Essays on AI, video, and the craft of building things, from an engineer who reads too much Dostoevsky.https://vinsonli.com/https://vinsonli.com/posts/what-would-it-cost-to-make-breaking-bad-with-ai/https://vinsonli.com/posts/what-would-it-cost-to-make-breaking-bad-with-ai/A line-by-line estimate for one 47-minute episode using this year's public video generation prices. Generating the footage is only part of the budget; performance and continuity remain much harder to get right.Fri, 11 Sep 2026 00:00:00 GMTvideoeconomicsfilmmakingmicrodramahttps://vinsonli.com/posts/taste-without-a-verifier/https://vinsonli.com/posts/taste-without-a-verifier/Coding agents somehow learned to build good-looking websites, although nothing checks whether a design is beautiful. What video editing can borrow from that: preference models, critics and loops.Fri, 21 Aug 2026 00:00:00 GMTtasteagentscreative-toolsvideohttps://vinsonli.com/posts/learning-like-a-baby/https://vinsonli.com/posts/learning-like-a-baby/Google DeepMind's Gemini Robotics ER 2 gives robots a better high-level brain. The part I still think nobody has built is the body-first learning underneath. Here's the research plan I'd run.Fri, 31 Jul 2026 00:00:00 GMTworld-modelsembodimentroboticscuriosityresearchhttps://vinsonli.com/posts/from-embeddings-to-ids-to-what-you-watch-next/https://vinsonli.com/posts/from-embeddings-to-ids-to-what-you-watch-next/Generative recommenders are going into production. How a video becomes an embedding, then a semantic ID, and how a Transformer uses those IDs to choose what you'll watch next.Thu, 18 Jun 2026 00:00:00 GMTrecommendationsemantic-idstransformervideohttps://vinsonli.com/posts/change-everything-except-the-motion/https://vinsonli.com/posts/change-everything-except-the-motion/At I/O this week, YouTube brought Gemini Omni to Shorts Remix: take a Short, change the characters, setting or style, and keep the motion and story. Why remix is the right first experience for video-to-video.Fri, 22 May 2026 00:00:00 GMTvideogenerative-modelsyoutuberemixcreatorshttps://vinsonli.com/posts/sora-is-shutting-down/https://vinsonli.com/posts/sora-is-shutting-down/OpenAI closed the Sora app this week, seven months after launching it as a social network, and the API ends in September. What went wrong, and what it says about where generative video belongs.Wed, 29 Apr 2026 00:00:00 GMTvideogenerative-modelsproductsocialhttps://vinsonli.com/posts/a-billion-dollars-for-jepa/https://vinsonli.com/posts/a-billion-dollars-for-jepa/Yann LeCun's AMI Labs raised $1.03 billion to build world models on the JEPA framework. Revisiting what I wrote in 2022, and the one thing I still think the plan is missing.Fri, 13 Mar 2026 00:00:00 GMTworld-modelsjepastartupsembodimenthttps://vinsonli.com/posts/the-model-plans-the-shots-now/https://vinsonli.com/posts/the-model-plans-the-shots-now/ByteDance's Seedance 2.0 generates multi-shot sequences with references for characters, props and sound, and triggered cease-and-desist letters within a day. What's left for the editor, and for rights holders.Tue, 17 Feb 2026 00:00:00 GMTvideogenerative-modelsfilmmakingcopyrighthttps://vinsonli.com/posts/a-weekend-with-the-world-api/https://vinsonli.com/posts/a-weekend-with-the-world-api/World Labs opened an API for Marble. I spent the weekend building a small workbench for rendering its worlds as Gaussian splats and authoring camera paths through them. What broke: scale, coordinates and drift.Sun, 25 Jan 2026 00:00:00 GMTworld-models3dgaussian-splattingside-projectshttps://vinsonli.com/posts/video-editing-isnt-a-knapsack-problem/https://vinsonli.com/posts/video-editing-isnt-a-knapsack-problem/I built an agent that edits my social videos end to end using knowledge of my personal taste. What that asks of the timeline, the feedback loop and the product.Tue, 16 Dec 2025 00:00:00 GMTvideoeditingagentscreative-toolshttps://vinsonli.com/posts/lecun-leaves-marble-ships/https://vinsonli.com/posts/lecun-leaves-marble-ships/In one week World Labs launched Marble, Yann LeCun confirmed he's leaving Meta to start a world-model company, and Google shipped Gemini 3. The world-model bet went public from three directions.Sun, 23 Nov 2025 00:00:00 GMTworld-modelsjepa3dindustryhttps://vinsonli.com/posts/sora-built-a-social-network/https://vinsonli.com/posts/sora-built-a-social-network/OpenAI launched Sora 2 as a TikTok-style app where every video is generated. The model is impressive. I'm less sure a feed of generations gives people a reason to keep watching.Mon, 06 Oct 2025 00:00:00 GMTvideogenerative-modelssocialproducthttps://vinsonli.com/posts/consistent-characters-are-the-unlock/https://vinsonli.com/posts/consistent-characters-are-the-unlock/Google's new image model, known everywhere as Nano Banana, keeps a person looking like themselves across edits and scenes. For AI storytelling, identity preservation matters more than image quality.Sun, 14 Sep 2025 00:00:00 GMTgenerative-modelsimagesstorytellingconsistencyhttps://vinsonli.com/posts/genie-3-remembers-where-you-painted-the-wall/https://vinsonli.com/posts/genie-3-remembers-where-you-painted-the-wall/DeepMind's Genie 3 generates interactive worlds in real time at 720p that stay consistent for minutes. Persistence is the test for a world model, and it just got much better.Sat, 09 Aug 2025 00:00:00 GMTworld-modelsdeepmindvideointeractivehttps://vinsonli.com/posts/sixty-episodes-one-minute-each/https://vinsonli.com/posts/sixty-episodes-one-minute-each/Microdramas are one of the fastest-growing forms of entertainment, and the format exists because of production cost. What happens when AI changes the cost?Sun, 27 Jul 2025 00:00:00 GMTvideomicrodramaeconomicsgenerative-modelshttps://vinsonli.com/posts/62-hours-of-robot-data/https://vinsonli.com/posts/62-hours-of-robot-data/Meta's V-JEPA 2 learns a world model from a million hours of video, then learns to plan robot actions from 62 hours of robot data. Promising, and still missing touch.Mon, 16 Jun 2025 00:00:00 GMTworld-modelsjeparoboticsvideohttps://vinsonli.com/posts/veo-3-has-sound/https://vinsonli.com/posts/veo-3-has-sound/Google's Veo 3 generates video with synchronized dialogue, sound effects and ambient audio. Sound turns clips into scenes. Also at I/O: a language model that writes by denoising.Sat, 24 May 2025 