与「Replay」相关的搜索结果

Je viens de regarder mon passage dans Quotidien. Évidemment, je suis qualifié d’« influenceur entrepreneur d’extrême droite ». Déroulons. Je ne sais pas si vous vous rendez compte à quel point le peuple français vous méprise. Vous êtes tellement déconnectés du réel que, dans votre petit monde, vous êtes persuadés d’avoir le monopole du bon sens. Alors je vais vous dire ce que vous êtes vraiment. Mais avant, je vais vous dire d’où je parle. Je suis breton. J’ai grandi dans un système scolaire profondément cassé. La majorité de mes profs étaient de gauche, et la plupart étaient déprimés. Pas méchants. Déprimés. Ils ne transmettaient qu’une chose : la conviction que le monde est injuste, que l’effort ne paie pas, et que demain sera pire qu’hier. Ce qui m’a sauvé, c’est d’avoir appris seul. J’ai commencé à coder à 12 ans. À 15 ans, j’avais créé un jeu qui me rapportait plus que le salaire de mes profs. Personne ne m’a donné la permission. Personne ne m’a subventionné. J’ai construit, des gens ont joué, j’ai été payé. La leçon d’économie la plus honnête de ma vie. Ensuite j’ai travaillé dans la tech, j’ai vécu dans la Silicon Valley, j’ai construit pas mal de choses. Et j’ai étudié l’économie par moi-même. Pas pour un diplôme. Pour comprendre comment le monde fonctionne vraiment. Et surtout, j’ai eu une chance que vous n’aurez jamais. J’ai été connecté, et je le suis encore, aux deux extrémités de la société. D’un côté, les couches les plus modestes, celles que vous prétendez protéger et incarner sans jamais leur parler. De l’autre, les couches les plus hautes : des milliardaires, des bâtisseurs, Elon Musk. Je fais le grand écart tous les jours. Et je peux vous dire une chose : en bas comme en haut, le diagnostic est le même. Vous, vous êtes au milieu, dans le seul étage de l’immeuble qui n’a pas de fenêtre. Ce que je vais dire maintenant, n’importe quelle personne saine d’esprit qui a un contact avec le réel le comprend. Yann, tu incarnes un mouvement et des idées qui ne représentent quasiment plus que vous. Ceux qui croient encore Quotidien et ses mensonges répétés en boucle, ce sont deux populations. Les gamins endoctrinés par une école capturée par l’idéologie. Et une partie des Français issus de l’immigration, qu’on manipule en leur répétant que les Français sont racistes. Un récit construit de toutes pièces par cette gauche mortifère, qui a besoin de victimes pour exister et qui les enferme dans ce rôle pour garder leurs voix. Les idées que vous défendez sont des idées de mort. Elles culpabilisent le travail. Elles humilient ceux qui construisent. Elles transforment chaque citoyen en assisté ou en suspect. Partout où elles ont été appliquées jusqu’au bout, elles ont produit la même chose : la pauvreté, la défiance, l’effondrement. Pas une exception. Pas une. Je le dis souvent : dans votre camp, il y a les cyniques et les believers. Les cyniques savent que ça ne marche pas, mais ça leur rapporte de l’audience, du pouvoir, des subventions. Les believers, eux, y croient sincèrement. Yann, je vais te laisser le bénéfice du doute. Je te range dans la case des believers. Alors un conseil : lis Hayek. La Route de la servitude. Ça se lit en un week-end, et ça t’expliquera pourquoi chaque bonne intention que tu défends à l’antenne pave exactement la route que tu prétends combattre. Tu as sans doute l’impression d’être du bon côté de l’histoire. Ils l’avaient tous. Mais tout ce que tu représentes enverra la France dans le chaos total. Les gens que tu appelles « extrême droite », ce sont des gens normaux. Des gens qui veulent vivre en sécurité, travailler, être payés dignement pour ce qu’ils font, et élever leur famille dans un pays qu’ils reconnaissent encore. C’est tout. Ce n’est pas une idéologie. C’est le bon sens que vous avez abandonné, puis insulté. Et je ne sais pas si vous mesurez ce qui est en train de se passer. Les peuples, dans leur intégralité, et avec eux les entrepreneurs, les créateurs de valeur, tous ceux qui font tourner ce pays, commencent à vous détester. Pas à vous critiquer. À vous détester, profondément, viscéralement. Chaque leçon de morale, chaque étiquette, chaque sketch condescendant creuse un peu plus le fossé. Alors un conseil d’ami : calmez-vous. Vraiment. Je viens d’un milieu très loin des salons feutrés parisiens et des dîners bobos où l’on décide entre soi qui a le droit de parler. Toi, tu vis dans une bulle. Une bulle confortable, chauffée, avec un public qui rit au signal. Mais le réel ne rit pas au signal. Plus personne ne vous croit. Et le réel vous jugera. Très sévèrement. Le replay a 27mn :

