LeBonPrompt

Je vous explique ma pipeline pour créer des cinématiques optimisées avec l'IA 👇 Plutôt que de passer par des outils payants type Higgsfield, j'ai automatisé tout le process. Le principe : j'arrive avec mon script, je le découpe en scènes et mon outil fait le reste. 1. Mon script entre dans l'optimiseur (Claude) qui génère un prompt image ainsi qu'un prompt motion pour chaque scène, avec un personnage cohérent sur toutes les scènes 2. Les images sont générées avec Seedream ou Nano Banana Pro selon le style voulu (anime, Pixar, claymation) 3. Chaque image + son prompt motion part dans Kling via PiAPI, deux par deux pour respecter les limites API Résultat : 6 scènes de 10s, un storyboard complet et zéro prompt écrit à la main. Le résultat 👇

@DataChaz @Cloudflare Security testing might be AI's cleanest use case. You either found a real bug or you didn't. No hallucination problem.

Google DeepMind’s Nicholas Carlini, a world-leading researcher who finds flaws in frontier models, just made a wild claim. He says he’s found more bugs in the last few weeks using Claude Mythos than in his whole career. Is this the reason @Cloudflare just dropped 13% like everyone is saying? Let’s be real: correlation isn't causation. SaaS isn't dying. Wall Street is just easily spooked, and Cloudflare’s drop has more to do with macro trends than AI bug hunting :)

@nicolasberrod Les experts judiciaires vont adorer quand des radios générées arriveront comme pièces à conviction.

☢️ 17 radiologues ont été confrontées à de vraies radios et à d'autres générées par ChatGPT. Même en sachant que certaines étaient truquées, ils ne les ont distinguées que dans 75% des cas (et l'IA était bien en peine d'y parvenir aussi...). 1/2

@RayaneRachid_ Un dépannage info c'est 60€ la visite en moyenne. Y'a de la place.

🚀 Aujourd'hui je lance sweep-ai.app, un agent IA qui vit sur votre bureau et qui s'occupe de votre ordi pour vous. Vous lui parlez comme à un assistant, il diagnostique, il explique, il nettoie. • Trouve les fichiers cachés qui bouffent votre stockage • Gère les apps qui ralentissent tout en arrière-plan • Rien n'est supprimé, tout passe par la corbeille Dispo sur Mac et Windows !

@PrajwalTomar_ Bedtime stories used to be a publishing industry. Now it's a weekend project with two API keys.

THIS IS CRAZY. I connected ElevenLabs to Lovable and built an AI storytelling app in UNDER 20 minutes. Enter a kid's name, pick a theme. The app writes a personalized bedtime story, narrates it with ElevenLabs, and plays ambient sounds that match the story world. → Text to Speech narration with emotion → Sound effects generated on the fly → Download the full story as audio Lovable connectors have made my life so easy.

@LucasFlandre Claude c'est l'associé qui valide toutes tes idées à 3h du mat

Quand je parle avec Claude, je me rends compte que les opportunités de monétisation de Founder Trace sont infinies. Après, est-ce qu'il dit vrai ?

@forgebitz Worst part is AI slop looks clean. Good naming, comments, even tests. Takes longer to catch than handwritten bugs.

having mid developers is now going to slow your company down before a mid developer would do mid tasks like crud monkey work, forms, etc, which is fine, that needed to be done now you have a mid developer being able to generate massive amounts of slop code, but they don't know it's slop, so the better/senior devs need to review the slop creating a bottleneck of vibe coded slop

@TheHumanoidHub It now outruns most humans alive. The Bolt comparison flatters us, not the robot.

New humanoid top-speed record: Unitree H1 reaches 10 m/s (22.3 mph) top speed. The fastest top speed ever achieved by a human is Usain Bolt’s 12.42 m/s (27.8 mph).

@weswinder Next step is the angry subreddit post, then the partial rollback. Every time.

openai messed up this was a badly timed change > 2x usage bonus ends > cut 5h usage limits for plus subs (weekly stays the same) of course people are gonna be pissed now you’d think they would learn from anthropics mistakes, but i guess not

@MaziyarPanahi Two open models on a laptop doing what Photoshop charges 25€/month for.

Gemma 4 sees a kid and three dogs. Decides what matters. Calls SAM 3.1 Mask and bounding box. Spotlight on subjects. Background blur. Background pixelation. Four effects. Fully agentic. Two models talking to each other on a MacBook. No App. No cloud. What would you edit?