00:00:00 GMTvideoaudiogenerative-modelsgooglehttps://vinsonli.com/posts/what-a-minute-of-ai-video-costs/https://vinsonli.com/posts/what-a-minute-of-ai-video-costs/Real per-second prices for Veo 2, Kling 2.0 and Jimeng this month, and why the retake rate matters more than the sticker price.Thu, 24 Apr 2025 00:00:00 GMTvideoeconomicsgenerative-modelschinahttps://vinsonli.com/posts/style-is-free-now-taste-isnt/https://vinsonli.com/posts/style-is-free-now-taste-isnt/GPT-4o's image generation turned the internet into Studio Ghibli for a week. When every style is one prompt away, the scarce thing is judgment.Sat, 29 Mar 2025 00:00:00 GMTgenerative-modelstastecreative-toolsimageshttps://vinsonli.com/posts/helix-and-my-golf-swing/https://vinsonli.com/posts/helix-and-my-golf-swing/Figure's Helix splits robot control into a slow model that understands and a fast one that moves. A golf swing works the same way, which is why thinking during it ruins it.Mon, 24 Feb 2025 00:00:00 GMTroboticscontrolgolfhumanoidshttps://vinsonli.com/posts/deepseek-and-the-price-of-intelligence/https://vinsonli.com/posts/deepseek-and-the-price-of-intelligence/A Chinese lab released a reasoning model that matches OpenAI's o1 and published how it did it. Nvidia lost about $600 billion in a day. Where the efficiency actually came from.Tue, 28 Jan 2025 00:00:00 GMTlanguagereasoningchinaeconomicshttps://vinsonli.com/posts/genie-2-and-world-labs-in-the-same-week/https://vinsonli.com/posts/genie-2-and-world-labs-in-the-same-week/DeepMind's Genie 2 turns one image into a playable 3D world, and Fei-Fei Li's World Labs turns one image into a 3D scene you can walk through. Worlds are the next medium after text, images and video.Sun, 08 Dec 2024 00:00:00 GMTworld-models3dgenerative-modelsdeepmindhttps://vinsonli.com/posts/describe-the-music-you-want/https://vinsonli.com/posts/describe-the-music-you-want/"Sad songs for a rainy drive" beats any genre taxonomy. Natural language is becoming the way people ask for music, and it changes what a music recommender has to understand.Tue, 19 Nov 2024 00:00:00 GMTmusicrecommendationlanguageproducthttps://vinsonli.com/posts/the-physics-nobel-went-to-neural-networks/https://vinsonli.com/posts/the-physics-nobel-went-to-neural-networks/Hopfield and Hinton won the physics prize, and Hassabis and Jumper shared chemistry for AlphaFold. A mechanical engineer's reaction to physics and AI converging.Sat, 12 Oct 2024 00:00:00 GMTaiphysicssciencehistoryhttps://vinsonli.com/posts/ais-bottleneck-is-a-power-plant/https://vinsonli.com/posts/ais-bottleneck-is-a-power-plant/Microsoft signed a 20-year deal to restart a reactor at Three Mile Island to power its data centers. The constraint on AI is moving from chips to electricity, and baseload nuclear is suddenly back.Tue, 24 Sep 2024 00:00:00 GMTenergyinfrastructureeconomicsnuclearhttps://vinsonli.com/posts/a-neural-net-runs-doom/https://vinsonli.com/posts/a-neural-net-runs-doom/Google Research's GameNGen simulates Doom in real time with a diffusion model: no game engine, just next-frame prediction conditioned on your inputs. Interactive video is a world model.Sat, 31 Aug 2024 00:00:00 GMTworld-modelsgamesdiffusionvideohttps://vinsonli.com/posts/search-plus-verification/https://vinsonli.com/posts/search-plus-verification/AlphaProof and AlphaGeometry 2 together solved four of six problems from this year's Math Olympiad, reaching silver-medal level. What makes a checkable answer so useful for training, and what happens when there isn't one.Mon, 29 Jul 2024 00:00:00 GMTreasoningdeepmindreinforcement-learningtastehttps://vinsonli.com/posts/kling-came-from-a-short-video-company/https://vinsonli.com/posts/kling-came-from-a-short-video-company/Kuaishou, TikTok's main rival in China, released a video model that rivals Sora's samples, and ordinary users in China can already try it. Video models get built by whoever has the video.Wed, 12 Jun 2024 00:00:00 GMTvideogenerative-modelschinashort-videohttps://vinsonli.com/posts/gpt-4o-hears-you-laugh/https://vinsonli.com/posts/gpt-4o-hears-you-laugh/OpenAI and Google both showed real-time multimodal assistants this week. Speech as a native modality changes what a conversation with a model is, and latency turns out to be the feature.Fri, 17 May 2024 00:00:00 GMTmultimodalvoiceagentsproducthttps://vinsonli.com/posts/music-generations-gpt-3-moment/https://vinsonli.com/posts/music-generations-gpt-3-moment/Suno v3 and Udio generate full songs with vocals and lyrics from a sentence, and some of them are good. What that means for artists, platforms and listeners.Sun, 14 Apr 2024 00:00:00 GMTmusicgenerative-modelscreatorshttps://vinsonli.com/posts/genie-learned-to-play-from-videos/https://vinsonli.com/posts/genie-learned-to-play-from-videos/DeepMind's Genie learned a controllable world model from 2D platformer videos with no action labels, by inferring eight latent actions on its own. The model learns the controls as well as the game.Tue, 05 Mar 2024 00:00:00 GMTworld-modelsvideodeepmindgameshttps://vinsonli.com/posts/world-simulator-is-doing-a-lot-of-work/https://vinsonli.com/posts/world-simulator-is-doing-a-lot-of-work/OpenAI's Sora generates minute-long videos that look astonishing. Its technical report calls it a world simulator. The samples show both why that's tempting and why it isn't true yet.Mon, 19 Feb 2024 00:00:00 GMTvideogenerative-modelsworld-modelsopenaihttps://vinsonli.com/posts/mobile-aloha-cooks-shrimp/https://vinsonli.com/posts/mobile-aloha-cooks-shrimp/Stanford's two-armed robot on a wheeled base learned to cook, wipe spills and call an elevator from about fifty demonstrations per task. Cheap hardware plus teleoperation is changing robot data.Tue, 09 Jan 2024 00:00:00 GMTroboticsimitation-learningdatahttps://vinsonli.com/posts/natively-multimodal/https://vinsonli.com/posts/natively-multimodal/Google announced Gemini, trained from the start on text, images, audio and video together. Why training on mixed modalities from day one matters more than bolting vision onto a language model.Tue, 12 Dec 2023 00:00:00 GMTmultimodalgeminigooglearchitecturehttps://vinsonli.com/posts/four-second-clips-cant-make-a-movie/https://vinsonli.com/posts/four-second-clips-cant-make-a-movie/Pika 1.0, Stable Video Diffusion and Runway's latest make beautiful short clips. What's