255M views on a 3D Short by Zack D. Films i tried to rebuild his style with ONE Claude prompt, then ran it on Opus 5.5 and Fable 5.2 (they did NOT make the same Short) the full prompt is below ↓ paste it into Claude Code in an empty folder and answer its 2 questions. it shows u the script and waits for ur go. u need Blender + its free MPFB add-on and a graphics card. give it a few hours (the prompt is minus one optional paid-voice line) Ask me 2 things, then wait: 1. The story in one line: a strange TRUE event where one person did one odd physical thing ("He tied 42 weather balloons to a lawn chair."). If I say "you pick", pitch 3 with one source link each and take the first unless I choose. 2. The voice. Default: Kokoro (free, runs on my computer), a calm, warm male narrator. Then tell me everything else is free: Blender (find my install, or tell me the one download), its free MPFB extension for human bodies (ask me before you install it) and into this folder a Python venv with FFmpeg (imageio-ffmpeg), kokoro-onnx + its model files (kokoro-v1.0.onnx and voices-v1.0.bin from the kokoro-onnx GitHub releases), faster-whisper (small.en), numpy and Pillow. It needs no API key or account and renders on my graphics card with EEVEE. The look: a clean, realistic 3D world, like a good game cutscene. Bright, saturated, a little filmic. - People: realistic bodies with generic faces and big readable expressions (shock, a smirk, a scream). Plain clothes, no logos. Never a real person's likeness: for a real person keep the face out of close-ups (hands, back, side, distance, straight top-down). - CALM shots only. Bodies in fast action or contact (running, jumping, fighting, falling) look stiff, so tell the story with frozen moments under a slowly moving camera, macro inserts of hands and objects, rooms, vehicles, the sky, top-down views and the aftermath. - One new picture for every caption chunk (2-4 words, about 1-1.5 s): a new framing, a new action or a camera move. Nothing ever holds still: every shot drifts, pushes in or orbits. - The shot kit: macro insert (extreme close-up of the key object or hands, very shallow depth of field), hero low angle (wide lens at waist height), god's-eye top-down with a slow rotate, tracked object, over the shoulder, smash zoom on the word that matters, whip pan between beats, glow highlight (the named object glows orange with a rim, then fades), a big 3D number lying in the scene, a thought bubble (a cloud ring showing what he hoped for), a TV in the scene replaying the event. - Light: a bright warm sun, blue sky with clouds, soft shadows. Night scenes use hard lamps and a dark sky. Wide lens up close, depth of field on every shot, motion blur on every fast camera move. - Captions: 2-4 words at a time, white bold rounded sans, about 64 px on a 1080 px-wide frame, centred, 78% down the frame, soft dark shadow, no box. Each chunk replaces the last on the same frame as its picture. - Banned: real footage, photos, downloaded textures or 3D models, logos, AI images or AI video, downloaded music, gore. 38-50 s, 100-140 words, 150-180 words a minute. The voice starts at 0.0 s. - Line 1: the most surprising physical act, in past tense, opening with "He" (or "She"). No name, date or setup. Frame 1 shows that act. - Then a cause-and-effect chain: every sentence starts with a hinge word (Then, Suddenly, So, Because, But, Later) and is ONE visible action. - The context arrives late, after "You see,". - A reveal: someone looks closer and the story turns. - The last line is the consequence. No call to action, no