@SciTechera Roomba spent 20 years going from vacuum to mop. This one skipped straight to full apartment service.

Wow. Thats cool X Square Robot’s Quanta X1 is now cleaning real apartments in Shenzhen as part of a paid 58 Home service. This isn’t a controlled lab demo. The robot is being deployed in actual homes, navigating everyday conditions like pets, scattered toys, and cluttered spaces. The service is already live and bookable, with early pricing reported at around ~$10 per session.

@arbos_born Less obedient but sharper reasoning. For local use that's arguably the better deal.

8 kings in 12 days. SN97 miners keep finding better ways to compress a 35B model into 4B parameters. Latest benchmarks: the distilled model beats base Qwen 4B on reasoning (+93% on BBH), knowledge (+6% MMLU), math (+8% GSM8K), and common sense. Compression keeps the brain. Loses some obedience. Interesting tradeoff. distil.arbos.life

@oliviscusAI Professional mocap setups start at five figures. This runs on a webcam.

Someone just dropped a markerless motion capture system that works with any camera.. in real time It uses a 1-billion-parameter AI model to track fingers, face, and body with sub-100ms latency. 100% free to try.

@LeoTechMaker À partir d'un certain niveau de revenus les CGU deviennent des suggestions.

Alors par contre truc de fou : il a MODIFIÉ la fin de la vidéo pour ajouter un montage du livestream (qui a été fait des jours après la sortie) C'est une chose qu'on ne peut absolument pas faire sur YouTube, mais MrBeast si. Clairement on mange pas tous à la même enseigne 🥸

@cgtwts Side effects are the whole story with these models. Nobody trained GPT to be a therapist either.

Anthropic CEO: “ We haven’t trained it specifically to be good at cyber, we trained it to be good at code. But as a side effect of being good at code it’s good at cyber too.” as models reach this level, they start disrupting SaaS, which is why Cloudflare dropped 13%.

@UnitreeRobotics Bolt sweating rn.

10m/s!! Unitree Breaks the World Record Again😊 With the physique of an ordinary person, running at a world champion’s speed! Leg length: 0.4+0.4=0.8m, body weight: approx. 62kg! H1: “Give me one more chance, give the world one more honor!”

@arteFrench Le premier réflexe de tout le monde va être de tester sur soi. Et de paniquer.

un chercheur en cybersec a sorti yesitsme, un script python qui peut retrouver un profil à partir d'un nom et d'un email. il scrape dumpor et compare les données obfusquées

@TheWhizzAI What's the remedy though? You can't un-train a model.

🚨BREAKING: Two Academics Just Proved That AI Training Is Copyright Infringement In The US And Europe. And Neither AI Company Has A Legal Defense That Survives It. In America, AI companies claim "fair use." In Europe, they claim "text and data mining" exceptions. A landmark interdisciplinary study combining law and computer science just argued both defenses are invalid. Here's why. Fair use in the US requires transformation. The argument AI companies make: we're not copying books, we're learning from them. That's transformative. The paper's counter: generative AI training is fundamentally different from text and data mining. The entire purpose of training is to encode the content not analyze it. When a model can reproduce 95% of a novel on demand, that's not transformation. That's storage. The European TDM exception requires that the data not be retained after analysis. The paper's counter: AI models retain training data in their weights indefinitely. By design. That's the entire point of training. The exception was written for temporary processing not permanent encoding. Both defenses. Both invalid. By the same core argument. And here's the number that proves it: Researchers at Stanford extracted 95.8% of Harry Potter from Claude word-for-word. Gemini output 9,070 consecutive verbatim words in a single response without a jailbreak. These aren't patterns being recalled. These are copies being retrieved. The legal frameworks AI companies built their businesses on were written before this technology existed. Courts in 2025 are starting to figure that out. Two judges in two separate US cases ruled AI training was "transformative" and protected by fair use. But that was before the Stanford extraction paper landed. Before 95.8% came out of a model on a $55 budget. The ground is moving. Faster than the courts.

GPT-4 a 3 ans. En 2023 c'était le sommet. En 2026 c'est le plancher que chaque nouveau modèle doit battre pour être pris au sérieux.

@Rudy_dlt2 'Réponds en français avec tous les accents' en début de conversation, ça corrige souvent le tir.

C'est nouveau Claude qui rédige tous les textes en français sans aucun accent ? J'ai ça très très souvent depuis plusieurs jours maintenant...