missing is the grammar of film: continuity, screen direction and eyelines.Tue, 28 Nov 2023 00:00:00 GMTvideogenerative-modelsfilmmakingcreative-toolshttps://vinsonli.com/posts/the-model-rewrites-your-prompt/https://vinsonli.com/posts/the-model-rewrites-your-prompt/DALL·E 3 in ChatGPT doesn't use what you typed. It has a language model write a long, detailed prompt for you. Creative tools are turning into agents that interpret intent.Thu, 19 Oct 2023 00:00:00 GMTgenerative-modelscreative-toolsagentsopenaihttps://vinsonli.com/posts/optimus-sorts-blocks-a-toddler-falls-down/https://vinsonli.com/posts/optimus-sorts-blocks-a-toddler-falls-down/Tesla's new video shows Optimus sorting colored blocks with a neural network trained end to end. Curated robot demos and messy child learning, and which one scales.Sat, 30 Sep 2023 00:00:00 GMTroboticshumanoidslearningembodimenthttps://vinsonli.com/posts/gaussian-splatting-killed-my-mesh-nostalgia/https://vinsonli.com/posts/gaussian-splatting-killed-my-mesh-nostalgia/3D Gaussian Splatting represents a scene as millions of fuzzy, colored blobs and renders it in real time at NeRF quality. It does this without a neural network.Thu, 24 Aug 2023 00:00:00 GMT3dgraphicsrepresentationgaussian-splattinghttps://vinsonli.com/posts/predicting-in-representation-space/https://vinsonli.com/posts/predicting-in-representation-space/Meta's I-JEPA is the first concrete result from LeCun's world model agenda. It learns image representations by predicting hidden regions in latent space, with no augmentations and no pixel reconstruction.Tue, 11 Jul 2023 00:00:00 GMTjepaself-supervised-learningrepresentationworld-modelshttps://vinsonli.com/posts/vision-pro-scans-your-face-to-make-you/https://vinsonli.com/posts/vision-pro-scans-your-face-to-make-you/Apple's headset builds a 3D 'Persona' of your face so you can appear on video calls while wearing it. Ten years after my thesis, 3D face capture ships inside a headset.Thu, 08 Jun 2023 00:00:00 GMTfaces3dvrappleavatarshttps://vinsonli.com/posts/recommendation-as-next-token-prediction/https://vinsonli.com/posts/recommendation-as-next-token-prediction/A new paper turns every item into a short code of semantic tokens, then has a Transformer generate the code of what you'll want next. The item vocabulary finally describes what things are.Sun, 21 May 2023 00:00:00 GMTrecommendationsemantic-idstransformerrepresentationhttps://vinsonli.com/posts/fake-drake/https://vinsonli.com/posts/fake-drake/An AI-generated song imitating Drake and The Weeknd got millions of plays before it was pulled. Voice is identity, and the industry needs consent and attribution systems, fast.Tue, 25 Apr 2023 00:00:00 GMTmusicgenerative-modelsvoiceethicshttps://vinsonli.com/posts/gpt-4-reads-the-picture/https://vinsonli.com/posts/gpt-4-reads-the-picture/OpenAI's GPT-4 scores near the top of the bar exam and can explain a joke in a photo. Describing a scene is useful, but I'm still unsure how far that gets it toward understanding the physical world.Fri, 17 Mar 2023 00:00:00 GMTlanguagemultimodalopenaigpt-4https://vinsonli.com/posts/control-beats-prompts/https://vinsonli.com/posts/control-beats-prompts/ControlNet lets you steer Stable Diffusion with a pose skeleton, a depth map or an edge sketch. Creators want to set the structure directly, and this gives them a way to.Mon, 20 Feb 2023 00:00:00 GMTgenerative-modelsdiffusioncreative-toolshttps://vinsonli.com/posts/text-to-music-is-a-representation-problem/https://vinsonli.com/posts/text-to-music-is-a-representation-problem/Google Research's MusicLM generates music from text descriptions. The interesting part is its stack of tokens: one for meaning, one for sound, one shared between music and words.Tue, 31 Jan 2023 00:00:00 GMTmusicgenerative-modelsaudiorepresentationhttps://vinsonli.com/posts/chatgpt-is-a-product-not-a-model/https://vinsonli.com/posts/chatgpt-is-a-product-not-a-model/A million people signed up in five days. The underlying model isn't new. What's new is the interface and the training to follow instructions, and that changes every software team.Tue, 06 Dec 2022 00:00:00 GMTlanguagechatgptproductopenaihttps://vinsonli.com/posts/optimus-walked-on-stage-watch-its-hands/https://vinsonli.com/posts/optimus-walked-on-stage-watch-its-hands/A year after the dancer in the suit, Tesla showed a real humanoid prototype. It walked slowly and waved. The legs get the attention. The hands are the harder and more important problem.Tue, 04 Oct 2022 00:00:00 GMTroboticshumanoidsteslahandshttps://vinsonli.com/posts/text-to-video-is-next/https://vinsonli.com/posts/text-to-video-is-next/Meta's Make-A-Video generates short clips from a sentence. They're five seconds long, low resolution and physically wrong in instructive ways. The hard parts are consistency, continuity and physics.Fri, 30 Sep 2022 00:00:00 GMTgenerative-modelsvideodiffusionhttps://vinsonli.com/posts/stable-diffusion-runs-on-my-own-computer/https://vinsonli.com/posts/stable-diffusion-runs-on-my-own-computer/Stability AI released the weights of a text-to-image model anyone can run on a consumer GPU. Open models change who gets to build, and what gets built.Sat, 27 Aug 2022 00:00:00 GMTgenerative-modelsdiffusionopen-sourcehttps://vinsonli.com/posts/lecuns-path-read-carefully/https://vinsonli.com/posts/lecuns-path-read-carefully/Yann LeCun's position paper argues that intelligence needs world models that predict in representation space, not pixels. Where I agree, and where I'd push back: bodies and hands.Sun, 03 Jul 2022 00:00:00 GMTworld-modelsjepaself-supervised-learningresearchhttps://vinsonli.com/posts/it-learned-minecraft-by-watching-youtube/https://vinsonli.com/posts/it-learned-minecraft-by-watching-youtube/OpenAI's VPT labeled 70,000 hours of Minecraft videos with the actions players took, using a small model trained on a little labeled data. Passive video became interaction data.Sun, 26 Jun 2022 00:00:00 GMTagentsvideolearningopenaihttps://vinsonli.com/posts/one-network-604-tasks/https://vinsonli.com/posts/one-network-604-tasks/DeepMind's Gato plays Atari, captions images, chats and stacks blocks with a real robot arm, all with the same weights. Turning actions into tokens is the interesting part.Mon, 16 May 2022 00:00:00 