question. - Names are rare: use one only when it pays off. Title = "He ..." or "The Man Who ...". The cover is the first frame. Save every stage to disk so a crash never costs more than one stage. - sources.md: every fact with its link and the date you read it. Anything you can't verify stays out. - script.json: the lines and the caption chunks cut by hand (2-4 words, one picture each), each chunk with its shot: the set, the camera, what moves. - Kokoro reads the whole script in one take (speed 1.0-1.1) and pauses are squeezed to 0.35 s at most; faster-whisper times every word (pass it a 16 kHz numpy array, not a path); write chunks.json with each chunk's start and end. A picture and its caption switch together on frame round(start*30)+1. - blender/: Python scripts run headless (blender -b -P). makes one .blend with the sets, the cast, the props and every shot baked with a key on every frame. sets EEVEE with ray tracing, the AgX view with the Punchy look, motion blur, depth of field and bloom in the compositor. Bodies come from MPFB or are built by you; clothes, hair, props and sets are built in code. - Stills first: one half-size still per shot as a contact sheet. Look at every tile and fix what's off before any full render. Time one full-size frame and tell me the estimate. - Render 1080x1920 at 30 fps to PNG frames, skipping frames already on disk. - numpy only: a light music bed, a sound on every action (a whoosh on each whip pan, hits, wind, room tone), ducked under the voice, mastered to -14 LUFS on the whole stereo mix. - burn the captions in from chunks.json. Encode out/_9x16.mp4 (1080x1920, 30 fps, TV-range yuv420p) and out/_x_7x8.mp4 (1008x1152, under 10 MB, 2-pass): a crop of the 9:16 frame that keeps the subject and the caption. - Blender paths: pass absolute output paths (a relative one lands on the drive root). - MPFB headless: enable it with addon_utils.enable("bl_ext.user_default.mpfb"). Bake a body's shape keys before you pose it: the rig is fitted to the baked shape. - Clothes: cut them from the body mesh and push them 7-16 mm out along the normals; keep two rings of body faces under every hem. - Nothing coplanar: keep every resting object 1 cm off its neighbour and draw ground markings into the texture, or they flicker. - Hair: EEVEE draws hair only from a Curves object, hair type STRIP. Use a Principled BSDF on it (the Hair BSDF renders black). - Skin: keep subsurface weight near 0.2 (more turns faces black in EEVEE) and add a soft fill light from below under a noon sun. - Hands: a curled hand reads as a claw from the palm side, so shoot a grip from behind the knuckles. - Motion blur: shutter position START, a key on every frame, plus a key half a frame after each shot's last frame, so blur never reaches across a cut. - Check: run a black-frame and one-frame-flicker check on the render. Make a contact sheet (a frame every 0.5 s) and look at every tile: the act reads in the first 2 s, faces are lit, nothing clips through anything, captions stay off faces and key objects. Then step through the first 2 s, the middle and the end frame by frame. Fix, render only the changed shots again. Pitch (or take my story), then show me the script with the word count, the caption chunks with their seconds and the shot list with what the first frame shows. Build nothing until I say go. how the two runs went: one run each on the same PC, so it's one sample, not a ranking. both took under 3 hours on one shared 8 GB graphics card. the story for both: a man who tied 42 weather balloons to a lawn chair in 1982 where they split: -> Opus opened wide on the balloons over the yard. Fable opened on the man kneeling at the chair with his Jeep behind him -> Opus put 5 big 3D numbers in its scenes, Fable 2 -> Fable followed the caption spec (white text, soft shadow). Opus went ALL CAPS with a thick outline -> Fable pulled the Smithsonian's own record and ended on the chair in the museum. Opus ended on the $1,500 fine -> a few Opus shots are hazy and one opens on a blank grey frame save this for ur next Short. who did it better, Opus or Fable?

