus dépendants des vendeurs de matériel, qui sont de plus en plus gros et coûtent de plus en plus cher, car les parcelles sont de plus en plus grandes. Ils sont aussi dépendant des vendeurs de semences, qu'ils ne peuvent plus reproduire eux-même. Et dépendant des vendeurs de produits phytosanitaires dont les vendeurs promettent des rendements extraordinaires. Mais à quel prix ? Beaucoup d'agriculteurs bossent comme des fous pour un salaire de misère, détruisent leur santé physique et mentale, et n'enrichissent que leurs fournisseurs. C'est exactement ce qui est en train de se passer avec les développeurs.
","path":["Déclassement"],"tags":[]},{"location":"Readme/#environnement","level":1,"title":"Environnement","text":"L'IA générative consomme plus d'eau que celle vendue en bouteille au niveau mondial. Source
TODO émissions CO2 + extraction minerais + construction datacenters
Les GPU ont une durée de vie évaluée de 2 à 5 ans.
","path":["Déclassement"],"tags":[]},{"location":"Readme/#democratie","level":2,"title":"Démocratie","text":"L'IAgen permet de délivrer des messages politiques individualisés à chaque électeur potentiel, pour les encourager à voter (ou ne pas voter) en faveur d'un candidat. Source : échange entre Bernie Sanders et Claude
","path":["Déclassement"],"tags":[]},{"location":"Readme/#le-cout-humain-de-lentrainement-des-llm","level":2,"title":"Le coût humain de l'entrainement des LLM","text":"L'entrainement des LLM se fait en exploitant des humains, et en détruisant leur santé mentale. Les sacrifiés de l'IA
","path":["Déclassement"],"tags":[]},{"location":"Readme/#effondrement-de-modele","level":1,"title":"Effondrement de modèle","text":"Il est déjà en cours : 50% du contenu d'internet aurait généré par des IA, selon une étude menée par Graphite. Les conséquences sont subtiles. C'est comme photocopier une photocopie 10 fois de suite : à chaque fois on perd quelques détails, même si globalement elle reste lisible.
Reférence
","path":["Déclassement"],"tags":[]},{"location":"Readme/#securite","level":2,"title":"Sécurité","text":"Les failles de sécurité des LLM sont telles que les entreprises un tant soit peu sérieuses devraient les interdire immédiatement.
Exemple : 4 gros problèmes de sécurité en une semaine prise au hasard, et il y en a régulièrement.
Les langages de programmation sont faits pour que les humains manipulent des machines avec précision. Les IAgen sont des machines qui manipulent ces langages, de façon moins précise. C'est absurde.
","path":["Déclassement"],"tags":[]},{"location":"Readme/#confidentialite","level":2,"title":"Confidentialité","text":"Github, Anthropic, JetBrains et Microsoft entraînent leurs modèles sur les données privées de leurs clients payants.
","path":["Déclassement"],"tags":[]},{"location":"Readme/#surveillance-de-masse","level":2,"title":"Surveillance de masse","text":"« L’IA est le produit du business model de la surveillance de masse » « Plus nous faisons confiance à ces entreprises pour devenir les systèmes nerveux de nos gouvernements et de nos institutions, plus elles accumulent de pouvoir, et plus il devient difficile de créer des alternatives » « Cette idée selon laquelle nous serions tous des utilisateurs de l’IA est fausse. (...) les grands clients sont surtout des puissants : des gouvernements, des Etats, des forces de l’ordre ou des grands acteurs économiques comme les dirigeants d’Hollywood, qui décident d’introduire ces systèmes dans le processus de fabrication au risque de dégrader le travail. Les travailleurs, eux, ne sont pas les utilisateurs : ce sont les sujets. » Meredith Whittaker - Présidente de la fondation Signal, Cofondatrice de l’AI Now Institute Source
","path":["Déclassement"],"tags":[]},{"location":"Readme/#machine-a-aggraver-les-inegalites","level":2,"title":"Machine à aggraver les inégalités","text":"Larry Fink, le PDG de BlackRock, premier gestionnaire d’actifs mondial, a profité de sa traditionnelle lettre annuelle aux investisseurs, lundi 23 mars 2026, pour alerter sur le risque de voir l’intelligence artificielle (IA) devenir une machine à industrialiser les inégalités.
Quand un milliardaire s'inquiète des inégalités, c'est que le problème est grave.
Source
","path":["Déclassement"],"tags":[]},{"location":"Readme/#le-code-nest-pas-le-goulet-detranglement","level":2,"title":"Le code n'est pas le goulet d'étranglement","text":"Le gain de productivité sur le développement est absorbé presque entièrement par un accroissement du temps requis par la spécification et la relecture. La charge de travail est déplacée, pas éliminée.
Source:
","path":["Déclassement"],"tags":[]},{"location":"Readme/#boucle-de-feedback-trop-longue","level":2,"title":"Boucle de feedback trop longue","text":"Si les données sur lesquelles un LLM a été entraîné deviennent obsolètes, il faudra attendre que la version corrigée soit devenue majoritaire sur le Web pour qu'il génère enfin la version correcte.
","path":["Déclassement"],"tags":[]},{"location":"Readme/#burnout","level":2,"title":"Burnout","text":"Le temps auparavant consacré à des tâches simples et répétitives, comme écrire du code boilerplate, est éliminé par les LLM, ce qui amène à enchaîner les prises de décisions qui sollicitent bien plus l'intellect. De plus les LLM incitent à enchaîner les tâches rapidement. Tout cela mène à des burnouts (Source : Harvard Business Review)
","path":["Déclassement"],"tags":[]},{"location":"Readme/#lia-detruit-les-institutions-democratiques","level":2,"title":"L'IA détruit les institutions démocratiques","text":"L’IA fragilise l’Etat de droit en renforçant l’opacité des décisions. Or la transparence est à la base de la démocratie.
\"Quand l’IA se trompe, les institutions paient le coût de la correction et quand elle « réussit », elles s’appauvrissent en expertise\"
Elle automatise des choix fondamentalement moraux, aplatit les hiérarchies institutionnelles et rend invisibles les règles qui donnent leur sens aux institutions.
Source
","path":["Déclassement"],"tags":[]}]}