GMTagentstransformerroboticsdeepmindhttps://vinsonli.com/posts/images-are-solved-ish-video-is-where-physics-lives/https://vinsonli.com/posts/images-are-solved-ish-video-is-where-physics-lives/DALL·E 2 generates images that look like real photos and paintings from a sentence. Why the jump to video is much harder than the jump from GANs to this.Tue, 12 Apr 2022 00:00:00 GMTgenerative-modelsdiffusionvideophysicshttps://vinsonli.com/posts/weve-been-undertraining/https://vinsonli.com/posts/weve-been-undertraining/DeepMind's Chinchilla paper says large language models have far too many parameters for the data they see. The fix is more data, and that raises a question about where it comes from.Thu, 31 Mar 2022 00:00:00 GMTlanguagescalingdeepminddatahttps://vinsonli.com/posts/metas-worst-day-was-tiktoks-best/https://vinsonli.com/posts/metas-worst-day-was-tiktoks-best/Meta lost over $230 billion in market value in one day after Facebook's first ever drop in daily users. The interest graph beat the social graph.Sat, 05 Feb 2022 00:00:00 GMTrecommendationshort-videosocialbusinesshttps://vinsonli.com/posts/nerf-went-from-hours-to-seconds/https://vinsonli.com/posts/nerf-went-from-hours-to-seconds/Nvidia's Instant NGP trains a neural radiance field in seconds with a multiresolution hash table. The representation mattered more than the network.Mon, 24 Jan 2022 00:00:00 GMT3dgraphicsrepresentationnerfhttps://vinsonli.com/posts/diffusion-is-going-to-eat-gans/https://vinsonli.com/posts/diffusion-is-going-to-eat-gans/Two papers this week, GLIDE and latent diffusion, make text-to-image generation with diffusion models look practical. What denoising actually learns, and why it beats the adversarial game.Tue, 28 Dec 2021 00:00:00 GMTgenerative-modelsdiffusionimageshttps://vinsonli.com/posts/facebook-deletes-a-billion-faceprints/https://vinsonli.com/posts/facebook-deletes-a-billion-faceprints/Meta is shutting down Facebook's face recognition system and deleting more than a billion face templates. An industry chapter closing, from someone who lived in it.Sat, 06 Nov 2021 00:00:00 GMTfacesprivacypolicyhttps://vinsonli.com/posts/the-metaverse-needs-faces-first/https://vinsonli.com/posts/the-metaverse-needs-faces-first/Facebook is now Meta and says it's building the metaverse. From someone who spent years on 3D faces: avatars that can't emote are just chat rooms with legs.Sat, 30 Oct 2021 00:00:00 GMTfacesvravatars3dhttps://vinsonli.com/posts/listening-sessions-are-sentences/https://vinsonli.com/posts/listening-sessions-are-sentences/Treat each song as a token and each listening session as a sentence, and recommendation starts to look like language modeling. What that framing gets right, and what it misses.Mon, 27 Sep 2021 00:00:00 GMTrecommendationtransformermusicsequence-modelshttps://vinsonli.com/posts/why-the-humanoid-form-matters/https://vinsonli.com/posts/why-the-humanoid-form-matters/Tesla's AI Day announced a humanoid robot and then showed a person dancing in a robot costume. The demo was a joke. The idea is right.Mon, 23 Aug 2021 00:00:00 GMTroboticsembodimentteslahumanoidshttps://vinsonli.com/posts/you-cant-skim-a-song/https://vinsonli.com/posts/you-cant-skim-a-song/Why music recommendation is a different problem from video or news recommendation: sequences, repetition, passive listening, and the difference between a skip and a mood.Sun, 18 Jul 2021 00:00:00 GMTmusicrecommendationproducthttps://vinsonli.com/posts/copilot-finished-my-function/https://vinsonli.com/posts/copilot-finished-my-function/GitHub's Copilot preview suggests whole functions inside the editor. Two days of using it on side projects, and what changes when the IDE becomes a conversation.Tue, 06 Jul 2021 00:00:00 GMTcodinglanguagetoolsopenaihttps://vinsonli.com/posts/joining-youtube-music/https://vinsonli.com/posts/joining-youtube-music/I've started as an engineering manager at YouTube Music. Why music, and why I think the next few years of music discovery will look very different.Wed, 26 May 2021 00:00:00 GMTcareermusicrecommendationyoutubehttps://vinsonli.com/posts/closing-two-chapters/https://vinsonli.com/posts/closing-two-chapters/I'm leaving TTT Studios after seven and a half years and Amanda AI after three. What I learned growing an engineering team from ten to fifty, and co-founding a company through a pandemic.Thu, 29 Apr 2021 00:00:00 GMTcareermanagementstartupsamanda-aihttps://vinsonli.com/posts/one-architecture-any-input/https://vinsonli.com/posts/one-architecture-any-input/DeepMind's Perceiver handles images, audio, video and point clouds with the same network, by cross-attending to a small latent array. Modalities stop needing their own models.Sun, 14 Mar 2021 00:00:00 GMTarchitecturemultimodaltransformerdeepmindhttps://vinsonli.com/posts/ten-years-between-albums/https://vinsonli.com/posts/ten-years-between-albums/万能青年旅店 released their second album after ten years. On listening to a band that sings about industrial decline in Hebei, and why art gets better once you can see the structure under it.Sun, 07 Feb 2021 00:00:00 GMTmusicchinaculturehttps://vinsonli.com/posts/pictures-and-words-in-the-same-space/https://vinsonli.com/posts/pictures-and-words-in-the-same-space/OpenAI's CLIP learns from 400 million image and caption pairs to put images and text in one embedding space. Zero-shot classification is the demo. Shared embeddings are the real story.Mon, 11 Jan 2021 00:00:00 GMTmultimodalrepresentationcomputer-visionlanguagehttps://vinsonli.com/posts/alphafold-and-learning-physics-from-data/https://vinsonli.com/posts/alphafold-and-learning-physics-from-data/DeepMind's AlphaFold 2 predicts protein structures about as accurately as experiments do. A fifty-year-old physics problem, solved mostly by learning from examples.Thu, 03 Dec 2020 00:00:00 GMTsciencedeepmindbiologyphysicshttps://vinsonli.com/posts/an-image-is-worth-16x16-words/https://vinsonli.com/posts/an-image-is-worth-16x16-words/A paper under review at ICLR cuts images into patches and feeds them to a plain Transformer. With enough data, it beats convolutional networks. One architecture for everything is getting closer.Mon, 09 Nov 2020 00:00:00 GMTcomputer-visiontransformerarchitecturedeep-learninghttps://vinsonli.com/posts/a-whole-scene-stored-inside-a-network/https://vinsonli.com/posts/a-whole-scene-stored-inside-a-network/NeRF represents a 3D scene as a small neural