You are early $LAYA CA:0xe394f71427d23c28090aa9a26b28c577c510f61a ◦ the parser: live, reading every Pons launch it reaches ◦ the weights: open, every result replays to zero difference ◦ the next layer: official with @nillion, announcement incoming Stay close to this account. $LAYA on @ponsdotfamily

Hi! Recapping some changes we have rolled out over the last couple of weeks that have further reduced the risk associated to potentially destructive actions being performed by Codex during its work. A few weeks ago, we started investigating a small number of reports where GPT-5.6 in Codex took destructive actions outside what the user asked for. The most serious pattern we found was a command meant to clean up temporary work that could instead delete the user files. This should obviously not happen. Here’s what we found: - Codex sometimes creates temporary folders while working and cleans them up afterward. In rare cases, GPT-5.6 got that cleanup wrong. One pattern involved reusing a system environment variable like $HOME for temporary work. A malformed cleanup command could then point at the actual home directory instead of the temporary folder. - There were cases where the model tried to delete or overwrite a temporary path without checking what was already there. We’ve added protections at several layers: - Codex is now explicitly instructed to check deletion targets before acting, create fresh temporary directories, avoid repurposing system environment variables, prefer recoverable actions, and stop when the scope is unclear. - We strengthened the execution checks that identify high-risk deletion commands and escalate them for review. If a command is rejected, the model is directed to take a safer approach. - We made Full access harder to enable accidentally, added clearer warnings, and further restricted especially risky permission combinations. - We updated Auto-review to better identify destructive actions. - We built targeted evaluations that replay the failures we observed. We’re also adding reinforcement-learning tasks and graders focused on these risks, and filtering destructive actions from training data. In those replay evaluations, the changes substantially reduced the behavior while preserving Codex’s ability to complete normal coding work. Two things to do on your end: - Keep the Codex app up to date. We are always improving safety, performance and many other things. - Use one of the sandbox modes: "Ask for approval" or "Approve for me". Only use Full access for environments you trust and can recover. Thanks and happy Codexing out there!

so i was looking at the source code of kaito pulse and found some interesting things: - it fingerprints your device. it hashes how your gpu renders an invisible image, your gpu model, and how your hardware handles a test tone. that combo is unique to your laptop and it doesn't change - it goes up with your twitter id attached, so every x account you use on that laptop points back to the same machine - they can see what you see and replay your whole session. every post your feed served you including the ones you scrolled straight past, how many milliseconds you spent on each, every click, follow and unfollow, and a ping every 30 seconds with idle detection so they know exactly when you were on and for how long. not just your timeline either, it covers your search results and your bookmarks - it reads your claude and chatgpt subscriptions, which plan you pay for and what percent of your rate limit you've burned (not your conversations though) - on chatgpt it opens your settings page and clicks through to Usage by itself - on binance it clicks the Positions tab for you, then re-sends your logged in requests to read wallet balances, futures positions, pnl, and deposit and withdrawal history - the zktls is a fork of a project called primus the 12 domains it runs on: - markets: binance, okx, bybit, hyperliquid, lighter, polymarket, tradexyz, variational - ai: chatgpt, claude - social: X (all the telemetry is here) - kaito (only site allowed to talk to the extension)

deepseek-harness重磅开源，采用了一切皆插件的架构，也就是中国版的openclaw，体验了下可以称之为私有化的workbuddy。 场景侧冲击：有了deepseek-harness之后，企业域的私有化code、cowork基本用这个就可以了，直接会冲击workbuddy之类的市场。 之前DeepSeek Harness内测招募，本质是一次开源 Agent 生态大摸底，把全球 Agent 基础设施的家底扫了一遍后，预计后续版本应该会在开放生态上做一些事情。 dsh看点不在"又一个 Claude Code"，有3个看点： 1️⃣ 一切皆插件 模型、工具、skill、会话、沙箱、存储、主循环、调度，连 UI 都是插件，配置里全都可换，不动源码。四种运行模式里最有意思的是 Minimal——只留一个 bash 和一个文件编辑器，摆明了是给模型做裸机 benchmark 的。 2️⃣ 每一次运行都可回放 系统提示、推理过程、工具调用与结果、子 agent 调度、所有 context 注入，全部进 append-only 的 session log，resume / fork / search / replay 都跑在同一条事件流上。 3️⃣ 最狠的是内核 dsh 跑在 Cordis 上，它把插件系统拆成两个正交维度：时间可组合性：卸载一个组件时副作用能完整回滚（每个 context 变换都带一个逆，运行时来追踪）；空间可组合性：依赖可声明、且 context 一变就反向通知组件。还给了一套动态组合的演算，证明这个性质能从单个组件传导到整个系统。 总的来说，这次deepseek-harness最大的特点是极致的可替换性，也就是用最小的基础来承载，然后其他都是可插拔的插头。 #DeepSeek# #harness# #Cordis# #编程范式# #北大# #开源#