network you can render from any viewpoint. No mesh, no triangles. My old thesis problem, answered from a different direction.Thu, 08 Oct 2020 00:00:00 GMT3dgraphicsrepresentationdeep-learninghttps://vinsonli.com/posts/short-video-starts-in-india/https://vinsonli.com/posts/short-video-starts-in-india/YouTube opened a beta of Shorts in India, where TikTok has been banned since June. Launch where the gap is, with the world's biggest music catalog behind you.Sat, 19 Sep 2020 00:00:00 GMTshort-videoyoutubeproductindiahttps://vinsonli.com/posts/reels-is-a-clone-thats-fine/https://vinsonli.com/posts/reels-is-a-clone-thats-fine/Instagram launched Reels the day before the TikTok executive order. The format is a commodity now, and the question is whose recommender learns fastest.Tue, 11 Aug 2020 00:00:00 GMTshort-videorecommendationproductsocialhttps://vinsonli.com/posts/gpt-3-and-the-react-component/https://vinsonli.com/posts/gpt-3-and-the-react-component/The demos flooding Twitter this week come from the same next-word objective as GPT-2, scaled a hundred times. Few-shot learning appeared on its own. Coding changes first.Tue, 21 Jul 2020 00:00:00 GMTlanguagegpt-3codingopenaihttps://vinsonli.com/posts/big-tech-steps-back-from-face-recognition/https://vinsonli.com/posts/big-tech-steps-back-from-face-recognition/In one week, three of the biggest companies in the field paused or ended police sales. Where I agree, and why a pause isn't a policy.Sat, 13 Jun 2020 00:00:00 GMTfacespolicyethicssurveillancehttps://vinsonli.com/posts/what-a-face-company-does-when-nobody-gathers/https://vinsonli.com/posts/what-a-face-company-does-when-nobody-gathers/Three months into the pandemic, with no conferences on the calendar. What we're trying, what we decided not to build, and how it feels to run a startup whose market disappeared.Tue, 19 May 2020 00:00:00 GMTamanda-aistartupscovidproducthttps://vinsonli.com/posts/travis-scott-played-a-video-game/https://vinsonli.com/posts/travis-scott-played-a-video-game/Fortnite's Astronomical event drew 12 million concurrent players. With every venue closed, a game became the biggest concert stage in the world.Tue, 28 Apr 2020 00:00:00 GMTmusicgamesvirtual-eventscovidhttps://vinsonli.com/posts/video-is-the-new-floor/https://vinsonli.com/posts/video-is-the-new-floor/Two weeks into working from home, every interaction is a video call. What that does to expectations for video, and what's missing from the tools.Fri, 27 Mar 2020 00:00:00 GMTvideocovidremote-workproducthttps://vinsonli.com/posts/mwc-is-cancelled/https://vinsonli.com/posts/mwc-is-cancelled/The world's biggest mobile conference was cancelled over the new coronavirus. A founder's worries, a few weeks in, with friends in Wuhan under lockdown.Fri, 14 Feb 2020 00:00:00 GMTamanda-aistartupscovideventshttps://vinsonli.com/posts/clearview-scraped-three-billion-faces/https://vinsonli.com/posts/clearview-scraped-three-billion-faces/The New York Times revealed a startup that matches any photo against billions of faces scraped from social media, and sells it to police. The worst version of my industry is now public.Tue, 21 Jan 2020 00:00:00 GMTfacesprivacysurveillanceamanda-aihttps://vinsonli.com/posts/the-decade-cameras-learned-to-see/https://vinsonli.com/posts/the-decade-cameras-learned-to-see/Looking back at 2010 to 2019 from the end of it: the moments that mattered to me, what I got right and wrong on this blog, and one bet for the 2020s.Mon, 30 Dec 2019 00:00:00 GMTretrospectiveaipredictionshttps://vinsonli.com/posts/muzero-learns-the-rules-it-isnt-given/https://vinsonli.com/posts/muzero-learns-the-rules-it-isnt-given/DeepMind's MuZero plays Go, chess, shogi and Atari at top level without being told the rules. It plans inside a model it learned, and the model only predicts what matters.Thu, 28 Nov 2019 00:00:00 GMTreinforcement-learningworld-modelsdeepmindplanninghttps://vinsonli.com/posts/200-seconds-vs-10000-years/https://vinsonli.com/posts/200-seconds-vs-10000-years/Google claims quantum supremacy with a 53-qubit chip. What the chip actually computed, why IBM disputes the number, and why there's still no workload.Sat, 26 Oct 2019 00:00:00 GMTquantumhardwarephysicsgooglehttps://vinsonli.com/posts/zao-put-your-face-in-a-movie/https://vinsonli.com/posts/zao-put-your-face-in-a-movie/A Chinese face-swap app went viral over the weekend and hit a privacy wall within days. China just ran the consumer deepfake experiment at scale.Tue, 03 Sep 2019 00:00:00 GMTfacesdeepfakeschinaprivacyhttps://vinsonli.com/posts/forty-different-lobbies/https://vinsonli.com/posts/forty-different-lobbies/A year of live events taught us that benchmark accuracy barely predicts real-world accuracy. What we changed in how we collect data and evaluate.Wed, 14 Aug 2019 00:00:00 GMTamanda-aimachine-learningdataoperationshttps://vinsonli.com/posts/the-embedding-table-is-the-model/https://vinsonli.com/posts/the-embedding-table-is-the-model/Facebook open-sourced DLRM, its deep learning recommendation model. It shows what big recommenders actually look like: mostly memory, with a small network on top.Sat, 20 Jul 2019 00:00:00 GMTrecommendationdeep-learninginfrastructurehttps://vinsonli.com/posts/datasets-have-a-consent-debt/https://vinsonli.com/posts/datasets-have-a-consent-debt/Microsoft quietly took down MS-Celeb-1M, ten million photos of a hundred thousand people. A lot of the face recognition industry was built on data nobody agreed to give.Tue, 11 Jun 2019 00:00:00 GMTfacesprivacydatasetsethicshttps://vinsonli.com/posts/san-francisco-banned-face-recognition/https://vinsonli.com/posts/san-francisco-banned-face-recognition/The ban covers city agencies, not private companies like ours. It still changed a lot of sales conversations this week. Where I think the line should be.Fri, 17 May 2019 00:00:00 GMTfacesprivacypolicyamanda-aihttps://vinsonli.com/posts/the-black-hole-photo-was-a-reconstruction-problem/https://vinsonli.com/posts/the-black-hole-photo-was-a-reconstruction-problem/The Event Horizon Telescope didn't take a picture in the ordinary sense. It filled in a mostly empty measurement with priors, and was careful about which ones.Sat, 13 Apr 2019 00:00:00 GMTimagingsciencecomputer-visionhttps://vinsonli.com/posts/the-bitter-lesson-read-by-a-hand-builder/https://vinsonli.com/posts/the-bitter-lesson-read-by-a-hand-builder/Rich Sutton says 70 years of AI research show that general methods plus computation beat human knowledge every time. He's right, and it stings. My one objection is about data.Mon, 18 Mar 2019 00:00:00 GMTairesearchessayshttps://vinsonli.com/posts/too-dangerous-to-release/https://vinsonli.com/posts/too-dangerous-to-release/OpenAI trained a much bigger language model and is holding back the full version. The unicorn story is impressive. What's actually dangerous is cheap, plausible text at scale.Tue, 19 Feb 2019 00:00:00 GMTlanguageai-safetyopenaigenerative-modelshttps://vinsonli.com/posts/alphastar-and-the-fairness-of-fast-hands/https://vinsonli.com/posts/alphastar-and-the-fairness-of-fast-hands/DeepMind's StarCraft agent beat two pros 10-0, then lost the one game where it had to move a camera like a person. What counts as intelligence when the body is different.Sun, 27 Jan 2019 00:00:00 GMTaigamesdeepmindembodimenthttps://vinsonli.com/posts/none-of-these-people-exist/https://vinsonli.com/posts/none-of-these-people-exist/Nvidia's StyleGAN generates photographic faces with control over pose, identity and freckles. A face company's view of faces becoming free.Wed, 19 Dec 2018 00:00:00 GMTfacesgenerative-modelsganshttps://vinsonli.com/posts/the-noisy-tv-problem/https://vinsonli.com/posts/the-noisy-tv-problem/A curious agent gets hypnotized by random static. OpenAI's Random Network Distillation fixes it and beats humans at Montezuma's Revenge. What curiosity should actually reward.Sat, 17 Nov 2018 00:00:00 GMTreinforcement-learningcuriosityexplorationhttps://vinsonli.com/posts/bert-reads-both-directions/https://vinsonli.com/posts/bert-reads-both-directions/Google's BERT beat almost every language benchmark by predicting hidden words using context on both sides. It's a representation model, which is a different thing from a text generator.Mon, 29 Oct 2018 00:00:00 GMTlanguagetransformerpretrainingrepresentationhttps://vinsonli.com/posts/thousands-of-faces-in-a-hotel-lobby/https://vinsonli.com/posts/thousands-of-faces-in-a-hotel-lobby/Our first live conference check-ins. Wait times dropped a lot. The model was the part that worked from day one. Lighting, queues and printers were not.Tue, 18 Sep 2018 00:00:00 GMTamanda-aistartupsfacesoperationshttps://vinsonli.com/posts/musically-is-now-tiktok/https://vinsonli.com/posts/musically-is-now-tiktok/ByteDance merged Musical.ly into TikTok this week. A short follow-up to what I wrote in November, and what the feed tells you about the product.Wed, 08 Aug 2018 00:00:00 GMTrecommendationshort-videochinaproducthttps://vinsonli.com/posts/a-robot-hand-with-a-hundred-years-of-practice/https://vinsonli.com/posts/a-robot-hand-with-a-hundred-years-of-practice/OpenAI's Dactyl learned to rotate a block with a human-like robot hand, entirely in simulation. Why hands, and why randomizing the simulator is the clever part.Tue, 31 Jul 2018 00:00:00 GMTroboticshandssimulationreinforcement-learninghttps://vinsonli.com/posts/read-everything-first-specialize-later/https://vinsonli.com/posts/read-everything-first-specialize-later/OpenAI trained a Transformer to predict the next word on thousands of books, then fine-tuned it on small tasks. Language is getting its ImageNet moment.Wed, 20 Jun 2018 00:00:00 GMTlanguagetransformerpretrainingdeep-learninghttps://vinsonli.com/posts/the-uber-car-saw-her-six-seconds-early/https://vinsonli.com/posts/the-uber-car-saw-her-six-seconds-early/The NTSB's preliminary report on the Tempe crash shows the sensors worked. Classification flickered, braking was disabled, and the human was supposed to be the backup.Sun, 27 May 2018 00:00:00 GMTself-drivingsafetysystemshttps://vinsonli.com/posts/faces-help-you-hear/https://vinsonli.com/posts/faces-help-you-hear/Google's Looking to Listen separates one voice from a crowd by watching the speaker's face. Modalities work better when they explain each other.Sat, 21 Apr 2018 00:00:00 GMTfacesaudiomultimodaldeep-learninghttps://vinsonli.com/posts/an-agent-that-dreams-its-own-racetrack/https://vinsonli.com/posts/an-agent-that-dreams-its-own-racetrack/Ha and Schmidhuber's World Models compresses what an agent sees, learns to predict what happens next, and trains a tiny controller inside its own dream. The most important paper I've read this year.Sat, 31 Mar 2018 00:00:00 GMTworld-modelsreinforcement-learninggenerative-modelshttps://vinsonli.com/posts/starting-a-face-recognition-company/https://vinsonli.com/posts/starting-a-face-recognition-company/I've co-founded a company that does face recognition on device. Why on device, why conference check-in first, and the lines we won't cross.Mon, 26 Feb 2018 00:00:00 GMTstartupsfacesprivacyamanda-aihttps://vinsonli.com/posts/can-a-network-have-taste/https://vinsonli.com/posts/can-a-network-have-taste/Google's NIMA predicts how people would rate a photo, as a distribution, not a single score. Modeling disagreement turns out to be the useful part.Mon, 22 Jan 2018 00:00:00 GMTaestheticsdeep-learningphotographytastehttps://vinsonli.com/posts/face-swaps-on-reddit/https://vinsonli.com/posts/face-swaps-on-reddit/Someone on Reddit is putting celebrities' faces into porn with a home GPU and open-source tools. How it works, and why consent has to be designed in now.Sat, 16 Dec 2017 00:00:00 GMTfacesdeepfakesethicsgenerative-modelshttps://vinsonli.com/posts/watch-the-recommendation-engine/https://vinsonli.com/posts/watch-the-recommendation-engine/ByteDance is buying Musical.ly. It looks like a Chinese news company buying an American teen app. It's really a recommendation engine buying a global audience.Tue, 14 Nov 2017 00:00:00 GMTrecommendationshort-videochinaproducthttps://vinsonli.com/posts/alphago-zero-threw-away-human-games/https://vinsonli.com/posts/alphago-zero-threw-away-human-games/Starting from random play, with no human data, it beat the version that beat Lee Sedol 100 games to 0 after three days. Human knowledge was a ceiling.Sat, 21 Oct 2017 00:00:00 GMTaireinforcement-learningdeepmindgohttps://vinsonli.com/posts/face-id-the-3d-face-is-consumer-hardware/https://vinsonli.com/posts/face-id-the-3d-face-is-consumer-hardware/The iPhone X projects 30,000 infrared dots onto your face to unlock the phone. Four years ago I was faking this in software from a single selfie.Wed, 13 Sep 2017 00:00:00 GMTfaces3dhardwareapplesecurityhttps://vinsonli.com/posts/nobody-taught-it-to-walk/https://vinsonli.com/posts/nobody-taught-it-to-walk/DeepMind's locomotion agents learned to run, jump and duck from a reward for moving forward and a varied course. The flailing arms are the point.Sat, 12 Aug 2017 00:00:00 GMTreinforcement-learningroboticssimulationembodimenthttps://vinsonli.com/posts/no-recurrence-no-convolution/https://vinsonli.com/posts/no-recurrence-no-convolution/A Google paper throws out the RNN and translates with attention alone. What attention actually computes, and why I think it goes beyond translation.Thu, 06 Jul 2017 00:00:00 GMTdeep-learninglanguagearchitecturetransformerhttps://vinsonli.com/posts/curiosity-is-a-prediction-error/https://vinsonli.com/posts/curiosity-is-a-prediction-error/A Berkeley paper gives agents an internal reward for being surprised by their own predictions. It learns Mario with no score at all, and it answers a question I asked two years ago.Sat, 17 Jun 2017 00:00:00 GMTreinforcement-learningcuriositylearninghttps://vinsonli.com/posts/ke-jie-in-wuzhen/https://vinsonli.com/posts/ke-jie-in-wuzhen/The world's top Go player lost 3-0 to AlphaGo in China this week, and Chinese viewers mostly couldn't watch it live. Why this match will matter more in China than the Lee Sedol one did.Sun, 28 May 2017 00:00:00 GMTaigochinadeepmindhttps://vinsonli.com/posts/neuralink-and-the-bandwidth-argument/https://vinsonli.com/posts/neuralink-and-the-bandwidth-argument/Elon Musk's new company wants a high-bandwidth link between brains and computers. The argument is about output speed. I think it's partly right, and too optimistic on timing.Sun, 23 Apr 2017 00:00:00 GMTbcihardwareneurosciencehttps://vinsonli.com/posts/breath-of-the-wild-is-a-physics-engine/https://vinsonli.com/posts/breath-of-the-wild-is-a-physics-engine/The new Zelda replaced scripted puzzles with a world that follows consistent rules. It's the best example I've seen of what a training environment for intelligence could look like.Sat, 18 Mar 2017 00:00:00 GMTgamesphysicssimulationlearninghttps://vinsonli.com/posts/the-camera-is-the-new-keyboard/https://vinsonli.com/posts/the-camera-is-the-new-keyboard/Snap filed to go public and calls itself a camera company. The phrase sounds like marketing, but it describes a real shift in how people put things into computers.Tue, 07 Feb 2017 00:00:00 GMTcamerasocialproductfaceshttps://vinsonli.com/posts/poker-fell/https://vinsonli.com/posts/poker-fell/CMU's Libratus beat four top professionals at heads-up no-limit hold'em over twenty days. Hidden information was supposed to protect humans a while longer.Tue, 31 Jan 2017 00:00:00 GMTaigamesgame-theoryhttps://vinsonli.com/posts/every-lab-is-building-a-playground/https://vinsonli.com/posts/every-lab-is-building-a-playground/DeepMind open-sourced its 3D lab and OpenAI released Universe in the same week. Environments are becoming the new datasets.Fri, 09 Dec 2016 00:00:00 GMTreinforcement-learningsimulationresearchhttps://vinsonli.com/posts/google-translate-got-better-overnight/https://vinsonli.com/posts/google-translate-got-better-overnight/Neural machine translation replaced phrase tables, starting with Chinese to English. Notes from someone who reads both, and why the zero-shot result is the bigger story.Sat, 19 Nov 2016 00:00:00 GMTtranslationlanguagedeep-learningchinesehttps://vinsonli.com/posts/vine-is-dead-short-video-isnt/https://vinsonli.com/posts/vine-is-dead-short-video-isnt/Twitter is shutting down Vine. It didn't lose because six-second videos were a bad idea.Sat, 29 Oct 2016 00:00:00 GMTvideosocialproductcreatorshttps://vinsonli.com/posts/16000-samples-a-second/https://vinsonli.com/posts/16000-samples-a-second/DeepMind's WaveNet generates raw audio one sample at a time, and its speech sounds far more human than anything before. How, and why it's so slow.Wed, 14 Sep 2016 00:00:00 GMTaudiogenerative-modelsdeepmindspeechhttps://vinsonli.com/posts/the-long-tail-of-driving/https://vinsonli.com/posts/the-long-tail-of-driving/Uber will put self-driving cars on Pittsburgh streets next month, six weeks after the first Autopilot fatality became public. The hard part of driving is the rare case.Sun, 21 Aug 2016 00:00:00 GMTself-drivingroboticssafetyhttps://vinsonli.com/posts/pokemon-go-is-ar-without-the-ar/https://vinsonli.com/posts/pokemon-go-is-ar-without-the-ar/The biggest augmented reality hit so far barely uses augmented reality. It's a map, built on years of data from another game.Wed, 13 Jul 2016 00:00:00 GMTargamesproductmapshttps://vinsonli.com/posts/still-caring-about-facial-conformation/https://vinsonli.com/posts/still-caring-about-facial-conformation/My last paper from grad school is out. Why I still think geometric priors for faces matter when every result now comes from deep learning.Thu, 30 Jun 2016 00:00:00 GMTfaces3dresearchhttps://vinsonli.com/posts/google-built-its-own-chip/https://vinsonli.com/posts/google-built-its-own-chip/The Tensor Processing Unit has been running in Google's data centers for a year. When a company designs custom silicon for a workload, the workload is here to stay.Sun, 22 May 2016 00:00:00 GMThardwaremachine-learninggoogleinfrastructurehttps://vinsonli.com/posts/chatbots-are-this-years-apps/https://vinsonli.com/posts/chatbots-are-this-years-apps/Facebook opened Messenger to bots at F8, two weeks after Microsoft's Tay learned racism from Twitter in a day. Why I'm telling clients to wait.Thu, 14 Apr 2016 00:00:00 GMTchatbotslanguageproductbusinesshttps://vinsonli.com/posts/move-37/https://vinsonli.com/posts/move-37/AlphaGo beat Lee Sedol four games to one. The move everyone is talking about, how the system found it, and why self-play is the part that matters.Wed, 16 Mar 2016 00:00:00 GMTaigoreinforcement-learningdeepmindhttps://vinsonli.com/posts/the-best-robotics-video-involves-a-hockey-stick/https://vinsonli.com/posts/the-best-robotics-video-involves-a-hockey-stick/Boston Dynamics' new Atlas gets shoved, loses its box and gets knocked flat, and keeps going. Why recovering is harder than walking.Fri, 26 Feb 2016 00:00:00 GMTroboticscontrolphysicshttps://vinsonli.com/posts/society-of-mind-as-an-engineering-manager/https://vinsonli.com/posts/society-of-mind-as-an-engineering-manager/Marvin Minsky died this week, the same week DeepMind announced a Go program that beat a professional. Two very different ideas of how minds get built.Fri, 29 Jan 2016 00:00:00 GMTaibooksmanagementgohttps://vinsonli.com/posts/152-layers-and-the-trick-is-learning-nothing/https://vinsonli.com/posts/152-layers-and-the-trick-is-learning-nothing/Microsoft Research's residual networks won ImageNet with a network eight times deeper than last year's. The idea behind it is almost too simple.Tue, 15 Dec 2015 00:00:00 GMTdeep-learningcomputer-visionarchitecturehttps://vinsonli.com/posts/tensorflow-is-open-source/https://vinsonli.com/posts/tensorflow-is-open-source/Google released its internal machine learning library. What changes for a thirty-person app studio that builds software for other companies.Thu, 12 Nov 2015 00:00:00 GMTmachine-learningtoolsengineeringbusinesshttps://vinsonli.com/posts/ac-dc-at-sixty/https://vinsonli.com/posts/ac-dc-at-sixty/Angus Young is sixty and still runs across the stage in a school uniform. Some thoughts after the show.Sat, 17 Oct 2015 00:00:00 GMTmusiclifehttps://vinsonli.com/posts/the-screen-feels-how-hard-you-press/https://vinsonli.com/posts/the-screen-feels-how-hard-you-press/3D Touch on the iPhone 6s, one year after I guessed pressure-sensitive phones were a couple of years away. How it measures force, and what it's for.Sat, 19 Sep 2015 00:00:00 GMThardwareinterfaceshapticsapplehttps://vinsonli.com/posts/van-gogh-is-a-gram-matrix/https://vinsonli.com/posts/van-gogh-is-a-gram-matrix/A new paper separates the content of an image from its style using a network trained for classification. How it works, and what it suggests about taste.Mon, 31 Aug 2015 00:00:00 GMTdeep-learningartstylegenerative-modelshttps://vinsonli.com/posts/human-djs-vs-the-skip-button/https://vinsonli.com/posts/human-djs-vs-the-skip-button/Apple Music launched with Beats 1 and human curators. Three weeks later, Spotify ships Discover Weekly. Two theories of music discovery.Fri, 24 Jul 2015 00:00:00 GMTmusicrecommendationproducthttps://vinsonli.com/posts/deepdream-sees-dogs-everywhere/https://vinsonli.com/posts/deepdream-sees-dogs-everywhere/Running a network in reverse to see what it learned. Why everything turns into dogs, and what that says about training data.Fri, 26 Jun 2015 00:00:00 GMTdeep-learningcomputer-visionvisualizationhttps://vinsonli.com/posts/google-photos-gives-away-what-i-built/https://vinsonli.com/posts/google-photos-gives-away-what-i-built/Unlimited storage and automatic face grouping, free. What happens to small companies when a platform ships their feature as a default.Sun, 31 May 2015 00:00:00 GMTproductfacesplatformsgooglehttps://vinsonli.com/posts/two-years-of-3d-faces-what-i-got-wrong/https://vinsonli.com/posts/two-years-of-3d-faces-what-i-got-wrong/Leaving the selfie-to-3D-face app I helped build out of my thesis. The technology mostly worked. The rest is the interesting part.Thu, 30 Apr 2015 00:00:00 GMTstartupsfacesproductcareerhttps://vinsonli.com/posts/the-dress-is-a-world-model-bug/https://vinsonli.com/posts/the-dress-is-a-world-model-bug/Half the internet sees white and gold, half sees blue and black. The disagreement is about lighting, and it says something about how vision works.Tue, 03 Mar 2015 00:00:00 GMTperceptionvisioncolorhttps://vinsonli.com/posts/zero-points-in-montezumas-revenge/https://vinsonli.com/posts/zero-points-in-montezumas-revenge/DeepMind's DQN paper is in Nature. What the network actually learns, and the game where it learns nothing at all.Fri, 27 Feb 2015 00:00:00 GMTreinforcement-learningdeep-learningdeepmindhttps://vinsonli.com/posts/hololens-maps-your-room-first/https://vinsonli.com/posts/hololens-maps-your-room-first/Microsoft's headset is mostly a sensing problem with a display attached. Why AR needs a model of the space before it can put anything in it.Tue, 27 Jan 2015 00:00:00 GMTarhardwarecomputer-visionhttps://vinsonli.com/posts/two-networks-arguing/https://vinsonli.com/posts/two-networks-arguing/The GAN paper from NIPS this year: a generator, a discriminator, and a learned idea of what counts as real.Thu, 18 Dec 2014 00:00:00 GMTdeep-learninggenerative-modelsfaceshttps://vinsonli.com/posts/a-computer-captioned-a-photo/https://vinsonli.com/posts/a-computer-captioned-a-photo/Google and Stanford both have networks that write sentences about images. How they work, and what they're actually learning.Sat, 22 Nov 2014 00:00:00 GMTcomputer-visiondeep-learninglanguagehttps://vinsonli.com/posts/what-a-light-field-actually-is/https://vinsonli.com/posts/what-a-light-field-actually-is/Magic Leap raised $542 million without showing anyone a product. A look at the optics they're promising, and why it's so hard to wear on your head.Sat, 25 Oct 2014 00:00:00 GMTaropticshardwarehttps://vinsonli.com/posts/the-watch-taps-your-wrist/https://vinsonli.com/posts/the-watch-taps-your-wrist/The part of the Apple Watch announcement I keep thinking about is a small mass on a spring, and what it says about how little computers use touch.Tue, 16 Sep 2014 00:00:00 GMThapticshardwareinterfacessensinghttps://vinsonli.com/posts/radial-basis-functions-for-people-who-dont-care/https://vinsonli.com/posts/radial-basis-functions-for-people-who-dont-care/I presented a facial animation paper this week. What it's about, explained with a rubber sheet, some pins and a steel plate.Tue, 26 Aug 2014 00:00:00 GMTfaces3dmathresearchhttps://vinsonli.com/posts/why-my-3d-faces-look-dead/https://vinsonli.com/posts/why-my-3d-faces-look-dead/Our app gets the shape of your face right and still looks like a wax museum. Most of the problem turned out to be motion.Sat, 19 Jul 2014 00:00:00 GMTfaces3danimationphysicsproducthttps://vinsonli.com/posts/a-video-is-not-a-stack-of-photos/https://vinsonli.com/posts/a-video-is-not-a-stack-of-photos/Two new papers on video classification, and why the network that only sees motion beat the one that sees the frames.Sun, 22 Jun 2014 00:00:00 GMTvideocomputer-visiondeep-learningoptical-flowhttps://vinsonli.com/posts/graduated-thesis-in-one-paragraph/https://vinsonli.com/posts/graduated-thesis-in-one-paragraph/What my master's thesis on 3D face meshes does, which part of it I think neural networks will replace soon, and which part I think they won't.Thu, 29 May 2014 00:00:00 GMTfaces3ddeep-learningcareer